diff options
58 files changed, 9745 insertions, 2812 deletions
diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000..5034363 --- /dev/null +++ b/.gitignore @@ -0,0 +1,2 @@ +__pycache__/ +build/
\ No newline at end of file diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index f40778a..3afbd2d 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -11,7 +11,7 @@ repos: #j.hooks: #j.- id: black -- repo: https://gitlab.com/pycqa/flake8 +- repo: https://github.com/pycqa/flake8 rev: 3.7.9 hooks: - id: flake8 diff --git a/CMakeLists.txt b/CMakeLists.txt index aaba60e..5f4f851 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1,6 +1,6 @@ #TOSA serialization library -# Copyright (c) 2020-2021, ARM Limited. +# Copyright (c) 2020-2023, ARM Limited. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -19,47 +19,89 @@ cmake_minimum_required(VERSION 3.13.4) project(TosaSerialization) -set(CMAKE_CXX_STANDARD 14 CACHE STRING "C++ standard to conform to") -set(CMAKE_CXX_STANDARD_REQUIRED YES) +set(CMAKE_CXX_STANDARD 17) +set(CMAKE_CXX_STANDARD_REQUIRED ON) set(CMAKE_VERBOSE_MAKEFILE ON) -include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include) -include_directories(${PROJECT_SOURCE_DIR}/third_party/flatbuffers/include) +option(BUILD_TESTS "Build test applications" ON) +option(FLATBUFFERS_ROOT "Location where the flatbuffers 'include' and 'lib' folders to be found" Off) + +include_directories(${PROJECT_SOURCE_DIR}/third_party/half/include) -# Turn off unnecessary flatbuffers targets -set(FLATBUFFERS_BUILD_TESTS OFF) -add_subdirectory(third_party/flatbuffers) +include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include) -add_library(tosa_serialization_lib STATIC +add_library(tosa_serialization_lib src/tosa_serialization_handler.cpp src/numpy_utils.cpp ) -target_link_libraries(tosa_serialization_lib PRIVATE flatbuffers) - -add_executable(serialization_read_write - test/src/serialization_read_write.cpp +# If flatbuffers is built externally just link it +if (TARGET flatbuffers) + target_link_libraries(tosa_serialization_lib PRIVATE flatbuffers) +else() + # Verify we have a valid flatbuffers include path. + # We will explicitly exclude the system include directories and only + # accept either a user supplied value or the local third_party/flatbuffers. + find_path(FLATBUFFERS_INCLUDE_PATH flatbuffers/flatbuffers.h + NO_DEFAULT_PATH + HINTS ${FLATBUFFERS_ROOT} ./third_party/flatbuffers + PATH_SUFFIXES include) + message(STATUS "Flatbuffers include located at: ${FLATBUFFERS_INCLUDE_PATH}") + include_directories(${FLATBUFFERS_INCLUDE_PATH}) + + # Next is the library. + # We will explicitly exclude the system lib directories and only accept + # either a user supplied value or the local third_party/flatbuffers. + find_library(FLATBUFFERS_LIBRARY + NAMES libflatbuffers.a flatbuffers + NO_DEFAULT_PATH + HINTS ${FLATBUFFERS_ROOT} ./third_party/flatbuffers + PATH_SUFFIXES lib) + + if(FLATBUFFERS_LIBRARY) + message(STATUS "Flatbuffers library located at: ${FLATBUFFERS_LIBRARY}") + target_link_libraries(tosa_serialization_lib PRIVATE ${FLATBUFFERS_LIBRARY}) + else() + # It's not there we treat third_party/flatbuffers as a sub project. + # In this case we'll need to build the downloaded source. + # Turn off unnecessary flatbuffers targets + set(FLATBUFFERS_BUILD_TESTS OFF) + add_subdirectory(third_party/flatbuffers) + target_link_libraries(tosa_serialization_lib PRIVATE flatbuffers) + endif() +endif() + +set(public_headers) +list(APPEND public_headers + include/attribute.h + include/attribute.def + include/float_utils.h + include/numpy_utils.h + include/tosa_generated.h + include/tosa_serialization_handler.h ) -target_link_libraries(serialization_read_write - tosa_serialization_lib -) +set_target_properties(tosa_serialization_lib PROPERTIES PUBLIC_HEADER "${public_headers}") -add_executable(serialization_npy_test - test/src/serialization_npy_test.cpp -) +# Optionally build test executables. +if (BUILD_TESTS) + add_executable(serialization_npy_test + test/src/serialization_npy_test.cpp + ) -target_link_libraries(serialization_npy_test - tosa_serialization_lib -) + target_link_libraries(serialization_npy_test + tosa_serialization_lib + ) +endif() set(TOSA_SERIALIZATION_LIB_CMAKE_DIR "${CMAKE_INSTALL_LIBDIR}/cmake/tosa_serialization_lib") - +# Follow GNU packaging norms for installation directory structure. +include(GNUInstallDirs) install( TARGETS tosa_serialization_lib EXPORT TosaSerializationLibTargets - ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR} - INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR} + PUBLIC_HEADER + ARCHIVE ) install(EXPORT TosaSerializationLibTargets @@ -158,3 +158,10 @@ numpy file, (4) format and (5) usage. # License The *TOSA Serialization Library* is licensed under Apache-2.0. + +## Third Party Projects + +- The [half](https://half.sourceforge.net/) library is licensed under MIT license. + +Other third party projects are referenced as sub-modules and as such, are licensed under the licenses stated in their projects. + diff --git a/include/attribute.def b/include/attribute.def index ea91869..723543e 100644 --- a/include/attribute.def +++ b/include/attribute.def @@ -1,5 +1,5 @@ -// Copyright (c) 2020-2021, ARM Limited. +// Copyright (c) 2020-2024, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -26,68 +26,56 @@ ...: variadic variables for more arguments, depending on NUM_ARGS_IN_ATTRIBUTES */ -DEF_ATTRIBUTE(Pool, 5, +DEF_ATTRIBUTE(Pool, 6, int32_t, V, pad, int32_t, V, kernel, int32_t, V, stride, int32_t, S, input_zp, - int32_t, S, output_zp) + int32_t, S, output_zp, + DType, S, acc_type) -DEF_ATTRIBUTE(Conv, 5, +DEF_ATTRIBUTE(Conv, 7, int32_t, V, pad, int32_t, V, stride, int32_t, V, dilation, int32_t, S, input_zp, - int32_t, S, weight_zp) + int32_t, S, weight_zp, + bool, S, local_bound, + DType, S, acc_type) -DEF_ATTRIBUTE(TransposeConv, 5, +DEF_ATTRIBUTE(TransposeConv, 7, int32_t, V, out_pad, int32_t, V, stride, int32_t, V, output_shape, int32_t, S, input_zp, - int32_t, S, weight_zp) + int32_t, S, weight_zp, + bool, S, local_bound, + DType, S, acc_type) -DEF_ATTRIBUTE(Pad, 3, - int32_t, V, padding, - int32_t, S, pad_const_int, - float, S, pad_const_fp) +DEF_ATTRIBUTE(Pad, 1, + uint8_t, V, pad_const) DEF_ATTRIBUTE(Axis, 1, int32_t, S, axis) -DEF_ATTRIBUTE(Reshape, 1, - int32_t, V, new_shape) - -DEF_ATTRIBUTE(Slice, 2, - int32_t, V, start, - int32_t, V, size) - -DEF_ATTRIBUTE(Tile, 1, - int32_t, V, multiples) - -DEF_ATTRIBUTE(Resize, 7, - int32_t, V, output_size, - int32_t, V, stride, - int32_t, V, offset, - int32_t, S, shift, - float, V, stride_fp, - float, V, offset_fp, +DEF_ATTRIBUTE(Resize, 4, + int16_t, V, scale, + int16_t, V, offset, + int16_t, V, border, ResizeMode, S, mode) -DEF_ATTRIBUTE(Clamp, 4, - int32_t, S, min_int, - int32_t, S, max_int, - float, S, min_fp, - float, S, max_fp) +DEF_ATTRIBUTE(Clamp, 2, + uint8_t, V, min_val, + uint8_t, V, max_val) DEF_ATTRIBUTE(Rescale, 7, int32_t, S, input_zp, int32_t, S, output_zp, - int32_t, V, multiplier, - int32_t, V, shift, bool, S, scale32, bool, S, double_round, - bool, S, per_channel) + bool, S, per_channel, + bool, S, input_unsigned, + bool, S, output_unsigned) DEF_ATTRIBUTE(Mul, 1, int32_t, S, shift) @@ -96,12 +84,12 @@ DEF_ATTRIBUTE(ArithmeticRightShift, 1, bool, S, round) DEF_ATTRIBUTE(CondIf, 2, - string, S, then_branch, - string, S, else_branch) + string, S, then_graph, + string, S, else_graph) DEF_ATTRIBUTE(WhileLoop, 2, - string, S, cond_branch, - string, S, body_branch) + string, S, cond_graph, + string, S, body_graph) DEF_ATTRIBUTE(Transpose, 1, int32_t, V, perms) @@ -120,3 +108,15 @@ DEF_ATTRIBUTE(FullyConnected, 2, DEF_ATTRIBUTE(Negate, 2, int32_t, S, input1_zp, int32_t, S, output_zp) + +DEF_ATTRIBUTE(Custom, 3, + string, S, operator_name, + string, S, domain_name, + uint8_t, V, implementation_attrs) + +DEF_ATTRIBUTE(FFT, 2, + bool, S, inverse, + bool, S, local_bound) + +DEF_ATTRIBUTE(RFFT, 1, + bool, S, local_bound) diff --git a/include/attribute.h b/include/attribute.h index 93f7bc4..d6c96a4 100644 --- a/include/attribute.h +++ b/include/attribute.h @@ -40,13 +40,43 @@ public: {} }; +inline int convertFlatbuffersU8toF32(const flatbuffers::Vector<uint8_t>& in, uint32_t out_size, std::vector<float>& out) +{ + out.clear(); + if (in.size() < out_size * sizeof(float)) + { + printf("convertFlatbuffersU8toF32(): uint8 Flatbuffers buffer size %u must be >= target size %ld\n", in.size(), + out_size * sizeof(float)); + return 1; + } + for (uint32_t i = 0; i < out_size; i++) + { + uint32_t byte0 = in[i * sizeof(float)]; + uint32_t byte1 = in[i * sizeof(float) + 1]; + uint32_t byte2 = in[i * sizeof(float) + 2]; + uint32_t byte3 = in[i * sizeof(float) + 3]; + uint32_t val_u32 = byte0 + (byte1 << 8) + (byte2 << 16) + (byte3 << 24); + float* val_fp32 = reinterpret_cast<float*>(&val_u32); + out.push_back(*val_fp32); + } + return 0; +} + #define DEF_ARGS_VER0_S_STR(V) _##V = p->V()->str(); #define DEF_ARGS_VER0_S_DEFAULT(V) _##V = p->V(); +#define DEF_ARGS_VER0_S_float_as_bytes(V) \ + { \ + std::vector<float> attr_vec; \ + if (p->V() && convertFlatbuffersU8toF32(*(p->V()), 1, attr_vec)) \ + assert(0 && "Failed to convert u8 buffer to f32"); \ + _##V = (!attr_vec.empty()) ? attr_vec[0] : 0.0f; \ + } #define DEF_ARGS_VER0_S_int32_t(V) DEF_ARGS_VER0_S_DEFAULT(V) -#define DEF_ARGS_VER0_S_float(V) DEF_ARGS_VER0_S_DEFAULT(V) +#define DEF_ARGS_VER0_S_float(V) DEF_ARGS_VER0_S_float_as_bytes(V) #define DEF_ARGS_VER0_S_bool(V) DEF_ARGS_VER0_S_DEFAULT(V) #define DEF_ARGS_VER0_S_ResizeMode(V) DEF_ARGS_VER0_S_DEFAULT(V) +#define DEF_ARGS_VER0_S_DType(V) DEF_ARGS_VER0_S_DEFAULT(V) #define DEF_ARGS_VER0_S_string(V) DEF_ARGS_VER0_S_STR(V) #define DEF_ARGS_VER0_S(T, V) DEF_ARGS_VER0_S_##T(V) @@ -100,6 +130,18 @@ public: DEF_ARGS_##VER(FALSE, T3, F3, V3) DEF_ARGS_##VER(FALSE, T4, F4, V4) DEF_ARGS_##VER(FALSE, T5, F5, V5) \ DEF_ARGS_##VER(FALSE, T6, F6, V6) +#define DEF_ARGS_8(VER, T0, F0, V0, T1, F1, V1, T2, F2, V2, T3, F3, V3, T4, F4, V4, T5, F5, V5, T6, F6, V6, T7, F7, \ + V7) \ + DEF_ARGS_##VER(TRUE, T0, F0, V0) DEF_ARGS_##VER(FALSE, T1, F1, V1) DEF_ARGS_##VER(FALSE, T2, F2, V2) \ + DEF_ARGS_##VER(FALSE, T3, F3, V3) DEF_ARGS_##VER(FALSE, T4, F4, V4) DEF_ARGS_##VER(FALSE, T5, F5, V5) \ + DEF_ARGS_##VER(FALSE, T6, F6, V6) DEF_ARGS_##VER(FALSE, T7, F7, V7) + +#define DEF_ARGS_9(VER, T0, F0, V0, T1, F1, V1, T2, F2, V2, T3, F3, V3, T4, F4, V4, T5, F5, V5, T6, F6, V6, T7, F7, \ + V7, T8, F8, V8) \ + DEF_ARGS_##VER(TRUE, T0, F0, V0) DEF_ARGS_##VER(FALSE, T1, F1, V1) DEF_ARGS_##VER(FALSE, T2, F2, V2) \ + DEF_ARGS_##VER(FALSE, T3, F3, V3) DEF_ARGS_##VER(FALSE, T4, F4, V4) DEF_ARGS_##VER(FALSE, T5, F5, V5) \ + DEF_ARGS_##VER(FALSE, T6, F6, V6) DEF_ARGS_##VER(FALSE, T7, F7, V7) DEF_ARGS_##VER(FALSE, T8, F8, V8) + #define DEF_VER0_VAR_DECL_PTR(NAME) const NAME* p = static_cast<const NAME*>(options); #define DEF_VER0_VAR_0(NAME) #define DEF_VER0_VAR_1(NAME) DEF_VER0_VAR_DECL_PTR(NAME) @@ -109,6 +151,8 @@ public: #define DEF_VER0_VAR_5(NAME) DEF_VER0_VAR_DECL_PTR(NAME) #define DEF_VER0_VAR_6(NAME) DEF_VER0_VAR_DECL_PTR(NAME) #define DEF_VER0_VAR_7(NAME) DEF_VER0_VAR_DECL_PTR(NAME) +#define DEF_VER0_VAR_8(NAME) DEF_VER0_VAR_DECL_PTR(NAME) +#define DEF_VER0_VAR_9(NAME) DEF_VER0_VAR_DECL_PTR(NAME) #define DEF_ATTRIBUTE(NAME, NUM_ARGS, ...) \ class Tosa##NAME##Attribute : public TosaAttributeBase \ @@ -144,6 +188,8 @@ public: #undef DEF_ARGS_5 #undef DEF_ARGS_6 #undef DEF_ARGS_7 +#undef DEF_ARGS_8 +#undef DEF_ARGS_9 #undef DEF_ARGS_VER0 #undef DEF_ARGS_VER1 #undef DEF_ARGS_VER2 @@ -153,6 +199,7 @@ public: #undef DEF_ARGS_VER0_S_float #undef DEF_ARGS_VER0_S_bool #undef DEF_ARGS_VER0_S_ResizeMode +#undef DEF_ARGS_VER0_S_DType #undef DEF_ARGS_VER0_S_string #undef DEF_ARGS_VER0_S_STR #undef DEF_ARGS_VER0_S_DEFAULT diff --git a/include/float_utils.h b/include/float_utils.h new file mode 100644 index 0000000..831ad74 --- /dev/null +++ b/include/float_utils.h @@ -0,0 +1,533 @@ +// Copyright (c) 2024, ARM Limited. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifndef TOSA_FLOAT_UTILS_H_ +#define TOSA_FLOAT_UTILS_H_ + +#include <algorithm> +#include <cstdint> +#include <limits> +#include <type_traits> +#if defined(__cpp_lib_bit_cast) +#include <bit> +#endif // defined(__cpp_lib_bit_cast) + +namespace tosa +{ + +namespace float_support +{ + +struct hidden +{}; + +#if defined(__cpp_lib_bit_cast) +#define BITCAST_CONSTEXPR constexpr inline + +constexpr inline int32_t get_bits(const float& f) +{ + return std::bit_cast<int32_t>(f); +} +constexpr inline float from_bits(const int32_t& i) +{ + return std::bit_cast<float>(i); +} + +#else +#define BITCAST_CONSTEXPR inline + +union ufloat32 +{ + constexpr ufloat32(const float& x) + : f(x) + {} + constexpr ufloat32(const int32_t& x) + : i(x) + {} + + float f; + int32_t i; +}; + +inline int32_t get_bits(const float& f) +{ + return ufloat32(f).i; +} +inline float from_bits(const int32_t& i) +{ + return ufloat32(i).f; +} +#endif + +} // namespace float_support + +template <typename storage_t, + size_t n_exp_bits, + bool has_nan, + bool with_denorm, + bool with_infinity, + std::enable_if_t<(n_exp_bits + 1 < sizeof(storage_t) * 8), bool> = true> +class float_t +{ + storage_t m_data = 0; + +public: + static constexpr size_t n_exponent_bits = n_exp_bits; + static constexpr size_t n_significand_bits = sizeof(storage_t) * 8 - 1 - n_exp_bits; + static constexpr int64_t exponent_bias = (1 << (n_exp_bits - 1)) - 1; + + /// \brief Construct a floating point type with the given bit + /// representation. + static constexpr float_t from_bits(storage_t bits) + { + return float_t(float_support::hidden(), bits); + } + + /// \brief Construct a float from the given sign, exponent and significand + /// bits. + static constexpr float_t from_bits(bool pm, storage_t e, storage_t s) + { + storage_t bits = pm ? 1 : 0; + + bits <<= n_exp_bits; + bits |= e; + + bits <<= n_significand_bits; + if (with_denorm || e) + bits |= s; + + return float_t(float_support::hidden(), bits); + } + + /// \brief (Hidden) Construct a float type from a given bit pattern + constexpr float_t(const float_support::hidden&, storage_t bits) + : m_data(bits) + {} + + constexpr float_t() + : m_data(0) + {} + constexpr float_t(const float_t& other) + : m_data(other.m_data) + {} + + /// \brief Cast to a different floating point representation. + template <typename other_storage_t, + size_t other_n_exp_bits, + bool other_has_nan, + bool other_has_denorm, + bool other_has_infinity> + constexpr inline + operator float_t<other_storage_t, other_n_exp_bits, other_has_nan, other_has_denorm, other_has_infinity>() const + { + using other_float_t = + float_t<other_storage_t, other_n_exp_bits, other_has_nan, other_has_denorm, other_has_infinity>; + + // Shortcut for types which are fundamentally similar (e.g., bf16 -> + // fp32) + if constexpr (n_exp_bits == other_n_exp_bits && sizeof(other_storage_t) >= sizeof(storage_t) && + has_nan == other_has_nan) + { + return other_float_t::from_bits(static_cast<other_storage_t>(m_data) + << (sizeof(other_storage_t) - sizeof(storage_t)) * 8); + } + + // Get initial values for the new floating point type + const bool sign_bit = m_data < 0; + int64_t new_exponent_bits = 0; + uint64_t new_significand = 0; + + if (is_nan() || is_infinity()) + { + new_exponent_bits = (1 << other_n_exp_bits) - 1; + + if (is_nan()) + { + if constexpr (other_has_infinity) + { + // Copy across the `not_quiet bit`; set the LSB. Don't + // attempt to copy across any of the rest of the payload. + new_significand = + 0x1 | (((significand() >> (n_significand_bits - 1)) & 1) << other_float_t::n_significand_bits); + } + else + { + new_significand = (1ul << other_float_t::n_significand_bits) - 1; + } + } + else if constexpr (!other_has_infinity) + { + new_significand = (1ul << other_float_t::n_significand_bits) - (other_has_nan ? 2 : 1); + } + } + else if (!is_zero()) + { + const int64_t this_exponent_bits = exponent_bits(); + { + constexpr int64_t exponent_rebias = other_float_t::exponent_bias - exponent_bias; + new_exponent_bits = std::max(this_exponent_bits + exponent_rebias, exponent_rebias + 1); + } + new_significand = this->significand() << (64 - n_significand_bits); + + // Normalise subnormals + if (this_exponent_bits == 0) + { + // Shift the most-significant 1 out of the magnitude to convert + // it to a significand. Modify the exponent accordingly. + uint8_t shift = __builtin_clzl(new_significand) + 1; + new_exponent_bits -= shift; + new_significand <<= shift; + } + + // Align the significand for the output type + uint32_t shift = 64 - other_float_t::n_significand_bits; + const bool other_is_subnormal = new_exponent_bits <= 0; + if (other_is_subnormal) + { + shift += 1 - new_exponent_bits; + new_exponent_bits = 0; + } + + const uint64_t shift_out = shift == 64 ? new_significand : new_significand & ((1ll << shift) - 1); + new_significand = shift == 64 ? 0 : new_significand >> shift; + + // Reinsert the most-significant-one if this is a subnormal in the + // output type. + new_significand |= (other_is_subnormal ? 1ll : 0) << (64 - shift); + + // Apply rounding based on the bits shifted out of the significand + const uint64_t shift_half = 1ll << (shift - 1); + if (shift_out > shift_half || (shift_out == shift_half && (new_significand & 1))) + { + new_significand += 1; + + // Handle the case that the significand overflowed due to + // rounding + constexpr uint64_t max_significand = (1ll << other_float_t::n_significand_bits) - 1; + if (new_significand > max_significand) + { + new_significand = 0; + new_exponent_bits++; + } + } + + // Saturate to infinity if the exponent is larger than can be + // represented in the output type. This can only occur if the size + // of the exponent of the new type is not greater than the exponent + // of the old type. + if constexpr (other_n_exp_bits <= n_exp_bits) + { + constexpr int64_t inf_exp_bits = (1ll << other_n_exp_bits) - 1; + if (new_exponent_bits >= inf_exp_bits) + { + new_exponent_bits = inf_exp_bits; + new_significand = + other_has_infinity ? 0 : (1ul << other_float_t::n_significand_bits) - (other_has_nan ? 2 : 1); + } + } + } + + return other_float_t::from_bits(sign_bit, new_exponent_bits, new_significand); + } + + /// \brief Convert from a 32-bit floating point value + BITCAST_CONSTEXPR + float_t(const float& f) + { + // If this format exactly represents the binary32 format then get + // the bits from the provided float; otherwise get a binary32 + // representation and then convert to this format. + if constexpr (represents_binary32()) + m_data = float_support::get_bits(f); + else + m_data = static_cast<float_t<storage_t, n_exp_bits, has_nan, with_denorm, with_infinity>>( + static_cast<float_t<int32_t, 8, true, true, true>>(f)) + .m_data; + } + + /// \brief Cast to a 32-bit floating point value + BITCAST_CONSTEXPR operator float() const + { + // If this format exactly represents the binary32 format then return + // a float; otherwise get a binary32 representation and then return + // a float. + if constexpr (represents_binary32()) + return float_support::from_bits(m_data); + else + return static_cast<float>(this->operator float_t<int32_t, 8, true, true, true>()); + } + + /// \brief Return whether this type represents the IEEE754 binary32 + /// format + constexpr static inline bool represents_binary32() + { + return std::is_same_v<storage_t, int32_t> && n_exp_bits == 8 && has_nan && with_denorm && with_infinity; + } + + constexpr auto operator-() const + { + return from_bits(m_data ^ (1ll << (sizeof(storage_t) * 8 - 1))); + } + + constexpr bool is_subnormal() const + { + return exponent_bits() == 0 && significand() != 0; + } + + constexpr bool is_zero() const + { + return exponent_bits() == 0 && significand() == 0; + } + + constexpr bool is_nan() const + { + return has_nan && (exponent_bits() == (1ul << n_exponent_bits) - 1) && + ((with_infinity && significand()) || + (!with_infinity && significand() == (1ul << n_significand_bits) - 1)); + } + + constexpr bool is_infinity() const + { + return with_infinity && ((exponent_bits() == (1ul << n_exponent_bits) - 1) && !significand()); + } + + constexpr inline const storage_t& bits() const + { + return m_data; + } + + /// \brief Get the exponent + constexpr inline int64_t exponent() const + { + return std::max<int64_t>(exponent_bits(), 1ul) - exponent_bias; + } + + /// \brief Get the bits from the exponent + constexpr inline uint64_t exponent_bits() const + { + constexpr uint64_t mask = (1ul << n_exp_bits) - 1; + return (m_data >> n_significand_bits) & mask; + } + + constexpr inline uint64_t significand() const + { + return m_data & ((1ul << n_significand_bits) - 1); + } + + constexpr inline bool operator==(const float_t& other) const + { + return !is_nan() && !other.is_nan() && ((is_zero() && other.is_zero()) || bits() == other.bits()); + } + + constexpr inline float_t& operator+=(const float_t& rhs) + { + this->m_data = static_cast<float_t>(static_cast<float>(*this) + static_cast<float>(rhs)).bits(); + return *this; + } +}; + +// This should probably be exported so we can use it elsewhere +#undef BITCAST_CONSTEXPR + +namespace float_support +{ + +// Pre-C++23 these can't be computed as constexpr, so have to hardcode them + +template <int> +struct digits10; // floor(log10(2) * (digits - 1) +template <int> +struct max_digits10; // ceil(log10(2) * digits + 1) +template <int> +struct min_exponent10; // floor(log10(2) * min_exponent) +template <int> +struct max_exponent10; // floor(log10(2) * max_exponent) + +template <> +struct digits10<8> +{ + constexpr static inline int value = 2; +}; + +template <> +struct max_digits10<8> +{ + constexpr static inline int value = 4; +}; + +template <> +struct digits10<10> +{ + constexpr static inline int value = 2; +}; + +template <> +struct max_digits10<10> +{ + constexpr static inline int value = 5; +}; + +template <> +struct digits10<24> +{ + constexpr static inline int value = 6; +}; + +template <> +struct max_digits10<24> +{ + constexpr static inline int value = 9; +}; + +template <> +struct min_exponent10<-13> +{ + constexpr static inline int value = -3; +}; + +template <> +struct max_exponent10<16> +{ + constexpr static inline int value = 4; +}; + +template <> +struct min_exponent10<-125> +{ + constexpr static inline int value = -37; +}; + +template <> +struct max_exponent10<128> +{ + constexpr static inline int value = 38; +}; + +template <int d> +inline constexpr int digits10_v = digits10<d>::value; +template <int d> +inline constexpr int max_digits10_v = max_digits10<d>::value; + +template <int e> +inline constexpr int min_exponent10_v = min_exponent10<e>::value; + +template <int e> +inline constexpr int max_exponent10_v = max_exponent10<e>::value; + +} // namespace float_support + +} // namespace tosa + +namespace std +{ + +template <typename storage_t, size_t n_exp_bits, bool has_nan, bool has_denorm, bool has_inf> +struct is_floating_point<tosa::float_t<storage_t, n_exp_bits, has_nan, has_denorm, has_inf>> + : std::integral_constant<bool, true> +{}; + +template <typename storage_t, size_t n_exp_bits, bool has_nan, bool with_denorm, bool with_inf> +class numeric_limits<tosa::float_t<storage_t, n_exp_bits, has_nan, with_denorm, with_inf>> +{ + using this_float_t = tosa::float_t<storage_t, n_exp_bits, has_nan, with_denorm, with_inf>; + +public: + static constexpr bool is_specialized = true; + + static constexpr auto min() noexcept + { + return this_float_t::from_bits(false, 1, 0); + } + + static constexpr auto max() noexcept + { + return this_float_t::from_bits(false, (1 << this_float_t::n_exponent_bits) - 2, + (1 << this_float_t::n_significand_bits) - 1); + } + + static constexpr auto lowest() noexcept + { + return -max(); + } + + static constexpr int digits = this_float_t::n_significand_bits + 1; + static constexpr int digits10 = tosa::float_support::digits10_v<digits>; + static constexpr int max_digits10 = tosa::float_support::max_digits10_v<digits>; + + static constexpr bool is_signed = true; + static constexpr bool is_integer = false; + static constexpr bool is_exact = false; + static constexpr int radix = 2; + + static constexpr auto epsilon() noexcept + { + return this_float_t::from_bits(false, this_float_t::exponent_bias - this_float_t::n_significand_bits, 0); + } + + static constexpr auto round_error() noexcept + { + return this_float_t::from_bits(0, this_float_t::exponent_bias - 1, 0); + } + + static constexpr int min_exponent = (1 - this_float_t::exponent_bias) + 1; + static constexpr int min_exponent10 = tosa::float_support::min_exponent10_v<min_exponent>; + static constexpr int max_exponent = this_float_t::exponent_bias + 1; + static constexpr int max_exponent10 = tosa::float_support::max_exponent10_v<max_exponent>; + + static constexpr bool has_infinity = with_inf; + static constexpr bool has_quiet_NaN = has_nan; + static constexpr bool has_signaling_NaN = true; + static constexpr float_denorm_style has_denorm = with_denorm ? denorm_present : denorm_absent; + static constexpr bool has_denorm_loss = false; + + static constexpr auto infinity() noexcept + { + if constexpr (with_inf) + { + return this_float_t::from_bits(false, (1 << this_float_t::n_exponent_bits) - 1, 0); + } + else + { + return this_float_t::from_bits(false, 0, 0); + } + } + + static constexpr auto quiet_NaN() noexcept + { + return this_float_t::from_bits(false, (1 << this_float_t::n_exponent_bits) - 1, + 1 << (this_float_t::n_significand_bits - 1) | 1); + } + + static constexpr auto signaling_NaN() noexcept + { + return this_float_t::from_bits(false, (1 << this_float_t::n_exponent_bits) - 1, 1); + } + + static constexpr auto denorm_min() noexcept + { + return this_float_t::from_bits(false, 0, 1); + } + + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = false; + static constexpr bool is_modulo = false; + + static constexpr bool traps = false; + static constexpr bool tinyness_before = false; + static constexpr float_round_style round_style = round_to_nearest; +}; + +} // namespace std + +#endif // TOSA_FLOAT_UTILS_H_ diff --git a/include/numpy_utils.h b/include/numpy_utils.h index c64bc17..60cf77e 100644 --- a/include/numpy_utils.h +++ b/include/numpy_utils.h @@ -1,5 +1,5 @@ -// Copyright (c) 2020-2021, ARM Limited. +// Copyright (c) 2020-2023, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -24,6 +24,8 @@ #include <cstring> #include <vector> +#include "half.hpp" + class NumpyUtilities { public: @@ -35,31 +37,89 @@ public: FILE_TYPE_MISMATCH, HEADER_PARSE_ERROR, BUFFER_SIZE_MISMATCH, + DATA_TYPE_NOT_SUPPORTED, }; - static NPError readFromNpyFile(const char* filename, const uint32_t elems, float* databuf); - - static NPError readFromNpyFile(const char* filename, const uint32_t elems, int32_t* databuf); - - static NPError readFromNpyFile(const char* filename, const uint32_t elems, int64_t* databuf); - - static NPError readFromNpyFile(const char* filename, const uint32_t elems, bool* databuf); - - static NPError writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const bool* databuf); - - static NPError writeToNpyFile(const char* filename, const uint32_t elems, const bool* databuf); - - static NPError writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const int32_t* databuf); - - static NPError writeToNpyFile(const char* filename, const uint32_t elems, const int32_t* databuf); + template <typename T> + static const char* getDTypeString(bool& is_bool) + { + is_bool = false; + if (std::is_same<T, bool>::value) + { + is_bool = true; + return "'|b1'"; + } + if (std::is_same<T, uint8_t>::value) + { + return "'|u1'"; + } + if (std::is_same<T, int8_t>::value) + { + return "'|i1'"; + } + if (std::is_same<T, uint16_t>::value) + { + return "'<u2'"; + } + if (std::is_same<T, int16_t>::value) + { + return "'<i2'"; + } + if (std::is_same<T, int32_t>::value) + { + return "'<i4'"; + } + if (std::is_same<T, int64_t>::value) + { + return "'<i8'"; + } + if (std::is_same<T, float>::value) + { + return "'<f4'"; + } + if (std::is_same<T, double>::value) + { + return "'<f8'"; + } + if (std::is_same<T, half_float::half>::value) + { + return "'<f2'"; + } + assert(false && "unsupported Dtype"); + }; - static NPError writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const int64_t* databuf); + template <typename T> + static NPError writeToNpyFile(const char* filename, const uint32_t elems, const T* databuf) + { + std::vector<int32_t> shape = { static_cast<int32_t>(elems) }; + return writeToNpyFile(filename, shape, databuf); + } - static NPError writeToNpyFile(const char* filename, const uint32_t elems, const int64_t* databuf); + template <typename T> + static NPError writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const T* databuf) + { + bool is_bool; + const char* dtype_str = getDTypeString<T>(is_bool); + return writeToNpyFileCommon(filename, dtype_str, sizeof(T), shape, databuf, is_bool); + } - static NPError writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const float* databuf); + template <typename T> + static NPError readFromNpyFile(const char* filename, const uint32_t elems, T* databuf) + { + bool is_bool; + const char* dtype_str = getDTypeString<T>(is_bool); + return readFromNpyFileCommon(filename, dtype_str, sizeof(T), elems, databuf, is_bool); + } - static NPError writeToNpyFile(const char* filename, const uint32_t elems, const float* databuf); + template <typename D, typename S> + static void copyBufferByElement(D* dest_buf, S* src_buf, int num) + { + static_assert(sizeof(D) >= sizeof(S), "The size of dest_buf must be equal to or larger than that of src_buf"); + for (int i = 0; i < num; ++i) + { + dest_buf[i] = src_buf[i]; + } + } private: static NPError writeToNpyFileCommon(const char* filename, @@ -75,7 +135,11 @@ private: void* databuf, bool bool_translate); static NPError checkNpyHeader(FILE* infile, const uint32_t elems, const char* dtype_str); + static NPError getHeader(FILE* infile, bool& is_signed, int& bit_length, char& byte_order); static NPError writeNpyHeader(FILE* outfile, const std::vector<int32_t>& shape, const char* dtype_str); }; +template <> +NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, int32_t* databuf); + #endif // _TOSA_NUMPY_UTILS_H diff --git a/include/tosa_generated.h b/include/tosa_generated.h index 023825d..20f6993 100644 --- a/include/tosa_generated.h +++ b/include/tosa_generated.h @@ -6,6 +6,13 @@ #include "flatbuffers/flatbuffers.h" +// Ensure the included flatbuffers.h is the same version as when this file was +// generated, otherwise it may not be compatible. +static_assert(FLATBUFFERS_VERSION_MAJOR == 23 && + FLATBUFFERS_VERSION_MINOR == 5 && + FLATBUFFERS_VERSION_REVISION == 26, + "Non-compatible flatbuffers version included"); + namespace tosa { struct PoolAttribute; @@ -23,15 +30,6 @@ struct PadAttributeBuilder; struct AxisAttribute; struct AxisAttributeBuilder; -struct ReshapeAttribute; -struct ReshapeAttributeBuilder; - -struct SliceAttribute; -struct SliceAttributeBuilder; - -struct TileAttribute; -struct TileAttributeBuilder; - struct ResizeAttribute; struct ResizeAttributeBuilder; @@ -68,6 +66,15 @@ struct FullyConnectedAttributeBuilder; struct NegateAttribute; struct NegateAttributeBuilder; +struct CustomAttribute; +struct CustomAttributeBuilder; + +struct FFTAttribute; +struct FFTAttributeBuilder; + +struct RFFTAttribute; +struct RFFTAttributeBuilder; + struct Version; struct VersionBuilder; @@ -80,10 +87,13 @@ struct TosaOperatorBuilder; struct TosaBasicBlock; struct TosaBasicBlockBuilder; +struct TosaRegion; +struct TosaRegionBuilder; + struct TosaGraph; struct TosaGraphBuilder; -enum DType { +enum DType : uint32_t { DType_UNKNOWN = 0, DType_BOOL = 1, DType_UINT8 = 2, @@ -92,13 +102,18 @@ enum DType { DType_INT16 = 5, DType_INT32 = 6, DType_INT48 = 7, - DType_FLOAT = 8, + DType_FP32 = 8, DType_UINT16 = 9, + DType_FP16 = 10, + DType_BF16 = 11, + DType_SHAPE = 12, + DType_FP8E4M3 = 13, + DType_FP8E5M2 = 14, DType_MIN = DType_UNKNOWN, - DType_MAX = DType_UINT16 + DType_MAX = DType_FP8E5M2 }; -inline const DType (&EnumValuesDType())[10] { +inline const DType (&EnumValuesDType())[15] { static const DType values[] = { DType_UNKNOWN, DType_BOOL, @@ -108,14 +123,19 @@ inline const DType (&EnumValuesDType())[10] { DType_INT16, DType_INT32, DType_INT48, - DType_FLOAT, - DType_UINT16 + DType_FP32, + DType_UINT16, + DType_FP16, + DType_BF16, + DType_SHAPE, + DType_FP8E4M3, + DType_FP8E5M2 }; return values; } inline const char * const *EnumNamesDType() { - static const char * const names[11] = { + static const char * const names[16] = { "UNKNOWN", "BOOL", "UINT8", @@ -124,20 +144,25 @@ inline const char * const *EnumNamesDType() { "INT16", "INT32", "INT48", - "FLOAT", + "FP32", "UINT16", + "FP16", + "BF16", + "SHAPE", + "FP8E4M3", + "FP8E5M2", nullptr }; return names; } inline const char *EnumNameDType(DType e) { - if (flatbuffers::IsOutRange(e, DType_UNKNOWN, DType_UINT16)) return ""; + if (::flatbuffers::IsOutRange(e, DType_UNKNOWN, DType_FP8E5M2)) return ""; const size_t index = static_cast<size_t>(e); return EnumNamesDType()[index]; } -enum ResizeMode { +enum ResizeMode : uint32_t { ResizeMode_UNKNOWN = 0, ResizeMode_NEAREST = 1, ResizeMode_BILINEAR = 2, @@ -165,12 +190,12 @@ inline const char * const *EnumNamesResizeMode() { } inline const char *EnumNameResizeMode(ResizeMode e) { - if (flatbuffers::IsOutRange(e, ResizeMode_UNKNOWN, ResizeMode_BILINEAR)) return ""; + if (::flatbuffers::IsOutRange(e, ResizeMode_UNKNOWN, ResizeMode_BILINEAR)) return ""; const size_t index = static_cast<size_t>(e); return EnumNamesResizeMode()[index]; } -enum Op { +enum Op : uint32_t { Op_UNKNOWN = 0, Op_ARGMAX = 1, Op_AVG_POOL2D = 2, @@ -240,11 +265,23 @@ enum Op { Op_CUSTOM = 66, Op_COND_IF = 67, Op_WHILE_LOOP = 68, + Op_FFT2D = 69, + Op_RFFT2D = 70, + Op_ERF = 71, + Op_DIM = 72, + Op_CONST_SHAPE = 73, + Op_CONCAT_SHAPE = 74, + Op_ADD_SHAPE = 75, + Op_SUB_SHAPE = 76, + Op_MUL_SHAPE = 77, + Op_DIV_SHAPE = 78, + Op_COS = 79, + Op_SIN = 80, Op_MIN = Op_UNKNOWN, - Op_MAX = Op_WHILE_LOOP + Op_MAX = Op_SIN }; -inline const Op (&EnumValuesOp())[69] { +inline const Op (&EnumValuesOp())[81] { static const Op values[] = { Op_UNKNOWN, Op_ARGMAX, @@ -314,13 +351,25 @@ inline const Op (&EnumValuesOp())[69] { Op_IDENTITY, Op_CUSTOM, Op_COND_IF, - Op_WHILE_LOOP + Op_WHILE_LOOP, + Op_FFT2D, + Op_RFFT2D, + Op_ERF, + Op_DIM, + Op_CONST_SHAPE, + Op_CONCAT_SHAPE, + Op_ADD_SHAPE, + Op_SUB_SHAPE, + Op_MUL_SHAPE, + Op_DIV_SHAPE, + Op_COS, + Op_SIN }; return values; } inline const char * const *EnumNamesOp() { - static const char * const names[70] = { + static const char * const names[82] = { "UNKNOWN", "ARGMAX", "AVG_POOL2D", @@ -390,41 +439,53 @@ inline const char * const *EnumNamesOp() { "CUSTOM", "COND_IF", "WHILE_LOOP", + "FFT2D", + "RFFT2D", + "ERF", + "DIM", + "CONST_SHAPE", + "CONCAT_SHAPE", + "ADD_SHAPE", + "SUB_SHAPE", + "MUL_SHAPE", + "DIV_SHAPE", + "COS", + "SIN", nullptr }; return names; } inline const char *EnumNameOp(Op e) { - if (flatbuffers::IsOutRange(e, Op_UNKNOWN, Op_WHILE_LOOP)) return ""; + if (::flatbuffers::IsOutRange(e, Op_UNKNOWN, Op_SIN)) return ""; const size_t index = static_cast<size_t>(e); return EnumNamesOp()[index]; } -enum Attribute { +enum Attribute : uint8_t { Attribute_NONE = 0, Attribute_PoolAttribute = 1, Attribute_ConvAttribute = 2, Attribute_TransposeConvAttribute = 3, Attribute_PadAttribute = 4, Attribute_AxisAttribute = 5, - Attribute_ReshapeAttribute = 6, - Attribute_SliceAttribute = 7, - Attribute_TileAttribute = 8, - Attribute_ResizeAttribute = 9, - Attribute_ClampAttribute = 10, - Attribute_RescaleAttribute = 11, - Attribute_MulAttribute = 12, - Attribute_ArithmeticRightShiftAttribute = 13, - Attribute_CondIfAttribute = 14, - Attribute_WhileLoopAttribute = 15, - Attribute_TransposeAttribute = 16, - Attribute_TableAttribute = 17, - Attribute_MatMulAttribute = 18, - Attribute_FullyConnectedAttribute = 19, - Attribute_NegateAttribute = 20, + Attribute_ResizeAttribute = 6, + Attribute_ClampAttribute = 7, + Attribute_RescaleAttribute = 8, + Attribute_MulAttribute = 9, + Attribute_ArithmeticRightShiftAttribute = 10, + Attribute_CondIfAttribute = 11, + Attribute_WhileLoopAttribute = 12, + Attribute_TransposeAttribute = 13, + Attribute_TableAttribute = 14, + Attribute_MatMulAttribute = 15, + Attribute_FullyConnectedAttribute = 16, + Attribute_NegateAttribute = 17, + Attribute_CustomAttribute = 18, + Attribute_FFTAttribute = 19, + Attribute_RFFTAttribute = 20, Attribute_MIN = Attribute_NONE, - Attribute_MAX = Attribute_NegateAttribute + Attribute_MAX = Attribute_RFFTAttribute }; inline const Attribute (&EnumValuesAttribute())[21] { @@ -435,9 +496,6 @@ inline const Attribute (&EnumValuesAttribute())[21] { Attribute_TransposeConvAttribute, Attribute_PadAttribute, Attribute_AxisAttribute, - Attribute_ReshapeAttribute, - Attribute_SliceAttribute, - Attribute_TileAttribute, Attribute_ResizeAttribute, Attribute_ClampAttribute, Attribute_RescaleAttribute, @@ -449,7 +507,10 @@ inline const Attribute (&EnumValuesAttribute())[21] { Attribute_TableAttribute, Attribute_MatMulAttribute, Attribute_FullyConnectedAttribute, - Attribute_NegateAttribute + Attribute_NegateAttribute, + Attribute_CustomAttribute, + Attribute_FFTAttribute, + Attribute_RFFTAttribute }; return values; } @@ -462,9 +523,6 @@ inline const char * const *EnumNamesAttribute() { "TransposeConvAttribute", "PadAttribute", "AxisAttribute", - "ReshapeAttribute", - "SliceAttribute", - "TileAttribute", "ResizeAttribute", "ClampAttribute", "RescaleAttribute", @@ -477,13 +535,16 @@ inline const char * const *EnumNamesAttribute() { "MatMulAttribute", "FullyConnectedAttribute", "NegateAttribute", + "CustomAttribute", + "FFTAttribute", + "RFFTAttribute", nullptr }; return names; } inline const char *EnumNameAttribute(Attribute e) { - if (flatbuffers::IsOutRange(e, Attribute_NONE, Attribute_NegateAttribute)) return ""; + if (::flatbuffers::IsOutRange(e, Attribute_NONE, Attribute_RFFTAttribute)) return ""; const size_t index = static_cast<size_t>(e); return EnumNamesAttribute()[index]; } @@ -512,18 +573,6 @@ template<> struct AttributeTraits<tosa::AxisAttribute> { static const Attribute enum_value = Attribute_AxisAttribute; }; -template<> struct AttributeTraits<tosa::ReshapeAttribute> { - static const Attribute enum_value = Attribute_ReshapeAttribute; -}; - -template<> struct AttributeTraits<tosa::SliceAttribute> { - static const Attribute enum_value = Attribute_SliceAttribute; -}; - -template<> struct AttributeTraits<tosa::TileAttribute> { - static const Attribute enum_value = Attribute_TileAttribute; -}; - template<> struct AttributeTraits<tosa::ResizeAttribute> { static const Attribute enum_value = Attribute_ResizeAttribute; }; @@ -572,26 +621,39 @@ template<> struct AttributeTraits<tosa::NegateAttribute> { static const Attribute enum_value = Attribute_NegateAttribute; }; -bool VerifyAttribute(flatbuffers::Verifier &verifier, const void *obj, Attribute type); -bool VerifyAttributeVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector<flatbuffers::Offset<void>> *values, const flatbuffers::Vector<uint8_t> *types); +template<> struct AttributeTraits<tosa::CustomAttribute> { + static const Attribute enum_value = Attribute_CustomAttribute; +}; + +template<> struct AttributeTraits<tosa::FFTAttribute> { + static const Attribute enum_value = Attribute_FFTAttribute; +}; + +template<> struct AttributeTraits<tosa::RFFTAttribute> { + static const Attribute enum_value = Attribute_RFFTAttribute; +}; + +bool VerifyAttribute(::flatbuffers::Verifier &verifier, const void *obj, Attribute type); +bool VerifyAttributeVector(::flatbuffers::Verifier &verifier, const ::flatbuffers::Vector<::flatbuffers::Offset<void>> *values, const ::flatbuffers::Vector<uint8_t> *types); -struct PoolAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct PoolAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef PoolAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_PAD = 4, VT_KERNEL = 6, VT_STRIDE = 8, VT_INPUT_ZP = 10, - VT_OUTPUT_ZP = 12 + VT_OUTPUT_ZP = 12, + VT_ACC_TYPE = 14 }; - const flatbuffers::Vector<int32_t> *pad() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_PAD); + const ::flatbuffers::Vector<int32_t> *pad() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_PAD); } - const flatbuffers::Vector<int32_t> *kernel() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_KERNEL); + const ::flatbuffers::Vector<int32_t> *kernel() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_KERNEL); } - const flatbuffers::Vector<int32_t> *stride() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_STRIDE); + const ::flatbuffers::Vector<int32_t> *stride() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_STRIDE); } int32_t input_zp() const { return GetField<int32_t>(VT_INPUT_ZP, 0); @@ -599,7 +661,10 @@ struct PoolAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t output_zp() const { return GetField<int32_t>(VT_OUTPUT_ZP, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + tosa::DType acc_type() const { + return static_cast<tosa::DType>(GetField<uint32_t>(VT_ACC_TYPE, 0)); + } + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_PAD) && verifier.VerifyVector(pad()) && @@ -607,23 +672,24 @@ struct PoolAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { verifier.VerifyVector(kernel()) && VerifyOffset(verifier, VT_STRIDE) && verifier.VerifyVector(stride()) && - VerifyField<int32_t>(verifier, VT_INPUT_ZP) && - VerifyField<int32_t>(verifier, VT_OUTPUT_ZP) && + VerifyField<int32_t>(verifier, VT_INPUT_ZP, 4) && + VerifyField<int32_t>(verifier, VT_OUTPUT_ZP, 4) && + VerifyField<uint32_t>(verifier, VT_ACC_TYPE, 4) && verifier.EndTable(); } }; struct PoolAttributeBuilder { typedef PoolAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_pad(flatbuffers::Offset<flatbuffers::Vector<int32_t>> pad) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_pad(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> pad) { fbb_.AddOffset(PoolAttribute::VT_PAD, pad); } - void add_kernel(flatbuffers::Offset<flatbuffers::Vector<int32_t>> kernel) { + void add_kernel(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> kernel) { fbb_.AddOffset(PoolAttribute::VT_KERNEL, kernel); } - void add_stride(flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride) { + void add_stride(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> stride) { fbb_.AddOffset(PoolAttribute::VT_STRIDE, stride); } void add_input_zp(int32_t input_zp) { @@ -632,26 +698,30 @@ struct PoolAttributeBuilder { void add_output_zp(int32_t output_zp) { fbb_.AddElement<int32_t>(PoolAttribute::VT_OUTPUT_ZP, output_zp, 0); } - explicit PoolAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + void add_acc_type(tosa::DType acc_type) { + fbb_.AddElement<uint32_t>(PoolAttribute::VT_ACC_TYPE, static_cast<uint32_t>(acc_type), 0); + } + explicit PoolAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - PoolAttributeBuilder &operator=(const PoolAttributeBuilder &); - flatbuffers::Offset<PoolAttribute> Finish() { + ::flatbuffers::Offset<PoolAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<PoolAttribute>(end); + auto o = ::flatbuffers::Offset<PoolAttribute>(end); return o; } }; -inline flatbuffers::Offset<PoolAttribute> CreatePoolAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> pad = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> kernel = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride = 0, +inline ::flatbuffers::Offset<PoolAttribute> CreatePoolAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> pad = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> kernel = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> stride = 0, int32_t input_zp = 0, - int32_t output_zp = 0) { + int32_t output_zp = 0, + tosa::DType acc_type = tosa::DType_UNKNOWN) { PoolAttributeBuilder builder_(_fbb); + builder_.add_acc_type(acc_type); builder_.add_output_zp(output_zp); builder_.add_input_zp(input_zp); builder_.add_stride(stride); @@ -660,13 +730,14 @@ inline flatbuffers::Offset<PoolAttribute> CreatePoolAttribute( return builder_.Finish(); } -inline flatbuffers::Offset<PoolAttribute> CreatePoolAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<PoolAttribute> CreatePoolAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const std::vector<int32_t> *pad = nullptr, const std::vector<int32_t> *kernel = nullptr, const std::vector<int32_t> *stride = nullptr, int32_t input_zp = 0, - int32_t output_zp = 0) { + int32_t output_zp = 0, + tosa::DType acc_type = tosa::DType_UNKNOWN) { auto pad__ = pad ? _fbb.CreateVector<int32_t>(*pad) : 0; auto kernel__ = kernel ? _fbb.CreateVector<int32_t>(*kernel) : 0; auto stride__ = stride ? _fbb.CreateVector<int32_t>(*stride) : 0; @@ -676,26 +747,29 @@ inline flatbuffers::Offset<PoolAttribute> CreatePoolAttributeDirect( kernel__, stride__, input_zp, - output_zp); + output_zp, + acc_type); } -struct ConvAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct ConvAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef ConvAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_PAD = 4, VT_STRIDE = 6, VT_DILATION = 8, VT_INPUT_ZP = 10, - VT_WEIGHT_ZP = 12 + VT_WEIGHT_ZP = 12, + VT_LOCAL_BOUND = 14, + VT_ACC_TYPE = 16 }; - const flatbuffers::Vector<int32_t> *pad() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_PAD); + const ::flatbuffers::Vector<int32_t> *pad() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_PAD); } - const flatbuffers::Vector<int32_t> *stride() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_STRIDE); + const ::flatbuffers::Vector<int32_t> *stride() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_STRIDE); } - const flatbuffers::Vector<int32_t> *dilation() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_DILATION); + const ::flatbuffers::Vector<int32_t> *dilation() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_DILATION); } int32_t input_zp() const { return GetField<int32_t>(VT_INPUT_ZP, 0); @@ -703,7 +777,13 @@ struct ConvAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t weight_zp() const { return GetField<int32_t>(VT_WEIGHT_ZP, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool local_bound() const { + return GetField<uint8_t>(VT_LOCAL_BOUND, 0) != 0; + } + tosa::DType acc_type() const { + return static_cast<tosa::DType>(GetField<uint32_t>(VT_ACC_TYPE, 0)); + } + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_PAD) && verifier.VerifyVector(pad()) && @@ -711,23 +791,25 @@ struct ConvAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { verifier.VerifyVector(stride()) && VerifyOffset(verifier, VT_DILATION) && verifier.VerifyVector(dilation()) && - VerifyField<int32_t>(verifier, VT_INPUT_ZP) && - VerifyField<int32_t>(verifier, VT_WEIGHT_ZP) && + VerifyField<int32_t>(verifier, VT_INPUT_ZP, 4) && + VerifyField<int32_t>(verifier, VT_WEIGHT_ZP, 4) && + VerifyField<uint8_t>(verifier, VT_LOCAL_BOUND, 1) && + VerifyField<uint32_t>(verifier, VT_ACC_TYPE, 4) && verifier.EndTable(); } }; struct ConvAttributeBuilder { typedef ConvAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_pad(flatbuffers::Offset<flatbuffers::Vector<int32_t>> pad) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_pad(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> pad) { fbb_.AddOffset(ConvAttribute::VT_PAD, pad); } - void add_stride(flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride) { + void add_stride(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> stride) { fbb_.AddOffset(ConvAttribute::VT_STRIDE, stride); } - void add_dilation(flatbuffers::Offset<flatbuffers::Vector<int32_t>> dilation) { + void add_dilation(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> dilation) { fbb_.AddOffset(ConvAttribute::VT_DILATION, dilation); } void add_input_zp(int32_t input_zp) { @@ -736,41 +818,52 @@ struct ConvAttributeBuilder { void add_weight_zp(int32_t weight_zp) { fbb_.AddElement<int32_t>(ConvAttribute::VT_WEIGHT_ZP, weight_zp, 0); } - explicit ConvAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + void add_local_bound(bool local_bound) { + fbb_.AddElement<uint8_t>(ConvAttribute::VT_LOCAL_BOUND, static_cast<uint8_t>(local_bound), 0); + } + void add_acc_type(tosa::DType acc_type) { + fbb_.AddElement<uint32_t>(ConvAttribute::VT_ACC_TYPE, static_cast<uint32_t>(acc_type), 0); + } + explicit ConvAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - ConvAttributeBuilder &operator=(const ConvAttributeBuilder &); - flatbuffers::Offset<ConvAttribute> Finish() { + ::flatbuffers::Offset<ConvAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<ConvAttribute>(end); + auto o = ::flatbuffers::Offset<ConvAttribute>(end); return o; } }; -inline flatbuffers::Offset<ConvAttribute> CreateConvAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> pad = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> dilation = 0, +inline ::flatbuffers::Offset<ConvAttribute> CreateConvAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> pad = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> stride = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> dilation = 0, int32_t input_zp = 0, - int32_t weight_zp = 0) { + int32_t weight_zp = 0, + bool local_bound = false, + tosa::DType acc_type = tosa::DType_UNKNOWN) { ConvAttributeBuilder builder_(_fbb); + builder_.add_acc_type(acc_type); builder_.add_weight_zp(weight_zp); builder_.add_input_zp(input_zp); builder_.add_dilation(dilation); builder_.add_stride(stride); builder_.add_pad(pad); + builder_.add_local_bound(local_bound); return builder_.Finish(); } -inline flatbuffers::Offset<ConvAttribute> CreateConvAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<ConvAttribute> CreateConvAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const std::vector<int32_t> *pad = nullptr, const std::vector<int32_t> *stride = nullptr, const std::vector<int32_t> *dilation = nullptr, int32_t input_zp = 0, - int32_t weight_zp = 0) { + int32_t weight_zp = 0, + bool local_bound = false, + tosa::DType acc_type = tosa::DType_UNKNOWN) { auto pad__ = pad ? _fbb.CreateVector<int32_t>(*pad) : 0; auto stride__ = stride ? _fbb.CreateVector<int32_t>(*stride) : 0; auto dilation__ = dilation ? _fbb.CreateVector<int32_t>(*dilation) : 0; @@ -780,26 +873,30 @@ inline flatbuffers::Offset<ConvAttribute> CreateConvAttributeDirect( stride__, dilation__, input_zp, - weight_zp); + weight_zp, + local_bound, + acc_type); } -struct TransposeConvAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TransposeConvAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TransposeConvAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_OUT_PAD = 4, VT_STRIDE = 6, VT_OUTPUT_SHAPE = 8, VT_INPUT_ZP = 10, - VT_WEIGHT_ZP = 12 + VT_WEIGHT_ZP = 12, + VT_LOCAL_BOUND = 14, + VT_ACC_TYPE = 16 }; - const flatbuffers::Vector<int32_t> *out_pad() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_OUT_PAD); + const ::flatbuffers::Vector<int32_t> *out_pad() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_OUT_PAD); } - const flatbuffers::Vector<int32_t> *stride() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_STRIDE); + const ::flatbuffers::Vector<int32_t> *stride() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_STRIDE); } - const flatbuffers::Vector<int32_t> *output_shape() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_OUTPUT_SHAPE); + const ::flatbuffers::Vector<int32_t> *output_shape() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_OUTPUT_SHAPE); } int32_t input_zp() const { return GetField<int32_t>(VT_INPUT_ZP, 0); @@ -807,7 +904,13 @@ struct TransposeConvAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Tab int32_t weight_zp() const { return GetField<int32_t>(VT_WEIGHT_ZP, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool local_bound() const { + return GetField<uint8_t>(VT_LOCAL_BOUND, 0) != 0; + } + tosa::DType acc_type() const { + return static_cast<tosa::DType>(GetField<uint32_t>(VT_ACC_TYPE, 0)); + } + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_OUT_PAD) && verifier.VerifyVector(out_pad()) && @@ -815,23 +918,25 @@ struct TransposeConvAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Tab verifier.VerifyVector(stride()) && VerifyOffset(verifier, VT_OUTPUT_SHAPE) && verifier.VerifyVector(output_shape()) && - VerifyField<int32_t>(verifier, VT_INPUT_ZP) && - VerifyField<int32_t>(verifier, VT_WEIGHT_ZP) && + VerifyField<int32_t>(verifier, VT_INPUT_ZP, 4) && + VerifyField<int32_t>(verifier, VT_WEIGHT_ZP, 4) && + VerifyField<uint8_t>(verifier, VT_LOCAL_BOUND, 1) && + VerifyField<uint32_t>(verifier, VT_ACC_TYPE, 4) && verifier.EndTable(); } }; struct TransposeConvAttributeBuilder { typedef TransposeConvAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_out_pad(flatbuffers::Offset<flatbuffers::Vector<int32_t>> out_pad) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_out_pad(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> out_pad) { fbb_.AddOffset(TransposeConvAttribute::VT_OUT_PAD, out_pad); } - void add_stride(flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride) { + void add_stride(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> stride) { fbb_.AddOffset(TransposeConvAttribute::VT_STRIDE, stride); } - void add_output_shape(flatbuffers::Offset<flatbuffers::Vector<int32_t>> output_shape) { + void add_output_shape(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> output_shape) { fbb_.AddOffset(TransposeConvAttribute::VT_OUTPUT_SHAPE, output_shape); } void add_input_zp(int32_t input_zp) { @@ -840,41 +945,52 @@ struct TransposeConvAttributeBuilder { void add_weight_zp(int32_t weight_zp) { fbb_.AddElement<int32_t>(TransposeConvAttribute::VT_WEIGHT_ZP, weight_zp, 0); } - explicit TransposeConvAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + void add_local_bound(bool local_bound) { + fbb_.AddElement<uint8_t>(TransposeConvAttribute::VT_LOCAL_BOUND, static_cast<uint8_t>(local_bound), 0); + } + void add_acc_type(tosa::DType acc_type) { + fbb_.AddElement<uint32_t>(TransposeConvAttribute::VT_ACC_TYPE, static_cast<uint32_t>(acc_type), 0); + } + explicit TransposeConvAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TransposeConvAttributeBuilder &operator=(const TransposeConvAttributeBuilder &); - flatbuffers::Offset<TransposeConvAttribute> Finish() { + ::flatbuffers::Offset<TransposeConvAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TransposeConvAttribute>(end); + auto o = ::flatbuffers::Offset<TransposeConvAttribute>(end); return o; } }; -inline flatbuffers::Offset<TransposeConvAttribute> CreateTransposeConvAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> out_pad = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> output_shape = 0, +inline ::flatbuffers::Offset<TransposeConvAttribute> CreateTransposeConvAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> out_pad = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> stride = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> output_shape = 0, int32_t input_zp = 0, - int32_t weight_zp = 0) { + int32_t weight_zp = 0, + bool local_bound = false, + tosa::DType acc_type = tosa::DType_UNKNOWN) { TransposeConvAttributeBuilder builder_(_fbb); + builder_.add_acc_type(acc_type); builder_.add_weight_zp(weight_zp); builder_.add_input_zp(input_zp); builder_.add_output_shape(output_shape); builder_.add_stride(stride); builder_.add_out_pad(out_pad); + builder_.add_local_bound(local_bound); return builder_.Finish(); } -inline flatbuffers::Offset<TransposeConvAttribute> CreateTransposeConvAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TransposeConvAttribute> CreateTransposeConvAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const std::vector<int32_t> *out_pad = nullptr, const std::vector<int32_t> *stride = nullptr, const std::vector<int32_t> *output_shape = nullptr, int32_t input_zp = 0, - int32_t weight_zp = 0) { + int32_t weight_zp = 0, + bool local_bound = false, + tosa::DType acc_type = tosa::DType_UNKNOWN) { auto out_pad__ = out_pad ? _fbb.CreateVector<int32_t>(*out_pad) : 0; auto stride__ = stride ? _fbb.CreateVector<int32_t>(*stride) : 0; auto output_shape__ = output_shape ? _fbb.CreateVector<int32_t>(*output_shape) : 0; @@ -884,86 +1000,64 @@ inline flatbuffers::Offset<TransposeConvAttribute> CreateTransposeConvAttributeD stride__, output_shape__, input_zp, - weight_zp); + weight_zp, + local_bound, + acc_type); } -struct PadAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct PadAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef PadAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_PADDING = 4, - VT_PAD_CONST_INT = 6, - VT_PAD_CONST_FP = 8 + VT_PAD_CONST = 4 }; - const flatbuffers::Vector<int32_t> *padding() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_PADDING); + const ::flatbuffers::Vector<uint8_t> *pad_const() const { + return GetPointer<const ::flatbuffers::Vector<uint8_t> *>(VT_PAD_CONST); } - int32_t pad_const_int() const { - return GetField<int32_t>(VT_PAD_CONST_INT, 0); - } - float pad_const_fp() const { - return GetField<float>(VT_PAD_CONST_FP, 0.0f); - } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_PADDING) && - verifier.VerifyVector(padding()) && - VerifyField<int32_t>(verifier, VT_PAD_CONST_INT) && - VerifyField<float>(verifier, VT_PAD_CONST_FP) && + VerifyOffset(verifier, VT_PAD_CONST) && + verifier.VerifyVector(pad_const()) && verifier.EndTable(); } }; struct PadAttributeBuilder { typedef PadAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_padding(flatbuffers::Offset<flatbuffers::Vector<int32_t>> padding) { - fbb_.AddOffset(PadAttribute::VT_PADDING, padding); - } - void add_pad_const_int(int32_t pad_const_int) { - fbb_.AddElement<int32_t>(PadAttribute::VT_PAD_CONST_INT, pad_const_int, 0); + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_pad_const(::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> pad_const) { + fbb_.AddOffset(PadAttribute::VT_PAD_CONST, pad_const); } - void add_pad_const_fp(float pad_const_fp) { - fbb_.AddElement<float>(PadAttribute::VT_PAD_CONST_FP, pad_const_fp, 0.0f); - } - explicit PadAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit PadAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - PadAttributeBuilder &operator=(const PadAttributeBuilder &); - flatbuffers::Offset<PadAttribute> Finish() { + ::flatbuffers::Offset<PadAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<PadAttribute>(end); + auto o = ::flatbuffers::Offset<PadAttribute>(end); return o; } }; -inline flatbuffers::Offset<PadAttribute> CreatePadAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> padding = 0, - int32_t pad_const_int = 0, - float pad_const_fp = 0.0f) { +inline ::flatbuffers::Offset<PadAttribute> CreatePadAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> pad_const = 0) { PadAttributeBuilder builder_(_fbb); - builder_.add_pad_const_fp(pad_const_fp); - builder_.add_pad_const_int(pad_const_int); - builder_.add_padding(padding); + builder_.add_pad_const(pad_const); return builder_.Finish(); } -inline flatbuffers::Offset<PadAttribute> CreatePadAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const std::vector<int32_t> *padding = nullptr, - int32_t pad_const_int = 0, - float pad_const_fp = 0.0f) { - auto padding__ = padding ? _fbb.CreateVector<int32_t>(*padding) : 0; +inline ::flatbuffers::Offset<PadAttribute> CreatePadAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const std::vector<uint8_t> *pad_const = nullptr) { + if (pad_const) { _fbb.ForceVectorAlignment(pad_const->size(), sizeof(uint8_t), 8); } + auto pad_const__ = pad_const ? _fbb.CreateVector<uint8_t>(*pad_const) : 0; return tosa::CreatePadAttribute( _fbb, - padding__, - pad_const_int, - pad_const_fp); + pad_const__); } -struct AxisAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct AxisAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef AxisAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_AXIS = 4 @@ -971,424 +1065,207 @@ struct AxisAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t axis() const { return GetField<int32_t>(VT_AXIS, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_AXIS) && + VerifyField<int32_t>(verifier, VT_AXIS, 4) && verifier.EndTable(); } }; struct AxisAttributeBuilder { typedef AxisAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_axis(int32_t axis) { fbb_.AddElement<int32_t>(AxisAttribute::VT_AXIS, axis, 0); } - explicit AxisAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit AxisAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - AxisAttributeBuilder &operator=(const AxisAttributeBuilder &); - flatbuffers::Offset<AxisAttribute> Finish() { + ::flatbuffers::Offset<AxisAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<AxisAttribute>(end); + auto o = ::flatbuffers::Offset<AxisAttribute>(end); return o; } }; -inline flatbuffers::Offset<AxisAttribute> CreateAxisAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<AxisAttribute> CreateAxisAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, int32_t axis = 0) { AxisAttributeBuilder builder_(_fbb); builder_.add_axis(axis); return builder_.Finish(); } -struct ReshapeAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef ReshapeAttributeBuilder Builder; - enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_NEW_SHAPE = 4 - }; - const flatbuffers::Vector<int32_t> *new_shape() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_NEW_SHAPE); - } - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_NEW_SHAPE) && - verifier.VerifyVector(new_shape()) && - verifier.EndTable(); - } -}; - -struct ReshapeAttributeBuilder { - typedef ReshapeAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_new_shape(flatbuffers::Offset<flatbuffers::Vector<int32_t>> new_shape) { - fbb_.AddOffset(ReshapeAttribute::VT_NEW_SHAPE, new_shape); - } - explicit ReshapeAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - ReshapeAttributeBuilder &operator=(const ReshapeAttributeBuilder &); - flatbuffers::Offset<ReshapeAttribute> Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<ReshapeAttribute>(end); - return o; - } -}; - -inline flatbuffers::Offset<ReshapeAttribute> CreateReshapeAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> new_shape = 0) { - ReshapeAttributeBuilder builder_(_fbb); - builder_.add_new_shape(new_shape); - return builder_.Finish(); -} - -inline flatbuffers::Offset<ReshapeAttribute> CreateReshapeAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const std::vector<int32_t> *new_shape = nullptr) { - auto new_shape__ = new_shape ? _fbb.CreateVector<int32_t>(*new_shape) : 0; - return tosa::CreateReshapeAttribute( - _fbb, - new_shape__); -} - -struct SliceAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef SliceAttributeBuilder Builder; - enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_START = 4, - VT_SIZE = 6 - }; - const flatbuffers::Vector<int32_t> *start() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_START); - } - const flatbuffers::Vector<int32_t> *size() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_SIZE); - } - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_START) && - verifier.VerifyVector(start()) && - VerifyOffset(verifier, VT_SIZE) && - verifier.VerifyVector(size()) && - verifier.EndTable(); - } -}; - -struct SliceAttributeBuilder { - typedef SliceAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_start(flatbuffers::Offset<flatbuffers::Vector<int32_t>> start) { - fbb_.AddOffset(SliceAttribute::VT_START, start); - } - void add_size(flatbuffers::Offset<flatbuffers::Vector<int32_t>> size) { - fbb_.AddOffset(SliceAttribute::VT_SIZE, size); - } - explicit SliceAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - SliceAttributeBuilder &operator=(const SliceAttributeBuilder &); - flatbuffers::Offset<SliceAttribute> Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<SliceAttribute>(end); - return o; - } -}; - -inline flatbuffers::Offset<SliceAttribute> CreateSliceAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> start = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> size = 0) { - SliceAttributeBuilder builder_(_fbb); - builder_.add_size(size); - builder_.add_start(start); - return builder_.Finish(); -} - -inline flatbuffers::Offset<SliceAttribute> CreateSliceAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const std::vector<int32_t> *start = nullptr, - const std::vector<int32_t> *size = nullptr) { - auto start__ = start ? _fbb.CreateVector<int32_t>(*start) : 0; - auto size__ = size ? _fbb.CreateVector<int32_t>(*size) : 0; - return tosa::CreateSliceAttribute( - _fbb, - start__, - size__); -} - -struct TileAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef TileAttributeBuilder Builder; - enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_MULTIPLES = 4 - }; - const flatbuffers::Vector<int32_t> *multiples() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_MULTIPLES); - } - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_MULTIPLES) && - verifier.VerifyVector(multiples()) && - verifier.EndTable(); - } -}; - -struct TileAttributeBuilder { - typedef TileAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_multiples(flatbuffers::Offset<flatbuffers::Vector<int32_t>> multiples) { - fbb_.AddOffset(TileAttribute::VT_MULTIPLES, multiples); - } - explicit TileAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - TileAttributeBuilder &operator=(const TileAttributeBuilder &); - flatbuffers::Offset<TileAttribute> Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TileAttribute>(end); - return o; - } -}; - -inline flatbuffers::Offset<TileAttribute> CreateTileAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> multiples = 0) { - TileAttributeBuilder builder_(_fbb); - builder_.add_multiples(multiples); - return builder_.Finish(); -} - -inline flatbuffers::Offset<TileAttribute> CreateTileAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const std::vector<int32_t> *multiples = nullptr) { - auto multiples__ = multiples ? _fbb.CreateVector<int32_t>(*multiples) : 0; - return tosa::CreateTileAttribute( - _fbb, - multiples__); -} - -struct ResizeAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct ResizeAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef ResizeAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_OUTPUT_SIZE = 4, - VT_STRIDE = 6, - VT_OFFSET = 8, - VT_SHIFT = 10, - VT_STRIDE_FP = 12, - VT_OFFSET_FP = 14, - VT_MODE = 16 + VT_SCALE = 4, + VT_OFFSET = 6, + VT_BORDER = 8, + VT_MODE = 10 }; - const flatbuffers::Vector<int32_t> *output_size() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_OUTPUT_SIZE); + const ::flatbuffers::Vector<int16_t> *scale() const { + return GetPointer<const ::flatbuffers::Vector<int16_t> *>(VT_SCALE); } - const flatbuffers::Vector<int32_t> *stride() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_STRIDE); + const ::flatbuffers::Vector<int16_t> *offset() const { + return GetPointer<const ::flatbuffers::Vector<int16_t> *>(VT_OFFSET); } - const flatbuffers::Vector<int32_t> *offset() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_OFFSET); - } - int32_t shift() const { - return GetField<int32_t>(VT_SHIFT, 0); - } - const flatbuffers::Vector<float> *stride_fp() const { - return GetPointer<const flatbuffers::Vector<float> *>(VT_STRIDE_FP); - } - const flatbuffers::Vector<float> *offset_fp() const { - return GetPointer<const flatbuffers::Vector<float> *>(VT_OFFSET_FP); + const ::flatbuffers::Vector<int16_t> *border() const { + return GetPointer<const ::flatbuffers::Vector<int16_t> *>(VT_BORDER); } tosa::ResizeMode mode() const { return static_cast<tosa::ResizeMode>(GetField<uint32_t>(VT_MODE, 0)); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_OUTPUT_SIZE) && - verifier.VerifyVector(output_size()) && - VerifyOffset(verifier, VT_STRIDE) && - verifier.VerifyVector(stride()) && + VerifyOffset(verifier, VT_SCALE) && + verifier.VerifyVector(scale()) && VerifyOffset(verifier, VT_OFFSET) && verifier.VerifyVector(offset()) && - VerifyField<int32_t>(verifier, VT_SHIFT) && - VerifyOffset(verifier, VT_STRIDE_FP) && - verifier.VerifyVector(stride_fp()) && - VerifyOffset(verifier, VT_OFFSET_FP) && - verifier.VerifyVector(offset_fp()) && - VerifyField<uint32_t>(verifier, VT_MODE) && + VerifyOffset(verifier, VT_BORDER) && + verifier.VerifyVector(border()) && + VerifyField<uint32_t>(verifier, VT_MODE, 4) && verifier.EndTable(); } }; struct ResizeAttributeBuilder { typedef ResizeAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_output_size(flatbuffers::Offset<flatbuffers::Vector<int32_t>> output_size) { - fbb_.AddOffset(ResizeAttribute::VT_OUTPUT_SIZE, output_size); - } - void add_stride(flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride) { - fbb_.AddOffset(ResizeAttribute::VT_STRIDE, stride); + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_scale(::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> scale) { + fbb_.AddOffset(ResizeAttribute::VT_SCALE, scale); } - void add_offset(flatbuffers::Offset<flatbuffers::Vector<int32_t>> offset) { + void add_offset(::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> offset) { fbb_.AddOffset(ResizeAttribute::VT_OFFSET, offset); } - void add_shift(int32_t shift) { - fbb_.AddElement<int32_t>(ResizeAttribute::VT_SHIFT, shift, 0); - } - void add_stride_fp(flatbuffers::Offset<flatbuffers::Vector<float>> stride_fp) { - fbb_.AddOffset(ResizeAttribute::VT_STRIDE_FP, stride_fp); - } - void add_offset_fp(flatbuffers::Offset<flatbuffers::Vector<float>> offset_fp) { - fbb_.AddOffset(ResizeAttribute::VT_OFFSET_FP, offset_fp); + void add_border(::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> border) { + fbb_.AddOffset(ResizeAttribute::VT_BORDER, border); } void add_mode(tosa::ResizeMode mode) { fbb_.AddElement<uint32_t>(ResizeAttribute::VT_MODE, static_cast<uint32_t>(mode), 0); } - explicit ResizeAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit ResizeAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - ResizeAttributeBuilder &operator=(const ResizeAttributeBuilder &); - flatbuffers::Offset<ResizeAttribute> Finish() { + ::flatbuffers::Offset<ResizeAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<ResizeAttribute>(end); + auto o = ::flatbuffers::Offset<ResizeAttribute>(end); return o; } }; -inline flatbuffers::Offset<ResizeAttribute> CreateResizeAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> output_size = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> stride = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> offset = 0, - int32_t shift = 0, - flatbuffers::Offset<flatbuffers::Vector<float>> stride_fp = 0, - flatbuffers::Offset<flatbuffers::Vector<float>> offset_fp = 0, +inline ::flatbuffers::Offset<ResizeAttribute> CreateResizeAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> scale = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> offset = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> border = 0, tosa::ResizeMode mode = tosa::ResizeMode_UNKNOWN) { ResizeAttributeBuilder builder_(_fbb); builder_.add_mode(mode); - builder_.add_offset_fp(offset_fp); - builder_.add_stride_fp(stride_fp); - builder_.add_shift(shift); + builder_.add_border(border); builder_.add_offset(offset); - builder_.add_stride(stride); - builder_.add_output_size(output_size); + builder_.add_scale(scale); return builder_.Finish(); } -inline flatbuffers::Offset<ResizeAttribute> CreateResizeAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const std::vector<int32_t> *output_size = nullptr, - const std::vector<int32_t> *stride = nullptr, - const std::vector<int32_t> *offset = nullptr, - int32_t shift = 0, - const std::vector<float> *stride_fp = nullptr, - const std::vector<float> *offset_fp = nullptr, +inline ::flatbuffers::Offset<ResizeAttribute> CreateResizeAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const std::vector<int16_t> *scale = nullptr, + const std::vector<int16_t> *offset = nullptr, + const std::vector<int16_t> *border = nullptr, tosa::ResizeMode mode = tosa::ResizeMode_UNKNOWN) { - auto output_size__ = output_size ? _fbb.CreateVector<int32_t>(*output_size) : 0; - auto stride__ = stride ? _fbb.CreateVector<int32_t>(*stride) : 0; - auto offset__ = offset ? _fbb.CreateVector<int32_t>(*offset) : 0; - auto stride_fp__ = stride_fp ? _fbb.CreateVector<float>(*stride_fp) : 0; - auto offset_fp__ = offset_fp ? _fbb.CreateVector<float>(*offset_fp) : 0; + auto scale__ = scale ? _fbb.CreateVector<int16_t>(*scale) : 0; + auto offset__ = offset ? _fbb.CreateVector<int16_t>(*offset) : 0; + auto border__ = border ? _fbb.CreateVector<int16_t>(*border) : 0; return tosa::CreateResizeAttribute( _fbb, - output_size__, - stride__, + scale__, offset__, - shift, - stride_fp__, - offset_fp__, + border__, mode); } -struct ClampAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct ClampAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef ClampAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_MIN_INT = 4, - VT_MAX_INT = 6, - VT_MIN_FP = 8, - VT_MAX_FP = 10 + VT_MIN_VAL = 4, + VT_MAX_VAL = 6 }; - int32_t min_int() const { - return GetField<int32_t>(VT_MIN_INT, 0); + const ::flatbuffers::Vector<uint8_t> *min_val() const { + return GetPointer<const ::flatbuffers::Vector<uint8_t> *>(VT_MIN_VAL); } - int32_t max_int() const { - return GetField<int32_t>(VT_MAX_INT, 0); + const ::flatbuffers::Vector<uint8_t> *max_val() const { + return GetPointer<const ::flatbuffers::Vector<uint8_t> *>(VT_MAX_VAL); } - float min_fp() const { - return GetField<float>(VT_MIN_FP, 0.0f); - } - float max_fp() const { - return GetField<float>(VT_MAX_FP, 0.0f); - } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_MIN_INT) && - VerifyField<int32_t>(verifier, VT_MAX_INT) && - VerifyField<float>(verifier, VT_MIN_FP) && - VerifyField<float>(verifier, VT_MAX_FP) && + VerifyOffset(verifier, VT_MIN_VAL) && + verifier.VerifyVector(min_val()) && + VerifyOffset(verifier, VT_MAX_VAL) && + verifier.VerifyVector(max_val()) && verifier.EndTable(); } }; struct ClampAttributeBuilder { typedef ClampAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_min_int(int32_t min_int) { - fbb_.AddElement<int32_t>(ClampAttribute::VT_MIN_INT, min_int, 0); - } - void add_max_int(int32_t max_int) { - fbb_.AddElement<int32_t>(ClampAttribute::VT_MAX_INT, max_int, 0); - } - void add_min_fp(float min_fp) { - fbb_.AddElement<float>(ClampAttribute::VT_MIN_FP, min_fp, 0.0f); + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_min_val(::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> min_val) { + fbb_.AddOffset(ClampAttribute::VT_MIN_VAL, min_val); } - void add_max_fp(float max_fp) { - fbb_.AddElement<float>(ClampAttribute::VT_MAX_FP, max_fp, 0.0f); + void add_max_val(::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> max_val) { + fbb_.AddOffset(ClampAttribute::VT_MAX_VAL, max_val); } - explicit ClampAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit ClampAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - ClampAttributeBuilder &operator=(const ClampAttributeBuilder &); - flatbuffers::Offset<ClampAttribute> Finish() { + ::flatbuffers::Offset<ClampAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<ClampAttribute>(end); + auto o = ::flatbuffers::Offset<ClampAttribute>(end); return o; } }; -inline flatbuffers::Offset<ClampAttribute> CreateClampAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - int32_t min_int = 0, - int32_t max_int = 0, - float min_fp = 0.0f, - float max_fp = 0.0f) { +inline ::flatbuffers::Offset<ClampAttribute> CreateClampAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> min_val = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> max_val = 0) { ClampAttributeBuilder builder_(_fbb); - builder_.add_max_fp(max_fp); - builder_.add_min_fp(min_fp); - builder_.add_max_int(max_int); - builder_.add_min_int(min_int); + builder_.add_max_val(max_val); + builder_.add_min_val(min_val); return builder_.Finish(); } -struct RescaleAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +inline ::flatbuffers::Offset<ClampAttribute> CreateClampAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const std::vector<uint8_t> *min_val = nullptr, + const std::vector<uint8_t> *max_val = nullptr) { + if (min_val) { _fbb.ForceVectorAlignment(min_val->size(), sizeof(uint8_t), 8); } + auto min_val__ = min_val ? _fbb.CreateVector<uint8_t>(*min_val) : 0; + if (max_val) { _fbb.ForceVectorAlignment(max_val->size(), sizeof(uint8_t), 8); } + auto max_val__ = max_val ? _fbb.CreateVector<uint8_t>(*max_val) : 0; + return tosa::CreateClampAttribute( + _fbb, + min_val__, + max_val__); +} + +struct RescaleAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef RescaleAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_INPUT_ZP = 4, VT_OUTPUT_ZP = 6, - VT_MULTIPLIER = 8, - VT_SHIFT = 10, - VT_SCALE32 = 12, - VT_DOUBLE_ROUND = 14, - VT_PER_CHANNEL = 16 + VT_SCALE32 = 8, + VT_DOUBLE_ROUND = 10, + VT_PER_CHANNEL = 12, + VT_INPUT_UNSIGNED = 14, + VT_OUTPUT_UNSIGNED = 16 }; int32_t input_zp() const { return GetField<int32_t>(VT_INPUT_ZP, 0); @@ -1396,12 +1273,6 @@ struct RescaleAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t output_zp() const { return GetField<int32_t>(VT_OUTPUT_ZP, 0); } - const flatbuffers::Vector<int32_t> *multiplier() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_MULTIPLIER); - } - const flatbuffers::Vector<int32_t> *shift() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_SHIFT); - } bool scale32() const { return GetField<uint8_t>(VT_SCALE32, 0) != 0; } @@ -1411,37 +1282,35 @@ struct RescaleAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { bool per_channel() const { return GetField<uint8_t>(VT_PER_CHANNEL, 0) != 0; } - bool Verify(flatbuffers::Verifier &verifier) const { + bool input_unsigned() const { + return GetField<uint8_t>(VT_INPUT_UNSIGNED, 0) != 0; + } + bool output_unsigned() const { + return GetField<uint8_t>(VT_OUTPUT_UNSIGNED, 0) != 0; + } + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_INPUT_ZP) && - VerifyField<int32_t>(verifier, VT_OUTPUT_ZP) && - VerifyOffset(verifier, VT_MULTIPLIER) && - verifier.VerifyVector(multiplier()) && - VerifyOffset(verifier, VT_SHIFT) && - verifier.VerifyVector(shift()) && - VerifyField<uint8_t>(verifier, VT_SCALE32) && - VerifyField<uint8_t>(verifier, VT_DOUBLE_ROUND) && - VerifyField<uint8_t>(verifier, VT_PER_CHANNEL) && + VerifyField<int32_t>(verifier, VT_INPUT_ZP, 4) && + VerifyField<int32_t>(verifier, VT_OUTPUT_ZP, 4) && + VerifyField<uint8_t>(verifier, VT_SCALE32, 1) && + VerifyField<uint8_t>(verifier, VT_DOUBLE_ROUND, 1) && + VerifyField<uint8_t>(verifier, VT_PER_CHANNEL, 1) && + VerifyField<uint8_t>(verifier, VT_INPUT_UNSIGNED, 1) && + VerifyField<uint8_t>(verifier, VT_OUTPUT_UNSIGNED, 1) && verifier.EndTable(); } }; struct RescaleAttributeBuilder { typedef RescaleAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_input_zp(int32_t input_zp) { fbb_.AddElement<int32_t>(RescaleAttribute::VT_INPUT_ZP, input_zp, 0); } void add_output_zp(int32_t output_zp) { fbb_.AddElement<int32_t>(RescaleAttribute::VT_OUTPUT_ZP, output_zp, 0); } - void add_multiplier(flatbuffers::Offset<flatbuffers::Vector<int32_t>> multiplier) { - fbb_.AddOffset(RescaleAttribute::VT_MULTIPLIER, multiplier); - } - void add_shift(flatbuffers::Offset<flatbuffers::Vector<int32_t>> shift) { - fbb_.AddOffset(RescaleAttribute::VT_SHIFT, shift); - } void add_scale32(bool scale32) { fbb_.AddElement<uint8_t>(RescaleAttribute::VT_SCALE32, static_cast<uint8_t>(scale32), 0); } @@ -1451,61 +1320,44 @@ struct RescaleAttributeBuilder { void add_per_channel(bool per_channel) { fbb_.AddElement<uint8_t>(RescaleAttribute::VT_PER_CHANNEL, static_cast<uint8_t>(per_channel), 0); } - explicit RescaleAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + void add_input_unsigned(bool input_unsigned) { + fbb_.AddElement<uint8_t>(RescaleAttribute::VT_INPUT_UNSIGNED, static_cast<uint8_t>(input_unsigned), 0); + } + void add_output_unsigned(bool output_unsigned) { + fbb_.AddElement<uint8_t>(RescaleAttribute::VT_OUTPUT_UNSIGNED, static_cast<uint8_t>(output_unsigned), 0); + } + explicit RescaleAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - RescaleAttributeBuilder &operator=(const RescaleAttributeBuilder &); - flatbuffers::Offset<RescaleAttribute> Finish() { + ::flatbuffers::Offset<RescaleAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<RescaleAttribute>(end); + auto o = ::flatbuffers::Offset<RescaleAttribute>(end); return o; } }; -inline flatbuffers::Offset<RescaleAttribute> CreateRescaleAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<RescaleAttribute> CreateRescaleAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, int32_t input_zp = 0, int32_t output_zp = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> multiplier = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> shift = 0, bool scale32 = false, bool double_round = false, - bool per_channel = false) { + bool per_channel = false, + bool input_unsigned = false, + bool output_unsigned = false) { RescaleAttributeBuilder builder_(_fbb); - builder_.add_shift(shift); - builder_.add_multiplier(multiplier); builder_.add_output_zp(output_zp); builder_.add_input_zp(input_zp); + builder_.add_output_unsigned(output_unsigned); + builder_.add_input_unsigned(input_unsigned); builder_.add_per_channel(per_channel); builder_.add_double_round(double_round); builder_.add_scale32(scale32); return builder_.Finish(); } -inline flatbuffers::Offset<RescaleAttribute> CreateRescaleAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - int32_t input_zp = 0, - int32_t output_zp = 0, - const std::vector<int32_t> *multiplier = nullptr, - const std::vector<int32_t> *shift = nullptr, - bool scale32 = false, - bool double_round = false, - bool per_channel = false) { - auto multiplier__ = multiplier ? _fbb.CreateVector<int32_t>(*multiplier) : 0; - auto shift__ = shift ? _fbb.CreateVector<int32_t>(*shift) : 0; - return tosa::CreateRescaleAttribute( - _fbb, - input_zp, - output_zp, - multiplier__, - shift__, - scale32, - double_round, - per_channel); -} - -struct MulAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct MulAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef MulAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_SHIFT = 4 @@ -1513,41 +1365,40 @@ struct MulAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t shift() const { return GetField<int32_t>(VT_SHIFT, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_SHIFT) && + VerifyField<int32_t>(verifier, VT_SHIFT, 4) && verifier.EndTable(); } }; struct MulAttributeBuilder { typedef MulAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_shift(int32_t shift) { fbb_.AddElement<int32_t>(MulAttribute::VT_SHIFT, shift, 0); } - explicit MulAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit MulAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - MulAttributeBuilder &operator=(const MulAttributeBuilder &); - flatbuffers::Offset<MulAttribute> Finish() { + ::flatbuffers::Offset<MulAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<MulAttribute>(end); + auto o = ::flatbuffers::Offset<MulAttribute>(end); return o; } }; -inline flatbuffers::Offset<MulAttribute> CreateMulAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<MulAttribute> CreateMulAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, int32_t shift = 0) { MulAttributeBuilder builder_(_fbb); builder_.add_shift(shift); return builder_.Finish(); } -struct ArithmeticRightShiftAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct ArithmeticRightShiftAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef ArithmeticRightShiftAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_ROUND = 4 @@ -1555,181 +1406,178 @@ struct ArithmeticRightShiftAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffe bool round() const { return GetField<uint8_t>(VT_ROUND, 0) != 0; } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<uint8_t>(verifier, VT_ROUND) && + VerifyField<uint8_t>(verifier, VT_ROUND, 1) && verifier.EndTable(); } }; struct ArithmeticRightShiftAttributeBuilder { typedef ArithmeticRightShiftAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_round(bool round) { fbb_.AddElement<uint8_t>(ArithmeticRightShiftAttribute::VT_ROUND, static_cast<uint8_t>(round), 0); } - explicit ArithmeticRightShiftAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit ArithmeticRightShiftAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - ArithmeticRightShiftAttributeBuilder &operator=(const ArithmeticRightShiftAttributeBuilder &); - flatbuffers::Offset<ArithmeticRightShiftAttribute> Finish() { + ::flatbuffers::Offset<ArithmeticRightShiftAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<ArithmeticRightShiftAttribute>(end); + auto o = ::flatbuffers::Offset<ArithmeticRightShiftAttribute>(end); return o; } }; -inline flatbuffers::Offset<ArithmeticRightShiftAttribute> CreateArithmeticRightShiftAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<ArithmeticRightShiftAttribute> CreateArithmeticRightShiftAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, bool round = false) { ArithmeticRightShiftAttributeBuilder builder_(_fbb); builder_.add_round(round); return builder_.Finish(); } -struct CondIfAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct CondIfAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef CondIfAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_THEN_BRANCH = 4, - VT_ELSE_BRANCH = 6 + VT_THEN_GRAPH = 4, + VT_ELSE_GRAPH = 6 }; - const flatbuffers::String *then_branch() const { - return GetPointer<const flatbuffers::String *>(VT_THEN_BRANCH); + const ::flatbuffers::String *then_graph() const { + return GetPointer<const ::flatbuffers::String *>(VT_THEN_GRAPH); } - const flatbuffers::String *else_branch() const { - return GetPointer<const flatbuffers::String *>(VT_ELSE_BRANCH); + const ::flatbuffers::String *else_graph() const { + return GetPointer<const ::flatbuffers::String *>(VT_ELSE_GRAPH); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_THEN_BRANCH) && - verifier.VerifyString(then_branch()) && - VerifyOffset(verifier, VT_ELSE_BRANCH) && - verifier.VerifyString(else_branch()) && + VerifyOffset(verifier, VT_THEN_GRAPH) && + verifier.VerifyString(then_graph()) && + VerifyOffset(verifier, VT_ELSE_GRAPH) && + verifier.VerifyString(else_graph()) && verifier.EndTable(); } }; struct CondIfAttributeBuilder { typedef CondIfAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_then_branch(flatbuffers::Offset<flatbuffers::String> then_branch) { - fbb_.AddOffset(CondIfAttribute::VT_THEN_BRANCH, then_branch); + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_then_graph(::flatbuffers::Offset<::flatbuffers::String> then_graph) { + fbb_.AddOffset(CondIfAttribute::VT_THEN_GRAPH, then_graph); } - void add_else_branch(flatbuffers::Offset<flatbuffers::String> else_branch) { - fbb_.AddOffset(CondIfAttribute::VT_ELSE_BRANCH, else_branch); + void add_else_graph(::flatbuffers::Offset<::flatbuffers::String> else_graph) { + fbb_.AddOffset(CondIfAttribute::VT_ELSE_GRAPH, else_graph); } - explicit CondIfAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit CondIfAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - CondIfAttributeBuilder &operator=(const CondIfAttributeBuilder &); - flatbuffers::Offset<CondIfAttribute> Finish() { + ::flatbuffers::Offset<CondIfAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<CondIfAttribute>(end); + auto o = ::flatbuffers::Offset<CondIfAttribute>(end); return o; } }; -inline flatbuffers::Offset<CondIfAttribute> CreateCondIfAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::String> then_branch = 0, - flatbuffers::Offset<flatbuffers::String> else_branch = 0) { +inline ::flatbuffers::Offset<CondIfAttribute> CreateCondIfAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::String> then_graph = 0, + ::flatbuffers::Offset<::flatbuffers::String> else_graph = 0) { CondIfAttributeBuilder builder_(_fbb); - builder_.add_else_branch(else_branch); - builder_.add_then_branch(then_branch); + builder_.add_else_graph(else_graph); + builder_.add_then_graph(then_graph); return builder_.Finish(); } -inline flatbuffers::Offset<CondIfAttribute> CreateCondIfAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const char *then_branch = nullptr, - const char *else_branch = nullptr) { - auto then_branch__ = then_branch ? _fbb.CreateString(then_branch) : 0; - auto else_branch__ = else_branch ? _fbb.CreateString(else_branch) : 0; +inline ::flatbuffers::Offset<CondIfAttribute> CreateCondIfAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const char *then_graph = nullptr, + const char *else_graph = nullptr) { + auto then_graph__ = then_graph ? _fbb.CreateString(then_graph) : 0; + auto else_graph__ = else_graph ? _fbb.CreateString(else_graph) : 0; return tosa::CreateCondIfAttribute( _fbb, - then_branch__, - else_branch__); + then_graph__, + else_graph__); } -struct WhileLoopAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct WhileLoopAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef WhileLoopAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_COND_BRANCH = 4, - VT_BODY_BRANCH = 6 + VT_COND_GRAPH = 4, + VT_BODY_GRAPH = 6 }; - const flatbuffers::String *cond_branch() const { - return GetPointer<const flatbuffers::String *>(VT_COND_BRANCH); + const ::flatbuffers::String *cond_graph() const { + return GetPointer<const ::flatbuffers::String *>(VT_COND_GRAPH); } - const flatbuffers::String *body_branch() const { - return GetPointer<const flatbuffers::String *>(VT_BODY_BRANCH); + const ::flatbuffers::String *body_graph() const { + return GetPointer<const ::flatbuffers::String *>(VT_BODY_GRAPH); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_COND_BRANCH) && - verifier.VerifyString(cond_branch()) && - VerifyOffset(verifier, VT_BODY_BRANCH) && - verifier.VerifyString(body_branch()) && + VerifyOffset(verifier, VT_COND_GRAPH) && + verifier.VerifyString(cond_graph()) && + VerifyOffset(verifier, VT_BODY_GRAPH) && + verifier.VerifyString(body_graph()) && verifier.EndTable(); } }; struct WhileLoopAttributeBuilder { typedef WhileLoopAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_cond_branch(flatbuffers::Offset<flatbuffers::String> cond_branch) { - fbb_.AddOffset(WhileLoopAttribute::VT_COND_BRANCH, cond_branch); + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_cond_graph(::flatbuffers::Offset<::flatbuffers::String> cond_graph) { + fbb_.AddOffset(WhileLoopAttribute::VT_COND_GRAPH, cond_graph); } - void add_body_branch(flatbuffers::Offset<flatbuffers::String> body_branch) { - fbb_.AddOffset(WhileLoopAttribute::VT_BODY_BRANCH, body_branch); + void add_body_graph(::flatbuffers::Offset<::flatbuffers::String> body_graph) { + fbb_.AddOffset(WhileLoopAttribute::VT_BODY_GRAPH, body_graph); } - explicit WhileLoopAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit WhileLoopAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - WhileLoopAttributeBuilder &operator=(const WhileLoopAttributeBuilder &); - flatbuffers::Offset<WhileLoopAttribute> Finish() { + ::flatbuffers::Offset<WhileLoopAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<WhileLoopAttribute>(end); + auto o = ::flatbuffers::Offset<WhileLoopAttribute>(end); return o; } }; -inline flatbuffers::Offset<WhileLoopAttribute> CreateWhileLoopAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::String> cond_branch = 0, - flatbuffers::Offset<flatbuffers::String> body_branch = 0) { +inline ::flatbuffers::Offset<WhileLoopAttribute> CreateWhileLoopAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::String> cond_graph = 0, + ::flatbuffers::Offset<::flatbuffers::String> body_graph = 0) { WhileLoopAttributeBuilder builder_(_fbb); - builder_.add_body_branch(body_branch); - builder_.add_cond_branch(cond_branch); + builder_.add_body_graph(body_graph); + builder_.add_cond_graph(cond_graph); return builder_.Finish(); } -inline flatbuffers::Offset<WhileLoopAttribute> CreateWhileLoopAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const char *cond_branch = nullptr, - const char *body_branch = nullptr) { - auto cond_branch__ = cond_branch ? _fbb.CreateString(cond_branch) : 0; - auto body_branch__ = body_branch ? _fbb.CreateString(body_branch) : 0; +inline ::flatbuffers::Offset<WhileLoopAttribute> CreateWhileLoopAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const char *cond_graph = nullptr, + const char *body_graph = nullptr) { + auto cond_graph__ = cond_graph ? _fbb.CreateString(cond_graph) : 0; + auto body_graph__ = body_graph ? _fbb.CreateString(body_graph) : 0; return tosa::CreateWhileLoopAttribute( _fbb, - cond_branch__, - body_branch__); + cond_graph__, + body_graph__); } -struct TransposeAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TransposeAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TransposeAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_PERMS = 4 }; - const flatbuffers::Vector<int32_t> *perms() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_PERMS); + const ::flatbuffers::Vector<int32_t> *perms() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_PERMS); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_PERMS) && verifier.VerifyVector(perms()) && @@ -1739,33 +1587,32 @@ struct TransposeAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { struct TransposeAttributeBuilder { typedef TransposeAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_perms(flatbuffers::Offset<flatbuffers::Vector<int32_t>> perms) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_perms(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> perms) { fbb_.AddOffset(TransposeAttribute::VT_PERMS, perms); } - explicit TransposeAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit TransposeAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TransposeAttributeBuilder &operator=(const TransposeAttributeBuilder &); - flatbuffers::Offset<TransposeAttribute> Finish() { + ::flatbuffers::Offset<TransposeAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TransposeAttribute>(end); + auto o = ::flatbuffers::Offset<TransposeAttribute>(end); return o; } }; -inline flatbuffers::Offset<TransposeAttribute> CreateTransposeAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> perms = 0) { +inline ::flatbuffers::Offset<TransposeAttribute> CreateTransposeAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> perms = 0) { TransposeAttributeBuilder builder_(_fbb); builder_.add_perms(perms); return builder_.Finish(); } -inline flatbuffers::Offset<TransposeAttribute> CreateTransposeAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TransposeAttribute> CreateTransposeAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const std::vector<int32_t> *perms = nullptr) { auto perms__ = perms ? _fbb.CreateVector<int32_t>(*perms) : 0; return tosa::CreateTransposeAttribute( @@ -1773,15 +1620,15 @@ inline flatbuffers::Offset<TransposeAttribute> CreateTransposeAttributeDirect( perms__); } -struct TableAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TableAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TableAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_TABLE = 4 }; - const flatbuffers::Vector<int16_t> *table() const { - return GetPointer<const flatbuffers::Vector<int16_t> *>(VT_TABLE); + const ::flatbuffers::Vector<int16_t> *table() const { + return GetPointer<const ::flatbuffers::Vector<int16_t> *>(VT_TABLE); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_TABLE) && verifier.VerifyVector(table()) && @@ -1791,33 +1638,32 @@ struct TableAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { struct TableAttributeBuilder { typedef TableAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_table(flatbuffers::Offset<flatbuffers::Vector<int16_t>> table) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_table(::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> table) { fbb_.AddOffset(TableAttribute::VT_TABLE, table); } - explicit TableAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit TableAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TableAttributeBuilder &operator=(const TableAttributeBuilder &); - flatbuffers::Offset<TableAttribute> Finish() { + ::flatbuffers::Offset<TableAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TableAttribute>(end); + auto o = ::flatbuffers::Offset<TableAttribute>(end); return o; } }; -inline flatbuffers::Offset<TableAttribute> CreateTableAttribute( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::Vector<int16_t>> table = 0) { +inline ::flatbuffers::Offset<TableAttribute> CreateTableAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::Vector<int16_t>> table = 0) { TableAttributeBuilder builder_(_fbb); builder_.add_table(table); return builder_.Finish(); } -inline flatbuffers::Offset<TableAttribute> CreateTableAttributeDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TableAttribute> CreateTableAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const std::vector<int16_t> *table = nullptr) { auto table__ = table ? _fbb.CreateVector<int16_t>(*table) : 0; return tosa::CreateTableAttribute( @@ -1825,7 +1671,7 @@ inline flatbuffers::Offset<TableAttribute> CreateTableAttributeDirect( table__); } -struct MatMulAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct MatMulAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef MatMulAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_A_ZP = 4, @@ -1837,38 +1683,37 @@ struct MatMulAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t b_zp() const { return GetField<int32_t>(VT_B_ZP, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_A_ZP) && - VerifyField<int32_t>(verifier, VT_B_ZP) && + VerifyField<int32_t>(verifier, VT_A_ZP, 4) && + VerifyField<int32_t>(verifier, VT_B_ZP, 4) && verifier.EndTable(); } }; struct MatMulAttributeBuilder { typedef MatMulAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_a_zp(int32_t a_zp) { fbb_.AddElement<int32_t>(MatMulAttribute::VT_A_ZP, a_zp, 0); } void add_b_zp(int32_t b_zp) { fbb_.AddElement<int32_t>(MatMulAttribute::VT_B_ZP, b_zp, 0); } - explicit MatMulAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit MatMulAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - MatMulAttributeBuilder &operator=(const MatMulAttributeBuilder &); - flatbuffers::Offset<MatMulAttribute> Finish() { + ::flatbuffers::Offset<MatMulAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<MatMulAttribute>(end); + auto o = ::flatbuffers::Offset<MatMulAttribute>(end); return o; } }; -inline flatbuffers::Offset<MatMulAttribute> CreateMatMulAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<MatMulAttribute> CreateMatMulAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, int32_t a_zp = 0, int32_t b_zp = 0) { MatMulAttributeBuilder builder_(_fbb); @@ -1877,7 +1722,7 @@ inline flatbuffers::Offset<MatMulAttribute> CreateMatMulAttribute( return builder_.Finish(); } -struct FullyConnectedAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct FullyConnectedAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef FullyConnectedAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_INPUT_ZP = 4, @@ -1889,38 +1734,37 @@ struct FullyConnectedAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Ta int32_t weight_zp() const { return GetField<int32_t>(VT_WEIGHT_ZP, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_INPUT_ZP) && - VerifyField<int32_t>(verifier, VT_WEIGHT_ZP) && + VerifyField<int32_t>(verifier, VT_INPUT_ZP, 4) && + VerifyField<int32_t>(verifier, VT_WEIGHT_ZP, 4) && verifier.EndTable(); } }; struct FullyConnectedAttributeBuilder { typedef FullyConnectedAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_input_zp(int32_t input_zp) { fbb_.AddElement<int32_t>(FullyConnectedAttribute::VT_INPUT_ZP, input_zp, 0); } void add_weight_zp(int32_t weight_zp) { fbb_.AddElement<int32_t>(FullyConnectedAttribute::VT_WEIGHT_ZP, weight_zp, 0); } - explicit FullyConnectedAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit FullyConnectedAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - FullyConnectedAttributeBuilder &operator=(const FullyConnectedAttributeBuilder &); - flatbuffers::Offset<FullyConnectedAttribute> Finish() { + ::flatbuffers::Offset<FullyConnectedAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<FullyConnectedAttribute>(end); + auto o = ::flatbuffers::Offset<FullyConnectedAttribute>(end); return o; } }; -inline flatbuffers::Offset<FullyConnectedAttribute> CreateFullyConnectedAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<FullyConnectedAttribute> CreateFullyConnectedAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, int32_t input_zp = 0, int32_t weight_zp = 0) { FullyConnectedAttributeBuilder builder_(_fbb); @@ -1929,7 +1773,7 @@ inline flatbuffers::Offset<FullyConnectedAttribute> CreateFullyConnectedAttribut return builder_.Finish(); } -struct NegateAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct NegateAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef NegateAttributeBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_INPUT1_ZP = 4, @@ -1941,38 +1785,37 @@ struct NegateAttribute FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { int32_t output_zp() const { return GetField<int32_t>(VT_OUTPUT_ZP, 0); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT_INPUT1_ZP) && - VerifyField<int32_t>(verifier, VT_OUTPUT_ZP) && + VerifyField<int32_t>(verifier, VT_INPUT1_ZP, 4) && + VerifyField<int32_t>(verifier, VT_OUTPUT_ZP, 4) && verifier.EndTable(); } }; struct NegateAttributeBuilder { typedef NegateAttribute Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_input1_zp(int32_t input1_zp) { fbb_.AddElement<int32_t>(NegateAttribute::VT_INPUT1_ZP, input1_zp, 0); } void add_output_zp(int32_t output_zp) { fbb_.AddElement<int32_t>(NegateAttribute::VT_OUTPUT_ZP, output_zp, 0); } - explicit NegateAttributeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit NegateAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - NegateAttributeBuilder &operator=(const NegateAttributeBuilder &); - flatbuffers::Offset<NegateAttribute> Finish() { + ::flatbuffers::Offset<NegateAttribute> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<NegateAttribute>(end); + auto o = ::flatbuffers::Offset<NegateAttribute>(end); return o; } }; -inline flatbuffers::Offset<NegateAttribute> CreateNegateAttribute( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<NegateAttribute> CreateNegateAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, int32_t input1_zp = 0, int32_t output_zp = 0) { NegateAttributeBuilder builder_(_fbb); @@ -1981,7 +1824,178 @@ inline flatbuffers::Offset<NegateAttribute> CreateNegateAttribute( return builder_.Finish(); } -struct Version FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct CustomAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { + typedef CustomAttributeBuilder Builder; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_OPERATOR_NAME = 4, + VT_DOMAIN_NAME = 6, + VT_IMPLEMENTATION_ATTRS = 8 + }; + const ::flatbuffers::String *operator_name() const { + return GetPointer<const ::flatbuffers::String *>(VT_OPERATOR_NAME); + } + const ::flatbuffers::String *domain_name() const { + return GetPointer<const ::flatbuffers::String *>(VT_DOMAIN_NAME); + } + const ::flatbuffers::Vector<uint8_t> *implementation_attrs() const { + return GetPointer<const ::flatbuffers::Vector<uint8_t> *>(VT_IMPLEMENTATION_ATTRS); + } + bool Verify(::flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_OPERATOR_NAME) && + verifier.VerifyString(operator_name()) && + VerifyOffset(verifier, VT_DOMAIN_NAME) && + verifier.VerifyString(domain_name()) && + VerifyOffset(verifier, VT_IMPLEMENTATION_ATTRS) && + verifier.VerifyVector(implementation_attrs()) && + verifier.EndTable(); + } +}; + +struct CustomAttributeBuilder { + typedef CustomAttribute Table; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_operator_name(::flatbuffers::Offset<::flatbuffers::String> operator_name) { + fbb_.AddOffset(CustomAttribute::VT_OPERATOR_NAME, operator_name); + } + void add_domain_name(::flatbuffers::Offset<::flatbuffers::String> domain_name) { + fbb_.AddOffset(CustomAttribute::VT_DOMAIN_NAME, domain_name); + } + void add_implementation_attrs(::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> implementation_attrs) { + fbb_.AddOffset(CustomAttribute::VT_IMPLEMENTATION_ATTRS, implementation_attrs); + } + explicit CustomAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ::flatbuffers::Offset<CustomAttribute> Finish() { + const auto end = fbb_.EndTable(start_); + auto o = ::flatbuffers::Offset<CustomAttribute>(end); + return o; + } +}; + +inline ::flatbuffers::Offset<CustomAttribute> CreateCustomAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::String> operator_name = 0, + ::flatbuffers::Offset<::flatbuffers::String> domain_name = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> implementation_attrs = 0) { + CustomAttributeBuilder builder_(_fbb); + builder_.add_implementation_attrs(implementation_attrs); + builder_.add_domain_name(domain_name); + builder_.add_operator_name(operator_name); + return builder_.Finish(); +} + +inline ::flatbuffers::Offset<CustomAttribute> CreateCustomAttributeDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const char *operator_name = nullptr, + const char *domain_name = nullptr, + const std::vector<uint8_t> *implementation_attrs = nullptr) { + auto operator_name__ = operator_name ? _fbb.CreateString(operator_name) : 0; + auto domain_name__ = domain_name ? _fbb.CreateString(domain_name) : 0; + auto implementation_attrs__ = implementation_attrs ? _fbb.CreateVector<uint8_t>(*implementation_attrs) : 0; + return tosa::CreateCustomAttribute( + _fbb, + operator_name__, + domain_name__, + implementation_attrs__); +} + +struct FFTAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { + typedef FFTAttributeBuilder Builder; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_INVERSE = 4, + VT_LOCAL_BOUND = 6 + }; + bool inverse() const { + return GetField<uint8_t>(VT_INVERSE, 0) != 0; + } + bool local_bound() const { + return GetField<uint8_t>(VT_LOCAL_BOUND, 0) != 0; + } + bool Verify(::flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField<uint8_t>(verifier, VT_INVERSE, 1) && + VerifyField<uint8_t>(verifier, VT_LOCAL_BOUND, 1) && + verifier.EndTable(); + } +}; + +struct FFTAttributeBuilder { + typedef FFTAttribute Table; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_inverse(bool inverse) { + fbb_.AddElement<uint8_t>(FFTAttribute::VT_INVERSE, static_cast<uint8_t>(inverse), 0); + } + void add_local_bound(bool local_bound) { + fbb_.AddElement<uint8_t>(FFTAttribute::VT_LOCAL_BOUND, static_cast<uint8_t>(local_bound), 0); + } + explicit FFTAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ::flatbuffers::Offset<FFTAttribute> Finish() { + const auto end = fbb_.EndTable(start_); + auto o = ::flatbuffers::Offset<FFTAttribute>(end); + return o; + } +}; + +inline ::flatbuffers::Offset<FFTAttribute> CreateFFTAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + bool inverse = false, + bool local_bound = false) { + FFTAttributeBuilder builder_(_fbb); + builder_.add_local_bound(local_bound); + builder_.add_inverse(inverse); + return builder_.Finish(); +} + +struct RFFTAttribute FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { + typedef RFFTAttributeBuilder Builder; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_LOCAL_BOUND = 4 + }; + bool local_bound() const { + return GetField<uint8_t>(VT_LOCAL_BOUND, 0) != 0; + } + bool Verify(::flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField<uint8_t>(verifier, VT_LOCAL_BOUND, 1) && + verifier.EndTable(); + } +}; + +struct RFFTAttributeBuilder { + typedef RFFTAttribute Table; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_local_bound(bool local_bound) { + fbb_.AddElement<uint8_t>(RFFTAttribute::VT_LOCAL_BOUND, static_cast<uint8_t>(local_bound), 0); + } + explicit RFFTAttributeBuilder(::flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ::flatbuffers::Offset<RFFTAttribute> Finish() { + const auto end = fbb_.EndTable(start_); + auto o = ::flatbuffers::Offset<RFFTAttribute>(end); + return o; + } +}; + +inline ::flatbuffers::Offset<RFFTAttribute> CreateRFFTAttribute( + ::flatbuffers::FlatBufferBuilder &_fbb, + bool local_bound = false) { + RFFTAttributeBuilder builder_(_fbb); + builder_.add_local_bound(local_bound); + return builder_.Finish(); +} + +struct Version FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef VersionBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT__MAJOR = 4, @@ -1990,60 +2004,59 @@ struct Version FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { VT__DRAFT = 10 }; int32_t _major() const { - return GetField<int32_t>(VT__MAJOR, 0); + return GetField<int32_t>(VT__MAJOR, -1); } int32_t _minor() const { - return GetField<int32_t>(VT__MINOR, 31); + return GetField<int32_t>(VT__MINOR, -1); } int32_t _patch() const { - return GetField<int32_t>(VT__PATCH, 0); + return GetField<int32_t>(VT__PATCH, -1); } bool _draft() const { return GetField<uint8_t>(VT__DRAFT, 1) != 0; } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<int32_t>(verifier, VT__MAJOR) && - VerifyField<int32_t>(verifier, VT__MINOR) && - VerifyField<int32_t>(verifier, VT__PATCH) && - VerifyField<uint8_t>(verifier, VT__DRAFT) && + VerifyField<int32_t>(verifier, VT__MAJOR, 4) && + VerifyField<int32_t>(verifier, VT__MINOR, 4) && + VerifyField<int32_t>(verifier, VT__PATCH, 4) && + VerifyField<uint8_t>(verifier, VT__DRAFT, 1) && verifier.EndTable(); } }; struct VersionBuilder { typedef Version Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add__major(int32_t _major) { - fbb_.AddElement<int32_t>(Version::VT__MAJOR, _major, 0); + fbb_.AddElement<int32_t>(Version::VT__MAJOR, _major, -1); } void add__minor(int32_t _minor) { - fbb_.AddElement<int32_t>(Version::VT__MINOR, _minor, 31); + fbb_.AddElement<int32_t>(Version::VT__MINOR, _minor, -1); } void add__patch(int32_t _patch) { - fbb_.AddElement<int32_t>(Version::VT__PATCH, _patch, 0); + fbb_.AddElement<int32_t>(Version::VT__PATCH, _patch, -1); } void add__draft(bool _draft) { fbb_.AddElement<uint8_t>(Version::VT__DRAFT, static_cast<uint8_t>(_draft), 1); } - explicit VersionBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit VersionBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - VersionBuilder &operator=(const VersionBuilder &); - flatbuffers::Offset<Version> Finish() { + ::flatbuffers::Offset<Version> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<Version>(end); + auto o = ::flatbuffers::Offset<Version>(end); return o; } }; -inline flatbuffers::Offset<Version> CreateVersion( - flatbuffers::FlatBufferBuilder &_fbb, - int32_t _major = 0, - int32_t _minor = 31, - int32_t _patch = 0, +inline ::flatbuffers::Offset<Version> CreateVersion( + ::flatbuffers::FlatBufferBuilder &_fbb, + int32_t _major = -1, + int32_t _minor = -1, + int32_t _patch = -1, bool _draft = true) { VersionBuilder builder_(_fbb); builder_.add__patch(_patch); @@ -2053,100 +2066,137 @@ inline flatbuffers::Offset<Version> CreateVersion( return builder_.Finish(); } -struct TosaTensor FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TosaTensor FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TosaTensorBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_NAME = 4, VT_SHAPE = 6, VT_TYPE = 8, - VT_DATA = 10 + VT_DATA = 10, + VT_VARIABLE = 12, + VT_IS_UNRANKED = 14, + VT_VARIABLE_NAME = 16 }; - const flatbuffers::String *name() const { - return GetPointer<const flatbuffers::String *>(VT_NAME); + const ::flatbuffers::String *name() const { + return GetPointer<const ::flatbuffers::String *>(VT_NAME); } - const flatbuffers::Vector<int32_t> *shape() const { - return GetPointer<const flatbuffers::Vector<int32_t> *>(VT_SHAPE); + const ::flatbuffers::Vector<int32_t> *shape() const { + return GetPointer<const ::flatbuffers::Vector<int32_t> *>(VT_SHAPE); } tosa::DType type() const { return static_cast<tosa::DType>(GetField<uint32_t>(VT_TYPE, 0)); } - const flatbuffers::Vector<uint8_t> *data() const { - return GetPointer<const flatbuffers::Vector<uint8_t> *>(VT_DATA); + const ::flatbuffers::Vector<uint8_t> *data() const { + return GetPointer<const ::flatbuffers::Vector<uint8_t> *>(VT_DATA); + } + bool variable() const { + return GetField<uint8_t>(VT_VARIABLE, 0) != 0; + } + bool is_unranked() const { + return GetField<uint8_t>(VT_IS_UNRANKED, 0) != 0; } - bool Verify(flatbuffers::Verifier &verifier) const { + const ::flatbuffers::String *variable_name() const { + return GetPointer<const ::flatbuffers::String *>(VT_VARIABLE_NAME); + } + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_NAME) && verifier.VerifyString(name()) && VerifyOffset(verifier, VT_SHAPE) && verifier.VerifyVector(shape()) && - VerifyField<uint32_t>(verifier, VT_TYPE) && + VerifyField<uint32_t>(verifier, VT_TYPE, 4) && VerifyOffset(verifier, VT_DATA) && verifier.VerifyVector(data()) && + VerifyField<uint8_t>(verifier, VT_VARIABLE, 1) && + VerifyField<uint8_t>(verifier, VT_IS_UNRANKED, 1) && + VerifyOffset(verifier, VT_VARIABLE_NAME) && + verifier.VerifyString(variable_name()) && verifier.EndTable(); } }; struct TosaTensorBuilder { typedef TosaTensor Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_name(flatbuffers::Offset<flatbuffers::String> name) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_name(::flatbuffers::Offset<::flatbuffers::String> name) { fbb_.AddOffset(TosaTensor::VT_NAME, name); } - void add_shape(flatbuffers::Offset<flatbuffers::Vector<int32_t>> shape) { + void add_shape(::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> shape) { fbb_.AddOffset(TosaTensor::VT_SHAPE, shape); } void add_type(tosa::DType type) { fbb_.AddElement<uint32_t>(TosaTensor::VT_TYPE, static_cast<uint32_t>(type), 0); } - void add_data(flatbuffers::Offset<flatbuffers::Vector<uint8_t>> data) { + void add_data(::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> data) { fbb_.AddOffset(TosaTensor::VT_DATA, data); } - explicit TosaTensorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + void add_variable(bool variable) { + fbb_.AddElement<uint8_t>(TosaTensor::VT_VARIABLE, static_cast<uint8_t>(variable), 0); + } + void add_is_unranked(bool is_unranked) { + fbb_.AddElement<uint8_t>(TosaTensor::VT_IS_UNRANKED, static_cast<uint8_t>(is_unranked), 0); + } + void add_variable_name(::flatbuffers::Offset<::flatbuffers::String> variable_name) { + fbb_.AddOffset(TosaTensor::VT_VARIABLE_NAME, variable_name); + } + explicit TosaTensorBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TosaTensorBuilder &operator=(const TosaTensorBuilder &); - flatbuffers::Offset<TosaTensor> Finish() { + ::flatbuffers::Offset<TosaTensor> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TosaTensor>(end); + auto o = ::flatbuffers::Offset<TosaTensor>(end); return o; } }; -inline flatbuffers::Offset<TosaTensor> CreateTosaTensor( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::String> name = 0, - flatbuffers::Offset<flatbuffers::Vector<int32_t>> shape = 0, +inline ::flatbuffers::Offset<TosaTensor> CreateTosaTensor( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::String> name = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<int32_t>> shape = 0, tosa::DType type = tosa::DType_UNKNOWN, - flatbuffers::Offset<flatbuffers::Vector<uint8_t>> data = 0) { + ::flatbuffers::Offset<::flatbuffers::Vector<uint8_t>> data = 0, + bool variable = false, + bool is_unranked = false, + ::flatbuffers::Offset<::flatbuffers::String> variable_name = 0) { TosaTensorBuilder builder_(_fbb); + builder_.add_variable_name(variable_name); builder_.add_data(data); builder_.add_type(type); builder_.add_shape(shape); builder_.add_name(name); + builder_.add_is_unranked(is_unranked); + builder_.add_variable(variable); return builder_.Finish(); } -inline flatbuffers::Offset<TosaTensor> CreateTosaTensorDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TosaTensor> CreateTosaTensorDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const char *name = nullptr, const std::vector<int32_t> *shape = nullptr, tosa::DType type = tosa::DType_UNKNOWN, - const std::vector<uint8_t> *data = nullptr) { + const std::vector<uint8_t> *data = nullptr, + bool variable = false, + bool is_unranked = false, + const char *variable_name = nullptr) { auto name__ = name ? _fbb.CreateString(name) : 0; auto shape__ = shape ? _fbb.CreateVector<int32_t>(*shape) : 0; if (data) { _fbb.ForceVectorAlignment(data->size(), sizeof(uint8_t), 8); } auto data__ = data ? _fbb.CreateVector<uint8_t>(*data) : 0; + auto variable_name__ = variable_name ? _fbb.CreateString(variable_name) : 0; return tosa::CreateTosaTensor( _fbb, name__, shape__, type, - data__); + data__, + variable, + is_unranked, + variable_name__); } -struct TosaOperator FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TosaOperator FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TosaOperatorBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_OP = 4, @@ -2180,15 +2230,6 @@ struct TosaOperator FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { const tosa::AxisAttribute *attribute_as_AxisAttribute() const { return attribute_type() == tosa::Attribute_AxisAttribute ? static_cast<const tosa::AxisAttribute *>(attribute()) : nullptr; } - const tosa::ReshapeAttribute *attribute_as_ReshapeAttribute() const { - return attribute_type() == tosa::Attribute_ReshapeAttribute ? static_cast<const tosa::ReshapeAttribute *>(attribute()) : nullptr; - } - const tosa::SliceAttribute *attribute_as_SliceAttribute() const { - return attribute_type() == tosa::Attribute_SliceAttribute ? static_cast<const tosa::SliceAttribute *>(attribute()) : nullptr; - } - const tosa::TileAttribute *attribute_as_TileAttribute() const { - return attribute_type() == tosa::Attribute_TileAttribute ? static_cast<const tosa::TileAttribute *>(attribute()) : nullptr; - } const tosa::ResizeAttribute *attribute_as_ResizeAttribute() const { return attribute_type() == tosa::Attribute_ResizeAttribute ? static_cast<const tosa::ResizeAttribute *>(attribute()) : nullptr; } @@ -2225,16 +2266,25 @@ struct TosaOperator FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { const tosa::NegateAttribute *attribute_as_NegateAttribute() const { return attribute_type() == tosa::Attribute_NegateAttribute ? static_cast<const tosa::NegateAttribute *>(attribute()) : nullptr; } - const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *inputs() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *>(VT_INPUTS); + const tosa::CustomAttribute *attribute_as_CustomAttribute() const { + return attribute_type() == tosa::Attribute_CustomAttribute ? static_cast<const tosa::CustomAttribute *>(attribute()) : nullptr; + } + const tosa::FFTAttribute *attribute_as_FFTAttribute() const { + return attribute_type() == tosa::Attribute_FFTAttribute ? static_cast<const tosa::FFTAttribute *>(attribute()) : nullptr; + } + const tosa::RFFTAttribute *attribute_as_RFFTAttribute() const { + return attribute_type() == tosa::Attribute_RFFTAttribute ? static_cast<const tosa::RFFTAttribute *>(attribute()) : nullptr; } - const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *outputs() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *>(VT_OUTPUTS); + const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *inputs() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *>(VT_INPUTS); } - bool Verify(flatbuffers::Verifier &verifier) const { + const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *outputs() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *>(VT_OUTPUTS); + } + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyField<uint32_t>(verifier, VT_OP) && - VerifyField<uint8_t>(verifier, VT_ATTRIBUTE_TYPE) && + VerifyField<uint32_t>(verifier, VT_OP, 4) && + VerifyField<uint8_t>(verifier, VT_ATTRIBUTE_TYPE, 1) && VerifyOffset(verifier, VT_ATTRIBUTE) && VerifyAttribute(verifier, attribute(), attribute_type()) && VerifyOffset(verifier, VT_INPUTS) && @@ -2267,18 +2317,6 @@ template<> inline const tosa::AxisAttribute *TosaOperator::attribute_as<tosa::Ax return attribute_as_AxisAttribute(); } -template<> inline const tosa::ReshapeAttribute *TosaOperator::attribute_as<tosa::ReshapeAttribute>() const { - return attribute_as_ReshapeAttribute(); -} - -template<> inline const tosa::SliceAttribute *TosaOperator::attribute_as<tosa::SliceAttribute>() const { - return attribute_as_SliceAttribute(); -} - -template<> inline const tosa::TileAttribute *TosaOperator::attribute_as<tosa::TileAttribute>() const { - return attribute_as_TileAttribute(); -} - template<> inline const tosa::ResizeAttribute *TosaOperator::attribute_as<tosa::ResizeAttribute>() const { return attribute_as_ResizeAttribute(); } @@ -2327,44 +2365,55 @@ template<> inline const tosa::NegateAttribute *TosaOperator::attribute_as<tosa:: return attribute_as_NegateAttribute(); } +template<> inline const tosa::CustomAttribute *TosaOperator::attribute_as<tosa::CustomAttribute>() const { + return attribute_as_CustomAttribute(); +} + +template<> inline const tosa::FFTAttribute *TosaOperator::attribute_as<tosa::FFTAttribute>() const { + return attribute_as_FFTAttribute(); +} + +template<> inline const tosa::RFFTAttribute *TosaOperator::attribute_as<tosa::RFFTAttribute>() const { + return attribute_as_RFFTAttribute(); +} + struct TosaOperatorBuilder { typedef TosaOperator Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; void add_op(tosa::Op op) { fbb_.AddElement<uint32_t>(TosaOperator::VT_OP, static_cast<uint32_t>(op), 0); } void add_attribute_type(tosa::Attribute attribute_type) { fbb_.AddElement<uint8_t>(TosaOperator::VT_ATTRIBUTE_TYPE, static_cast<uint8_t>(attribute_type), 0); } - void add_attribute(flatbuffers::Offset<void> attribute) { + void add_attribute(::flatbuffers::Offset<void> attribute) { fbb_.AddOffset(TosaOperator::VT_ATTRIBUTE, attribute); } - void add_inputs(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> inputs) { + void add_inputs(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> inputs) { fbb_.AddOffset(TosaOperator::VT_INPUTS, inputs); } - void add_outputs(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> outputs) { + void add_outputs(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> outputs) { fbb_.AddOffset(TosaOperator::VT_OUTPUTS, outputs); } - explicit TosaOperatorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit TosaOperatorBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TosaOperatorBuilder &operator=(const TosaOperatorBuilder &); - flatbuffers::Offset<TosaOperator> Finish() { + ::flatbuffers::Offset<TosaOperator> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TosaOperator>(end); + auto o = ::flatbuffers::Offset<TosaOperator>(end); return o; } }; -inline flatbuffers::Offset<TosaOperator> CreateTosaOperator( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TosaOperator> CreateTosaOperator( + ::flatbuffers::FlatBufferBuilder &_fbb, tosa::Op op = tosa::Op_UNKNOWN, tosa::Attribute attribute_type = tosa::Attribute_NONE, - flatbuffers::Offset<void> attribute = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> inputs = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> outputs = 0) { + ::flatbuffers::Offset<void> attribute = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> inputs = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> outputs = 0) { TosaOperatorBuilder builder_(_fbb); builder_.add_outputs(outputs); builder_.add_inputs(inputs); @@ -2374,15 +2423,15 @@ inline flatbuffers::Offset<TosaOperator> CreateTosaOperator( return builder_.Finish(); } -inline flatbuffers::Offset<TosaOperator> CreateTosaOperatorDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TosaOperator> CreateTosaOperatorDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, tosa::Op op = tosa::Op_UNKNOWN, tosa::Attribute attribute_type = tosa::Attribute_NONE, - flatbuffers::Offset<void> attribute = 0, - const std::vector<flatbuffers::Offset<flatbuffers::String>> *inputs = nullptr, - const std::vector<flatbuffers::Offset<flatbuffers::String>> *outputs = nullptr) { - auto inputs__ = inputs ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*inputs) : 0; - auto outputs__ = outputs ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*outputs) : 0; + ::flatbuffers::Offset<void> attribute = 0, + const std::vector<::flatbuffers::Offset<::flatbuffers::String>> *inputs = nullptr, + const std::vector<::flatbuffers::Offset<::flatbuffers::String>> *outputs = nullptr) { + auto inputs__ = inputs ? _fbb.CreateVector<::flatbuffers::Offset<::flatbuffers::String>>(*inputs) : 0; + auto outputs__ = outputs ? _fbb.CreateVector<::flatbuffers::Offset<::flatbuffers::String>>(*outputs) : 0; return tosa::CreateTosaOperator( _fbb, op, @@ -2392,7 +2441,7 @@ inline flatbuffers::Offset<TosaOperator> CreateTosaOperatorDirect( outputs__); } -struct TosaBasicBlock FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TosaBasicBlock FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TosaBasicBlockBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_NAME = 4, @@ -2401,22 +2450,22 @@ struct TosaBasicBlock FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { VT_INPUTS = 10, VT_OUTPUTS = 12 }; - const flatbuffers::String *name() const { - return GetPointer<const flatbuffers::String *>(VT_NAME); + const ::flatbuffers::String *name() const { + return GetPointer<const ::flatbuffers::String *>(VT_NAME); } - const flatbuffers::Vector<flatbuffers::Offset<tosa::TosaOperator>> *operators() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<tosa::TosaOperator>> *>(VT_OPERATORS); + const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaOperator>> *operators() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaOperator>> *>(VT_OPERATORS); } - const flatbuffers::Vector<flatbuffers::Offset<tosa::TosaTensor>> *tensors() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<tosa::TosaTensor>> *>(VT_TENSORS); + const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaTensor>> *tensors() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaTensor>> *>(VT_TENSORS); } - const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *inputs() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *>(VT_INPUTS); + const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *inputs() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *>(VT_INPUTS); } - const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *outputs() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>> *>(VT_OUTPUTS); + const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *outputs() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>> *>(VT_OUTPUTS); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_NAME) && verifier.VerifyString(name()) && @@ -2438,42 +2487,41 @@ struct TosaBasicBlock FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { struct TosaBasicBlockBuilder { typedef TosaBasicBlock Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_name(flatbuffers::Offset<flatbuffers::String> name) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_name(::flatbuffers::Offset<::flatbuffers::String> name) { fbb_.AddOffset(TosaBasicBlock::VT_NAME, name); } - void add_operators(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<tosa::TosaOperator>>> operators) { + void add_operators(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaOperator>>> operators) { fbb_.AddOffset(TosaBasicBlock::VT_OPERATORS, operators); } - void add_tensors(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<tosa::TosaTensor>>> tensors) { + void add_tensors(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaTensor>>> tensors) { fbb_.AddOffset(TosaBasicBlock::VT_TENSORS, tensors); } - void add_inputs(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> inputs) { + void add_inputs(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> inputs) { fbb_.AddOffset(TosaBasicBlock::VT_INPUTS, inputs); } - void add_outputs(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> outputs) { + void add_outputs(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> outputs) { fbb_.AddOffset(TosaBasicBlock::VT_OUTPUTS, outputs); } - explicit TosaBasicBlockBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit TosaBasicBlockBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TosaBasicBlockBuilder &operator=(const TosaBasicBlockBuilder &); - flatbuffers::Offset<TosaBasicBlock> Finish() { + ::flatbuffers::Offset<TosaBasicBlock> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TosaBasicBlock>(end); + auto o = ::flatbuffers::Offset<TosaBasicBlock>(end); return o; } }; -inline flatbuffers::Offset<TosaBasicBlock> CreateTosaBasicBlock( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<flatbuffers::String> name = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<tosa::TosaOperator>>> operators = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<tosa::TosaTensor>>> tensors = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> inputs = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>> outputs = 0) { +inline ::flatbuffers::Offset<TosaBasicBlock> CreateTosaBasicBlock( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::String> name = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaOperator>>> operators = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaTensor>>> tensors = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> inputs = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<::flatbuffers::String>>> outputs = 0) { TosaBasicBlockBuilder builder_(_fbb); builder_.add_outputs(outputs); builder_.add_inputs(inputs); @@ -2483,18 +2531,18 @@ inline flatbuffers::Offset<TosaBasicBlock> CreateTosaBasicBlock( return builder_.Finish(); } -inline flatbuffers::Offset<TosaBasicBlock> CreateTosaBasicBlockDirect( - flatbuffers::FlatBufferBuilder &_fbb, +inline ::flatbuffers::Offset<TosaBasicBlock> CreateTosaBasicBlockDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, const char *name = nullptr, - const std::vector<flatbuffers::Offset<tosa::TosaOperator>> *operators = nullptr, - const std::vector<flatbuffers::Offset<tosa::TosaTensor>> *tensors = nullptr, - const std::vector<flatbuffers::Offset<flatbuffers::String>> *inputs = nullptr, - const std::vector<flatbuffers::Offset<flatbuffers::String>> *outputs = nullptr) { + const std::vector<::flatbuffers::Offset<tosa::TosaOperator>> *operators = nullptr, + const std::vector<::flatbuffers::Offset<tosa::TosaTensor>> *tensors = nullptr, + const std::vector<::flatbuffers::Offset<::flatbuffers::String>> *inputs = nullptr, + const std::vector<::flatbuffers::Offset<::flatbuffers::String>> *outputs = nullptr) { auto name__ = name ? _fbb.CreateString(name) : 0; - auto operators__ = operators ? _fbb.CreateVector<flatbuffers::Offset<tosa::TosaOperator>>(*operators) : 0; - auto tensors__ = tensors ? _fbb.CreateVector<flatbuffers::Offset<tosa::TosaTensor>>(*tensors) : 0; - auto inputs__ = inputs ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*inputs) : 0; - auto outputs__ = outputs ? _fbb.CreateVector<flatbuffers::Offset<flatbuffers::String>>(*outputs) : 0; + auto operators__ = operators ? _fbb.CreateVector<::flatbuffers::Offset<tosa::TosaOperator>>(*operators) : 0; + auto tensors__ = tensors ? _fbb.CreateVector<::flatbuffers::Offset<tosa::TosaTensor>>(*tensors) : 0; + auto inputs__ = inputs ? _fbb.CreateVector<::flatbuffers::Offset<::flatbuffers::String>>(*inputs) : 0; + auto outputs__ = outputs ? _fbb.CreateVector<::flatbuffers::Offset<::flatbuffers::String>>(*outputs) : 0; return tosa::CreateTosaBasicBlock( _fbb, name__, @@ -2504,73 +2552,139 @@ inline flatbuffers::Offset<TosaBasicBlock> CreateTosaBasicBlockDirect( outputs__); } -struct TosaGraph FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { +struct TosaRegion FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { + typedef TosaRegionBuilder Builder; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NAME = 4, + VT_BLOCKS = 6 + }; + const ::flatbuffers::String *name() const { + return GetPointer<const ::flatbuffers::String *>(VT_NAME); + } + const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaBasicBlock>> *blocks() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaBasicBlock>> *>(VT_BLOCKS); + } + bool Verify(::flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_NAME) && + verifier.VerifyString(name()) && + VerifyOffset(verifier, VT_BLOCKS) && + verifier.VerifyVector(blocks()) && + verifier.VerifyVectorOfTables(blocks()) && + verifier.EndTable(); + } +}; + +struct TosaRegionBuilder { + typedef TosaRegion Table; + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_name(::flatbuffers::Offset<::flatbuffers::String> name) { + fbb_.AddOffset(TosaRegion::VT_NAME, name); + } + void add_blocks(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaBasicBlock>>> blocks) { + fbb_.AddOffset(TosaRegion::VT_BLOCKS, blocks); + } + explicit TosaRegionBuilder(::flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ::flatbuffers::Offset<TosaRegion> Finish() { + const auto end = fbb_.EndTable(start_); + auto o = ::flatbuffers::Offset<TosaRegion>(end); + return o; + } +}; + +inline ::flatbuffers::Offset<TosaRegion> CreateTosaRegion( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<::flatbuffers::String> name = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaBasicBlock>>> blocks = 0) { + TosaRegionBuilder builder_(_fbb); + builder_.add_blocks(blocks); + builder_.add_name(name); + return builder_.Finish(); +} + +inline ::flatbuffers::Offset<TosaRegion> CreateTosaRegionDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + const char *name = nullptr, + const std::vector<::flatbuffers::Offset<tosa::TosaBasicBlock>> *blocks = nullptr) { + auto name__ = name ? _fbb.CreateString(name) : 0; + auto blocks__ = blocks ? _fbb.CreateVector<::flatbuffers::Offset<tosa::TosaBasicBlock>>(*blocks) : 0; + return tosa::CreateTosaRegion( + _fbb, + name__, + blocks__); +} + +struct TosaGraph FLATBUFFERS_FINAL_CLASS : private ::flatbuffers::Table { typedef TosaGraphBuilder Builder; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { VT_VERSION = 4, - VT_BLOCKS = 6 + VT_REGIONS = 6 }; const tosa::Version *version() const { return GetPointer<const tosa::Version *>(VT_VERSION); } - const flatbuffers::Vector<flatbuffers::Offset<tosa::TosaBasicBlock>> *blocks() const { - return GetPointer<const flatbuffers::Vector<flatbuffers::Offset<tosa::TosaBasicBlock>> *>(VT_BLOCKS); + const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaRegion>> *regions() const { + return GetPointer<const ::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaRegion>> *>(VT_REGIONS); } - bool Verify(flatbuffers::Verifier &verifier) const { + bool Verify(::flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_VERSION) && + VerifyOffsetRequired(verifier, VT_VERSION) && verifier.VerifyTable(version()) && - VerifyOffset(verifier, VT_BLOCKS) && - verifier.VerifyVector(blocks()) && - verifier.VerifyVectorOfTables(blocks()) && + VerifyOffset(verifier, VT_REGIONS) && + verifier.VerifyVector(regions()) && + verifier.VerifyVectorOfTables(regions()) && verifier.EndTable(); } }; struct TosaGraphBuilder { typedef TosaGraph Table; - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_version(flatbuffers::Offset<tosa::Version> version) { + ::flatbuffers::FlatBufferBuilder &fbb_; + ::flatbuffers::uoffset_t start_; + void add_version(::flatbuffers::Offset<tosa::Version> version) { fbb_.AddOffset(TosaGraph::VT_VERSION, version); } - void add_blocks(flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<tosa::TosaBasicBlock>>> blocks) { - fbb_.AddOffset(TosaGraph::VT_BLOCKS, blocks); + void add_regions(::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaRegion>>> regions) { + fbb_.AddOffset(TosaGraph::VT_REGIONS, regions); } - explicit TosaGraphBuilder(flatbuffers::FlatBufferBuilder &_fbb) + explicit TosaGraphBuilder(::flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); } - TosaGraphBuilder &operator=(const TosaGraphBuilder &); - flatbuffers::Offset<TosaGraph> Finish() { + ::flatbuffers::Offset<TosaGraph> Finish() { const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset<TosaGraph>(end); + auto o = ::flatbuffers::Offset<TosaGraph>(end); + fbb_.Required(o, TosaGraph::VT_VERSION); return o; } }; -inline flatbuffers::Offset<TosaGraph> CreateTosaGraph( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<tosa::Version> version = 0, - flatbuffers::Offset<flatbuffers::Vector<flatbuffers::Offset<tosa::TosaBasicBlock>>> blocks = 0) { +inline ::flatbuffers::Offset<TosaGraph> CreateTosaGraph( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<tosa::Version> version = 0, + ::flatbuffers::Offset<::flatbuffers::Vector<::flatbuffers::Offset<tosa::TosaRegion>>> regions = 0) { TosaGraphBuilder builder_(_fbb); - builder_.add_blocks(blocks); + builder_.add_regions(regions); builder_.add_version(version); return builder_.Finish(); } -inline flatbuffers::Offset<TosaGraph> CreateTosaGraphDirect( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset<tosa::Version> version = 0, - const std::vector<flatbuffers::Offset<tosa::TosaBasicBlock>> *blocks = nullptr) { - auto blocks__ = blocks ? _fbb.CreateVector<flatbuffers::Offset<tosa::TosaBasicBlock>>(*blocks) : 0; +inline ::flatbuffers::Offset<TosaGraph> CreateTosaGraphDirect( + ::flatbuffers::FlatBufferBuilder &_fbb, + ::flatbuffers::Offset<tosa::Version> version = 0, + const std::vector<::flatbuffers::Offset<tosa::TosaRegion>> *regions = nullptr) { + auto regions__ = regions ? _fbb.CreateVector<::flatbuffers::Offset<tosa::TosaRegion>>(*regions) : 0; return tosa::CreateTosaGraph( _fbb, version, - blocks__); + regions__); } -inline bool VerifyAttribute(flatbuffers::Verifier &verifier, const void *obj, Attribute type) { +inline bool VerifyAttribute(::flatbuffers::Verifier &verifier, const void *obj, Attribute type) { switch (type) { case Attribute_NONE: { return true; @@ -2595,18 +2709,6 @@ inline bool VerifyAttribute(flatbuffers::Verifier &verifier, const void *obj, At auto ptr = reinterpret_cast<const tosa::AxisAttribute *>(obj); return verifier.VerifyTable(ptr); } - case Attribute_ReshapeAttribute: { - auto ptr = reinterpret_cast<const tosa::ReshapeAttribute *>(obj); - return verifier.VerifyTable(ptr); - } - case Attribute_SliceAttribute: { - auto ptr = reinterpret_cast<const tosa::SliceAttribute *>(obj); - return verifier.VerifyTable(ptr); - } - case Attribute_TileAttribute: { - auto ptr = reinterpret_cast<const tosa::TileAttribute *>(obj); - return verifier.VerifyTable(ptr); - } case Attribute_ResizeAttribute: { auto ptr = reinterpret_cast<const tosa::ResizeAttribute *>(obj); return verifier.VerifyTable(ptr); @@ -2655,14 +2757,26 @@ inline bool VerifyAttribute(flatbuffers::Verifier &verifier, const void *obj, At auto ptr = reinterpret_cast<const tosa::NegateAttribute *>(obj); return verifier.VerifyTable(ptr); } + case Attribute_CustomAttribute: { + auto ptr = reinterpret_cast<const tosa::CustomAttribute *>(obj); + return verifier.VerifyTable(ptr); + } + case Attribute_FFTAttribute: { + auto ptr = reinterpret_cast<const tosa::FFTAttribute *>(obj); + return verifier.VerifyTable(ptr); + } + case Attribute_RFFTAttribute: { + auto ptr = reinterpret_cast<const tosa::RFFTAttribute *>(obj); + return verifier.VerifyTable(ptr); + } default: return true; } } -inline bool VerifyAttributeVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector<flatbuffers::Offset<void>> *values, const flatbuffers::Vector<uint8_t> *types) { +inline bool VerifyAttributeVector(::flatbuffers::Verifier &verifier, const ::flatbuffers::Vector<::flatbuffers::Offset<void>> *values, const ::flatbuffers::Vector<uint8_t> *types) { if (!values || !types) return !values && !types; if (values->size() != types->size()) return false; - for (flatbuffers::uoffset_t i = 0; i < values->size(); ++i) { + for (::flatbuffers::uoffset_t i = 0; i < values->size(); ++i) { if (!VerifyAttribute( verifier, values->Get(i), types->GetEnum<Attribute>(i))) { return false; @@ -2672,11 +2786,11 @@ inline bool VerifyAttributeVector(flatbuffers::Verifier &verifier, const flatbuf } inline const tosa::TosaGraph *GetTosaGraph(const void *buf) { - return flatbuffers::GetRoot<tosa::TosaGraph>(buf); + return ::flatbuffers::GetRoot<tosa::TosaGraph>(buf); } inline const tosa::TosaGraph *GetSizePrefixedTosaGraph(const void *buf) { - return flatbuffers::GetSizePrefixedRoot<tosa::TosaGraph>(buf); + return ::flatbuffers::GetSizePrefixedRoot<tosa::TosaGraph>(buf); } inline const char *TosaGraphIdentifier() { @@ -2684,17 +2798,22 @@ inline const char *TosaGraphIdentifier() { } inline bool TosaGraphBufferHasIdentifier(const void *buf) { - return flatbuffers::BufferHasIdentifier( + return ::flatbuffers::BufferHasIdentifier( buf, TosaGraphIdentifier()); } +inline bool SizePrefixedTosaGraphBufferHasIdentifier(const void *buf) { + return ::flatbuffers::BufferHasIdentifier( + buf, TosaGraphIdentifier(), true); +} + inline bool VerifyTosaGraphBuffer( - flatbuffers::Verifier &verifier) { + ::flatbuffers::Verifier &verifier) { return verifier.VerifyBuffer<tosa::TosaGraph>(TosaGraphIdentifier()); } inline bool VerifySizePrefixedTosaGraphBuffer( - flatbuffers::Verifier &verifier) { + ::flatbuffers::Verifier &verifier) { return verifier.VerifySizePrefixedBuffer<tosa::TosaGraph>(TosaGraphIdentifier()); } @@ -2703,14 +2822,14 @@ inline const char *TosaGraphExtension() { } inline void FinishTosaGraphBuffer( - flatbuffers::FlatBufferBuilder &fbb, - flatbuffers::Offset<tosa::TosaGraph> root) { + ::flatbuffers::FlatBufferBuilder &fbb, + ::flatbuffers::Offset<tosa::TosaGraph> root) { fbb.Finish(root, TosaGraphIdentifier()); } inline void FinishSizePrefixedTosaGraphBuffer( - flatbuffers::FlatBufferBuilder &fbb, - flatbuffers::Offset<tosa::TosaGraph> root) { + ::flatbuffers::FlatBufferBuilder &fbb, + ::flatbuffers::Offset<tosa::TosaGraph> root) { fbb.FinishSizePrefixed(root, TosaGraphIdentifier()); } diff --git a/include/tosa_serialization_handler.h b/include/tosa_serialization_handler.h index 53dcf1a..f5f9e58 100644 --- a/include/tosa_serialization_handler.h +++ b/include/tosa_serialization_handler.h @@ -1,5 +1,5 @@ -// Copyright (c) 2020-2021, ARM Limited. +// Copyright (c) 2020-2024, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -18,6 +18,7 @@ #include "attribute.h" #include "flatbuffers/idl.h" #include "flatbuffers/util.h" +#include "float_utils.h" #include "numpy_utils.h" #include "tosa_generated.h" #include <cstdint> @@ -27,7 +28,7 @@ // Keep version number in sync with the version default value with schema/tosa.fbs #define TOSA_VERSION_MAJOR 0 -#define TOSA_VERSION_MINOR 31 +#define TOSA_VERSION_MINOR 100 #define TOSA_VERSION_PATCH 0 #define TOSA_VERSION_DRAFT true #define TENSOR_BUFFER_FORCE_ALIGNMENT 8 @@ -57,7 +58,7 @@ struct TosaVersion enum class compat_t { COMPLETELY_COMPATIBLE, - PARTIALLY_COMPATIBLE, + BACKWARD_COMPATIBLE, NOT_COMPATIBLE }; @@ -86,17 +87,53 @@ struct TosaVersion return str; } - compat_t is_compatible(const TosaVersion& rhs) const + static bool less_than(const TosaVersion& version1, const TosaVersion& version2) { - if (rhs._major == _major && rhs._minor == _minor) + if (version1._major < version2._major) { - if (rhs._patch == _patch && rhs._draft == _draft) + return true; + } + else if (version1._major == version2._major) + { + if (version1._minor < version2._minor) { - return TosaVersion::compat_t::COMPLETELY_COMPATIBLE; + return true; } - else + else if (version1._minor == version2._minor) { - return TosaVersion::compat_t::PARTIALLY_COMPATIBLE; + if (version1._patch < version2._patch) + { + return true; + } + else if (version1._patch == version2._patch) + { + if (version1._draft == true && version2._draft == false) + { + return true; + } + } + } + } + return false; + } + + static TosaVersion::compat_t is_compatible(const TosaVersion& tosa_fb_version, + const TosaVersion& serializer_version) + { + bool major_match = (serializer_version._major == tosa_fb_version._major); + bool minor_match = (serializer_version._minor == tosa_fb_version._minor); + bool patch_match = (serializer_version._patch == tosa_fb_version._patch); + bool draft_match = (serializer_version._draft == tosa_fb_version._draft); + + if (major_match && minor_match && patch_match && draft_match) + return TosaVersion::compat_t::COMPLETELY_COMPATIBLE; + + // We currently support backward compatibility starting from 0.100.0 + if ((tosa_fb_version._major == 0 && tosa_fb_version._minor >= 100) || (tosa_fb_version._major > 0)) + { + if (less_than(tosa_fb_version, serializer_version)) + { + return TosaVersion::compat_t::BACKWARD_COMPATIBLE; } } return TosaVersion::compat_t::NOT_COMPATIBLE; @@ -112,11 +149,17 @@ public: TosaSerializationTensor(const flatbuffers::String* name, const flatbuffers::Vector<int32_t>* shape, DType dtype, - const flatbuffers::Vector<uint8_t>* data); + const flatbuffers::Vector<uint8_t>* data, + const bool variable = false, + const bool is_unranked = false, + const flatbuffers::String* variable_name = NULL); TosaSerializationTensor(const std::string& name, const std::vector<int32_t>& shape, DType dtype, - const std::vector<uint8_t>& data); + const std::vector<uint8_t>& data, + const bool variable = false, + const bool is_unranked = false, + const std::string& variable_name = ""); TosaSerializationTensor(); ~TosaSerializationTensor(); @@ -129,14 +172,26 @@ public: { return _shape; } - DType GetDtype() + DType GetDtype() const { return _dtype; } + bool GetVariable() const + { + return _variable; + } const std::vector<uint8_t>& GetData() const { return _data; } + bool GetIsUnranked() const + { + return _is_unranked; + } + const std::string GetVariableName() const + { + return _variable_name; + } // modifier void SetDtype(DType dtype) @@ -155,12 +210,28 @@ public: { _data = std::move(data); } + void SetIsUnranked(const bool value) + { + _is_unranked = value; + } + void SetDimSize(size_t dim, uint32_t new_size) + { + if (dim >= _shape.size()) + { + printf("dim is out of bound\n"); + assert(0); + } + _shape[dim] = new_size; + } private: DType _dtype; /* data type enumeration, see tosa_isa_generated.h */ std::vector<int32_t> _shape; /* shape of the tensor */ std::string _name; /* name of the tensor, used for solving dependency */ + bool _variable; /* is this a variable tensor */ std::vector<uint8_t> _data; /* data array */ + bool _is_unranked; /* whether this is an unranked tensor */ + std::string _variable_name; /* name for variable tensors */ }; class TosaSerializationOperator @@ -216,11 +287,13 @@ class TosaSerializationBasicBlock public: // constructor and destructor TosaSerializationBasicBlock(const std::string& name, + const std::string& region_name, const std::vector<TosaSerializationOperator*>& operators, const std::vector<TosaSerializationTensor*>& tensors, const std::vector<std::string>& inputs, const std::vector<std::string>& outputs); TosaSerializationBasicBlock(std::string&& name, + std::string&& region_name, std::vector<TosaSerializationOperator*>&& operators, std::vector<TosaSerializationTensor*>&& tensors, std::vector<std::string>&& inputs, @@ -232,10 +305,15 @@ public: { return _name; } + std::string GetRegionName() const + { + return _region_name; + } std::vector<TosaSerializationOperator*>& GetOperators() { return _operators; } + std::vector<TosaSerializationTensor*>& GetTensors() { return _tensors; @@ -259,19 +337,59 @@ public: { return _inputs; } + std::vector<std::string>& GetOutputs() { return _outputs; } private: - std::string _name; /* name of basic block */ + std::string _name; /* name of basic block */ + std::string _region_name; std::vector<TosaSerializationOperator*> _operators; /* TosaSerializationOperator list */ std::vector<TosaSerializationTensor*> _tensors; /* TosaSerializationTensor list */ std::vector<std::string> _inputs; /* array of string to specify block inputs */ std::vector<std::string> _outputs; /* array of string to specify block outputs */ }; +class TosaSerializationRegion +{ +public: + // constructor and desctructor + TosaSerializationRegion(const std::string& name, const std::vector<TosaSerializationBasicBlock*>& blocks); + TosaSerializationRegion(const std::string&& name, const std::vector<TosaSerializationBasicBlock*>&& blocks); + ~TosaSerializationRegion(); + + // accessors + std::string GetName() const + { + return this->_name; + } + + std::vector<TosaSerializationBasicBlock*>& GetBlocks() + { + return this->_blocks; + } + + TosaSerializationBasicBlock* GetBlockByName(std::string name) + { + TosaSerializationBasicBlock* result = nullptr; + for (auto block : GetBlocks()) + { + if (block->GetName() == name) + { + result = block; + break; + } + } + return result; + } + +private: + std::string _name; /* name of basic block */ + std::vector<TosaSerializationBasicBlock*> _blocks; /* TosaSerializationBasicBlock list */ +}; + /* * this is a helper class for writing/reading Tosa ISA * supported format: .tosa (flatbuffer), .json @@ -294,7 +412,12 @@ public: tosa_err_t LoadFileSchema(const char* schema_filename); // data format conversion. little-endian. + static tosa_err_t ConvertBF16toU8(const std::vector<float>& in, std::vector<uint8_t>& out); + static tosa_err_t ConvertFP8E4M3toU8(const std::vector<float>& in, std::vector<uint8_t>& out); + static tosa_err_t ConvertFP8E5M2toU8(const std::vector<float>& in, std::vector<uint8_t>& out); + static tosa_err_t ConvertF16toU8(const std::vector<float>& in, std::vector<uint8_t>& out); static tosa_err_t ConvertF32toU8(const std::vector<float>& in, std::vector<uint8_t>& out); + static tosa_err_t ConvertI64toU8(const std::vector<int64_t>& in, std::vector<uint8_t>& out); static tosa_err_t ConvertI48toU8(const std::vector<int64_t>& in, std::vector<uint8_t>& out); static tosa_err_t ConvertI32toU8(const std::vector<int32_t>& in, std::vector<uint8_t>& out); static tosa_err_t ConvertI16toU8(const std::vector<int16_t>& in, std::vector<uint8_t>& out); @@ -302,7 +425,13 @@ public: static tosa_err_t ConvertI4toU8(const std::vector<int8_t>& in, std::vector<uint8_t>& out); static tosa_err_t ConvertBooltoU8(const std::vector<bool>& in, std::vector<uint8_t>& out); + static tosa_err_t ConvertU8toBF16(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<float>& out); + static tosa_err_t ConvertU8toFP8E4M3(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<float>& out); + static tosa_err_t ConvertU8toFP8E5M2(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<float>& out); + static tosa_err_t + ConvertU8toF16(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<half_float::half>& out); static tosa_err_t ConvertU8toF32(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<float>& out); + static tosa_err_t ConvertU8toI64(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<int64_t>& out); static tosa_err_t ConvertU8toI48(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<int64_t>& out); static tosa_err_t ConvertU8toI32(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<int32_t>& out); static tosa_err_t ConvertU8toI16(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<int16_t>& out); @@ -310,6 +439,8 @@ public: static tosa_err_t ConvertU8toI4(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<int8_t>& out); static tosa_err_t ConvertU8toBool(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<bool>& out); + static void ForceAlignTensorData(std::vector<uint8_t>& buf); + // version const TosaVersion& GetVersion() { @@ -317,39 +448,29 @@ public: } // accessor - std::vector<TosaSerializationBasicBlock*>& GetBlocks() + std::vector<TosaSerializationRegion*>& GetRegions() { - return _blocks; + return _regions; } - TosaSerializationBasicBlock* GetBlockByName(std::string name) + TosaSerializationRegion* GetMainRegion() { - TosaSerializationBasicBlock* result = nullptr; - for (auto block : GetBlocks()) + return _regions[0]; + } + + TosaSerializationRegion* GetRegionByName(std::string name) + { + TosaSerializationRegion* result = nullptr; + for (auto region : GetRegions()) { - if (block->GetName() == name) + if (region->GetName() == name) { - result = block; + result = region; break; } } return result; } - TosaSerializationBasicBlock* GetMainBlock() - { - TosaSerializationBasicBlock* main_block = GetBlockByName(std::string("main")); - assert(main_block); - return main_block; - } - - std::vector<std::string>& GetInputs() - { - return GetMainBlock()->GetInputs(); - } - std::vector<std::string>& GetOutputs() - { - return GetMainBlock()->GetOutputs(); - } bool GetSchemaLoaded() const { @@ -360,14 +481,13 @@ protected: tosa_err_t Clear(); tosa_err_t Deserialize(const uint8_t* buf); tosa_err_t Serialize(); - TosaVersion ParseTosaSchemaVersion(std::string schema); private: - TosaVersion _version; /* version struct */ - flatbuffers::FlatBufferBuilder _builder; /* flatbuffer builder */ - flatbuffers::Parser _parser; /* flatbuffer parser, used for json parsing */ - std::vector<TosaSerializationBasicBlock*> _blocks; /* array structure to store all TosaSerializationBasicBlock */ - bool _schemaLoaded; /* is the schema properly loaded? */ + TosaVersion _version; /* version struct */ + flatbuffers::FlatBufferBuilder _builder; /* flatbuffer builder */ + flatbuffers::Parser _parser; /* flatbuffer parser, used for json parsing */ + std::vector<TosaSerializationRegion*> _regions; /* array structure to store all TosaSerializationRegion */ + bool _schemaLoaded; /* is the schema properly loaded? */ }; } // namespace tosa diff --git a/pyproject.toml b/pyproject.toml new file mode 100644 index 0000000..6cfb8f6 --- /dev/null +++ b/pyproject.toml @@ -0,0 +1,43 @@ +# Copyright (c) 2023, ARM Limited. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# +# Description: +# Packaging for the TOSA serialization lib +# + +[build-system] +requires = [ + "setuptools>=42", + "wheel", + "setuptools_scm[toml]>=6.0" +] +build-backend = "setuptools.build_meta" + +[project] +name="tosa_serialization_lib" +description="TOSA serialization library" +authors=[ { "name" = "Arm Ltd" } ] +license={ "file" = "LICENSE.txt" } +dynamic = ["version"] + +[tool.setuptools] +packages = [ "tosa", "serializer" ] + +[tool.setuptools.package-dir] +tosa = "python/tosa" +serializer = "python/serializer" + +[tool.setuptools_scm] +version_scheme = "no-guess-dev" diff --git a/python/serializer/tosa_serializer.py b/python/serializer/tosa_serializer.py index ec1c12d..bbfb37e 100644 --- a/python/serializer/tosa_serializer.py +++ b/python/serializer/tosa_serializer.py @@ -1,4 +1,4 @@ -# Copyright (c) 2020-2022, ARM Limited. +# Copyright (c) 2020-2024, ARM Limited. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -13,6 +13,7 @@ # limitations under the License. import os +import serializer.tosa_serializer as ts import json import flatbuffers import numpy as np @@ -20,6 +21,7 @@ import struct from enum import IntEnum, unique from tosa import ( TosaGraph, + TosaRegion, TosaBasicBlock, TosaTensor, TosaOperator, @@ -30,7 +32,7 @@ import tosa.Op as TosaOp # Keep version number in sync with the version default value with schema/tosa.fbs TOSA_VERSION_MAJOR = 0 -TOSA_VERSION_MINOR = 31 +TOSA_VERSION_MINOR = 100 TOSA_VERSION_PATCH = 0 TOSA_VERSION_DRAFT = True TOSA_VERSION = [ @@ -56,8 +58,13 @@ DTypeNames = [ "INT16", "INT32", "INT48", - "FLOAT", + "FP32", "UINT16", + "FP16", + "BF16", + "SHAPE", + "FP8E4M3", + "FP8E5M2", ] ByteMask = np.uint64(0xFF) @@ -90,6 +97,7 @@ class TosaSerializerUnion: self.bools = [] self.floats = [] self.strings = [] + self.int16vecs = [] self.intvecs = [] self.fpvecs = [] @@ -106,6 +114,9 @@ class TosaSerializerUnion: for fcn, val in self.intvecs: intVecList.append((fcn, TosaSerializer.serializeInt32Vec(builder, val))) + for fcn, val in self.int16vecs: + intVecList.append((fcn, TosaSerializer.serializeInt16Vec(builder, val))) + for fcn, val in self.fpvecs: fpVecList.append((fcn, TosaSerializer.serializeFpVec(builder, val))) @@ -141,7 +152,15 @@ class TosaSerializerAttribute(TosaSerializerUnion): def __init__(self): super().__init__() - def PoolAttribute(self, kernel, stride, pad, input_zp, output_zp): + def PoolAttribute( + self, + kernel, + stride, + pad, + input_zp, + output_zp, + acc_type, + ): from tosa import PoolAttribute as a, Attribute self.utype = Attribute.Attribute().PoolAttribute @@ -152,8 +171,11 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.intvecs.append((a.AddStride, stride)) self.ints.append((a.AddInputZp, input_zp)) self.ints.append((a.AddOutputZp, output_zp)) + self.ints.append((a.AddAccType, acc_type)) - def ConvAttribute(self, pad, stride, dilation, input_zp, weight_zp): + def ConvAttribute( + self, pad, stride, dilation, input_zp, weight_zp, local_bound, acc_type + ): from tosa import ConvAttribute as a, Attribute self.utype = Attribute.Attribute().ConvAttribute @@ -164,8 +186,12 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.intvecs.append((a.AddDilation, dilation)) self.ints.append((a.AddInputZp, input_zp)) self.ints.append((a.AddWeightZp, weight_zp)) + self.bools.append((a.AddLocalBound, local_bound)) + self.ints.append((a.AddAccType, acc_type)) - def TransposeConvAttribute(self, outpad, stride, output_shape, input_zp, weight_zp): + def TransposeConvAttribute( + self, outpad, stride, output_shape, input_zp, weight_zp, local_bound, acc_type + ): from tosa import TransposeConvAttribute as a, Attribute self.utype = Attribute.Attribute().TransposeConvAttribute @@ -176,16 +202,21 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.intvecs.append((a.AddOutputShape, output_shape)) self.ints.append((a.AddInputZp, input_zp)) self.ints.append((a.AddWeightZp, weight_zp)) + self.bools.append((a.AddLocalBound, local_bound)) + self.ints.append((a.AddAccType, acc_type)) - def PadAttribute(self, padding, pad_const_int, pad_const_fp): + def PadAttribute(self, serializer_builder, pad_const_val_as_bytes): from tosa import PadAttribute as a, Attribute self.utype = Attribute.Attribute().PadAttribute self.optFcns = (a.Start, a.End) - self.intvecs.append((a.AddPadding, padding)) - self.ints.append((a.AddPadConstInt, pad_const_int)) - self.floats.append((a.AddPadConstFp, pad_const_fp)) + # serialize pad_const_val_as_bytes as uint8 vector + serialized_pad_const_val = ts.TosaSerializer.serializeUint8Vec( + serializer_builder, pad_const_val_as_bytes + ) + + self.floats.append((a.AddPadConst, serialized_pad_const_val)) def AxisAttribute(self, axis): from tosa import AxisAttribute as a, Attribute @@ -195,61 +226,43 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.ints.append((a.AddAxis, axis)) - def ReshapeAttribute(self, new_shape): - from tosa import ReshapeAttribute as a, Attribute - - self.utype = Attribute.Attribute().ReshapeAttribute - self.optFcns = (a.Start, a.End) - - self.intvecs.append((a.AddNewShape, new_shape)) - - def SliceAttribute(self, start, size): - from tosa import SliceAttribute as a, Attribute - - self.utype = Attribute.Attribute().SliceAttribute - self.optFcns = (a.Start, a.End) - - self.intvecs.append((a.AddStart, start)) - self.intvecs.append((a.AddSize, size)) - - def TileAttribute(self, multiples): - from tosa import TileAttribute as a, Attribute - - self.utype = Attribute.Attribute().TileAttribute - self.optFcns = (a.Start, a.End) - - self.intvecs.append((a.AddMultiples, multiples)) - - def ResizeAttribute( - self, output_size, stride, offset, shift, stride_fp, offset_fp, mode - ): + def ResizeAttribute(self, scale, offset, border, mode): from tosa import ResizeAttribute as a, Attribute self.utype = Attribute.Attribute().ResizeAttribute self.optFcns = (a.Start, a.End) - self.intvecs.append((a.AddOutputSize, output_size)) - self.intvecs.append((a.AddStride, stride)) - self.intvecs.append((a.AddOffset, offset)) - self.ints.append((a.AddShift, shift)) - self.fpvecs.append((a.AddStrideFp, stride_fp)) - self.fpvecs.append((a.AddOffsetFp, offset_fp)) + self.int16vecs.append((a.AddScale, scale)) + self.int16vecs.append((a.AddOffset, offset)) + self.int16vecs.append((a.AddBorder, border)) self.ints.append((a.AddMode, mode)) - def ClampAttribute(self, minint, maxint, minfp, maxfp): + def ClampAttribute(self, serializer_builder, min_val_as_bytes, max_val_as_bytes): from tosa import ClampAttribute as a, Attribute self.utype = Attribute.Attribute().ClampAttribute self.optFcns = (a.Start, a.End) - self.ints.append((a.AddMinInt, minint)) - self.ints.append((a.AddMaxInt, maxint)) + # min/max float attributes serialized as uint8 vectors + serialized_min_val = ts.TosaSerializer.serializeUint8Vec( + serializer_builder, min_val_as_bytes + ) + serialized_max_val = ts.TosaSerializer.serializeUint8Vec( + serializer_builder, max_val_as_bytes + ) - self.ints.append((a.AddMinFp, minfp)) - self.ints.append((a.AddMaxFp, maxfp)) + self.floats.append((a.AddMinVal, serialized_min_val)) + self.floats.append((a.AddMaxVal, serialized_max_val)) def RescaleAttribute( - self, input_zp, output_zp, multiplier, shift, scale32, double_round, per_channel + self, + input_zp, + output_zp, + scale32, + double_round, + per_channel, + input_unsigned, + output_unsigned, ): from tosa import RescaleAttribute as a, Attribute @@ -258,11 +271,11 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.ints.append((a.AddInputZp, input_zp)) self.ints.append((a.AddOutputZp, output_zp)) - self.intvecs.append((a.AddMultiplier, multiplier)) - self.intvecs.append((a.AddShift, shift)) self.bools.append((a.AddScale32, scale32)) self.bools.append((a.AddDoubleRound, double_round)) self.bools.append((a.AddPerChannel, per_channel)) + self.bools.append((a.AddInputUnsigned, input_unsigned)) + self.bools.append((a.AddOutputUnsigned, output_unsigned)) def MulAttribute(self, shift): from tosa import MulAttribute as a, Attribute @@ -283,23 +296,23 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.bools.append((a.AddRound, round)) - def CondIfAttribute(self, then_branch, else_branch): + def CondIfAttribute(self, then_graph, else_graph): from tosa import CondIfAttribute as a, Attribute self.utype = Attribute.Attribute().CondIfAttribute self.optFcns = (a.Start, a.End) - self.strings.append((a.AddThenBranch, then_branch)) - self.strings.append((a.AddElseBranch, else_branch)) + self.strings.append((a.AddThenGraph, then_graph)) + self.strings.append((a.AddElseGraph, else_graph)) - def WhileLoopAttribute(self, cond_branch, body_branch): + def WhileLoopAttribute(self, cond_graph, body_graph): from tosa import WhileLoopAttribute as a, Attribute self.utype = Attribute.Attribute().WhileLoopAttribute self.optFcns = (a.Start, a.End) - self.strings.append((a.AddCondBranch, cond_branch)) - self.strings.append((a.AddBodyBranch, body_branch)) + self.strings.append((a.AddCondGraph, cond_graph)) + self.strings.append((a.AddBodyGraph, body_graph)) def TransposeAttribute(self, perms): from tosa import TransposeAttribute as a, Attribute @@ -315,7 +328,7 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.utype = Attribute.Attribute().TableAttribute self.optFcns = (a.Start, a.End) - self.intvecs.append((a.AddTable, table)) + self.int16vecs.append((a.AddTable, table)) def MatMulAttribute(self, A_zp, B_zp): from tosa import MatMulAttribute as a, Attribute @@ -344,6 +357,23 @@ class TosaSerializerAttribute(TosaSerializerUnion): self.ints.append((a.AddInput1Zp, input1_zp)) self.ints.append((a.AddOutputZp, output_zp)) + def FFTAttribute(self, inverse, local_bound): + from tosa import FFTAttribute as a, Attribute + + self.utype = Attribute.Attribute().FFTAttribute + self.optFcns = (a.Start, a.End) + + self.bools.append((a.AddInverse, inverse)) + self.bools.append((a.AddLocalBound, local_bound)) + + def RFFTAttribute(self, local_bound): + from tosa import RFFTAttribute as a, Attribute + + self.utype = Attribute.Attribute().RFFTAttribute + self.optFcns = (a.Start, a.End) + + self.bools.append((a.AddLocalBound, local_bound)) + class TosaSerializerTensor: def __init__( @@ -363,15 +393,30 @@ class TosaSerializerTensor: self.shape = shape self.dtype = dtype + if ( + dtype == DType.FP32 + or dtype == DType.BF16 + or dtype == DType.FP8E4M3 + or dtype == DType.FP8E5M2 + ): + fntype = np.float32 + elif dtype == DType.FP16: + fntype = np.float16 + else: + fntype = int + if isinstance(data, np.ndarray): - data = data.flatten().astype(int).tolist() - data = list(map(int, data)) - self.data = data + data = data.flatten().astype(fntype).tolist() + data = list(map(fntype, data)) elif isinstance(data, list): - data = list(map(int, data)) - self.data = data + data = list(map(fntype, data)) + elif data is not None: + # Assume data is rank 0 data type + data = list(map(fntype, [data])) else: - self.data = None + data = None + + self.data = data # Filename for placeholder tensors. These get generated by the test generation # process and are written to disk, but are considered input tensors by the @@ -381,12 +426,12 @@ class TosaSerializerTensor: self.placeholderFilename = placeholderFilename def __str__(self): - str = "TosaSerializerTensor name: {} shape: {} dtype: {}".format( + concatString = "TosaSerializerTensor name: {} shape: {} dtype: {}".format( self.name, self.shape, DTypeNames[self.dtype], ) - return str + return concatString def setDtype(self, dtype): self.dtype = dtype @@ -395,60 +440,7 @@ class TosaSerializerTensor: fb_name = builder.CreateString(self.name) fb_shapes = TosaSerializer.serializeInt32Vec(builder, self.shape) if self.data: - u8_data = list() - # little endianess - if self.dtype == DType.BOOL: - for val in self.data: - val_u8 = np.uint8(val) - u8_data.append(val_u8) - elif self.dtype == DType.INT4: - in_size = len(self.data) - out_size = (in_size + 1) // 2 - for i in range(out_size): - val_0 = self.data[2 * i] - if (2 * i + 1) < in_size: - val_1 = self.data[2 * i + 1] - else: - val_1 = 0 - val_i8 = (val_0 & 0xF) | ((val_1 & 0xF) << 4) - val_u8 = np.uint8(val_i8) - u8_data.append(val_u8) - elif self.dtype == DType.INT8: - for val in self.data: - val_u8 = np.uint8(val) - u8_data.append(val_u8) - elif self.dtype == DType.INT16: - for val in self.data: - val_u16 = np.uint16(val) - b0 = val_u16 & ByteMask - b1 = (val_u16 >> np.uint16(8)) & ByteMask - u8_data.extend([b0, b1]) - elif self.dtype == DType.INT32: - for val in self.data: - val_u32 = np.uint32(val) - b0 = val_u32 & ByteMask - b1 = (val_u32 >> np.uint32(8)) & ByteMask - b2 = (val_u32 >> np.uint32(16)) & ByteMask - b3 = (val_u32 >> np.uint32(24)) & ByteMask - u8_data.extend([b0, b1, b2, b3]) - elif self.dtype == DType.INT48: - for val in self.data: - val_u64 = np.uint64(val) - b0 = val_u64 & ByteMask - b1 = (val_u64 >> np.uint64(8)) & ByteMask - b2 = (val_u64 >> np.uint64(16)) & ByteMask - b3 = (val_u64 >> np.uint64(24)) & ByteMask - b4 = (val_u64 >> np.uint64(32)) & ByteMask - b5 = (val_u64 >> np.uint64(40)) & ByteMask - u8_data.extend([b0, b1, b2, b3, b4, b5]) - elif self.dtype == DType.FLOAT: - for val in self.data: - b = struct.pack("!f", val) - u8_data.extend([b[3], b[2], b[1], b[0]]) - else: - raise Exception( - "unsupported data type {}".format(DTypeNames[self.dtype]) - ) + u8_data = TosaSerializer.convertDataToUint8Vec(self.dtype, self.data) fb_data = TosaSerializer.serializeUint8Vec(builder, u8_data) TosaTensor.Start(builder) @@ -469,14 +461,14 @@ class TosaSerializerOperator: self.outputs = TosaSerializer.toList(outputs) def __str__(self): - str = "Op {}\n----\n".format(self.op) + concatString = "Op {}\n----\n".format(self.op) for i in self.inputs: - str = str + " Input: {}\n".format(i) + concatString = concatString + " Input: {}\n".format(i) for o in self.outputs: - str = str + " Output: {}\n".format(o) + concatString = concatString + " Output: {}\n".format(o) - return str + return concatString def serialize(self, builder): fb_inputs = TosaSerializer.serializeStrVec( @@ -561,41 +553,39 @@ class TosaSerializerBasicBlock: return TosaBasicBlock.End(builder) +# How CONSTs are treated in the flatbuffer @unique -class TensorDir(IntEnum): - PLACEHOLDER = 0 - CONST = 1 - INTERMEDIATE = 2 - RESULT = 3 +class ConstMode(IntEnum): + EMBED = 0 + EMBED_DUMP = 1 + INPUTS = 2 -class TosaSerializer: - def __init__(self, pathPrefix): - self.add_compat_methods() - # Get the global TOSA version if not already defined - - self.builder = flatbuffers.Builder(0) - +class TosaSerializerRegion: + def __init__(self, name, pathPrefix, constMode=ConstMode.EMBED): + self.name = name self.basicBlocks = [] - self.startBasicBlock("main") - self.pathPrefix = pathPrefix - - # Indicies used for adding/naming tensors self.currInputIdx = 0 self.currConstIdx = 0 self.currLayerIdx = 1 self.currResultIdx = 0 + self.pathPrefix = pathPrefix + self.constMode = constMode - # Is this an illegal test that is expected to fail? - self.expectedReturnCode = 0 - self.expectedFailure = False - self.expectedFailureDesc = "" + def addBasicBlock(self, name): + self.currBasicBlock = TosaSerializerBasicBlock(name) + self.basicBlocks.append(self.currBasicBlock) - def __str__(self): - str = "" - for bb in self.basicBlocks: - str = str + bb.__str__() - return str + def serialize(self, builder): + fb_name = builder.CreateString(self.name) + fbv_basicBlocks = TosaSerializer.serializeObjVec( + builder, self.basicBlocks, TosaRegion.StartBlocksVector + ) + + TosaRegion.Start(builder) + TosaRegion.AddName(builder, fb_name) + TosaRegion.AddBlocks(builder, fbv_basicBlocks) + return TosaRegion.End(builder) def addPlaceholder(self, shape, dtype, vals): if not self.currBasicBlock: @@ -614,21 +604,42 @@ class TosaSerializer: return tens - def addConst(self, shape, dtype, vals): + def addConst(self, shape, dtype, vals, name=None): if not self.currBasicBlock: raise Exception("addTensor called without valid basic block") - name = "const-{}".format(self.currInputIdx) - self.currInputIdx = self.currInputIdx + 1 + if name is None: + name = "const-{}".format(self.currInputIdx) + self.currInputIdx = self.currInputIdx + 1 - tens = self.currBasicBlock.addTensor(name, shape, dtype, vals) + if self.constMode == ConstMode.INPUTS: + # Save const as input file + filename = "{}.npy".format(name) + tensor_vals = None + self.currBasicBlock.addInput(name) + else: + # Embed const in flatbuffer + filename = None + tensor_vals = vals + + tens = self.currBasicBlock.addTensor(name, shape, dtype, tensor_vals, filename) # Add the operator now - self.currBasicBlock.addOperator(TosaOp.Op().CONST, [], name) + if dtype == DType.SHAPE: + self.currBasicBlock.addOperator(TosaOp.Op().CONST_SHAPE, [], name) + else: + self.currBasicBlock.addOperator(TosaOp.Op().CONST, [], name) + + # Save the const data to file for debug or as input files + if vals is not None and self.constMode in [ + ConstMode.EMBED_DUMP, + ConstMode.INPUTS, + ]: + filename = "{}.npy".format(name) + np.save(os.path.join(self.pathPrefix, filename), vals, False) return tens def addIntermediate(self, shape, dtype): - if not self.currBasicBlock: raise Exception("addTensor called without valid basic block") @@ -640,7 +651,13 @@ class TosaSerializer: return tens def addInputTensor(self, tensor): - self.currBasicBlock.addTensor(tensor.name, tensor.shape, tensor.dtype) + self.currBasicBlock.addTensor( + tensor.name, + tensor.shape, + tensor.dtype, + tensor.data, + tensor.placeholderFilename, + ) self.currBasicBlock.addInput(tensor.name) def addOutputTensor(self, tensor): @@ -658,7 +675,6 @@ class TosaSerializer: return tens def addOperator(self, op, inputs, outputs, attributes=None): - if op == TosaOp.Op().CONST: raise Exception("Use addConstTensor() to add CONST ops") @@ -669,6 +685,62 @@ class TosaSerializer: attributes, ) + +@unique +class TensorDir(IntEnum): + PLACEHOLDER = 0 + CONST = 1 + INTERMEDIATE = 2 + RESULT = 3 + + +class TosaSerializer: + def __init__(self, pathPrefix, constMode=ConstMode.EMBED): + self.builder = flatbuffers.Builder(0) + + # Enables inspection of constant data outside of graph + self.constMode = constMode + + self.regions = [] + self.startRegion("main", pathPrefix) + + self.currRegion.addBasicBlock("main") + + # Is this an illegal test that is expected to fail? + self.expectedReturnCode = 0 + self.expectedFailure = False + self.expectedFailureDesc = "" + + def __str__(self): + concatString = "" + for region in self.regions: + concatString = concatString + str(region) + return concatString + + def addPlaceholder(self, shape, dtype, vals): + return self.currRegion.addPlaceholder(shape, dtype, vals) + + def addConst(self, shape, dtype, vals, name=None): + return self.currRegion.addConst(shape, dtype, vals, name) + + def addIntermediate(self, shape, dtype): + return self.currRegion.addIntermediate(shape, dtype) + + def addInputTensor(self, tensor): + self.currRegion.addInputTensor(tensor) + + def addOutputTensor(self, tensor): + self.currRegion.addOutputTensor(tensor) + + def addOutput(self, shape, dtype): + return self.currRegion.addOutput(shape, dtype) + + def addOperator(self, op, inputs, outputs, attributes=None): + return self.currRegion.addOperator(op, inputs, outputs, attributes) + + def addBasicBlock(self, name): + self.currRegion.addBasicBlock(name) + def setExpectedReturnCode(self, val, fail, desc=""): self.expectedReturnCode = val @@ -680,19 +752,19 @@ class TosaSerializer: builder = self.builder Version.Start(builder) - Version.Add_major(builder, TOSA_VERSION[0]) - Version.Add_minor(builder, TOSA_VERSION[1]) - Version.Add_patch(builder, TOSA_VERSION[2]) - Version.Add_draft(builder, TOSA_VERSION[3]) + Version.Add_Major(builder, TOSA_VERSION[0]) + Version.Add_Minor(builder, TOSA_VERSION[1]) + Version.Add_Patch(builder, TOSA_VERSION[2]) + Version.Add_Draft(builder, TOSA_VERSION[3]) version = Version.End(builder) - fbv_bb = TosaSerializer.serializeObjVec( - builder, self.basicBlocks, TosaGraph.StartBlocksVector + fbv_region = TosaSerializer.serializeObjVec( + builder, self.regions, TosaGraph.StartRegionsVector ) TosaGraph.Start(builder) TosaGraph.AddVersion(builder, version) - TosaGraph.AddBlocks(builder, fbv_bb) + TosaGraph.AddRegions(builder, fbv_region) graph = TosaGraph.End(builder) self.builder.Finish(graph, TOSA_GRAPH_IDENTIFIER) @@ -709,16 +781,17 @@ class TosaSerializer: ofm_name = [] ofm_file = [] - for b in self.basicBlocks: - if b.name == "main": - for i in b.inputs: - ifm_name.append(i) - ifm_file.append(b.tensors[i].placeholderFilename) - for o in b.outputs: - ofm_name.append(o) - # Make up an OFM filename here. One isn't generated until the - # reference tool is run, so any name is a good name - ofm_file.append("ref-{}.npy".format(o)) + for region in self.regions: + for block in region.basicBlocks: + if block and block.name == "main": + for i in block.inputs: + ifm_name.append(i) + ifm_file.append(block.tensors[i].placeholderFilename) + for o in block.outputs: + ofm_name.append(o) + # Make up an OFM filename here. One isn't generated until the + # reference tool is run, so any name is a good name + ofm_file.append("ref-{}.npy".format(o)) test_desc["ifm_name"] = ifm_name test_desc["ifm_file"] = ifm_file @@ -731,9 +804,9 @@ class TosaSerializer: return json.dumps(test_desc, indent=" ") - def startBasicBlock(self, name): - self.currBasicBlock = TosaSerializerBasicBlock(name) - self.basicBlocks.append(self.currBasicBlock) + def startRegion(self, name, pathPrefix): + self.currRegion = TosaSerializerRegion(name, pathPrefix, self.constMode) + self.regions.append(self.currRegion) @staticmethod def serializeStrVec(builder, vec, start_fcn): @@ -757,6 +830,16 @@ class TosaSerializer: return builder.EndVector(len(vec)) @staticmethod + def serializeInt16Vec(builder, vec): + builder.StartVector(2, len(vec), 4) + for v in vec[::-1]: + builder.PrependInt16(v) + try: + return builder.EndVector() + except TypeError: + return builder.EndVector(len(vec)) + + @staticmethod def serializeInt32Vec(builder, vec): builder.StartVector(4, len(vec), 4) for v in vec[::-1]: @@ -797,369 +880,104 @@ class TosaSerializer: else: return [val] - # Remove when switching to flatbuffers 2.0 - # contains a mapping of the deprecated 1.12 method to the 2.0 version - - def add_compat_methods(self): - - from tosa import ArithmeticRightShiftAttribute - - if not hasattr(ArithmeticRightShiftAttribute, "Start"): - ArithmeticRightShiftAttribute.Start = ( - ArithmeticRightShiftAttribute.ArithmeticRightShiftAttributeStart - ) - ArithmeticRightShiftAttribute.AddRound = ( - ArithmeticRightShiftAttribute.ArithmeticRightShiftAttributeAddRound - ) - ArithmeticRightShiftAttribute.End = ( - ArithmeticRightShiftAttribute.ArithmeticRightShiftAttributeEnd - ) - from tosa import AxisAttribute - - if not hasattr(AxisAttribute, "Start"): - AxisAttribute.Start = AxisAttribute.AxisAttributeStart - AxisAttribute.AddAxis = AxisAttribute.AxisAttributeAddAxis - AxisAttribute.End = AxisAttribute.AxisAttributeEnd - from tosa import ClampAttribute - - if not hasattr(ClampAttribute, "Start"): - ClampAttribute.Start = ClampAttribute.ClampAttributeStart - ClampAttribute.AddMinInt = ClampAttribute.ClampAttributeAddMinInt - ClampAttribute.AddMaxInt = ClampAttribute.ClampAttributeAddMaxInt - ClampAttribute.AddMinFp = ClampAttribute.ClampAttributeAddMinFp - ClampAttribute.AddMaxFp = ClampAttribute.ClampAttributeAddMaxFp - ClampAttribute.End = ClampAttribute.ClampAttributeEnd - from tosa import CondIfAttribute - - if not hasattr(CondIfAttribute, "Start"): - CondIfAttribute.Start = CondIfAttribute.CondIfAttributeStart - CondIfAttribute.AddThenBranch = CondIfAttribute.CondIfAttributeAddThenBranch - CondIfAttribute.AddElseBranch = CondIfAttribute.CondIfAttributeAddElseBranch - CondIfAttribute.End = CondIfAttribute.CondIfAttributeEnd - from tosa import ConvAttribute - - if not hasattr(ConvAttribute, "Start"): - ConvAttribute.Start = ConvAttribute.ConvAttributeStart - ConvAttribute.AddPad = ConvAttribute.ConvAttributeAddPad - ConvAttribute.StartPadVector = ConvAttribute.ConvAttributeStartPadVector - ConvAttribute.AddStride = ConvAttribute.ConvAttributeAddStride - ConvAttribute.StartStrideVector = ( - ConvAttribute.ConvAttributeStartStrideVector - ) - ConvAttribute.AddDilation = ConvAttribute.ConvAttributeAddDilation - ConvAttribute.StartDilationVector = ( - ConvAttribute.ConvAttributeStartDilationVector - ) - ConvAttribute.AddInputZp = ConvAttribute.ConvAttributeAddInputZp - ConvAttribute.AddWeightZp = ConvAttribute.ConvAttributeAddWeightZp - ConvAttribute.End = ConvAttribute.ConvAttributeEnd - from tosa import FullyConnectedAttribute - - if not hasattr(FullyConnectedAttribute, "Start"): - FullyConnectedAttribute.Start = ( - FullyConnectedAttribute.FullyConnectedAttributeStart - ) - FullyConnectedAttribute.AddInputZp = ( - FullyConnectedAttribute.FullyConnectedAttributeAddInputZp - ) - FullyConnectedAttribute.AddWeightZp = ( - FullyConnectedAttribute.FullyConnectedAttributeAddWeightZp - ) - FullyConnectedAttribute.End = ( - FullyConnectedAttribute.FullyConnectedAttributeEnd - ) - from tosa import MatMulAttribute - - if not hasattr(MatMulAttribute, "Start"): - MatMulAttribute.Start = MatMulAttribute.MatMulAttributeStart - MatMulAttribute.AddAZp = MatMulAttribute.MatMulAttributeAddAZp - MatMulAttribute.AddBZp = MatMulAttribute.MatMulAttributeAddBZp - MatMulAttribute.End = MatMulAttribute.MatMulAttributeEnd - from tosa import PoolAttribute - - if not hasattr(PoolAttribute, "Start"): - PoolAttribute.Start = PoolAttribute.PoolAttributeStart - PoolAttribute.AddPad = PoolAttribute.PoolAttributeAddPad - PoolAttribute.StartPadVector = PoolAttribute.PoolAttributeStartPadVector - PoolAttribute.AddKernel = PoolAttribute.PoolAttributeAddKernel - PoolAttribute.StartKernelVector = ( - PoolAttribute.PoolAttributeStartKernelVector - ) - PoolAttribute.AddStride = PoolAttribute.PoolAttributeAddStride - PoolAttribute.StartStrideVector = ( - PoolAttribute.PoolAttributeStartStrideVector - ) - PoolAttribute.AddInputZp = PoolAttribute.PoolAttributeAddInputZp - PoolAttribute.AddOutputZp = PoolAttribute.PoolAttributeAddOutputZp - PoolAttribute.End = PoolAttribute.PoolAttributeEnd - from tosa import MulAttribute - - if not hasattr(MulAttribute, "Start"): - MulAttribute.Start = MulAttribute.MulAttributeStart - MulAttribute.AddShift = MulAttribute.MulAttributeAddShift - MulAttribute.End = MulAttribute.MulAttributeEnd - from tosa import PadAttribute - - if not hasattr(PadAttribute, "Start"): - PadAttribute.Start = PadAttribute.PadAttributeStart - PadAttribute.AddPadding = PadAttribute.PadAttributeAddPadding - PadAttribute.StartPaddingVector = ( - PadAttribute.PadAttributeStartPaddingVector - ) - PadAttribute.AddPadConstInt = PadAttribute.PadAttributeAddPadConstInt - PadAttribute.AddPadConstFp = PadAttribute.PadAttributeAddPadConstFp - PadAttribute.End = PadAttribute.PadAttributeEnd - from tosa import PoolAttribute - - if not hasattr(PoolAttribute, "Start"): - PoolAttribute.Start = PoolAttribute.PoolAttributeStart - PoolAttribute.AddPad = PoolAttribute.PoolAttributeAddPad - PoolAttribute.StartPadVector = PoolAttribute.PoolAttributeStartPadVector - PoolAttribute.AddKernel = PoolAttribute.PoolAttributeAddKernel - PoolAttribute.StartKernelVector = ( - PoolAttribute.PoolAttributeStartKernelVector - ) - PoolAttribute.AddStride = PoolAttribute.PoolAttributeAddStride - PoolAttribute.StartStrideVector = ( - PoolAttribute.PoolAttributeStartStrideVector - ) - PoolAttribute.AddInputZp = PoolAttribute.PoolAttributeAddInputZp - PoolAttribute.AddOutputZp = PoolAttribute.PoolAttributeAddOutputZp - PoolAttribute.End = PoolAttribute.PoolAttributeEnd - from tosa import RescaleAttribute - - if not hasattr(RescaleAttribute, "Start"): - RescaleAttribute.Start = RescaleAttribute.RescaleAttributeStart - RescaleAttribute.AddInputZp = RescaleAttribute.RescaleAttributeAddInputZp - RescaleAttribute.AddOutputZp = RescaleAttribute.RescaleAttributeAddOutputZp - RescaleAttribute.AddMultiplier = ( - RescaleAttribute.RescaleAttributeAddMultiplier - ) - RescaleAttribute.StartMultiplierVector = ( - RescaleAttribute.RescaleAttributeStartMultiplierVector - ) - RescaleAttribute.AddShift = RescaleAttribute.RescaleAttributeAddShift - RescaleAttribute.StartShiftVector = ( - RescaleAttribute.RescaleAttributeStartShiftVector - ) - RescaleAttribute.AddScale32 = RescaleAttribute.RescaleAttributeAddScale32 - RescaleAttribute.AddDoubleRound = ( - RescaleAttribute.RescaleAttributeAddDoubleRound - ) - RescaleAttribute.AddPerChannel = ( - RescaleAttribute.RescaleAttributeAddPerChannel - ) - RescaleAttribute.End = RescaleAttribute.RescaleAttributeEnd - from tosa import ReshapeAttribute - - if not hasattr(ReshapeAttribute, "Start"): - ReshapeAttribute.Start = ReshapeAttribute.ReshapeAttributeStart - ReshapeAttribute.AddNewShape = ReshapeAttribute.ReshapeAttributeAddNewShape - ReshapeAttribute.StartNewShapeVector = ( - ReshapeAttribute.ReshapeAttributeStartNewShapeVector - ) - ReshapeAttribute.End = ReshapeAttribute.ReshapeAttributeEnd - from tosa import ResizeAttribute - - if not hasattr(ResizeAttribute, "Start"): - ResizeAttribute.Start = ResizeAttribute.ResizeAttributeStart - ResizeAttribute.AddOutputSize = ResizeAttribute.ResizeAttributeAddOutputSize - ResizeAttribute.StartOutputSizeVector = ( - ResizeAttribute.ResizeAttributeStartOutputSizeVector - ) - ResizeAttribute.AddStride = ResizeAttribute.ResizeAttributeAddStride - ResizeAttribute.StartStrideVector = ( - ResizeAttribute.ResizeAttributeStartStrideVector - ) - ResizeAttribute.AddOffset = ResizeAttribute.ResizeAttributeAddOffset - ResizeAttribute.StartOffsetVector = ( - ResizeAttribute.ResizeAttributeStartOffsetVector - ) - ResizeAttribute.AddShift = ResizeAttribute.ResizeAttributeAddShift - ResizeAttribute.AddStrideFp = ResizeAttribute.ResizeAttributeAddStrideFp - ResizeAttribute.StartStrideFpVector = ( - ResizeAttribute.ResizeAttributeStartStrideFpVector - ) - ResizeAttribute.AddOffsetFp = ResizeAttribute.ResizeAttributeAddOffsetFp - ResizeAttribute.StartOffsetFpVector = ( - ResizeAttribute.ResizeAttributeStartOffsetFpVector - ) - ResizeAttribute.AddMode = ResizeAttribute.ResizeAttributeAddMode - ResizeAttribute.End = ResizeAttribute.ResizeAttributeEnd - from tosa import SliceAttribute - - if not hasattr(SliceAttribute, "Start"): - SliceAttribute.Start = SliceAttribute.SliceAttributeStart - SliceAttribute.AddStart = SliceAttribute.SliceAttributeAddStart - SliceAttribute.StartStartVector = ( - SliceAttribute.SliceAttributeStartStartVector - ) - SliceAttribute.AddSize = SliceAttribute.SliceAttributeAddSize - SliceAttribute.StartSizeVector = ( - SliceAttribute.SliceAttributeStartSizeVector - ) - SliceAttribute.End = SliceAttribute.SliceAttributeEnd - from tosa import TableAttribute - - if not hasattr(TableAttribute, "Start"): - TableAttribute.Start = TableAttribute.TableAttributeStart - TableAttribute.AddTable = TableAttribute.TableAttributeAddTable - TableAttribute.StartTableVector = ( - TableAttribute.TableAttributeStartTableVector - ) - TableAttribute.End = TableAttribute.TableAttributeEnd - from tosa import TileAttribute - - if not hasattr(TileAttribute, "Start"): - TileAttribute.Start = TileAttribute.TileAttributeStart - TileAttribute.AddMultiples = TileAttribute.TileAttributeAddMultiples - TileAttribute.StartMultiplesVector = ( - TileAttribute.TileAttributeStartMultiplesVector - ) - TileAttribute.End = TileAttribute.TileAttributeEnd - from tosa import TosaBasicBlock - - if not hasattr(TosaBasicBlock, "Start"): - TosaBasicBlock.Start = TosaBasicBlock.TosaBasicBlockStart - TosaBasicBlock.AddName = TosaBasicBlock.TosaBasicBlockAddName - TosaBasicBlock.AddOperators = TosaBasicBlock.TosaBasicBlockAddOperators - TosaBasicBlock.StartOperatorsVector = ( - TosaBasicBlock.TosaBasicBlockStartOperatorsVector - ) - TosaBasicBlock.AddTensors = TosaBasicBlock.TosaBasicBlockAddTensors - TosaBasicBlock.StartTensorsVector = ( - TosaBasicBlock.TosaBasicBlockStartTensorsVector - ) - TosaBasicBlock.AddInputs = TosaBasicBlock.TosaBasicBlockAddInputs - TosaBasicBlock.StartInputsVector = ( - TosaBasicBlock.TosaBasicBlockStartInputsVector - ) - TosaBasicBlock.AddOutputs = TosaBasicBlock.TosaBasicBlockAddOutputs - TosaBasicBlock.StartOutputsVector = ( - TosaBasicBlock.TosaBasicBlockStartOutputsVector - ) - TosaBasicBlock.End = TosaBasicBlock.TosaBasicBlockEnd - from tosa import TosaGraph - - if not hasattr(TosaGraph, "Start"): - TosaGraph.Start = TosaGraph.TosaGraphStart - TosaGraph.AddVersion = TosaGraph.TosaGraphAddVersion - TosaGraph.AddBlocks = TosaGraph.TosaGraphAddBlocks - TosaGraph.StartBlocksVector = TosaGraph.TosaGraphStartBlocksVector - TosaGraph.End = TosaGraph.TosaGraphEnd - from tosa import TosaOperator - - if not hasattr(TosaOperator, "Start"): - TosaOperator.Start = TosaOperator.TosaOperatorStart - TosaOperator.AddOp = TosaOperator.TosaOperatorAddOp - TosaOperator.AddAttributeType = TosaOperator.TosaOperatorAddAttributeType - TosaOperator.AddAttribute = TosaOperator.TosaOperatorAddAttribute - TosaOperator.AddInputs = TosaOperator.TosaOperatorAddInputs - TosaOperator.StartInputsVector = TosaOperator.TosaOperatorStartInputsVector - TosaOperator.AddOutputs = TosaOperator.TosaOperatorAddOutputs - TosaOperator.StartOutputsVector = ( - TosaOperator.TosaOperatorStartOutputsVector - ) - TosaOperator.End = TosaOperator.TosaOperatorEnd - from tosa import TosaTensor - - if not hasattr(TosaTensor, "Start"): - TosaTensor.Start = TosaTensor.TosaTensorStart - TosaTensor.AddName = TosaTensor.TosaTensorAddName - TosaTensor.AddShape = TosaTensor.TosaTensorAddShape - TosaTensor.StartShapeVector = TosaTensor.TosaTensorStartShapeVector - TosaTensor.AddType = TosaTensor.TosaTensorAddType - TosaTensor.AddData = TosaTensor.TosaTensorAddData - TosaTensor.StartDataVector = TosaTensor.TosaTensorStartDataVector - TosaTensor.End = TosaTensor.TosaTensorEnd - from tosa import TransposeAttribute - - if not hasattr(TransposeAttribute, "Start"): - TransposeAttribute.Start = TransposeAttribute.TransposeAttributeStart - TransposeAttribute.AddPerms = TransposeAttribute.TransposeAttributeAddPerms - TransposeAttribute.StartPermsVector = ( - TransposeAttribute.TransposeAttributeStartPermsVector - ) - TransposeAttribute.End = TransposeAttribute.TransposeAttributeEnd - from tosa import TransposeConvAttribute - - if not hasattr(TransposeConvAttribute, "Start"): - TransposeConvAttribute.Start = ( - TransposeConvAttribute.TransposeConvAttributeStart - ) - TransposeConvAttribute.AddOutPad = ( - TransposeConvAttribute.TransposeConvAttributeAddOutPad - ) - TransposeConvAttribute.StartOutPadVector = ( - TransposeConvAttribute.TransposeConvAttributeStartOutPadVector - ) - TransposeConvAttribute.AddStride = ( - TransposeConvAttribute.TransposeConvAttributeAddStride - ) - TransposeConvAttribute.StartStrideVector = ( - TransposeConvAttribute.TransposeConvAttributeStartStrideVector - ) - TransposeConvAttribute.AddOutputShape = ( - TransposeConvAttribute.TransposeConvAttributeAddOutputShape - ) - TransposeConvAttribute.StartOutputShapeVector = ( - TransposeConvAttribute.TransposeConvAttributeStartOutputShapeVector - ) - TransposeConvAttribute.AddInputZp = ( - TransposeConvAttribute.TransposeConvAttributeAddInputZp - ) - TransposeConvAttribute.AddWeightZp = ( - TransposeConvAttribute.TransposeConvAttributeAddWeightZp - ) - TransposeConvAttribute.End = ( - TransposeConvAttribute.TransposeConvAttributeEnd - ) - from tosa import Version - - if not hasattr(Version, "Start"): - Version.Start = Version.VersionStart - Version.Add_major = Version.VersionAdd_major - Version.Add_minor = Version.VersionAdd_minor - Version.Add_patch = Version.VersionAdd_patch - Version.Add_draft = Version.VersionAdd_draft - Version.End = Version.VersionEnd - from tosa import MatMulAttribute - - if not hasattr(MatMulAttribute, "Start"): - MatMulAttribute.Start = MatMulAttribute.MatMulAttributeStart - MatMulAttribute.AddAZp = MatMulAttribute.MatMulAttributeAddAZp - MatMulAttribute.AddBZp = MatMulAttribute.MatMulAttributeAddBZp - MatMulAttribute.End = MatMulAttribute.MatMulAttributeEnd - from tosa import FullyConnectedAttribute - - if not hasattr(FullyConnectedAttribute, "Start"): - FullyConnectedAttribute.Start = ( - FullyConnectedAttribute.FullyConnectedAttributeStart - ) - FullyConnectedAttribute.AddInputZp = ( - FullyConnectedAttribute.FullyConnectedAttributeAddInputZp - ) - FullyConnectedAttribute.AddWeightZp = ( - FullyConnectedAttribute.FullyConnectedAttributeAddWeightZp - ) - FullyConnectedAttribute.End = ( - FullyConnectedAttribute.FullyConnectedAttributeEnd - ) - from tosa import NegateAttribute - - if not hasattr(NegateAttribute, "Start"): - NegateAttribute.Start = NegateAttribute.NegateAttributeStart - NegateAttribute.AddInput1Zp = NegateAttribute.NegateAttributeAddInput1Zp - NegateAttribute.AddOutputZp = NegateAttribute.NegateAttributeAddOutputZp - NegateAttribute.End = NegateAttribute.NegateAttributeEnd - from tosa import WhileLoopAttribute - - if not hasattr(WhileLoopAttribute, "Start"): - WhileLoopAttribute.Start = WhileLoopAttribute.WhileLoopAttributeStart - WhileLoopAttribute.AddCondBranch = ( - WhileLoopAttribute.WhileLoopAttributeAddCondBranch - ) - WhileLoopAttribute.AddBodyBranch = ( - WhileLoopAttribute.WhileLoopAttributeAddBodyBranch - ) - WhileLoopAttribute.End = WhileLoopAttribute.WhileLoopAttributeEnd + @staticmethod + def convertDataToUint8Vec(dtype, data): + u8_data = list() + # little endianess + if dtype == DType.BOOL: + for val in data: + val_u8 = np.uint8(val) + u8_data.append(val_u8) + elif dtype == DType.INT4: + in_size = len(data) + out_size = (in_size + 1) // 2 + for i in range(out_size): + val_0 = data[2 * i] + if (2 * i + 1) < in_size: + val_1 = data[2 * i + 1] + else: + val_1 = 0 + val_i8 = (val_0 & 0xF) | ((val_1 & 0xF) << 4) + val_u8 = np.uint8(val_i8) + u8_data.append(val_u8) + elif dtype == DType.INT8: + for val in data: + val_u8 = np.array(val).astype(dtype=np.uint8) + u8_data.append(val_u8) + elif dtype == DType.INT16: + for val in data: + val_u16 = np.array(val).astype(dtype=np.uint16) + b0 = val_u16 & ByteMask + b1 = (val_u16 >> np.uint16(8)) & ByteMask + u8_data.extend([b0, b1]) + elif dtype == DType.INT32: + for val in data: + val_u32 = np.array(val).astype(dtype=np.uint32) + b0 = val_u32 & ByteMask + b1 = (val_u32 >> np.uint32(8)) & ByteMask + b2 = (val_u32 >> np.uint32(16)) & ByteMask + b3 = (val_u32 >> np.uint32(24)) & ByteMask + u8_data.extend([b0, b1, b2, b3]) + elif dtype == DType.INT48: + for val in data: + val_u64 = np.uint64(val) + b0 = val_u64 & ByteMask + b1 = (val_u64 >> np.uint64(8)) & ByteMask + b2 = (val_u64 >> np.uint64(16)) & ByteMask + b3 = (val_u64 >> np.uint64(24)) & ByteMask + b4 = (val_u64 >> np.uint64(32)) & ByteMask + b5 = (val_u64 >> np.uint64(40)) & ByteMask + u8_data.extend([b0, b1, b2, b3, b4, b5]) + elif dtype == DType.SHAPE: + for val in data: + val_u64 = np.uint64(val) + b0 = val_u64 & ByteMask + b1 = (val_u64 >> np.uint64(8)) & ByteMask + b2 = (val_u64 >> np.uint64(16)) & ByteMask + b3 = (val_u64 >> np.uint64(24)) & ByteMask + b4 = (val_u64 >> np.uint64(32)) & ByteMask + b5 = (val_u64 >> np.uint64(40)) & ByteMask + b6 = (val_u64 >> np.uint64(48)) & ByteMask + b7 = (val_u64 >> np.uint64(56)) & ByteMask + u8_data.extend([b0, b1, b2, b3, b4, b5, b6, b7]) + elif dtype == DType.FP16: + np_arr = np.array(data, dtype=np.float16) + u8_data.extend(np_arr.view(np.uint8)) + elif dtype == DType.FP32: + # for val in data: + # b = struct.pack("!f", val) + # u8_data.extend([b[3], b[2], b[1], b[0]]) + np_arr = np.array(data, dtype=np.float32) + u8_data.extend(np_arr.view(np.uint8)) + elif dtype == DType.BF16: + for val in data: + # convert val to little endian byte arrays b + b = struct.pack("<f", val) + # val => [ b[3], b[2], b[1], b[0] ] + # keep only most significant 2 bytes for bf16 + # in little endian ordering + u8_data.extend([b[2], b[3]]) + elif dtype == DType.FP8E4M3: + for val in data: + # convert val to fp8_bits then to single byte + f32_as_int = struct.unpack(">L", struct.pack(">f", val))[0] + f32_bits = f"{f32_as_int:032b}" + fp8_bits = f32_bits[0] + f32_bits[1:5] + f32_bits[9:12] + fp8_bytes = int(fp8_bits, 2).to_bytes(1, byteorder="little") + u8_data.extend(fp8_bytes) + elif dtype == DType.FP8E5M2: + for val in data: + # convert val to fp8_bits then to single byte + f32_as_int = struct.unpack(">L", struct.pack(">f", val))[0] + f32_bits = f"{f32_as_int:032b}" + fp8_bits = f32_bits[0] + f32_bits[1:6] + f32_bits[9:11] + fp8_bytes = int(fp8_bits, 2).to_bytes(1, byteorder="little") + u8_data.extend(fp8_bytes) + elif dtype == TosaDType.DType: + # Serialize DType enum data as uint8 bytes + for val in data: + np_arr = np.array(data, dtype=np.uint32) + u8_data.extend(np_arr.view(np.uint8)) + else: + raise Exception("unsupported data type {}".format(DTypeNames[dtype])) + return u8_data diff --git a/python/tosa/ArithmeticRightShiftAttribute.py b/python/tosa/ArithmeticRightShiftAttribute.py index 226dc0e..f2aaeb5 100644 --- a/python/tosa/ArithmeticRightShiftAttribute.py +++ b/python/tosa/ArithmeticRightShiftAttribute.py @@ -10,13 +10,17 @@ class ArithmeticRightShiftAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsArithmeticRightShiftAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = ArithmeticRightShiftAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsArithmeticRightShiftAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def ArithmeticRightShiftAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -31,6 +35,20 @@ class ArithmeticRightShiftAttribute(object): return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) return False -def ArithmeticRightShiftAttributeStart(builder): builder.StartObject(1) -def ArithmeticRightShiftAttributeAddRound(builder, round): builder.PrependBoolSlot(0, round, 0) -def ArithmeticRightShiftAttributeEnd(builder): return builder.EndObject() +def ArithmeticRightShiftAttributeStart(builder): + builder.StartObject(1) + +def Start(builder): + ArithmeticRightShiftAttributeStart(builder) + +def ArithmeticRightShiftAttributeAddRound(builder, round): + builder.PrependBoolSlot(0, round, 0) + +def AddRound(builder, round): + ArithmeticRightShiftAttributeAddRound(builder, round) + +def ArithmeticRightShiftAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return ArithmeticRightShiftAttributeEnd(builder) diff --git a/python/tosa/Attribute.py b/python/tosa/Attribute.py index 166de8e..6abdcfb 100644 --- a/python/tosa/Attribute.py +++ b/python/tosa/Attribute.py @@ -9,19 +9,18 @@ class Attribute(object): TransposeConvAttribute = 3 PadAttribute = 4 AxisAttribute = 5 - ReshapeAttribute = 6 - SliceAttribute = 7 - TileAttribute = 8 - ResizeAttribute = 9 - ClampAttribute = 10 - RescaleAttribute = 11 - MulAttribute = 12 - ArithmeticRightShiftAttribute = 13 - CondIfAttribute = 14 - WhileLoopAttribute = 15 - TransposeAttribute = 16 - TableAttribute = 17 - MatMulAttribute = 18 - FullyConnectedAttribute = 19 - NegateAttribute = 20 - + ResizeAttribute = 6 + ClampAttribute = 7 + RescaleAttribute = 8 + MulAttribute = 9 + ArithmeticRightShiftAttribute = 10 + CondIfAttribute = 11 + WhileLoopAttribute = 12 + TransposeAttribute = 13 + TableAttribute = 14 + MatMulAttribute = 15 + FullyConnectedAttribute = 16 + NegateAttribute = 17 + CustomAttribute = 18 + FFTAttribute = 19 + RFFTAttribute = 20 diff --git a/python/tosa/AxisAttribute.py b/python/tosa/AxisAttribute.py index 30876b7..7ce4a63 100644 --- a/python/tosa/AxisAttribute.py +++ b/python/tosa/AxisAttribute.py @@ -10,13 +10,17 @@ class AxisAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsAxisAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = AxisAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsAxisAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def AxisAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -31,6 +35,20 @@ class AxisAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def AxisAttributeStart(builder): builder.StartObject(1) -def AxisAttributeAddAxis(builder, axis): builder.PrependInt32Slot(0, axis, 0) -def AxisAttributeEnd(builder): return builder.EndObject() +def AxisAttributeStart(builder): + builder.StartObject(1) + +def Start(builder): + AxisAttributeStart(builder) + +def AxisAttributeAddAxis(builder, axis): + builder.PrependInt32Slot(0, axis, 0) + +def AddAxis(builder, axis): + AxisAttributeAddAxis(builder, axis) + +def AxisAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return AxisAttributeEnd(builder) diff --git a/python/tosa/ClampAttribute.py b/python/tosa/ClampAttribute.py index 066dd4f..6a41498 100644 --- a/python/tosa/ClampAttribute.py +++ b/python/tosa/ClampAttribute.py @@ -10,13 +10,17 @@ class ClampAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsClampAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = ClampAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsClampAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def ClampAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -25,36 +29,91 @@ class ClampAttribute(object): self._tab = flatbuffers.table.Table(buf, pos) # ClampAttribute - def MinInt(self): + def MinVal(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: - return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) + a = self._tab.Vector(o) + return self._tab.Get(flatbuffers.number_types.Uint8Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 1)) return 0 # ClampAttribute - def MaxInt(self): + def MinValAsNumpy(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + if o != 0: + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Uint8Flags, o) + return 0 + + # ClampAttribute + def MinValLength(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + if o != 0: + return self._tab.VectorLen(o) + return 0 + + # ClampAttribute + def MinValIsNone(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + return o == 0 + + # ClampAttribute + def MaxVal(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: - return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) + a = self._tab.Vector(o) + return self._tab.Get(flatbuffers.number_types.Uint8Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 1)) return 0 # ClampAttribute - def MinFp(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) + def MaxValAsNumpy(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: - return self._tab.Get(flatbuffers.number_types.Float32Flags, o + self._tab.Pos) - return 0.0 + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Uint8Flags, o) + return 0 # ClampAttribute - def MaxFp(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) + def MaxValLength(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: - return self._tab.Get(flatbuffers.number_types.Float32Flags, o + self._tab.Pos) - return 0.0 - -def ClampAttributeStart(builder): builder.StartObject(4) -def ClampAttributeAddMinInt(builder, minInt): builder.PrependInt32Slot(0, minInt, 0) -def ClampAttributeAddMaxInt(builder, maxInt): builder.PrependInt32Slot(1, maxInt, 0) -def ClampAttributeAddMinFp(builder, minFp): builder.PrependFloat32Slot(2, minFp, 0.0) -def ClampAttributeAddMaxFp(builder, maxFp): builder.PrependFloat32Slot(3, maxFp, 0.0) -def ClampAttributeEnd(builder): return builder.EndObject() + return self._tab.VectorLen(o) + return 0 + + # ClampAttribute + def MaxValIsNone(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) + return o == 0 + +def ClampAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + ClampAttributeStart(builder) + +def ClampAttributeAddMinVal(builder, minVal): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(minVal), 0) + +def AddMinVal(builder, minVal): + ClampAttributeAddMinVal(builder, minVal) + +def ClampAttributeStartMinValVector(builder, numElems): + return builder.StartVector(1, numElems, 1) + +def StartMinValVector(builder, numElems: int) -> int: + return ClampAttributeStartMinValVector(builder, numElems) + +def ClampAttributeAddMaxVal(builder, maxVal): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(maxVal), 0) + +def AddMaxVal(builder, maxVal): + ClampAttributeAddMaxVal(builder, maxVal) + +def ClampAttributeStartMaxValVector(builder, numElems): + return builder.StartVector(1, numElems, 1) + +def StartMaxValVector(builder, numElems: int) -> int: + return ClampAttributeStartMaxValVector(builder, numElems) + +def ClampAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return ClampAttributeEnd(builder) diff --git a/python/tosa/CondIfAttribute.py b/python/tosa/CondIfAttribute.py index 57e5cb7..8f2aa9b 100644 --- a/python/tosa/CondIfAttribute.py +++ b/python/tosa/CondIfAttribute.py @@ -10,13 +10,17 @@ class CondIfAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsCondIfAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = CondIfAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsCondIfAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def CondIfAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -25,20 +29,39 @@ class CondIfAttribute(object): self._tab = flatbuffers.table.Table(buf, pos) # CondIfAttribute - def ThenBranch(self): + def ThenGraph(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: return self._tab.String(o + self._tab.Pos) return None # CondIfAttribute - def ElseBranch(self): + def ElseGraph(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: return self._tab.String(o + self._tab.Pos) return None -def CondIfAttributeStart(builder): builder.StartObject(2) -def CondIfAttributeAddThenBranch(builder, thenBranch): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(thenBranch), 0) -def CondIfAttributeAddElseBranch(builder, elseBranch): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(elseBranch), 0) -def CondIfAttributeEnd(builder): return builder.EndObject() +def CondIfAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + CondIfAttributeStart(builder) + +def CondIfAttributeAddThenGraph(builder, thenGraph): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(thenGraph), 0) + +def AddThenGraph(builder, thenGraph): + CondIfAttributeAddThenGraph(builder, thenGraph) + +def CondIfAttributeAddElseGraph(builder, elseGraph): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(elseGraph), 0) + +def AddElseGraph(builder, elseGraph): + CondIfAttributeAddElseGraph(builder, elseGraph) + +def CondIfAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return CondIfAttributeEnd(builder) diff --git a/python/tosa/ConvAttribute.py b/python/tosa/ConvAttribute.py index 8244ea5..dfa75dc 100644 --- a/python/tosa/ConvAttribute.py +++ b/python/tosa/ConvAttribute.py @@ -10,13 +10,17 @@ class ConvAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsConvAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = ConvAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsConvAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def ConvAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -119,13 +123,88 @@ class ConvAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def ConvAttributeStart(builder): builder.StartObject(5) -def ConvAttributeAddPad(builder, pad): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(pad), 0) -def ConvAttributeStartPadVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ConvAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) -def ConvAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ConvAttributeAddDilation(builder, dilation): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(dilation), 0) -def ConvAttributeStartDilationVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ConvAttributeAddInputZp(builder, inputZp): builder.PrependInt32Slot(3, inputZp, 0) -def ConvAttributeAddWeightZp(builder, weightZp): builder.PrependInt32Slot(4, weightZp, 0) -def ConvAttributeEnd(builder): return builder.EndObject() + # ConvAttribute + def LocalBound(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + + # ConvAttribute + def AccType(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16)) + if o != 0: + return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos) + return 0 + +def ConvAttributeStart(builder): + builder.StartObject(7) + +def Start(builder): + ConvAttributeStart(builder) + +def ConvAttributeAddPad(builder, pad): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(pad), 0) + +def AddPad(builder, pad): + ConvAttributeAddPad(builder, pad) + +def ConvAttributeStartPadVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartPadVector(builder, numElems: int) -> int: + return ConvAttributeStartPadVector(builder, numElems) + +def ConvAttributeAddStride(builder, stride): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) + +def AddStride(builder, stride): + ConvAttributeAddStride(builder, stride) + +def ConvAttributeStartStrideVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartStrideVector(builder, numElems: int) -> int: + return ConvAttributeStartStrideVector(builder, numElems) + +def ConvAttributeAddDilation(builder, dilation): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(dilation), 0) + +def AddDilation(builder, dilation): + ConvAttributeAddDilation(builder, dilation) + +def ConvAttributeStartDilationVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartDilationVector(builder, numElems: int) -> int: + return ConvAttributeStartDilationVector(builder, numElems) + +def ConvAttributeAddInputZp(builder, inputZp): + builder.PrependInt32Slot(3, inputZp, 0) + +def AddInputZp(builder, inputZp): + ConvAttributeAddInputZp(builder, inputZp) + +def ConvAttributeAddWeightZp(builder, weightZp): + builder.PrependInt32Slot(4, weightZp, 0) + +def AddWeightZp(builder, weightZp): + ConvAttributeAddWeightZp(builder, weightZp) + +def ConvAttributeAddLocalBound(builder, localBound): + builder.PrependBoolSlot(5, localBound, 0) + +def AddLocalBound(builder, localBound): + ConvAttributeAddLocalBound(builder, localBound) + +def ConvAttributeAddAccType(builder, accType): + builder.PrependUint32Slot(6, accType, 0) + +def AddAccType(builder, accType): + ConvAttributeAddAccType(builder, accType) + +def ConvAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return ConvAttributeEnd(builder) diff --git a/python/tosa/CustomAttribute.py b/python/tosa/CustomAttribute.py new file mode 100644 index 0000000..db35dca --- /dev/null +++ b/python/tosa/CustomAttribute.py @@ -0,0 +1,106 @@ +# automatically generated by the FlatBuffers compiler, do not modify + +# namespace: tosa + +import flatbuffers +from flatbuffers.compat import import_numpy +np = import_numpy() + +class CustomAttribute(object): + __slots__ = ['_tab'] + + @classmethod + def GetRootAs(cls, buf, offset=0): + n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) + x = CustomAttribute() + x.Init(buf, n + offset) + return x + + @classmethod + def GetRootAsCustomAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod + def CustomAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): + return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) + + # CustomAttribute + def Init(self, buf, pos): + self._tab = flatbuffers.table.Table(buf, pos) + + # CustomAttribute + def OperatorName(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + if o != 0: + return self._tab.String(o + self._tab.Pos) + return None + + # CustomAttribute + def DomainName(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) + if o != 0: + return self._tab.String(o + self._tab.Pos) + return None + + # CustomAttribute + def ImplementationAttrs(self, j): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) + if o != 0: + a = self._tab.Vector(o) + return self._tab.Get(flatbuffers.number_types.Uint8Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 1)) + return 0 + + # CustomAttribute + def ImplementationAttrsAsNumpy(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) + if o != 0: + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Uint8Flags, o) + return 0 + + # CustomAttribute + def ImplementationAttrsLength(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) + if o != 0: + return self._tab.VectorLen(o) + return 0 + + # CustomAttribute + def ImplementationAttrsIsNone(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) + return o == 0 + +def CustomAttributeStart(builder): + builder.StartObject(3) + +def Start(builder): + CustomAttributeStart(builder) + +def CustomAttributeAddOperatorName(builder, operatorName): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(operatorName), 0) + +def AddOperatorName(builder, operatorName): + CustomAttributeAddOperatorName(builder, operatorName) + +def CustomAttributeAddDomainName(builder, domainName): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(domainName), 0) + +def AddDomainName(builder, domainName): + CustomAttributeAddDomainName(builder, domainName) + +def CustomAttributeAddImplementationAttrs(builder, implementationAttrs): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(implementationAttrs), 0) + +def AddImplementationAttrs(builder, implementationAttrs): + CustomAttributeAddImplementationAttrs(builder, implementationAttrs) + +def CustomAttributeStartImplementationAttrsVector(builder, numElems): + return builder.StartVector(1, numElems, 1) + +def StartImplementationAttrsVector(builder, numElems: int) -> int: + return CustomAttributeStartImplementationAttrsVector(builder, numElems) + +def CustomAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return CustomAttributeEnd(builder) diff --git a/python/tosa/DType.py b/python/tosa/DType.py index ad9048b..e585cb9 100644 --- a/python/tosa/DType.py +++ b/python/tosa/DType.py @@ -11,6 +11,10 @@ class DType(object): INT16 = 5 INT32 = 6 INT48 = 7 - FLOAT = 8 + FP32 = 8 UINT16 = 9 - + FP16 = 10 + BF16 = 11 + SHAPE = 12 + FP8E4M3 = 13 + FP8E5M2 = 14 diff --git a/python/tosa/FFTAttribute.py b/python/tosa/FFTAttribute.py new file mode 100644 index 0000000..d1624c2 --- /dev/null +++ b/python/tosa/FFTAttribute.py @@ -0,0 +1,67 @@ +# automatically generated by the FlatBuffers compiler, do not modify + +# namespace: tosa + +import flatbuffers +from flatbuffers.compat import import_numpy +np = import_numpy() + +class FFTAttribute(object): + __slots__ = ['_tab'] + + @classmethod + def GetRootAs(cls, buf, offset=0): + n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) + x = FFTAttribute() + x.Init(buf, n + offset) + return x + + @classmethod + def GetRootAsFFTAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod + def FFTAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): + return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) + + # FFTAttribute + def Init(self, buf, pos): + self._tab = flatbuffers.table.Table(buf, pos) + + # FFTAttribute + def Inverse(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + + # FFTAttribute + def LocalBound(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + +def FFTAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + FFTAttributeStart(builder) + +def FFTAttributeAddInverse(builder, inverse): + builder.PrependBoolSlot(0, inverse, 0) + +def AddInverse(builder, inverse): + FFTAttributeAddInverse(builder, inverse) + +def FFTAttributeAddLocalBound(builder, localBound): + builder.PrependBoolSlot(1, localBound, 0) + +def AddLocalBound(builder, localBound): + FFTAttributeAddLocalBound(builder, localBound) + +def FFTAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return FFTAttributeEnd(builder) diff --git a/python/tosa/FullyConnectedAttribute.py b/python/tosa/FullyConnectedAttribute.py index 62b480d..8854503 100644 --- a/python/tosa/FullyConnectedAttribute.py +++ b/python/tosa/FullyConnectedAttribute.py @@ -10,13 +10,17 @@ class FullyConnectedAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsFullyConnectedAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = FullyConnectedAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsFullyConnectedAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def FullyConnectedAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -38,7 +42,26 @@ class FullyConnectedAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def FullyConnectedAttributeStart(builder): builder.StartObject(2) -def FullyConnectedAttributeAddInputZp(builder, inputZp): builder.PrependInt32Slot(0, inputZp, 0) -def FullyConnectedAttributeAddWeightZp(builder, weightZp): builder.PrependInt32Slot(1, weightZp, 0) -def FullyConnectedAttributeEnd(builder): return builder.EndObject() +def FullyConnectedAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + FullyConnectedAttributeStart(builder) + +def FullyConnectedAttributeAddInputZp(builder, inputZp): + builder.PrependInt32Slot(0, inputZp, 0) + +def AddInputZp(builder, inputZp): + FullyConnectedAttributeAddInputZp(builder, inputZp) + +def FullyConnectedAttributeAddWeightZp(builder, weightZp): + builder.PrependInt32Slot(1, weightZp, 0) + +def AddWeightZp(builder, weightZp): + FullyConnectedAttributeAddWeightZp(builder, weightZp) + +def FullyConnectedAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return FullyConnectedAttributeEnd(builder) diff --git a/python/tosa/MatMulAttribute.py b/python/tosa/MatMulAttribute.py index 601f13f..325428a 100644 --- a/python/tosa/MatMulAttribute.py +++ b/python/tosa/MatMulAttribute.py @@ -10,13 +10,17 @@ class MatMulAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsMatMulAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = MatMulAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsMatMulAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def MatMulAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -38,7 +42,26 @@ class MatMulAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def MatMulAttributeStart(builder): builder.StartObject(2) -def MatMulAttributeAddAZp(builder, aZp): builder.PrependInt32Slot(0, aZp, 0) -def MatMulAttributeAddBZp(builder, bZp): builder.PrependInt32Slot(1, bZp, 0) -def MatMulAttributeEnd(builder): return builder.EndObject() +def MatMulAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + MatMulAttributeStart(builder) + +def MatMulAttributeAddAZp(builder, aZp): + builder.PrependInt32Slot(0, aZp, 0) + +def AddAZp(builder, aZp): + MatMulAttributeAddAZp(builder, aZp) + +def MatMulAttributeAddBZp(builder, bZp): + builder.PrependInt32Slot(1, bZp, 0) + +def AddBZp(builder, bZp): + MatMulAttributeAddBZp(builder, bZp) + +def MatMulAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return MatMulAttributeEnd(builder) diff --git a/python/tosa/MulAttribute.py b/python/tosa/MulAttribute.py index 79be4d3..236d3f5 100644 --- a/python/tosa/MulAttribute.py +++ b/python/tosa/MulAttribute.py @@ -10,13 +10,17 @@ class MulAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsMulAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = MulAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsMulAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def MulAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -31,6 +35,20 @@ class MulAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def MulAttributeStart(builder): builder.StartObject(1) -def MulAttributeAddShift(builder, shift): builder.PrependInt32Slot(0, shift, 0) -def MulAttributeEnd(builder): return builder.EndObject() +def MulAttributeStart(builder): + builder.StartObject(1) + +def Start(builder): + MulAttributeStart(builder) + +def MulAttributeAddShift(builder, shift): + builder.PrependInt32Slot(0, shift, 0) + +def AddShift(builder, shift): + MulAttributeAddShift(builder, shift) + +def MulAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return MulAttributeEnd(builder) diff --git a/python/tosa/NegateAttribute.py b/python/tosa/NegateAttribute.py index 24a57dc..eae46f5 100644 --- a/python/tosa/NegateAttribute.py +++ b/python/tosa/NegateAttribute.py @@ -10,13 +10,17 @@ class NegateAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsNegateAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = NegateAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsNegateAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def NegateAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -38,7 +42,26 @@ class NegateAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def NegateAttributeStart(builder): builder.StartObject(2) -def NegateAttributeAddInput1Zp(builder, input1Zp): builder.PrependInt32Slot(0, input1Zp, 0) -def NegateAttributeAddOutputZp(builder, outputZp): builder.PrependInt32Slot(1, outputZp, 0) -def NegateAttributeEnd(builder): return builder.EndObject() +def NegateAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + NegateAttributeStart(builder) + +def NegateAttributeAddInput1Zp(builder, input1Zp): + builder.PrependInt32Slot(0, input1Zp, 0) + +def AddInput1Zp(builder, input1Zp): + NegateAttributeAddInput1Zp(builder, input1Zp) + +def NegateAttributeAddOutputZp(builder, outputZp): + builder.PrependInt32Slot(1, outputZp, 0) + +def AddOutputZp(builder, outputZp): + NegateAttributeAddOutputZp(builder, outputZp) + +def NegateAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return NegateAttributeEnd(builder) diff --git a/python/tosa/Op.py b/python/tosa/Op.py index 181e457..35b2b80 100644 --- a/python/tosa/Op.py +++ b/python/tosa/Op.py @@ -72,4 +72,15 @@ class Op(object): CUSTOM = 66 COND_IF = 67 WHILE_LOOP = 68 - + FFT2D = 69 + RFFT2D = 70 + ERF = 71 + DIM = 72 + CONST_SHAPE = 73 + CONCAT_SHAPE = 74 + ADD_SHAPE = 75 + SUB_SHAPE = 76 + MUL_SHAPE = 77 + DIV_SHAPE = 78 + COS = 79 + SIN = 80 diff --git a/python/tosa/PadAttribute.py b/python/tosa/PadAttribute.py index 0875481..301bf17 100644 --- a/python/tosa/PadAttribute.py +++ b/python/tosa/PadAttribute.py @@ -10,13 +10,17 @@ class PadAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsPadAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = PadAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsPadAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def PadAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -25,49 +29,52 @@ class PadAttribute(object): self._tab = flatbuffers.table.Table(buf, pos) # PadAttribute - def Padding(self, j): + def PadConst(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) + return self._tab.Get(flatbuffers.number_types.Uint8Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 1)) return 0 # PadAttribute - def PaddingAsNumpy(self): + def PadConstAsNumpy(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Uint8Flags, o) return 0 # PadAttribute - def PaddingLength(self): + def PadConstLength(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: return self._tab.VectorLen(o) return 0 # PadAttribute - def PaddingIsNone(self): + def PadConstIsNone(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) return o == 0 - # PadAttribute - def PadConstInt(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) - if o != 0: - return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) - return 0 +def PadAttributeStart(builder): + builder.StartObject(1) - # PadAttribute - def PadConstFp(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) - if o != 0: - return self._tab.Get(flatbuffers.number_types.Float32Flags, o + self._tab.Pos) - return 0.0 - -def PadAttributeStart(builder): builder.StartObject(3) -def PadAttributeAddPadding(builder, padding): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(padding), 0) -def PadAttributeStartPaddingVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def PadAttributeAddPadConstInt(builder, padConstInt): builder.PrependInt32Slot(1, padConstInt, 0) -def PadAttributeAddPadConstFp(builder, padConstFp): builder.PrependFloat32Slot(2, padConstFp, 0.0) -def PadAttributeEnd(builder): return builder.EndObject() +def Start(builder): + PadAttributeStart(builder) + +def PadAttributeAddPadConst(builder, padConst): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(padConst), 0) + +def AddPadConst(builder, padConst): + PadAttributeAddPadConst(builder, padConst) + +def PadAttributeStartPadConstVector(builder, numElems): + return builder.StartVector(1, numElems, 1) + +def StartPadConstVector(builder, numElems: int) -> int: + return PadAttributeStartPadConstVector(builder, numElems) + +def PadAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return PadAttributeEnd(builder) diff --git a/python/tosa/PoolAttribute.py b/python/tosa/PoolAttribute.py index 8b6903e..c13e038 100644 --- a/python/tosa/PoolAttribute.py +++ b/python/tosa/PoolAttribute.py @@ -10,13 +10,17 @@ class PoolAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsPoolAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = PoolAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsPoolAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def PoolAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -119,13 +123,75 @@ class PoolAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def PoolAttributeStart(builder): builder.StartObject(5) -def PoolAttributeAddPad(builder, pad): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(pad), 0) -def PoolAttributeStartPadVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def PoolAttributeAddKernel(builder, kernel): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(kernel), 0) -def PoolAttributeStartKernelVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def PoolAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) -def PoolAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def PoolAttributeAddInputZp(builder, inputZp): builder.PrependInt32Slot(3, inputZp, 0) -def PoolAttributeAddOutputZp(builder, outputZp): builder.PrependInt32Slot(4, outputZp, 0) -def PoolAttributeEnd(builder): return builder.EndObject() + # PoolAttribute + def AccType(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) + if o != 0: + return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos) + return 0 + +def PoolAttributeStart(builder): + builder.StartObject(6) + +def Start(builder): + PoolAttributeStart(builder) + +def PoolAttributeAddPad(builder, pad): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(pad), 0) + +def AddPad(builder, pad): + PoolAttributeAddPad(builder, pad) + +def PoolAttributeStartPadVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartPadVector(builder, numElems: int) -> int: + return PoolAttributeStartPadVector(builder, numElems) + +def PoolAttributeAddKernel(builder, kernel): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(kernel), 0) + +def AddKernel(builder, kernel): + PoolAttributeAddKernel(builder, kernel) + +def PoolAttributeStartKernelVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartKernelVector(builder, numElems: int) -> int: + return PoolAttributeStartKernelVector(builder, numElems) + +def PoolAttributeAddStride(builder, stride): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) + +def AddStride(builder, stride): + PoolAttributeAddStride(builder, stride) + +def PoolAttributeStartStrideVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartStrideVector(builder, numElems: int) -> int: + return PoolAttributeStartStrideVector(builder, numElems) + +def PoolAttributeAddInputZp(builder, inputZp): + builder.PrependInt32Slot(3, inputZp, 0) + +def AddInputZp(builder, inputZp): + PoolAttributeAddInputZp(builder, inputZp) + +def PoolAttributeAddOutputZp(builder, outputZp): + builder.PrependInt32Slot(4, outputZp, 0) + +def AddOutputZp(builder, outputZp): + PoolAttributeAddOutputZp(builder, outputZp) + +def PoolAttributeAddAccType(builder, accType): + builder.PrependUint32Slot(5, accType, 0) + +def AddAccType(builder, accType): + PoolAttributeAddAccType(builder, accType) + +def PoolAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return PoolAttributeEnd(builder) diff --git a/python/tosa/RFFTAttribute.py b/python/tosa/RFFTAttribute.py new file mode 100644 index 0000000..7f76024 --- /dev/null +++ b/python/tosa/RFFTAttribute.py @@ -0,0 +1,54 @@ +# automatically generated by the FlatBuffers compiler, do not modify + +# namespace: tosa + +import flatbuffers +from flatbuffers.compat import import_numpy +np = import_numpy() + +class RFFTAttribute(object): + __slots__ = ['_tab'] + + @classmethod + def GetRootAs(cls, buf, offset=0): + n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) + x = RFFTAttribute() + x.Init(buf, n + offset) + return x + + @classmethod + def GetRootAsRFFTAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod + def RFFTAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): + return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) + + # RFFTAttribute + def Init(self, buf, pos): + self._tab = flatbuffers.table.Table(buf, pos) + + # RFFTAttribute + def LocalBound(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + +def RFFTAttributeStart(builder): + builder.StartObject(1) + +def Start(builder): + RFFTAttributeStart(builder) + +def RFFTAttributeAddLocalBound(builder, localBound): + builder.PrependBoolSlot(0, localBound, 0) + +def AddLocalBound(builder, localBound): + RFFTAttributeAddLocalBound(builder, localBound) + +def RFFTAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return RFFTAttributeEnd(builder) diff --git a/python/tosa/RescaleAttribute.py b/python/tosa/RescaleAttribute.py index 8ba68aa..12e7ced 100644 --- a/python/tosa/RescaleAttribute.py +++ b/python/tosa/RescaleAttribute.py @@ -10,13 +10,17 @@ class RescaleAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsRescaleAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = RescaleAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsRescaleAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def RescaleAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -39,88 +43,90 @@ class RescaleAttribute(object): return 0 # RescaleAttribute - def Multiplier(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 - - # RescaleAttribute - def MultiplierAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) - return 0 - - # RescaleAttribute - def MultiplierLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) - if o != 0: - return self._tab.VectorLen(o) - return 0 - - # RescaleAttribute - def MultiplierIsNone(self): + def Scale32(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) - return o == 0 - - # RescaleAttribute - def Shift(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 - - # RescaleAttribute - def ShiftAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) - return 0 + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False # RescaleAttribute - def ShiftLength(self): + def DoubleRound(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) if o != 0: - return self._tab.VectorLen(o) - return 0 - - # RescaleAttribute - def ShiftIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) - return o == 0 + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False # RescaleAttribute - def Scale32(self): + def PerChannel(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) if o != 0: return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) return False # RescaleAttribute - def DoubleRound(self): + def InputUnsigned(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) if o != 0: return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) return False # RescaleAttribute - def PerChannel(self): + def OutputUnsigned(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16)) if o != 0: return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) return False -def RescaleAttributeStart(builder): builder.StartObject(7) -def RescaleAttributeAddInputZp(builder, inputZp): builder.PrependInt32Slot(0, inputZp, 0) -def RescaleAttributeAddOutputZp(builder, outputZp): builder.PrependInt32Slot(1, outputZp, 0) -def RescaleAttributeAddMultiplier(builder, multiplier): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(multiplier), 0) -def RescaleAttributeStartMultiplierVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def RescaleAttributeAddShift(builder, shift): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(shift), 0) -def RescaleAttributeStartShiftVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def RescaleAttributeAddScale32(builder, scale32): builder.PrependBoolSlot(4, scale32, 0) -def RescaleAttributeAddDoubleRound(builder, doubleRound): builder.PrependBoolSlot(5, doubleRound, 0) -def RescaleAttributeAddPerChannel(builder, perChannel): builder.PrependBoolSlot(6, perChannel, 0) -def RescaleAttributeEnd(builder): return builder.EndObject() +def RescaleAttributeStart(builder): + builder.StartObject(7) + +def Start(builder): + RescaleAttributeStart(builder) + +def RescaleAttributeAddInputZp(builder, inputZp): + builder.PrependInt32Slot(0, inputZp, 0) + +def AddInputZp(builder, inputZp): + RescaleAttributeAddInputZp(builder, inputZp) + +def RescaleAttributeAddOutputZp(builder, outputZp): + builder.PrependInt32Slot(1, outputZp, 0) + +def AddOutputZp(builder, outputZp): + RescaleAttributeAddOutputZp(builder, outputZp) + +def RescaleAttributeAddScale32(builder, scale32): + builder.PrependBoolSlot(2, scale32, 0) + +def AddScale32(builder, scale32): + RescaleAttributeAddScale32(builder, scale32) + +def RescaleAttributeAddDoubleRound(builder, doubleRound): + builder.PrependBoolSlot(3, doubleRound, 0) + +def AddDoubleRound(builder, doubleRound): + RescaleAttributeAddDoubleRound(builder, doubleRound) + +def RescaleAttributeAddPerChannel(builder, perChannel): + builder.PrependBoolSlot(4, perChannel, 0) + +def AddPerChannel(builder, perChannel): + RescaleAttributeAddPerChannel(builder, perChannel) + +def RescaleAttributeAddInputUnsigned(builder, inputUnsigned): + builder.PrependBoolSlot(5, inputUnsigned, 0) + +def AddInputUnsigned(builder, inputUnsigned): + RescaleAttributeAddInputUnsigned(builder, inputUnsigned) + +def RescaleAttributeAddOutputUnsigned(builder, outputUnsigned): + builder.PrependBoolSlot(6, outputUnsigned, 0) + +def AddOutputUnsigned(builder, outputUnsigned): + RescaleAttributeAddOutputUnsigned(builder, outputUnsigned) + +def RescaleAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return RescaleAttributeEnd(builder) diff --git a/python/tosa/ReshapeAttribute.py b/python/tosa/ReshapeAttribute.py deleted file mode 100644 index 73b1ee8..0000000 --- a/python/tosa/ReshapeAttribute.py +++ /dev/null @@ -1,57 +0,0 @@ -# automatically generated by the FlatBuffers compiler, do not modify - -# namespace: tosa - -import flatbuffers -from flatbuffers.compat import import_numpy -np = import_numpy() - -class ReshapeAttribute(object): - __slots__ = ['_tab'] - - @classmethod - def GetRootAsReshapeAttribute(cls, buf, offset): - n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) - x = ReshapeAttribute() - x.Init(buf, n + offset) - return x - - @classmethod - def ReshapeAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): - return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) - - # ReshapeAttribute - def Init(self, buf, pos): - self._tab = flatbuffers.table.Table(buf, pos) - - # ReshapeAttribute - def NewShape(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 - - # ReshapeAttribute - def NewShapeAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) - return 0 - - # ReshapeAttribute - def NewShapeLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - return self._tab.VectorLen(o) - return 0 - - # ReshapeAttribute - def NewShapeIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - return o == 0 - -def ReshapeAttributeStart(builder): builder.StartObject(1) -def ReshapeAttributeAddNewShape(builder, newShape): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(newShape), 0) -def ReshapeAttributeStartNewShapeVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ReshapeAttributeEnd(builder): return builder.EndObject() diff --git a/python/tosa/ResizeAttribute.py b/python/tosa/ResizeAttribute.py index 1ed2dc0..96bfa56 100644 --- a/python/tosa/ResizeAttribute.py +++ b/python/tosa/ResizeAttribute.py @@ -10,13 +10,17 @@ class ResizeAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsResizeAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = ResizeAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsResizeAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def ResizeAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -25,165 +29,143 @@ class ResizeAttribute(object): self._tab = flatbuffers.table.Table(buf, pos) # ResizeAttribute - def OutputSize(self, j): + def Scale(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) + return self._tab.Get(flatbuffers.number_types.Int16Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 2)) return 0 # ResizeAttribute - def OutputSizeAsNumpy(self): + def ScaleAsNumpy(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int16Flags, o) return 0 # ResizeAttribute - def OutputSizeLength(self): + def ScaleLength(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: return self._tab.VectorLen(o) return 0 # ResizeAttribute - def OutputSizeIsNone(self): + def ScaleIsNone(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) return o == 0 # ResizeAttribute - def Stride(self, j): + def Offset(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) + return self._tab.Get(flatbuffers.number_types.Int16Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 2)) return 0 # ResizeAttribute - def StrideAsNumpy(self): + def OffsetAsNumpy(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int16Flags, o) return 0 # ResizeAttribute - def StrideLength(self): + def OffsetLength(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: return self._tab.VectorLen(o) return 0 # ResizeAttribute - def StrideIsNone(self): + def OffsetIsNone(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) return o == 0 # ResizeAttribute - def Offset(self, j): + def Border(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) if o != 0: a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) + return self._tab.Get(flatbuffers.number_types.Int16Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 2)) return 0 # ResizeAttribute - def OffsetAsNumpy(self): + def BorderAsNumpy(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) + return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int16Flags, o) return 0 # ResizeAttribute - def OffsetLength(self): + def BorderLength(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) if o != 0: return self._tab.VectorLen(o) return 0 # ResizeAttribute - def OffsetIsNone(self): + def BorderIsNone(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) return o == 0 # ResizeAttribute - def Shift(self): + def Mode(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) if o != 0: - return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) + return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos) return 0 - # ResizeAttribute - def StrideFp(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Float32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 +def ResizeAttributeStart(builder): + builder.StartObject(4) - # ResizeAttribute - def StrideFpAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Float32Flags, o) - return 0 +def Start(builder): + ResizeAttributeStart(builder) - # ResizeAttribute - def StrideFpLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) - if o != 0: - return self._tab.VectorLen(o) - return 0 +def ResizeAttributeAddScale(builder, scale): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(scale), 0) - # ResizeAttribute - def StrideFpIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) - return o == 0 +def AddScale(builder, scale): + ResizeAttributeAddScale(builder, scale) - # ResizeAttribute - def OffsetFp(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Float32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 +def ResizeAttributeStartScaleVector(builder, numElems): + return builder.StartVector(2, numElems, 2) - # ResizeAttribute - def OffsetFpAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Float32Flags, o) - return 0 +def StartScaleVector(builder, numElems: int) -> int: + return ResizeAttributeStartScaleVector(builder, numElems) - # ResizeAttribute - def OffsetFpLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) - if o != 0: - return self._tab.VectorLen(o) - return 0 +def ResizeAttributeAddOffset(builder, offset): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(offset), 0) - # ResizeAttribute - def OffsetFpIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) - return o == 0 +def AddOffset(builder, offset): + ResizeAttributeAddOffset(builder, offset) - # ResizeAttribute - def Mode(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16)) - if o != 0: - return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos) - return 0 +def ResizeAttributeStartOffsetVector(builder, numElems): + return builder.StartVector(2, numElems, 2) + +def StartOffsetVector(builder, numElems: int) -> int: + return ResizeAttributeStartOffsetVector(builder, numElems) + +def ResizeAttributeAddBorder(builder, border): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(border), 0) + +def AddBorder(builder, border): + ResizeAttributeAddBorder(builder, border) + +def ResizeAttributeStartBorderVector(builder, numElems): + return builder.StartVector(2, numElems, 2) + +def StartBorderVector(builder, numElems: int) -> int: + return ResizeAttributeStartBorderVector(builder, numElems) + +def ResizeAttributeAddMode(builder, mode): + builder.PrependUint32Slot(3, mode, 0) + +def AddMode(builder, mode): + ResizeAttributeAddMode(builder, mode) + +def ResizeAttributeEnd(builder): + return builder.EndObject() -def ResizeAttributeStart(builder): builder.StartObject(7) -def ResizeAttributeAddOutputSize(builder, outputSize): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(outputSize), 0) -def ResizeAttributeStartOutputSizeVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ResizeAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) -def ResizeAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ResizeAttributeAddOffset(builder, offset): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(offset), 0) -def ResizeAttributeStartOffsetVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ResizeAttributeAddShift(builder, shift): builder.PrependInt32Slot(3, shift, 0) -def ResizeAttributeAddStrideFp(builder, strideFp): builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(strideFp), 0) -def ResizeAttributeStartStrideFpVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ResizeAttributeAddOffsetFp(builder, offsetFp): builder.PrependUOffsetTRelativeSlot(5, flatbuffers.number_types.UOffsetTFlags.py_type(offsetFp), 0) -def ResizeAttributeStartOffsetFpVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def ResizeAttributeAddMode(builder, mode): builder.PrependUint32Slot(6, mode, 0) -def ResizeAttributeEnd(builder): return builder.EndObject() +def End(builder): + return ResizeAttributeEnd(builder) diff --git a/python/tosa/ResizeMode.py b/python/tosa/ResizeMode.py index 65bcd5d..388ecda 100644 --- a/python/tosa/ResizeMode.py +++ b/python/tosa/ResizeMode.py @@ -6,4 +6,3 @@ class ResizeMode(object): UNKNOWN = 0 NEAREST = 1 BILINEAR = 2 - diff --git a/python/tosa/SliceAttribute.py b/python/tosa/SliceAttribute.py deleted file mode 100644 index d3f6073..0000000 --- a/python/tosa/SliceAttribute.py +++ /dev/null @@ -1,86 +0,0 @@ -# automatically generated by the FlatBuffers compiler, do not modify - -# namespace: tosa - -import flatbuffers -from flatbuffers.compat import import_numpy -np = import_numpy() - -class SliceAttribute(object): - __slots__ = ['_tab'] - - @classmethod - def GetRootAsSliceAttribute(cls, buf, offset): - n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) - x = SliceAttribute() - x.Init(buf, n + offset) - return x - - @classmethod - def SliceAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): - return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) - - # SliceAttribute - def Init(self, buf, pos): - self._tab = flatbuffers.table.Table(buf, pos) - - # SliceAttribute - def Start(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 - - # SliceAttribute - def StartAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) - return 0 - - # SliceAttribute - def StartLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - return self._tab.VectorLen(o) - return 0 - - # SliceAttribute - def StartIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - return o == 0 - - # SliceAttribute - def Size(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 - - # SliceAttribute - def SizeAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) - return 0 - - # SliceAttribute - def SizeLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) - if o != 0: - return self._tab.VectorLen(o) - return 0 - - # SliceAttribute - def SizeIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) - return o == 0 - -def SliceAttributeStart(builder): builder.StartObject(2) -def SliceAttributeAddStart(builder, start): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(start), 0) -def SliceAttributeStartStartVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def SliceAttributeAddSize(builder, size): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(size), 0) -def SliceAttributeStartSizeVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def SliceAttributeEnd(builder): return builder.EndObject() diff --git a/python/tosa/TableAttribute.py b/python/tosa/TableAttribute.py index 49a5c9a..6caa1f2 100644 --- a/python/tosa/TableAttribute.py +++ b/python/tosa/TableAttribute.py @@ -10,13 +10,17 @@ class TableAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTableAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TableAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTableAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TableAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -51,7 +55,26 @@ class TableAttribute(object): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) return o == 0 -def TableAttributeStart(builder): builder.StartObject(1) -def TableAttributeAddTable(builder, table): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(table), 0) -def TableAttributeStartTableVector(builder, numElems): return builder.StartVector(2, numElems, 2) -def TableAttributeEnd(builder): return builder.EndObject() +def TableAttributeStart(builder): + builder.StartObject(1) + +def Start(builder): + TableAttributeStart(builder) + +def TableAttributeAddTable(builder, table): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(table), 0) + +def AddTable(builder, table): + TableAttributeAddTable(builder, table) + +def TableAttributeStartTableVector(builder, numElems): + return builder.StartVector(2, numElems, 2) + +def StartTableVector(builder, numElems: int) -> int: + return TableAttributeStartTableVector(builder, numElems) + +def TableAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return TableAttributeEnd(builder) diff --git a/python/tosa/TileAttribute.py b/python/tosa/TileAttribute.py deleted file mode 100644 index 03dd0fb..0000000 --- a/python/tosa/TileAttribute.py +++ /dev/null @@ -1,57 +0,0 @@ -# automatically generated by the FlatBuffers compiler, do not modify - -# namespace: tosa - -import flatbuffers -from flatbuffers.compat import import_numpy -np = import_numpy() - -class TileAttribute(object): - __slots__ = ['_tab'] - - @classmethod - def GetRootAsTileAttribute(cls, buf, offset): - n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) - x = TileAttribute() - x.Init(buf, n + offset) - return x - - @classmethod - def TileAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): - return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) - - # TileAttribute - def Init(self, buf, pos): - self._tab = flatbuffers.table.Table(buf, pos) - - # TileAttribute - def Multiples(self, j): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - a = self._tab.Vector(o) - return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4)) - return 0 - - # TileAttribute - def MultiplesAsNumpy(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o) - return 0 - - # TileAttribute - def MultiplesLength(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - if o != 0: - return self._tab.VectorLen(o) - return 0 - - # TileAttribute - def MultiplesIsNone(self): - o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) - return o == 0 - -def TileAttributeStart(builder): builder.StartObject(1) -def TileAttributeAddMultiples(builder, multiples): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(multiples), 0) -def TileAttributeStartMultiplesVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TileAttributeEnd(builder): return builder.EndObject() diff --git a/python/tosa/TosaBasicBlock.py b/python/tosa/TosaBasicBlock.py index 1c93c63..b31f455 100644 --- a/python/tosa/TosaBasicBlock.py +++ b/python/tosa/TosaBasicBlock.py @@ -10,13 +10,17 @@ class TosaBasicBlock(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTosaBasicBlock(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TosaBasicBlock() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTosaBasicBlock(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TosaBasicBlockBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -121,14 +125,68 @@ class TosaBasicBlock(object): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) return o == 0 -def TosaBasicBlockStart(builder): builder.StartObject(5) -def TosaBasicBlockAddName(builder, name): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0) -def TosaBasicBlockAddOperators(builder, operators): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(operators), 0) -def TosaBasicBlockStartOperatorsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaBasicBlockAddTensors(builder, tensors): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(tensors), 0) -def TosaBasicBlockStartTensorsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaBasicBlockAddInputs(builder, inputs): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(inputs), 0) -def TosaBasicBlockStartInputsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaBasicBlockAddOutputs(builder, outputs): builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(outputs), 0) -def TosaBasicBlockStartOutputsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaBasicBlockEnd(builder): return builder.EndObject() +def TosaBasicBlockStart(builder): + builder.StartObject(5) + +def Start(builder): + TosaBasicBlockStart(builder) + +def TosaBasicBlockAddName(builder, name): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0) + +def AddName(builder, name): + TosaBasicBlockAddName(builder, name) + +def TosaBasicBlockAddOperators(builder, operators): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(operators), 0) + +def AddOperators(builder, operators): + TosaBasicBlockAddOperators(builder, operators) + +def TosaBasicBlockStartOperatorsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartOperatorsVector(builder, numElems: int) -> int: + return TosaBasicBlockStartOperatorsVector(builder, numElems) + +def TosaBasicBlockAddTensors(builder, tensors): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(tensors), 0) + +def AddTensors(builder, tensors): + TosaBasicBlockAddTensors(builder, tensors) + +def TosaBasicBlockStartTensorsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartTensorsVector(builder, numElems: int) -> int: + return TosaBasicBlockStartTensorsVector(builder, numElems) + +def TosaBasicBlockAddInputs(builder, inputs): + builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(inputs), 0) + +def AddInputs(builder, inputs): + TosaBasicBlockAddInputs(builder, inputs) + +def TosaBasicBlockStartInputsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartInputsVector(builder, numElems: int) -> int: + return TosaBasicBlockStartInputsVector(builder, numElems) + +def TosaBasicBlockAddOutputs(builder, outputs): + builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(outputs), 0) + +def AddOutputs(builder, outputs): + TosaBasicBlockAddOutputs(builder, outputs) + +def TosaBasicBlockStartOutputsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartOutputsVector(builder, numElems: int) -> int: + return TosaBasicBlockStartOutputsVector(builder, numElems) + +def TosaBasicBlockEnd(builder): + return builder.EndObject() + +def End(builder): + return TosaBasicBlockEnd(builder) diff --git a/python/tosa/TosaGraph.py b/python/tosa/TosaGraph.py index eceffdb..84b51a7 100644 --- a/python/tosa/TosaGraph.py +++ b/python/tosa/TosaGraph.py @@ -10,13 +10,17 @@ class TosaGraph(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTosaGraph(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TosaGraph() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTosaGraph(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TosaGraphBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -36,32 +40,56 @@ class TosaGraph(object): return None # TosaGraph - def Blocks(self, j): + def Regions(self, j): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: x = self._tab.Vector(o) x += flatbuffers.number_types.UOffsetTFlags.py_type(j) * 4 x = self._tab.Indirect(x) - from tosa.TosaBasicBlock import TosaBasicBlock - obj = TosaBasicBlock() + from tosa.TosaRegion import TosaRegion + obj = TosaRegion() obj.Init(self._tab.Bytes, x) return obj return None # TosaGraph - def BlocksLength(self): + def RegionsLength(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: return self._tab.VectorLen(o) return 0 # TosaGraph - def BlocksIsNone(self): + def RegionsIsNone(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) return o == 0 -def TosaGraphStart(builder): builder.StartObject(2) -def TosaGraphAddVersion(builder, version): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(version), 0) -def TosaGraphAddBlocks(builder, blocks): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(blocks), 0) -def TosaGraphStartBlocksVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaGraphEnd(builder): return builder.EndObject() +def TosaGraphStart(builder): + builder.StartObject(2) + +def Start(builder): + TosaGraphStart(builder) + +def TosaGraphAddVersion(builder, version): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(version), 0) + +def AddVersion(builder, version): + TosaGraphAddVersion(builder, version) + +def TosaGraphAddRegions(builder, regions): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(regions), 0) + +def AddRegions(builder, regions): + TosaGraphAddRegions(builder, regions) + +def TosaGraphStartRegionsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartRegionsVector(builder, numElems: int) -> int: + return TosaGraphStartRegionsVector(builder, numElems) + +def TosaGraphEnd(builder): + return builder.EndObject() + +def End(builder): + return TosaGraphEnd(builder) diff --git a/python/tosa/TosaOperator.py b/python/tosa/TosaOperator.py index fd11f76..2b889ad 100644 --- a/python/tosa/TosaOperator.py +++ b/python/tosa/TosaOperator.py @@ -10,13 +10,17 @@ class TosaOperator(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTosaOperator(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TosaOperator() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTosaOperator(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TosaOperatorBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -88,12 +92,56 @@ class TosaOperator(object): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) return o == 0 -def TosaOperatorStart(builder): builder.StartObject(5) -def TosaOperatorAddOp(builder, op): builder.PrependUint32Slot(0, op, 0) -def TosaOperatorAddAttributeType(builder, attributeType): builder.PrependUint8Slot(1, attributeType, 0) -def TosaOperatorAddAttribute(builder, attribute): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(attribute), 0) -def TosaOperatorAddInputs(builder, inputs): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(inputs), 0) -def TosaOperatorStartInputsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaOperatorAddOutputs(builder, outputs): builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(outputs), 0) -def TosaOperatorStartOutputsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaOperatorEnd(builder): return builder.EndObject() +def TosaOperatorStart(builder): + builder.StartObject(5) + +def Start(builder): + TosaOperatorStart(builder) + +def TosaOperatorAddOp(builder, op): + builder.PrependUint32Slot(0, op, 0) + +def AddOp(builder, op): + TosaOperatorAddOp(builder, op) + +def TosaOperatorAddAttributeType(builder, attributeType): + builder.PrependUint8Slot(1, attributeType, 0) + +def AddAttributeType(builder, attributeType): + TosaOperatorAddAttributeType(builder, attributeType) + +def TosaOperatorAddAttribute(builder, attribute): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(attribute), 0) + +def AddAttribute(builder, attribute): + TosaOperatorAddAttribute(builder, attribute) + +def TosaOperatorAddInputs(builder, inputs): + builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(inputs), 0) + +def AddInputs(builder, inputs): + TosaOperatorAddInputs(builder, inputs) + +def TosaOperatorStartInputsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartInputsVector(builder, numElems: int) -> int: + return TosaOperatorStartInputsVector(builder, numElems) + +def TosaOperatorAddOutputs(builder, outputs): + builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(outputs), 0) + +def AddOutputs(builder, outputs): + TosaOperatorAddOutputs(builder, outputs) + +def TosaOperatorStartOutputsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartOutputsVector(builder, numElems: int) -> int: + return TosaOperatorStartOutputsVector(builder, numElems) + +def TosaOperatorEnd(builder): + return builder.EndObject() + +def End(builder): + return TosaOperatorEnd(builder) diff --git a/python/tosa/TosaRegion.py b/python/tosa/TosaRegion.py new file mode 100644 index 0000000..7fd6e3c --- /dev/null +++ b/python/tosa/TosaRegion.py @@ -0,0 +1,91 @@ +# automatically generated by the FlatBuffers compiler, do not modify + +# namespace: tosa + +import flatbuffers +from flatbuffers.compat import import_numpy +np = import_numpy() + +class TosaRegion(object): + __slots__ = ['_tab'] + + @classmethod + def GetRootAs(cls, buf, offset=0): + n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) + x = TosaRegion() + x.Init(buf, n + offset) + return x + + @classmethod + def GetRootAsTosaRegion(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod + def TosaRegionBufferHasIdentifier(cls, buf, offset, size_prefixed=False): + return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) + + # TosaRegion + def Init(self, buf, pos): + self._tab = flatbuffers.table.Table(buf, pos) + + # TosaRegion + def Name(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) + if o != 0: + return self._tab.String(o + self._tab.Pos) + return None + + # TosaRegion + def Blocks(self, j): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) + if o != 0: + x = self._tab.Vector(o) + x += flatbuffers.number_types.UOffsetTFlags.py_type(j) * 4 + x = self._tab.Indirect(x) + from tosa.TosaBasicBlock import TosaBasicBlock + obj = TosaBasicBlock() + obj.Init(self._tab.Bytes, x) + return obj + return None + + # TosaRegion + def BlocksLength(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) + if o != 0: + return self._tab.VectorLen(o) + return 0 + + # TosaRegion + def BlocksIsNone(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) + return o == 0 + +def TosaRegionStart(builder): + builder.StartObject(2) + +def Start(builder): + TosaRegionStart(builder) + +def TosaRegionAddName(builder, name): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0) + +def AddName(builder, name): + TosaRegionAddName(builder, name) + +def TosaRegionAddBlocks(builder, blocks): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(blocks), 0) + +def AddBlocks(builder, blocks): + TosaRegionAddBlocks(builder, blocks) + +def TosaRegionStartBlocksVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartBlocksVector(builder, numElems: int) -> int: + return TosaRegionStartBlocksVector(builder, numElems) + +def TosaRegionEnd(builder): + return builder.EndObject() + +def End(builder): + return TosaRegionEnd(builder) diff --git a/python/tosa/TosaTensor.py b/python/tosa/TosaTensor.py index a6f609d..3fb9f86 100644 --- a/python/tosa/TosaTensor.py +++ b/python/tosa/TosaTensor.py @@ -10,13 +10,17 @@ class TosaTensor(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTosaTensor(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TosaTensor() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTosaTensor(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TosaTensorBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -92,11 +96,89 @@ class TosaTensor(object): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) return o == 0 -def TosaTensorStart(builder): builder.StartObject(4) -def TosaTensorAddName(builder, name): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0) -def TosaTensorAddShape(builder, shape): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(shape), 0) -def TosaTensorStartShapeVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TosaTensorAddType(builder, type): builder.PrependUint32Slot(2, type, 0) -def TosaTensorAddData(builder, data): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(data), 0) -def TosaTensorStartDataVector(builder, numElems): return builder.StartVector(1, numElems, 1) -def TosaTensorEnd(builder): return builder.EndObject() + # TosaTensor + def Variable(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + + # TosaTensor + def IsUnranked(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + + # TosaTensor + def VariableName(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16)) + if o != 0: + return self._tab.String(o + self._tab.Pos) + return None + +def TosaTensorStart(builder): + builder.StartObject(7) + +def Start(builder): + TosaTensorStart(builder) + +def TosaTensorAddName(builder, name): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0) + +def AddName(builder, name): + TosaTensorAddName(builder, name) + +def TosaTensorAddShape(builder, shape): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(shape), 0) + +def AddShape(builder, shape): + TosaTensorAddShape(builder, shape) + +def TosaTensorStartShapeVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartShapeVector(builder, numElems: int) -> int: + return TosaTensorStartShapeVector(builder, numElems) + +def TosaTensorAddType(builder, type): + builder.PrependUint32Slot(2, type, 0) + +def AddType(builder, type): + TosaTensorAddType(builder, type) + +def TosaTensorAddData(builder, data): + builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(data), 0) + +def AddData(builder, data): + TosaTensorAddData(builder, data) + +def TosaTensorStartDataVector(builder, numElems): + return builder.StartVector(1, numElems, 1) + +def StartDataVector(builder, numElems: int) -> int: + return TosaTensorStartDataVector(builder, numElems) + +def TosaTensorAddVariable(builder, variable): + builder.PrependBoolSlot(4, variable, 0) + +def AddVariable(builder, variable): + TosaTensorAddVariable(builder, variable) + +def TosaTensorAddIsUnranked(builder, isUnranked): + builder.PrependBoolSlot(5, isUnranked, 0) + +def AddIsUnranked(builder, isUnranked): + TosaTensorAddIsUnranked(builder, isUnranked) + +def TosaTensorAddVariableName(builder, variableName): + builder.PrependUOffsetTRelativeSlot(6, flatbuffers.number_types.UOffsetTFlags.py_type(variableName), 0) + +def AddVariableName(builder, variableName): + TosaTensorAddVariableName(builder, variableName) + +def TosaTensorEnd(builder): + return builder.EndObject() + +def End(builder): + return TosaTensorEnd(builder) diff --git a/python/tosa/TransposeAttribute.py b/python/tosa/TransposeAttribute.py index 82e6b3a..71cfdf0 100644 --- a/python/tosa/TransposeAttribute.py +++ b/python/tosa/TransposeAttribute.py @@ -10,13 +10,17 @@ class TransposeAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTransposeAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TransposeAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTransposeAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TransposeAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -51,7 +55,26 @@ class TransposeAttribute(object): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) return o == 0 -def TransposeAttributeStart(builder): builder.StartObject(1) -def TransposeAttributeAddPerms(builder, perms): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(perms), 0) -def TransposeAttributeStartPermsVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TransposeAttributeEnd(builder): return builder.EndObject() +def TransposeAttributeStart(builder): + builder.StartObject(1) + +def Start(builder): + TransposeAttributeStart(builder) + +def TransposeAttributeAddPerms(builder, perms): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(perms), 0) + +def AddPerms(builder, perms): + TransposeAttributeAddPerms(builder, perms) + +def TransposeAttributeStartPermsVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartPermsVector(builder, numElems: int) -> int: + return TransposeAttributeStartPermsVector(builder, numElems) + +def TransposeAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return TransposeAttributeEnd(builder) diff --git a/python/tosa/TransposeConvAttribute.py b/python/tosa/TransposeConvAttribute.py index 8ca5ba7..e5397a8 100644 --- a/python/tosa/TransposeConvAttribute.py +++ b/python/tosa/TransposeConvAttribute.py @@ -10,13 +10,17 @@ class TransposeConvAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsTransposeConvAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = TransposeConvAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsTransposeConvAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def TransposeConvAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -119,13 +123,88 @@ class TransposeConvAttribute(object): return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) return 0 -def TransposeConvAttributeStart(builder): builder.StartObject(5) -def TransposeConvAttributeAddOutPad(builder, outPad): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(outPad), 0) -def TransposeConvAttributeStartOutPadVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TransposeConvAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) -def TransposeConvAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TransposeConvAttributeAddOutputShape(builder, outputShape): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(outputShape), 0) -def TransposeConvAttributeStartOutputShapeVector(builder, numElems): return builder.StartVector(4, numElems, 4) -def TransposeConvAttributeAddInputZp(builder, inputZp): builder.PrependInt32Slot(3, inputZp, 0) -def TransposeConvAttributeAddWeightZp(builder, weightZp): builder.PrependInt32Slot(4, weightZp, 0) -def TransposeConvAttributeEnd(builder): return builder.EndObject() + # TransposeConvAttribute + def LocalBound(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14)) + if o != 0: + return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) + return False + + # TransposeConvAttribute + def AccType(self): + o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16)) + if o != 0: + return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos) + return 0 + +def TransposeConvAttributeStart(builder): + builder.StartObject(7) + +def Start(builder): + TransposeConvAttributeStart(builder) + +def TransposeConvAttributeAddOutPad(builder, outPad): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(outPad), 0) + +def AddOutPad(builder, outPad): + TransposeConvAttributeAddOutPad(builder, outPad) + +def TransposeConvAttributeStartOutPadVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartOutPadVector(builder, numElems: int) -> int: + return TransposeConvAttributeStartOutPadVector(builder, numElems) + +def TransposeConvAttributeAddStride(builder, stride): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0) + +def AddStride(builder, stride): + TransposeConvAttributeAddStride(builder, stride) + +def TransposeConvAttributeStartStrideVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartStrideVector(builder, numElems: int) -> int: + return TransposeConvAttributeStartStrideVector(builder, numElems) + +def TransposeConvAttributeAddOutputShape(builder, outputShape): + builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(outputShape), 0) + +def AddOutputShape(builder, outputShape): + TransposeConvAttributeAddOutputShape(builder, outputShape) + +def TransposeConvAttributeStartOutputShapeVector(builder, numElems): + return builder.StartVector(4, numElems, 4) + +def StartOutputShapeVector(builder, numElems: int) -> int: + return TransposeConvAttributeStartOutputShapeVector(builder, numElems) + +def TransposeConvAttributeAddInputZp(builder, inputZp): + builder.PrependInt32Slot(3, inputZp, 0) + +def AddInputZp(builder, inputZp): + TransposeConvAttributeAddInputZp(builder, inputZp) + +def TransposeConvAttributeAddWeightZp(builder, weightZp): + builder.PrependInt32Slot(4, weightZp, 0) + +def AddWeightZp(builder, weightZp): + TransposeConvAttributeAddWeightZp(builder, weightZp) + +def TransposeConvAttributeAddLocalBound(builder, localBound): + builder.PrependBoolSlot(5, localBound, 0) + +def AddLocalBound(builder, localBound): + TransposeConvAttributeAddLocalBound(builder, localBound) + +def TransposeConvAttributeAddAccType(builder, accType): + builder.PrependUint32Slot(6, accType, 0) + +def AddAccType(builder, accType): + TransposeConvAttributeAddAccType(builder, accType) + +def TransposeConvAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return TransposeConvAttributeEnd(builder) diff --git a/python/tosa/Version.py b/python/tosa/Version.py index 06c3ba3..369f356 100644 --- a/python/tosa/Version.py +++ b/python/tosa/Version.py @@ -10,13 +10,17 @@ class Version(object): __slots__ = ['_tab'] @classmethod - def GetRootAsVersion(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = Version() x.Init(buf, n + offset) return x @classmethod + def GetRootAsVersion(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def VersionBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -25,36 +29,65 @@ class Version(object): self._tab = flatbuffers.table.Table(buf, pos) # Version - def _major(self): + def _Major(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) - return 0 + return -1 # Version - def _minor(self): + def _Minor(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) - return 31 + return -1 # Version - def _patch(self): + def _Patch(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8)) if o != 0: return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos) - return 0 + return -1 # Version - def _draft(self): + def _Draft(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10)) if o != 0: return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos)) return True -def VersionStart(builder): builder.StartObject(4) -def VersionAdd_major(builder, Major): builder.PrependInt32Slot(0, Major, 0) -def VersionAdd_minor(builder, Minor): builder.PrependInt32Slot(1, Minor, 31) -def VersionAdd_patch(builder, Patch): builder.PrependInt32Slot(2, Patch, 0) -def VersionAdd_draft(builder, Draft): builder.PrependBoolSlot(3, Draft, 1) -def VersionEnd(builder): return builder.EndObject() +def VersionStart(builder): + builder.StartObject(4) + +def Start(builder): + VersionStart(builder) + +def VersionAdd_Major(builder, _Major): + builder.PrependInt32Slot(0, _Major, -1) + +def Add_Major(builder, _Major): + VersionAdd_Major(builder, _Major) + +def VersionAdd_Minor(builder, _Minor): + builder.PrependInt32Slot(1, _Minor, -1) + +def Add_Minor(builder, _Minor): + VersionAdd_Minor(builder, _Minor) + +def VersionAdd_Patch(builder, _Patch): + builder.PrependInt32Slot(2, _Patch, -1) + +def Add_Patch(builder, _Patch): + VersionAdd_Patch(builder, _Patch) + +def VersionAdd_Draft(builder, _Draft): + builder.PrependBoolSlot(3, _Draft, 1) + +def Add_Draft(builder, _Draft): + VersionAdd_Draft(builder, _Draft) + +def VersionEnd(builder): + return builder.EndObject() + +def End(builder): + return VersionEnd(builder) diff --git a/python/tosa/WhileLoopAttribute.py b/python/tosa/WhileLoopAttribute.py index 1e18bca..5148c26 100644 --- a/python/tosa/WhileLoopAttribute.py +++ b/python/tosa/WhileLoopAttribute.py @@ -10,13 +10,17 @@ class WhileLoopAttribute(object): __slots__ = ['_tab'] @classmethod - def GetRootAsWhileLoopAttribute(cls, buf, offset): + def GetRootAs(cls, buf, offset=0): n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset) x = WhileLoopAttribute() x.Init(buf, n + offset) return x @classmethod + def GetRootAsWhileLoopAttribute(cls, buf, offset=0): + """This method is deprecated. Please switch to GetRootAs.""" + return cls.GetRootAs(buf, offset) + @classmethod def WhileLoopAttributeBufferHasIdentifier(cls, buf, offset, size_prefixed=False): return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x4F\x53\x41", size_prefixed=size_prefixed) @@ -25,20 +29,39 @@ class WhileLoopAttribute(object): self._tab = flatbuffers.table.Table(buf, pos) # WhileLoopAttribute - def CondBranch(self): + def CondGraph(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4)) if o != 0: return self._tab.String(o + self._tab.Pos) return None # WhileLoopAttribute - def BodyBranch(self): + def BodyGraph(self): o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6)) if o != 0: return self._tab.String(o + self._tab.Pos) return None -def WhileLoopAttributeStart(builder): builder.StartObject(2) -def WhileLoopAttributeAddCondBranch(builder, condBranch): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(condBranch), 0) -def WhileLoopAttributeAddBodyBranch(builder, bodyBranch): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(bodyBranch), 0) -def WhileLoopAttributeEnd(builder): return builder.EndObject() +def WhileLoopAttributeStart(builder): + builder.StartObject(2) + +def Start(builder): + WhileLoopAttributeStart(builder) + +def WhileLoopAttributeAddCondGraph(builder, condGraph): + builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(condGraph), 0) + +def AddCondGraph(builder, condGraph): + WhileLoopAttributeAddCondGraph(builder, condGraph) + +def WhileLoopAttributeAddBodyGraph(builder, bodyGraph): + builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(bodyGraph), 0) + +def AddBodyGraph(builder, bodyGraph): + WhileLoopAttributeAddBodyGraph(builder, bodyGraph) + +def WhileLoopAttributeEnd(builder): + return builder.EndObject() + +def End(builder): + return WhileLoopAttributeEnd(builder) diff --git a/schema/tosa.fbs b/schema/tosa.fbs index 57440ef..79b83b1 100644 --- a/schema/tosa.fbs +++ b/schema/tosa.fbs @@ -1,5 +1,5 @@ -// Copyright (c) 2020-2021, ARM Limited. +// Copyright (c) 2020-2024, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -20,6 +20,9 @@ file_identifier "TOSA"; // File extension of any written files. file_extension "tosa"; +// NOTE: New values added to the schema should be placed +// at the end of the list in order to keep schema stable. + enum DType:uint32 { UNKNOWN = 0, BOOL, @@ -29,8 +32,13 @@ enum DType:uint32 { INT16, INT32, INT48, - FLOAT, + FP32, UINT16, + FP16, + BF16, + SHAPE, + FP8E4M3, + FP8E5M2, } enum ResizeMode:uint32 { @@ -41,8 +49,6 @@ enum ResizeMode:uint32 { enum Op:uint32 { UNKNOWN = 0, - - // Tensor Operator ARGMAX, AVG_POOL2D, CONV2D, @@ -52,14 +58,10 @@ enum Op:uint32 { MATMUL, MAX_POOL2D, TRANSPOSE_CONV2D, - - // Activation CLAMP, RESERVED, SIGMOID, TANH, - - // Elementwise-Binary ADD, ARITHMETIC_RIGHT_SHIFT, BITWISE_AND, @@ -77,8 +79,6 @@ enum Op:uint32 { POW, SUB, TABLE, - - // Elementwise-Unary ABS, BITWISE_NOT, CEIL, @@ -90,24 +90,16 @@ enum Op:uint32 { NEGATE, RECIPROCAL, RSQRT, - - // Elementwise-Ternary SELECT, - - // Logical EQUAL, GREATER, GREATER_EQUAL, - - // Reduction REDUCE_ANY, REDUCE_ALL, REDUCE_MAX, REDUCE_MIN, REDUCE_PRODUCT, REDUCE_SUM, - - // Data layout operation CONCAT, PAD, RESHAPE, @@ -115,28 +107,28 @@ enum Op:uint32 { SLICE, TILE, TRANSPOSE, - - // Gather/scatter operation GATHER, SCATTER, - - // Image RESIZE, - - // Type conversion CAST, RESCALE, - - // Data Nodes CONST, IDENTITY, - - // Custom operations CUSTOM, - - // Control flow operators COND_IF, WHILE_LOOP, + FFT2D, + RFFT2D, + ERF, + DIM, + CONST_SHAPE, + CONCAT_SHAPE, + ADD_SHAPE, + SUB_SHAPE, + MUL_SHAPE, + DIV_SHAPE, + COS, + SIN, } union Attribute { @@ -145,9 +137,6 @@ union Attribute { TransposeConvAttribute, PadAttribute, AxisAttribute, - ReshapeAttribute, - SliceAttribute, - TileAttribute, ResizeAttribute, ClampAttribute, RescaleAttribute, @@ -159,7 +148,10 @@ union Attribute { TableAttribute, MatMulAttribute, FullyConnectedAttribute, - NegateAttribute + NegateAttribute, + CustomAttribute, + FFTAttribute, + RFFTAttribute, } table PoolAttribute { @@ -168,6 +160,7 @@ table PoolAttribute { stride: [int32]; input_zp: int32; output_zp: int32; + acc_type: DType; } table ConvAttribute { @@ -176,6 +169,8 @@ table ConvAttribute { dilation: [int32]; input_zp: int32; weight_zp: int32; + local_bound: bool; + acc_type: DType; } table TransposeConvAttribute { @@ -184,56 +179,38 @@ table TransposeConvAttribute { output_shape: [int32]; input_zp: int32; weight_zp: int32; + local_bound: bool; + acc_type: DType; } table PadAttribute { - padding: [int32]; - pad_const_int: int32; - pad_const_fp: float; + pad_const: [ubyte] (force_align: 8); } table AxisAttribute { axis: int32; } -table ReshapeAttribute { - new_shape: [int32]; -} - -table SliceAttribute { - start: [int32]; - size: [int32]; -} - -table TileAttribute { - multiples: [int32]; -} - table ResizeAttribute { - output_size: [int32]; - stride: [int32]; - offset: [int32]; - shift: int32; - stride_fp: [float]; - offset_fp: [float]; + scale: [int16]; + offset: [int16]; + border: [int16]; mode: ResizeMode; } table ClampAttribute { - min_int: int32; - max_int: int32; - min_fp: float; - max_fp: float; + min_val: [ubyte] (force_align: 8); + max_val: [ubyte] (force_align: 8); } table RescaleAttribute { input_zp: int32; output_zp: int32; - multiplier: [int32]; - shift: [int32]; scale32: bool; double_round: bool; per_channel: bool; + input_unsigned: bool; + output_unsigned: bool; } table MulAttribute { @@ -245,13 +222,13 @@ table ArithmeticRightShiftAttribute { } table CondIfAttribute { - then_branch: string; - else_branch: string; + then_graph: string; + else_graph: string; } table WhileLoopAttribute { - cond_branch: string; - body_branch: string; + cond_graph: string; + body_graph: string; } table TransposeAttribute { @@ -277,10 +254,25 @@ table NegateAttribute { output_zp: int32; } +table CustomAttribute { + operator_name:string; + domain_name:string; + implementation_attrs:[ubyte]; +} + +table FFTAttribute { + inverse: bool; + local_bound: bool; +} + +table RFFTAttribute { + local_bound: bool; +} + table Version { - _major: int32 = 0; - _minor: int32 = 31; - _patch: int32 = 0; + _major: int32 = -1; + _minor: int32 = -1; + _patch: int32 = -1; _draft: bool = true; } @@ -289,11 +281,14 @@ table TosaTensor { shape:[int32]; // shape of the tensor type:DType; // data type of the tensor data: [ubyte] (force_align: 8); // raw data array if it's a constant tensor. + variable: bool; // is this a variable tensor + is_unranked: bool; // whether this is an unranked tensor + variable_name:string; // name for variable attribute } table TosaOperator { op:Op; // operator enum - attribute: Attribute; // union structure. operator attribute + attribute:Attribute; // union structure. operator attribute inputs:[string]; // list of input tensor names outputs:[string]; // list of output tensor names } @@ -306,9 +301,14 @@ table TosaBasicBlock { outputs:[string]; // name of graph outputs } +table TosaRegion { + name:string; // name of region + blocks:[TosaBasicBlock]; // basic blocks array +} + table TosaGraph { - version: Version; - blocks:[TosaBasicBlock]; // basic blocks array + version:Version (required); + regions:[TosaRegion]; // regions array } root_type TosaGraph; diff --git a/setup.py b/setup.py new file mode 100644 index 0000000..3008a9e --- /dev/null +++ b/setup.py @@ -0,0 +1,18 @@ +# Copyright (c) 2023, ARM Limited. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Setup script for backward compatibility.""" +from setuptools import setup + +setup() diff --git a/src/numpy_utils.cpp b/src/numpy_utils.cpp index 80c680f..e4171d7 100644 --- a/src/numpy_utils.cpp +++ b/src/numpy_utils.cpp @@ -1,5 +1,5 @@ -// Copyright (c) 2020-2021, ARM Limited. +// Copyright (c) 2020-2024, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -14,6 +14,9 @@ // limitations under the License. #include "numpy_utils.h" +#include "half.hpp" +#include <algorithm> +#include <memory> // Magic NUMPY header static const char NUMPY_HEADER_STR[] = "\x93NUMPY\x1\x0\x76\x0{"; @@ -21,28 +24,81 @@ static const int NUMPY_HEADER_SZ = 128; // Maximum shape dimensions supported static const int NUMPY_MAX_DIMS_SUPPORTED = 10; -NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, bool* databuf) -{ - const char dtype_str[] = "'|b1'"; - return readFromNpyFileCommon(filename, dtype_str, 1, elems, databuf, true); -} - +// This is an entry function for reading 8-/16-/32-bit npy file. +template <> NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, int32_t* databuf) { - const char dtype_str[] = "'<i4'"; - return readFromNpyFileCommon(filename, dtype_str, sizeof(int32_t), elems, databuf, false); -} + FILE* infile = nullptr; + NPError rc = HEADER_PARSE_ERROR; + assert(filename); + assert(databuf); -NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, int64_t* databuf) -{ - const char dtype_str[] = "'<i8'"; - return readFromNpyFileCommon(filename, dtype_str, sizeof(int64_t), elems, databuf, false); -} + infile = fopen(filename, "rb"); + if (!infile) + { + return FILE_NOT_FOUND; + } -NumpyUtilities::NPError NumpyUtilities::readFromNpyFile(const char* filename, const uint32_t elems, float* databuf) -{ - const char dtype_str[] = "'<f4'"; - return readFromNpyFileCommon(filename, dtype_str, sizeof(float), elems, databuf, false); + bool is_signed = false; + int length_per_byte = 0; + char byte_order; + rc = getHeader(infile, is_signed, length_per_byte, byte_order); + if (rc != NO_ERROR) + return rc; + + switch (length_per_byte) + { + case 1: + if (is_signed) + { + int8_t* tmp_buf = new int8_t[elems]; + rc = readFromNpyFile<int8_t>(filename, elems, tmp_buf); + copyBufferByElement(databuf, tmp_buf, elems); + delete[] tmp_buf; + } + else + { + uint8_t* tmp_buf = new uint8_t[elems]; + rc = readFromNpyFile<uint8_t>(filename, elems, tmp_buf); + copyBufferByElement(databuf, tmp_buf, elems); + delete[] tmp_buf; + } + break; + case 2: + if (is_signed) + { + int16_t* tmp_buf = new int16_t[elems]; + rc = readFromNpyFile<int16_t>(filename, elems, tmp_buf); + copyBufferByElement(databuf, tmp_buf, elems); + delete[] tmp_buf; + } + else + { + uint16_t* tmp_buf = new uint16_t[elems]; + rc = readFromNpyFile<uint16_t>(filename, elems, tmp_buf); + copyBufferByElement(databuf, tmp_buf, elems); + delete[] tmp_buf; + } + break; + case 4: + if (is_signed) + { + bool is_bool; + const char* dtype_str = getDTypeString<int32_t>(is_bool); + rc = readFromNpyFileCommon(filename, dtype_str, sizeof(int32_t), elems, databuf, is_bool); + } + else + { + // uint32, not supported + rc = DATA_TYPE_NOT_SUPPORTED; + } + break; + default: + return DATA_TYPE_NOT_SUPPORTED; + break; + } + + return rc; } NumpyUtilities::NPError NumpyUtilities::readFromNpyFileCommon(const char* filename, @@ -101,6 +157,46 @@ NumpyUtilities::NPError NumpyUtilities::readFromNpyFileCommon(const char* filena return rc; } +NumpyUtilities::NPError NumpyUtilities::getHeader(FILE* infile, bool& is_signed, int& bit_length, char& byte_order) +{ + char buf[NUMPY_HEADER_SZ + 1]; + NPError rc = NO_ERROR; + assert(infile); + + if (fread(buf, NUMPY_HEADER_SZ, 1, infile) != 1) + { + return HEADER_PARSE_ERROR; + } + + // Validate the numpy magic number + if (memcmp(buf, NUMPY_HEADER_STR, sizeof(NUMPY_HEADER_STR) - 1)) + { + return HEADER_PARSE_ERROR; + } + + std::string dic_string(buf, NUMPY_HEADER_SZ); + + std::string desc_str("descr':"); + size_t offset = dic_string.find(desc_str); + if (offset == std::string::npos) + return HEADER_PARSE_ERROR; + + offset += desc_str.size() + 1; + // Skip whitespace and the opening ' + while (offset < dic_string.size() && (std::isspace(dic_string[offset]) || dic_string[offset] == '\'')) + offset++; + // Check for overflow + if (offset + 2 > dic_string.size()) + return HEADER_PARSE_ERROR; + + byte_order = dic_string[offset]; + is_signed = dic_string[offset + 1] == 'u' ? false : true; + bit_length = (int)dic_string[offset + 2] - '0'; + + rewind(infile); + return rc; +} + NumpyUtilities::NPError NumpyUtilities::checkNpyHeader(FILE* infile, const uint32_t elems, const char* dtype_str) { char buf[NUMPY_HEADER_SZ + 1]; @@ -253,60 +349,6 @@ NumpyUtilities::NPError NumpyUtilities::checkNpyHeader(FILE* infile, const uint3 return rc; } -NumpyUtilities::NPError NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const bool* databuf) -{ - std::vector<int32_t> shape = { (int32_t)elems }; - return writeToNpyFile(filename, shape, databuf); -} - -NumpyUtilities::NPError - NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const bool* databuf) -{ - const char dtype_str[] = "'|b1'"; - return writeToNpyFileCommon(filename, dtype_str, 1, shape, databuf, true); // bools written as size 1 -} - -NumpyUtilities::NPError - NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const int32_t* databuf) -{ - std::vector<int32_t> shape = { (int32_t)elems }; - return writeToNpyFile(filename, shape, databuf); -} - -NumpyUtilities::NPError - NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const int32_t* databuf) -{ - const char dtype_str[] = "'<i4'"; - return writeToNpyFileCommon(filename, dtype_str, sizeof(int32_t), shape, databuf, false); -} - -NumpyUtilities::NPError - NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const int64_t* databuf) -{ - std::vector<int32_t> shape = { (int32_t)elems }; - return writeToNpyFile(filename, shape, databuf); -} - -NumpyUtilities::NPError - NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const int64_t* databuf) -{ - const char dtype_str[] = "'<i8'"; - return writeToNpyFileCommon(filename, dtype_str, sizeof(int64_t), shape, databuf, false); -} - -NumpyUtilities::NPError NumpyUtilities::writeToNpyFile(const char* filename, const uint32_t elems, const float* databuf) -{ - std::vector<int32_t> shape = { (int32_t)elems }; - return writeToNpyFile(filename, shape, databuf); -} - -NumpyUtilities::NPError - NumpyUtilities::writeToNpyFile(const char* filename, const std::vector<int32_t>& shape, const float* databuf) -{ - const char dtype_str[] = "'<f4'"; - return writeToNpyFileCommon(filename, dtype_str, sizeof(float), shape, databuf, false); -} - NumpyUtilities::NPError NumpyUtilities::writeToNpyFileCommon(const char* filename, const char* dtype_str, const size_t elementsize, @@ -390,12 +432,11 @@ NumpyUtilities::NPError // Output the format dictionary // Hard-coded for I32 for now - headerPos += - snprintf(header + headerPos, NUMPY_HEADER_SZ - headerPos, "'descr': %s, 'fortran_order': False, 'shape': (%d,", - dtype_str, shape.empty() ? 1 : shape[0]); + headerPos += snprintf(header + headerPos, NUMPY_HEADER_SZ - headerPos, + "'descr': %s, 'fortran_order': False, 'shape': (", dtype_str); - // Remainder of shape array - for (i = 1; i < shape.size(); i++) + // Add shape contents (if any - as this will be empty for rank 0) + for (i = 0; i < shape.size(); i++) { headerPos += snprintf(header + headerPos, NUMPY_HEADER_SZ - headerPos, " %d,", shape[i]); } diff --git a/src/tosa_serialization_handler.cpp b/src/tosa_serialization_handler.cpp index 3a0ce43..85625cd 100644 --- a/src/tosa_serialization_handler.cpp +++ b/src/tosa_serialization_handler.cpp @@ -1,5 +1,5 @@ -// Copyright (c) 2020-2021, ARM Limited. +// Copyright (c) 2020-2024, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -14,18 +14,28 @@ // limitations under the License. #include "tosa_serialization_handler.h" +#include "half.hpp" #include <iostream> using namespace tosa; +using fp8e4m3 = tosa::float_t<int8_t, 4, true, true, false>; +using fp8e5m2 = tosa::float_t<int8_t, 5, true, true, true>; + TosaSerializationTensor::TosaSerializationTensor(const flatbuffers::String* name, const flatbuffers::Vector<int32_t>* shape, DType dtype, - const flatbuffers::Vector<uint8_t>* data) + const flatbuffers::Vector<uint8_t>* data, + const bool variable, + const bool is_unranked, + const flatbuffers::String* variable_name) { - _dtype = dtype; - - std::copy(shape->begin(), shape->end(), std::back_inserter(_shape)); + _dtype = dtype; + _variable = variable; + if (shape) + { + std::copy(shape->begin(), shape->end(), std::back_inserter(_shape)); + } assert(name); _name = name->str(); @@ -34,23 +44,37 @@ TosaSerializationTensor::TosaSerializationTensor(const flatbuffers::String* name { std::copy(data->begin(), data->end(), std::back_inserter(_data)); } + _is_unranked = is_unranked; + + if (variable_name) + { + _variable_name = variable_name->str(); + } } TosaSerializationTensor::TosaSerializationTensor(const std::string& name, const std::vector<int32_t>& shape, DType dtype, - const std::vector<uint8_t>& data) + const std::vector<uint8_t>& data, + const bool variable, + const bool is_unranked, + const std::string& variable_name) { - _dtype = dtype; - _shape = shape; - _name = name; - _data = data; + _dtype = dtype; + _variable = variable; + _shape = shape; + _name = name; + _data = data; + _is_unranked = is_unranked; + _variable_name = variable_name; } TosaSerializationTensor::TosaSerializationTensor() { - _dtype = DType_UNKNOWN; - _name = "UNKNOWN"; + _dtype = DType_UNKNOWN; + _variable = false; + _name = "UNKNOWN"; + _is_unranked = false; } TosaSerializationTensor::~TosaSerializationTensor() @@ -112,29 +136,33 @@ TosaSerializationOperator::~TosaSerializationOperator() } TosaSerializationBasicBlock::TosaSerializationBasicBlock(const std::string& name, + const std::string& region_name, const std::vector<TosaSerializationOperator*>& operators, const std::vector<TosaSerializationTensor*>& tensors, const std::vector<std::string>& inputs, const std::vector<std::string>& outputs) { - _name = name; - _operators = operators; - _tensors = tensors; - _inputs = inputs; - _outputs = outputs; + _name = name; + _region_name = region_name; + _operators = operators; + _tensors = tensors; + _inputs = inputs; + _outputs = outputs; } TosaSerializationBasicBlock::TosaSerializationBasicBlock(std::string&& name, + std::string&& region_name, std::vector<TosaSerializationOperator*>&& operators, std::vector<TosaSerializationTensor*>&& tensors, std::vector<std::string>&& inputs, std::vector<std::string>&& outputs) { - _name = std::move(name); - _operators = std::move(operators); - _tensors = std::move(tensors); - _inputs = std::move(inputs); - _outputs = std::move(outputs); + _name = std::move(name); + _region_name = std::move(region_name); + _operators = std::move(operators); + _tensors = std::move(tensors); + _inputs = std::move(inputs); + _outputs = std::move(outputs); } TosaSerializationBasicBlock::~TosaSerializationBasicBlock() @@ -152,65 +180,38 @@ TosaSerializationBasicBlock::~TosaSerializationBasicBlock() } } -TosaSerializationHandler::TosaSerializationHandler() +TosaSerializationRegion::TosaSerializationRegion(const std::string& name, + const std::vector<TosaSerializationBasicBlock*>& blocks) { - _schemaLoaded = false; - _version = TosaVersion(TOSA_VERSION_MAJOR, TOSA_VERSION_MINOR, TOSA_VERSION_PATCH, TOSA_VERSION_DRAFT); + _name = name; + _blocks = blocks; } -TosaSerializationHandler::~TosaSerializationHandler() +TosaSerializationRegion::TosaSerializationRegion(const std::string&& name, + const std::vector<TosaSerializationBasicBlock*>&& blocks) { - Clear(); // deallocate all basic blocks + _name = std::move(name); + _blocks = std::move(blocks); } -TosaVersion TosaSerializationHandler::ParseTosaSchemaVersion(std::string schema) +TosaSerializationRegion::~TosaSerializationRegion() { - // Parse all 4 version fields in schema file - static const char* keywords[4] = { "major: int32 = ", "minor: int32 = ", "patch: int32 = ", "draft: bool = " }; - string keyword_str[4]; - size_t search_pos = 0; - size_t keyword_pos; - size_t semicolon_pos; - // parse integer field first - for (int32_t i = 0; i < 4; i++) - { - keyword_pos = schema.find(keywords[i], search_pos); - if (keyword_pos == std::string::npos) - { - printf("ERROR: can't find keyword \"%s\" in schema\n", keywords[i]); - assert(0); - } - semicolon_pos = schema.find(';', keyword_pos); - if (keyword_pos == std::string::npos) - { - printf("ERROR: can't find ';' in schema\n"); - assert(0); - } - keyword_str[i] = - schema.substr(keyword_pos + strlen(keywords[i]), semicolon_pos - keyword_pos - strlen(keywords[i])); - search_pos = semicolon_pos; - } - - int32_t schema_major = 0; - int32_t schema_minor = 0; - int32_t schema_patch = 0; - bool schema_draft = false; - try - { - schema_major = stoi(keyword_str[0]); - schema_minor = stoi(keyword_str[1]); - schema_patch = stoi(keyword_str[2]); - schema_draft = (keyword_str[3] == "true") ? true : false; - } - catch (std::invalid_argument& e) + // deallocate all blocks + for (auto block : GetBlocks()) { - printf("ERROR: fail at stoi(): %s\n", e.what()); - assert(0); + delete block; // ~TosaSerializationBasicBlock() } +} - TosaVersion schema_version(schema_major, schema_minor, schema_patch, schema_draft); +TosaSerializationHandler::TosaSerializationHandler() +{ + _schemaLoaded = false; + _version = TosaVersion(TOSA_VERSION_MAJOR, TOSA_VERSION_MINOR, TOSA_VERSION_PATCH, TOSA_VERSION_DRAFT); +} - return schema_version; +TosaSerializationHandler::~TosaSerializationHandler() +{ + Clear(); // deallocate all basic blocks } tosa_err_t TosaSerializationHandler::LoadFileSchema(const char* schema_filename) @@ -227,23 +228,6 @@ tosa_err_t TosaSerializationHandler::LoadFileSchema(const char* schema_filename) ok = _parser.Parse(schema.c_str()); - TosaVersion schema_version = ParseTosaSchemaVersion(schema); - - TosaVersion::compat_t is_compat = schema_version.is_compatible(GetVersion()); - switch (is_compat) - { - case TosaVersion::compat_t::COMPLETELY_COMPATIBLE: - break; - case TosaVersion::compat_t::PARTIALLY_COMPATIBLE: - printf("WARNING: Schema flatbuffer version %s is partially compatible with serializer version %s\n", - schema_version.to_string().c_str(), GetVersion().to_string().c_str()); - break; - case TosaVersion::compat_t::NOT_COMPATIBLE: - printf("ERROR: Schema flatbuffer version %s is not compatible with serializer version %s\n", - schema_version.to_string().c_str(), GetVersion().to_string().c_str()); - return TOSA_VERSION_MISMATCH; - } - if (!ok) { printf("Error parsing ISA schema file: %s\n", schema_filename); @@ -308,7 +292,7 @@ tosa_err_t TosaSerializationHandler::SaveFileJson(const char* filename) uint8_t* buf = _builder.GetBufferPointer(); - if (!GenerateText(_parser, buf, &jsongen)) + if (GenText(_parser, buf, &jsongen)) { printf("Couldn't serialize parsed data to JSON!\n"); return TOSA_FILE_ERROR; @@ -399,11 +383,11 @@ tosa_err_t TosaSerializationHandler::SaveFileTosaFlatbuffer(const char* filename tosa_err_t TosaSerializationHandler::Clear() { // deallocate all basic blocks - for (auto bb : GetBlocks()) + for (auto region : GetRegions()) { - delete bb; + delete region; } - _blocks.clear(); + _regions.clear(); return TOSA_OK; } @@ -416,20 +400,13 @@ tosa_err_t TosaSerializationHandler::Deserialize(const uint8_t* buf) } auto fb_tosa_graph = GetTosaGraph(buf); auto fb_tosa_version = fb_tosa_graph->version(); - auto fb_tosa_blocks = fb_tosa_graph->blocks(); - - std::vector<std::string> operator_inputs_container; - std::vector<std::string> operator_outputs_container; - - std::vector<TosaSerializationOperator*> block_operators_container; - std::vector<TosaSerializationTensor*> block_tensors_container; - std::vector<std::string> block_inputs_container; - std::vector<std::string> block_outputs_container; + auto fb_tosa_regions = fb_tosa_graph->regions(); TosaAttributeBase* typed_attribute = NULL; TosaSerializationOperator* new_operator = NULL; TosaSerializationBasicBlock* new_block = NULL; TosaSerializationTensor* new_tensor = NULL; + TosaSerializationRegion* new_region = NULL; // erase container Clear(); @@ -437,226 +414,241 @@ tosa_err_t TosaSerializationHandler::Deserialize(const uint8_t* buf) TosaVersion read_version(fb_tosa_version->_major(), fb_tosa_version->_minor(), fb_tosa_version->_patch(), fb_tosa_version->_draft()); - TosaVersion::compat_t is_compat = read_version.is_compatible(GetVersion()); + TosaVersion::compat_t is_compat = TosaVersion::is_compatible(read_version, GetVersion()); switch (is_compat) { case TosaVersion::compat_t::COMPLETELY_COMPATIBLE: break; - case TosaVersion::compat_t::PARTIALLY_COMPATIBLE: - printf("WARNING: Read flatbuffer version %s is partially compatible with serializer version %s\n", + case TosaVersion::compat_t::BACKWARD_COMPATIBLE: + printf("WARNING: Different Tosa flatbuffer and serializer versions detected. Read Tosa flatbuffer version " + "%s is backward " + "compatible with serializer version %s\n", read_version.to_string().c_str(), GetVersion().to_string().c_str()); break; case TosaVersion::compat_t::NOT_COMPATIBLE: - printf("ERROR: Read flatbuffer version %s is not compatible with serializer version %s\n", + printf("ERROR: Read Tosa flatbuffer version %s is not compatible with serializer version %s\n", read_version.to_string().c_str(), GetVersion().to_string().c_str()); return TOSA_VERSION_MISMATCH; } - for (size_t i = 0; i < fb_tosa_blocks->size(); i++) + for (size_t i = 0; i < fb_tosa_regions->size(); i++) { - auto curr_block = fb_tosa_blocks->Get(i); + auto curr_region = fb_tosa_regions->Get(i); + auto region_name = curr_region->name()->str(); + auto fb_tosa_blocks = curr_region->blocks(); - auto block_name = curr_block->name()->str(); + new_region = new TosaSerializationRegion(curr_region->name()->str(), {}); + this->GetRegions().push_back(new_region); - auto fb_tosa_operators = curr_block->operators(); - block_operators_container.clear(); - for (size_t j = 0; j < fb_tosa_operators->size(); j++) + for (size_t i = 0; i < fb_tosa_blocks->size(); i++) { - auto curr_operator = fb_tosa_operators->Get(j); + std::vector<TosaSerializationOperator*> block_operators_container; + std::vector<TosaSerializationTensor*> block_tensors_container; + std::vector<std::string> block_inputs_container; + std::vector<std::string> block_outputs_container; - auto operator_op = curr_operator->op(); - auto attribute_type = curr_operator->attribute_type(); - auto attribute = curr_operator->attribute(); + auto curr_block = fb_tosa_blocks->Get(i); - // input tensors - auto operator_inputs = curr_operator->inputs(); - operator_inputs_container.clear(); - if (operator_inputs) + auto block_name = curr_block->name()->str(); + + auto fb_tosa_operators = curr_block->operators(); + for (size_t j = 0; j < fb_tosa_operators->size(); j++) { - for (size_t k = 0; k < operator_inputs->size(); k++) + auto curr_operator = fb_tosa_operators->Get(j); + + auto operator_op = curr_operator->op(); + auto attribute_type = curr_operator->attribute_type(); + auto attribute = curr_operator->attribute(); + + std::vector<std::string> operator_inputs_container; + std::vector<std::string> operator_outputs_container; + + // input tensors + auto operator_inputs = curr_operator->inputs(); + if (operator_inputs) { - auto curr_input = operator_inputs->Get(k); - operator_inputs_container.push_back(curr_input->str()); + for (size_t k = 0; k < operator_inputs->size(); k++) + { + auto curr_input = operator_inputs->Get(k); + operator_inputs_container.push_back(curr_input->str()); + } } - } - // output tensors - auto operator_outputs = curr_operator->outputs(); - operator_outputs_container.clear(); - if (operator_outputs) - { - for (size_t k = 0; k < operator_outputs->size(); k++) + // output tensors + auto operator_outputs = curr_operator->outputs(); + if (operator_outputs) { - auto curr_output = operator_outputs->Get(k); - operator_outputs_container.push_back(curr_output->str()); + for (size_t k = 0; k < operator_outputs->size(); k++) + { + auto curr_output = operator_outputs->Get(k); + operator_outputs_container.push_back(curr_output->str()); + } } - } - switch (attribute_type) - { - case Attribute_NONE: - typed_attribute = new TosaNoneAttribute(); - break; + switch (attribute_type) + { + case Attribute_NONE: + typed_attribute = new TosaNoneAttribute(); + break; #define DEF_ATTRIBUTE(NAME, ...) \ case Attribute_##NAME##Attribute: \ typed_attribute = new Tosa##NAME##Attribute(attribute); \ break; #include "attribute.def" #undef DEF_ATTRIBUTE - default: - printf("TosaSerializationHandler::Deserialize(): Attribute %s not implemented yet\n", - EnumNamesAttribute()[attribute_type]); - return TOSA_INTERNAL_ERROR; + default: + printf("TosaSerializationHandler::Deserialize(): Attribute %s not implemented yet\n", + EnumNamesAttribute()[attribute_type]); + return TOSA_INTERNAL_ERROR; + } + + new_operator = new TosaSerializationOperator(operator_op, attribute_type, typed_attribute, + operator_inputs_container, operator_outputs_container); + if (new_operator) + { + block_operators_container.push_back(new_operator); + } + else + { + return TOSA_MEMORY_ERROR; + } + + if (typed_attribute) + delete typed_attribute; } - new_operator = new TosaSerializationOperator(operator_op, attribute_type, typed_attribute, - operator_inputs_container, operator_outputs_container); - if (new_operator) + auto block_inputs = curr_block->inputs(); + auto block_outputs = curr_block->outputs(); + + for (size_t j = 0; j < block_inputs->size(); j++) { - block_operators_container.push_back(new_operator); + auto curr_block_input = block_inputs->Get(j); + block_inputs_container.push_back(curr_block_input->str()); } - else + for (size_t j = 0; j < block_outputs->size(); j++) { - return TOSA_MEMORY_ERROR; + auto curr_block_output = block_outputs->Get(j); + block_outputs_container.push_back(curr_block_output->str()); } - if (typed_attribute) - delete typed_attribute; - } - - auto fb_tosa_tensors = curr_block->tensors(); - block_tensors_container.clear(); - for (size_t j = 0; j < fb_tosa_tensors->size(); j++) - { - auto curr_tensor = fb_tosa_tensors->Get(j); - - auto tensor_name = curr_tensor->name(); - auto tensor_shape = curr_tensor->shape(); - auto tensor_type = curr_tensor->type(); - auto tensor_data = curr_tensor->data(); - - new_tensor = new TosaSerializationTensor(tensor_name, tensor_shape, tensor_type, tensor_data); - if (new_tensor) + auto fb_tosa_tensors = curr_block->tensors(); + for (size_t j = 0; j < fb_tosa_tensors->size(); j++) + { + auto curr_tensor = fb_tosa_tensors->Get(j); + + auto tensor_name = curr_tensor->name(); + auto tensor_shape = curr_tensor->shape(); + auto tensor_type = curr_tensor->type(); + auto tensor_variable = curr_tensor->variable(); + auto tensor_data = curr_tensor->data(); + auto tensor_is_unranked = curr_tensor->is_unranked(); + auto tensor_variable_name = curr_tensor->variable_name(); + + new_tensor = new TosaSerializationTensor(tensor_name, tensor_shape, tensor_type, tensor_data, + tensor_variable, tensor_is_unranked, tensor_variable_name); + if (new_tensor) + { + block_tensors_container.push_back(new_tensor); + } + else + { + return TOSA_MEMORY_ERROR; + } + } + new_block = new TosaSerializationBasicBlock(block_name, region_name, block_operators_container, + block_tensors_container, block_inputs_container, + block_outputs_container); + if (new_block) { - block_tensors_container.push_back(new_tensor); + new_region->GetBlocks().push_back(new_block); } else { return TOSA_MEMORY_ERROR; } - } - - auto block_inputs = curr_block->inputs(); - auto block_outputs = curr_block->outputs(); - - block_inputs_container.clear(); - block_outputs_container.clear(); - - for (size_t j = 0; j < block_inputs->size(); j++) - { - auto curr_block_input = block_inputs->Get(j); - block_inputs_container.push_back(curr_block_input->str()); - } - for (size_t j = 0; j < block_outputs->size(); j++) - { - auto curr_block_output = block_outputs->Get(j); - block_outputs_container.push_back(curr_block_output->str()); - } - - new_block = new TosaSerializationBasicBlock(block_name, block_operators_container, block_tensors_container, - block_inputs_container, block_outputs_container); - if (new_block) - { - this->GetBlocks().push_back(new_block); - } - else - { - return TOSA_MEMORY_ERROR; - } + } // end block for_loop } return TOSA_OK; } -tosa_err_t TosaSerializationHandler::Serialize() +std::vector<uint8_t> float_to_u8_helper(float f_in) { - std::vector<flatbuffers::Offset<TosaBasicBlock>> fboffset_blocks; - - std::vector<flatbuffers::Offset<TosaOperator>> fboffset_block_operators; - std::vector<flatbuffers::Offset<TosaTensor>> fboffset_block_tensors; - std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_block_inputs; - std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_block_outputs; + // Push back a single float value to the buffer with *NO PADDING* + // Therefore ConvertF32toU8 function not used + std::vector<uint8_t> u8_out; + uint32_t* val_u32 = reinterpret_cast<uint32_t*>(&f_in); + u8_out.push_back(*val_u32 & 0xFF); + u8_out.push_back((*val_u32 >> 8) & 0xFF); + u8_out.push_back((*val_u32 >> 16) & 0xFF); + u8_out.push_back((*val_u32 >> 24) & 0xFF); + return u8_out; +} - std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_operator_inputs; - std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_operator_outputs; +tosa_err_t TosaSerializationHandler::Serialize() +{ + // regions + std::vector<flatbuffers::Offset<TosaRegion>> fboffset_regions; // translate TosaFlatbufferOperator to flatbuffers::Offset<TosaOperator> - for (auto block : GetBlocks()) + for (auto region : GetRegions()) { - fboffset_block_operators.clear(); - fboffset_block_tensors.clear(); - fboffset_block_inputs.clear(); - fboffset_block_outputs.clear(); - - auto block_name = _builder.CreateString(block->GetName().c_str()); - - for (auto tensor_str : block->GetInputs()) - { - auto tensor_name = _builder.CreateString(tensor_str.c_str()); - fboffset_block_inputs.push_back(tensor_name); - } - - for (auto tensor_str : block->GetOutputs()) - { - auto tensor_name = _builder.CreateString(tensor_str.c_str()); - fboffset_block_outputs.push_back(tensor_name); - } - - auto fb_block_inputs = _builder.CreateVector(fboffset_block_inputs); - auto fb_block_outputs = _builder.CreateVector(fboffset_block_outputs); - - for (auto op : block->GetOperators()) + std::vector<flatbuffers::Offset<TosaBasicBlock>> fboffset_blocks; + for (auto block : region->GetBlocks()) { - fboffset_operator_inputs.clear(); - fboffset_operator_outputs.clear(); - - auto operator_op = op->GetOp(); - auto attribute_type = op->GetAttributeType(); - - for (auto tensor_str : op->GetInputTensorNames()) + std::vector<flatbuffers::Offset<TosaOperator>> fboffset_block_operators; + std::vector<flatbuffers::Offset<TosaTensor>> fboffset_block_tensors; + std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_block_inputs; + std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_block_outputs; + auto block_name = _builder.CreateString(block->GetName().c_str()); + for (auto tensor_str : block->GetInputs()) { auto tensor_name = _builder.CreateString(tensor_str.c_str()); - fboffset_operator_inputs.push_back(tensor_name); + fboffset_block_inputs.push_back(tensor_name); } - - for (auto tensor_str : op->GetOutputTensorNames()) + for (auto tensor_str : block->GetOutputs()) { auto tensor_name = _builder.CreateString(tensor_str.c_str()); - fboffset_operator_outputs.push_back(tensor_name); + fboffset_block_outputs.push_back(tensor_name); } - - auto fb_operator_inputs = _builder.CreateVector(fboffset_operator_inputs); - auto fb_operator_outputs = _builder.CreateVector(fboffset_operator_outputs); - - flatbuffers::Offset<void> fb_attribute; - switch (attribute_type) + auto fb_block_inputs = _builder.CreateVector(fboffset_block_inputs); + auto fb_block_outputs = _builder.CreateVector(fboffset_block_outputs); + for (auto op : block->GetOperators()) { - case Attribute_NONE: - fb_attribute = 0; - break; - + std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_operator_inputs; + std::vector<flatbuffers::Offset<flatbuffers::String>> fboffset_operator_outputs; + auto operator_op = op->GetOp(); + auto attribute_type = op->GetAttributeType(); + for (auto tensor_str : op->GetInputTensorNames()) + { + auto tensor_name = _builder.CreateString(tensor_str.c_str()); + fboffset_operator_inputs.push_back(tensor_name); + } + for (auto tensor_str : op->GetOutputTensorNames()) + { + auto tensor_name = _builder.CreateString(tensor_str.c_str()); + fboffset_operator_outputs.push_back(tensor_name); + } + auto fb_operator_inputs = _builder.CreateVector(fboffset_operator_inputs); + auto fb_operator_outputs = _builder.CreateVector(fboffset_operator_outputs); + flatbuffers::Offset<void> fb_attribute; + switch (attribute_type) + { + case Attribute_NONE: + fb_attribute = 0; + break; #define DEF_ARGS_S_STR(NAME, V) , _builder.CreateString(reinterpret_cast<Tosa##NAME*>(op->GetAttribute())->V().c_str()) +#define DEF_ARGS_S_FP_as_U8(NAME, V) \ + , _builder.CreateVector<uint8_t>(float_to_u8_helper(reinterpret_cast<Tosa##NAME*>(op->GetAttribute())->V())) #define DEF_ARGS_S_DEFAULT(NAME, V) , reinterpret_cast<Tosa##NAME*>(op->GetAttribute())->V() - #define DEF_ARGS_S_int32_t(NAME, V) DEF_ARGS_S_DEFAULT(NAME, V) -#define DEF_ARGS_S_float(NAME, V) DEF_ARGS_S_DEFAULT(NAME, V) +#define DEF_ARGS_S_float(NAME, V) DEF_ARGS_S_FP_as_U8(NAME, V) #define DEF_ARGS_S_bool(NAME, V) DEF_ARGS_S_DEFAULT(NAME, V) #define DEF_ARGS_S_ResizeMode(NAME, V) DEF_ARGS_S_DEFAULT(NAME, V) +#define DEF_ARGS_S_DType(NAME, V) DEF_ARGS_S_DEFAULT(NAME, V) #define DEF_ARGS_S_string(NAME, V) DEF_ARGS_S_STR(NAME, V) - #define DEF_ARGS_S(NAME, T, V) DEF_ARGS_S_##T(NAME, V) #define DEF_ARGS_V(NAME, T, V) , _builder.CreateVector<T>(reinterpret_cast<Tosa##NAME*>(op->GetAttribute())->V()) - #define DEF_ARGS_1(NAME, T0, F0, V0) DEF_ARGS_##F0(NAME, T0, V0) #define DEF_ARGS_2(NAME, T0, F0, V0, T1, F1, V1) DEF_ARGS_##F0(NAME, T0, V0) DEF_ARGS_##F1(NAME, T1, V1) #define DEF_ARGS_3(NAME, T0, F0, V0, T1, F1, V1, T2, F2, V2) \ @@ -672,11 +664,20 @@ tosa_err_t TosaSerializationHandler::Serialize() #define DEF_ARGS_7(NAME, T0, F0, V0, T1, F1, V1, T2, F2, V2, T3, F3, V3, T4, F4, V4, T5, F5, V5, T6, F6, V6) \ DEF_ARGS_##F0(NAME, T0, V0) DEF_ARGS_##F1(NAME, T1, V1) DEF_ARGS_##F2(NAME, T2, V2) DEF_ARGS_##F3(NAME, T3, V3) \ DEF_ARGS_##F4(NAME, T4, V4) DEF_ARGS_##F5(NAME, T5, V5) DEF_ARGS_##F6(NAME, T6, V6) +#define DEF_ARGS_8(NAME, T0, F0, V0, T1, F1, V1, T2, F2, V2, T3, F3, V3, T4, F4, V4, T5, F5, V5, T6, F6, V6, T7, F7, \ + V7) \ + DEF_ARGS_##F0(NAME, T0, V0) DEF_ARGS_##F1(NAME, T1, V1) DEF_ARGS_##F2(NAME, T2, V2) DEF_ARGS_##F3(NAME, T3, V3) \ + DEF_ARGS_##F4(NAME, T4, V4) DEF_ARGS_##F5(NAME, T5, V5) DEF_ARGS_##F6(NAME, T6, V6) \ + DEF_ARGS_##F7(NAME, T7, V7) +#define DEF_ARGS_9(NAME, T0, F0, V0, T1, F1, V1, T2, F2, V2, T3, F3, V3, T4, F4, V4, T5, F5, V5, T6, F6, V6, T7, F7, \ + V7, T8, F8, V8) \ + DEF_ARGS_##F0(NAME, T0, V0) DEF_ARGS_##F1(NAME, T1, V1) DEF_ARGS_##F2(NAME, T2, V2) DEF_ARGS_##F3(NAME, T3, V3) \ + DEF_ARGS_##F4(NAME, T4, V4) DEF_ARGS_##F5(NAME, T5, V5) DEF_ARGS_##F6(NAME, T6, V6) \ + DEF_ARGS_##F7(NAME, T7, V7) DEF_ARGS_##F8(NAME, T8, V8) #define DEF_ATTRIBUTE(NAME, NUM_ARGS, ...) \ case Attribute_##NAME##Attribute: \ fb_attribute = Create##NAME##Attribute(_builder DEF_ARGS_##NUM_ARGS(NAME##Attribute, __VA_ARGS__)).Union(); \ break; - #include "attribute.def" #undef DEF_ATTRIBUTE #undef DEF_ARGS_1 @@ -692,53 +693,56 @@ tosa_err_t TosaSerializationHandler::Serialize() #undef DEF_ARGS_S_float #undef DEF_ARGS_S_bool #undef DEF_ARGS_S_ResizeMode +#undef DEF_ARGS_S_DType #undef DEF_ARGS_S_string #undef DEF_ARGS_S_STR #undef DEF_ARGS_S_DEFAULT - default: - printf("TosaSerializationHandler::Serialize(): Attribute %s not implemented yet\n", - EnumNamesAttribute()[attribute_type]); - return TOSA_INTERNAL_ERROR; + default: + printf("TosaSerializationHandler::Serialize(): Attribute %s not implemented yet\n", + EnumNamesAttribute()[attribute_type]); + return TOSA_INTERNAL_ERROR; + } + auto fboffset_operator = CreateTosaOperator(_builder, operator_op, attribute_type, fb_attribute, + fb_operator_inputs, fb_operator_outputs); + fboffset_block_operators.push_back(fboffset_operator); } + auto fb_block_operators = _builder.CreateVector(fboffset_block_operators); + for (auto tensor : block->GetTensors()) + { + auto tensor_name = _builder.CreateString(tensor->GetName().c_str()); + auto tensor_shape = _builder.CreateVector(tensor->GetShape()); + auto tensor_dtype = tensor->GetDtype(); + bool tensor_variable = tensor->GetVariable(); + auto tensor_data = _builder.CreateVector(tensor->GetData()); + auto tensor_is_unranked = tensor->GetIsUnranked(); + auto tensor_variable_name = _builder.CreateString(tensor->GetVariableName().c_str()); + auto fboffset_tensor = CreateTosaTensor(_builder, tensor_name, tensor_shape, tensor_dtype, tensor_data, + tensor_variable, tensor_is_unranked, tensor_variable_name); + fboffset_block_tensors.push_back(fboffset_tensor); + } + auto fb_block_tensors = _builder.CreateVector(fboffset_block_tensors); + auto fboffset_block = CreateTosaBasicBlock(_builder, block_name, fb_block_operators, fb_block_tensors, + fb_block_inputs, fb_block_outputs); + fboffset_blocks.push_back(fboffset_block); + } // end block for_loop + auto fb_blocks = _builder.CreateVector(fboffset_blocks); - auto fboffset_operator = CreateTosaOperator(_builder, operator_op, attribute_type, fb_attribute, - fb_operator_inputs, fb_operator_outputs); - fboffset_block_operators.push_back(fboffset_operator); - } - - auto fb_block_operators = _builder.CreateVector(fboffset_block_operators); - - for (auto tensor : block->GetTensors()) - { - - auto tensor_name = _builder.CreateString(tensor->GetName().c_str()); - auto tensor_shape = _builder.CreateVector(tensor->GetShape()); - auto tensor_dtype = tensor->GetDtype(); - auto tensor_data = _builder.CreateVector(tensor->GetData()); - - auto fboffset_tensor = CreateTosaTensor(_builder, tensor_name, tensor_shape, tensor_dtype, tensor_data); - fboffset_block_tensors.push_back(fboffset_tensor); - } - - auto fb_block_tensors = _builder.CreateVector(fboffset_block_tensors); - - auto fboffset_block = CreateTosaBasicBlock(_builder, block_name, fb_block_operators, fb_block_tensors, - fb_block_inputs, fb_block_outputs); - fboffset_blocks.push_back(fboffset_block); - } + auto region_name = _builder.CreateString(region->GetName().c_str()); + auto fboffset_region = CreateTosaRegion(_builder, region_name, fb_blocks); + fboffset_regions.push_back(fboffset_region); + } // end region for_loop - auto fb_blocks = _builder.CreateVector(fboffset_blocks); + auto fb_regions = _builder.CreateVector(fboffset_regions); auto fb_version = CreateVersion(_builder, TOSA_VERSION_MAJOR, TOSA_VERSION_MINOR, TOSA_VERSION_PATCH, TOSA_VERSION_DRAFT); - - auto fb_graph = CreateTosaGraph(_builder, fb_version, fb_blocks); + auto fb_graph = CreateTosaGraph(_builder, fb_version, fb_regions); _builder.Finish(fb_graph, TosaGraphIdentifier()); return TOSA_OK; } -void zero_pad(std::vector<uint8_t>& buf) +void TosaSerializationHandler::ForceAlignTensorData(std::vector<uint8_t>& buf) { while ((buf.size() % TENSOR_BUFFER_FORCE_ALIGNMENT) != 0) { @@ -746,6 +750,66 @@ void zero_pad(std::vector<uint8_t>& buf) } } +tosa_err_t TosaSerializationHandler::ConvertBF16toU8(const std::vector<float>& in, std::vector<uint8_t>& out) +{ + // Note: Converts fp32->bf16 by ignoring the least significant 16 bits + out.clear(); + for (auto val : in) + { + uint32_t* val_u32 = reinterpret_cast<uint32_t*>(&val); + uint8_t f32_byte2 = (*val_u32 >> 16) & 0xFF; + uint8_t f32_byte3 = (*val_u32 >> 24) & 0xFF; + // little endian: byte2 followed by byte3 + out.push_back(f32_byte2); + out.push_back(f32_byte3); + } + ForceAlignTensorData(out); + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertFP8E4M3toU8(const std::vector<float>& in, std::vector<uint8_t>& out) +{ + // Note: Converts fp32->FP8E4M3 before converting to unint8_t + out.clear(); + for (auto val : in) + { + auto f8 = static_cast<fp8e4m3>(val); + uint8_t b8 = f8.bits(); + out.push_back(b8); + } + ForceAlignTensorData(out); + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertFP8E5M2toU8(const std::vector<float>& in, std::vector<uint8_t>& out) +{ + // Note: Converts fp32->FP8E5M2 before converting to uint8_t + out.clear(); + for (auto val : in) + { + auto f8 = static_cast<fp8e5m2>(val); + uint8_t b8 = f8.bits(); + out.push_back(b8); + } + ForceAlignTensorData(out); + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertF16toU8(const std::vector<float>& in, std::vector<uint8_t>& out) +{ + // Note: Converts fp32->fp16 before converting to uint8_t + out.clear(); + for (auto val : in) + { + half_float::half val_f16 = half_float::half_cast<half_float::half, float>(val); + uint16_t* val_u16 = reinterpret_cast<uint16_t*>(&val_f16); + out.push_back(*val_u16 & 0xFF); + out.push_back((*val_u16 >> 8) & 0xFF); + } + ForceAlignTensorData(out); + return TOSA_OK; +} + tosa_err_t TosaSerializationHandler::ConvertF32toU8(const std::vector<float>& in, std::vector<uint8_t>& out) { out.clear(); @@ -757,7 +821,26 @@ tosa_err_t TosaSerializationHandler::ConvertF32toU8(const std::vector<float>& in out.push_back((*val_u32 >> 16) & 0xFF); out.push_back((*val_u32 >> 24) & 0xFF); } - zero_pad(out); + ForceAlignTensorData(out); + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertI64toU8(const std::vector<int64_t>& in, std::vector<uint8_t>& out) +{ + out.clear(); + for (auto val : in) + { + uint64_t* val_u64 = reinterpret_cast<uint64_t*>(&val); + out.push_back(*val_u64 & 0xFF); + out.push_back((*val_u64 >> 8) & 0xFF); + out.push_back((*val_u64 >> 16) & 0xFF); + out.push_back((*val_u64 >> 24) & 0xFF); + out.push_back((*val_u64 >> 32) & 0xFF); + out.push_back((*val_u64 >> 40) & 0xFF); + out.push_back((*val_u64 >> 48) & 0xFF); + out.push_back((*val_u64 >> 56) & 0xFF); + } + ForceAlignTensorData(out); return TOSA_OK; } @@ -774,7 +857,7 @@ tosa_err_t TosaSerializationHandler::ConvertI48toU8(const std::vector<int64_t>& out.push_back((*val_u64 >> 32) & 0xFF); out.push_back((*val_u64 >> 40) & 0xFF); } - zero_pad(out); + ForceAlignTensorData(out); return TOSA_OK; } @@ -789,7 +872,7 @@ tosa_err_t TosaSerializationHandler::ConvertI32toU8(const std::vector<int32_t>& out.push_back((*val_u32 >> 16) & 0xFF); out.push_back((*val_u32 >> 24) & 0xFF); } - zero_pad(out); + ForceAlignTensorData(out); return TOSA_OK; } @@ -802,7 +885,7 @@ tosa_err_t TosaSerializationHandler::ConvertI16toU8(const std::vector<int16_t>& out.push_back(*val_u16 & 0xFF); out.push_back((*val_u16 >> 8) & 0xFF); } - zero_pad(out); + ForceAlignTensorData(out); return TOSA_OK; } @@ -814,7 +897,7 @@ tosa_err_t TosaSerializationHandler::ConvertI8toU8(const std::vector<int8_t>& in uint8_t* val_u8 = reinterpret_cast<uint8_t*>(&val); out.push_back(*val_u8); } - zero_pad(out); + ForceAlignTensorData(out); return TOSA_OK; } @@ -845,7 +928,7 @@ tosa_err_t TosaSerializationHandler::ConvertI4toU8(const std::vector<int8_t>& in uint8_t val_u8 = static_cast<uint8_t>(val_packed); out.push_back(val_u8); } - zero_pad(out); + ForceAlignTensorData(out); return TOSA_OK; } @@ -857,7 +940,105 @@ tosa_err_t TosaSerializationHandler::ConvertBooltoU8(const std::vector<bool>& in uint8_t val_u8 = val; out.push_back(val_u8); } - zero_pad(out); + ForceAlignTensorData(out); + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertU8toBF16(const std::vector<uint8_t>& in, + uint32_t out_size, + std::vector<float>& out) +{ + // Note: bf16 values returned in fp32 type + out.clear(); + if (in.size() < out_size * sizeof(int16_t)) + { + printf("TosaSerializationHandler::ConvertU8toBF16(): uint8 buffer size %ld must >= target size %ld\n", + in.size(), out_size * sizeof(int16_t)); + return TOSA_USER_ERROR; + } + + for (uint32_t i = 0; i < out_size; i++) + { + uint32_t f32_byte2 = in[i * sizeof(int16_t)]; + uint32_t f32_byte3 = in[i * sizeof(int16_t) + 1]; + uint32_t val_u32 = (f32_byte2 << 16) + (f32_byte3 << 24); + + // Reinterpret u32 bytes as fp32 + float val_f32 = *(float*)&val_u32; + out.push_back(val_f32); + } + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertU8toFP8E4M3(const std::vector<uint8_t>& in, + uint32_t out_size, + std::vector<float>& out) +{ + // Note: FP8E4M3 values returned in fp32 type + out.clear(); + if (in.size() < out_size * sizeof(int8_t)) + { + printf("TosaSerializationHandler::ConvertU8toF16(): uint8 buffer size %ld must >= target size %ld\n", in.size(), + out_size * sizeof(int8_t)); + return TOSA_USER_ERROR; + } + + for (uint32_t i = 0; i < out_size; i++) + { + int8_t bits = static_cast<int8_t>(in[i * sizeof(int8_t)]); + auto f8 = fp8e4m3::from_bits(bits); + float val_f32 = static_cast<float>(f8); + out.push_back(val_f32); + } + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertU8toFP8E5M2(const std::vector<uint8_t>& in, + uint32_t out_size, + std::vector<float>& out) +{ + // Note: FP8E5M2 values returned in fp32 type + out.clear(); + if (in.size() < out_size * sizeof(int8_t)) + { + printf("TosaSerializationHandler::ConvertU8toF16(): uint8 buffer size %ld must >= target size %ld\n", in.size(), + out_size * sizeof(int8_t)); + return TOSA_USER_ERROR; + } + + for (uint32_t i = 0; i < out_size; i++) + { + int8_t bits = static_cast<int8_t>(in[i * sizeof(int8_t)]); + auto f8 = fp8e5m2::from_bits(bits); + float val_f32 = static_cast<float>(f8); + out.push_back(val_f32); + } + return TOSA_OK; +} + +tosa_err_t TosaSerializationHandler::ConvertU8toF16(const std::vector<uint8_t>& in, + uint32_t out_size, + std::vector<half_float::half>& out) +{ + // Note: fp16 values returned in fp32 type + out.clear(); + if (in.size() < out_size * sizeof(int16_t)) + { + printf("TosaSerializationHandler::ConvertU8toF16(): uint8 buffer size %ld must >= target size %ld\n", in.size(), + out_size * sizeof(int16_t)); + return TOSA_USER_ERROR; + } + + for (uint32_t i = 0; i < out_size; i++) + { + uint16_t f16_byte0 = in[i * sizeof(int16_t)]; + uint16_t f16_byte1 = in[i * sizeof(int16_t) + 1]; + uint16_t val_u16 = f16_byte0 + (f16_byte1 << 8); + + // Reinterpret u16 byte as fp16 then convert to fp32 + half_float::half val_f16 = *(half_float::half*)&val_u16; + out.push_back(val_f16); + } return TOSA_OK; } @@ -884,6 +1065,35 @@ tosa_err_t return TOSA_OK; } +tosa_err_t TosaSerializationHandler::ConvertU8toI64(const std::vector<uint8_t>& in, + uint32_t out_size, + std::vector<int64_t>& out) +{ + out.clear(); + if (in.size() < out_size * sizeof(int64_t)) + { + printf("TosaSerializationHandler::ConvertU8toI64(): uint8 buffer size %ld must >= target size %ld\n", in.size(), + out_size * sizeof(int64_t)); + return TOSA_USER_ERROR; + } + for (uint32_t i = 0; i < out_size; i++) + { + uint64_t byte0 = in[i * sizeof(int64_t)]; + uint64_t byte1 = in[i * sizeof(int64_t) + 1]; + uint64_t byte2 = in[i * sizeof(int64_t) + 2]; + uint64_t byte3 = in[i * sizeof(int64_t) + 3]; + uint64_t byte4 = in[i * sizeof(int64_t) + 4]; + uint64_t byte5 = in[i * sizeof(int64_t) + 5]; + uint64_t byte6 = in[i * sizeof(int64_t) + 6]; + uint64_t byte7 = in[i * sizeof(int64_t) + 7]; + uint64_t val_u64 = byte0 + (byte1 << 8) + (byte2 << 16) + (byte3 << 24) + (byte4 << 32) + (byte5 << 40) + + (byte6 << 48) + (byte7 << 56); + int64_t* val_i64 = reinterpret_cast<int64_t*>(&val_u64); + out.push_back(*val_i64); + } + return TOSA_OK; +} + tosa_err_t TosaSerializationHandler::ConvertU8toI48(const std::vector<uint8_t>& in, uint32_t out_size, std::vector<int64_t>& out) diff --git a/test/scripts/test_npy_fileio.py b/test/scripts/test_npy_fileio.py index e0a6f5d..272c124 100755 --- a/test/scripts/test_npy_fileio.py +++ b/test/scripts/test_npy_fileio.py @@ -122,7 +122,7 @@ def main(): xunit_suite = xunit_result.create_suite("basic_serialization") max_size = 128 - datatypes = ["int32", "int64", "float", "bool"] + datatypes = ["int32", "int64", "float", "bool", "double"] random.seed(args.seed) failed = 0 diff --git a/test/scripts/test_serialization.py b/test/scripts/test_serialization.py deleted file mode 100755 index 834bc1d..0000000 --- a/test/scripts/test_serialization.py +++ /dev/null @@ -1,197 +0,0 @@ -#!/usr/bin/env python3 - -# Copyright (c) 2021, ARM Limited. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -""" Simple test script which uses serialization_read_write to copy tosa files. It -uses flatc to convert to json for comparison since the binary files may -differ. """ - -import argparse -import filecmp -import random -import shlex -import subprocess -from datetime import datetime -from enum import IntEnum, unique -from pathlib import Path -from xunit.xunit import xunit_results, xunit_test - - -@unique -class TestResult(IntEnum): - PASS = 0 - COMMAND_ERROR = 1 - MISMATCH = 2 - SKIPPED = 3 - - -def parseArgs(): - baseDir = (Path(__file__).parent / "../..").resolve() - buildDir = (baseDir / "build").resolve() - parser = argparse.ArgumentParser() - parser.add_argument( - "-t", - "--testdir", - dest="test", - type=str, - required=True, - help="Directory of tosa files to verify", - ) - parser.add_argument( - "--flatc", - default=str(buildDir / "third_party/flatbuffers/flatc"), - help="location of flatc compiler", - ) - parser.add_argument( - "-s", - "--schema", - default=str(baseDir / "schema/tosa.fbs"), - help="location of schema file", - ) - parser.add_argument( - "-c", - "--cmd", - default=str(buildDir / "serialization_read_write"), - help="Command to read/write test file", - ) - parser.add_argument( - "-v", "--verbose", action="store_true", help="verbose", default=False - ) - parser.add_argument( - "--xunit-file", default="result.xml", help="xunit result output file" - ) - args = parser.parse_args() - - # check that required files exist - if not Path(args.flatc).exists(): - print("flatc not found at location " + args.flatc) - parser.print_help() - exit(1) - if not Path(args.cmd).exists(): - print("command not found at location " + args.cmd) - parser.print_help() - exit(1) - if not Path(args.schema).exists(): - print("schema not found at location " + args.schema) - parser.print_help() - exit(1) - return args - - -def run_sh_command(full_cmd, verbose=False, capture_output=False): - """Utility function to run an external command. Optionally return captured - stdout/stderr""" - - # Quote the command line for printing - full_cmd_esc = [shlex.quote(x) for x in full_cmd] - - if verbose: - print("### Running {}".format(" ".join(full_cmd_esc))) - - if capture_output: - rc = subprocess.run(full_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE) - if rc.returncode != 0: - print(rc.stdout.decode("utf-8")) - print(rc.stderr.decode("utf-8")) - raise Exception( - "Error running command: {}.\n{}".format( - " ".join(full_cmd_esc), rc.stderr.decode("utf-8") - ) - ) - return (rc.stdout, rc.stderr) - else: - rc = subprocess.run(full_cmd) - if rc.returncode != 0: - raise Exception("Error running command: {}".format(" ".join(full_cmd_esc))) - - -def runTest(args, testfile): - start_time = datetime.now() - result = TestResult.PASS - message = "" - - target = Path(f"serialization_script_output-{random.randint(0,10000)}.tosa") - source_json = Path(testfile.stem + ".json") - target_json = Path(target.stem + ".json") - - # Remove any previous files - if target.exists(): - target.unlink() - if source_json.exists(): - source_json.unlink() - if target_json.exists(): - target_json.unlink() - - try: - cmd = [args.cmd, str(testfile), str(target)] - run_sh_command(cmd, args.verbose) - # Create result json - cmd = [args.flatc, "--json", "--raw-binary", args.schema, "--", str(target)] - run_sh_command(cmd, args.verbose) - # Create source json - cmd = [args.flatc, "--json", "--raw-binary", args.schema, "--", str(testfile)] - run_sh_command(cmd, args.verbose) - if not filecmp.cmp(str(target_json), str(source_json), False): - print("Failed to compare files on " + str(testfile)) - result = TestResult.MISMATCH - # Cleanup generated files - source_json.unlink() - target_json.unlink() - target.unlink() - - except Exception as e: - message = str(e) - result = TestResult.COMMAND_ERROR - end_time = datetime.now() - return result, message, end_time - start_time - - -def getTestFiles(dir): - files = Path(dir).glob("**/*.tosa") - return files - - -def main(): - args = parseArgs() - testfiles = getTestFiles(args.test) - - suitename = "basic_serialization" - classname = "copy_test" - - xunit_result = xunit_results() - xunit_suite = xunit_result.create_suite("basic_serialization") - - failed = 0 - count = 0 - for test in testfiles: - count = count + 1 - (result, message, time_delta) = runTest(args, test) - xt = xunit_test(str(test), f"{suitename}.{classname}") - xt.time = str( - float(time_delta.seconds) + (float(time_delta.microseconds) * 1e-6) - ) - if result == TestResult.PASS: - pass - else: - xt.failed(message) - failed = failed + 1 - xunit_suite.tests.append(xt) - - xunit_result.write_results(args.xunit_file) - print(f"Total tests run: {count} failures: {failed}") - - -if __name__ == "__main__": - exit(main()) diff --git a/test/scripts/testfiles/test.tosa b/test/scripts/testfiles/test.tosa Binary files differdeleted file mode 100644 index 3b4ca56..0000000 --- a/test/scripts/testfiles/test.tosa +++ /dev/null diff --git a/test/src/serialization_npy_test.cpp b/test/src/serialization_npy_test.cpp index 27ec464..24e3aff 100644 --- a/test/src/serialization_npy_test.cpp +++ b/test/src/serialization_npy_test.cpp @@ -1,4 +1,4 @@ -// Copyright (c) 2021, ARM Limited. +// Copyright (c) 2021,2024, ARM Limited. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -37,7 +37,7 @@ int test_int_type(std::vector<int32_t> shape, std::default_random_engine& gen, s } auto buffer = std::make_unique<T[]>(total_size); - for (int i = 0; i < total_size; i++) + for (size_t i = 0; i < total_size; i++) { buffer[i] = gen_data(gen); } @@ -76,7 +76,46 @@ int test_float_type(std::vector<int32_t> shape, std::default_random_engine& gen, } auto buffer = std::make_unique<T[]>(total_size); - for (int i = 0; i < total_size; i++) + for (size_t i = 0; i < total_size; i++) + { + buffer[i] = gen_data(gen); + } + + NumpyUtilities::NPError err = NumpyUtilities::writeToNpyFile(filename.c_str(), shape, buffer.get()); + if (err != NumpyUtilities::NO_ERROR) + { + std::cout << "Error writing file, code " << err << std::endl; + return 1; + } + + auto read_buffer = std::make_unique<T[]>(total_size); + err = NumpyUtilities::readFromNpyFile(filename.c_str(), total_size, read_buffer.get()); + if (err != NumpyUtilities::NO_ERROR) + { + std::cout << "Error reading file, code " << err << std::endl; + return 1; + } + if (memcmp(buffer.get(), read_buffer.get(), total_size * sizeof(T))) + { + std::cout << "Miscompare" << std::endl; + return 1; + } + return 0; +} + +template <class T> +int test_double_type(std::vector<int32_t> shape, std::default_random_engine& gen, std::string& filename) +{ + size_t total_size = 1; + std::uniform_real_distribution<T> gen_data(std::numeric_limits<T>::min(), std::numeric_limits<T>::max()); + + for (auto i : shape) + { + total_size *= i; + } + + auto buffer = std::make_unique<T[]>(total_size); + for (size_t i = 0; i < total_size; i++) { buffer[i] = gen_data(gen); } @@ -114,7 +153,7 @@ int test_bool_type(std::vector<int32_t> shape, std::default_random_engine& gen, } auto buffer = std::make_unique<bool[]>(total_size); - for (int i = 0; i < total_size; i++) + for (size_t i = 0; i < total_size; i++) { buffer[i] = (gen_data(gen)) ? true : false; } @@ -144,15 +183,13 @@ int test_bool_type(std::vector<int32_t> shape, std::default_random_engine& gen, int main(int argc, char** argv) { - size_t total_size = 1; - int32_t seed = 1; + int32_t seed = 1; std::string str_type; std::string str_shape; std::string filename = "npytest.npy"; std::vector<int32_t> shape; - bool verbose = false; int opt; - while ((opt = getopt(argc, argv, "d:f:s:t:v")) != -1) + while ((opt = getopt(argc, argv, "d:f:s:t:")) != -1) { switch (opt) { @@ -168,9 +205,6 @@ int main(int argc, char** argv) case 't': str_shape = optarg; break; - case 'v': - verbose = true; - break; default: std::cerr << "Invalid argument" << std::endl; break; @@ -193,7 +227,6 @@ int main(int argc, char** argv) break; int val = stoi(substr, &pos, 0); assert(val); - total_size *= val; shape.push_back(val); } @@ -212,6 +245,10 @@ int main(int argc, char** argv) { return test_float_type<float>(shape, gen, filename); } + else if (str_type == "double") + { + return test_double_type<double>(shape, gen, filename); + } else if (str_type == "bool") { return test_bool_type(shape, gen, filename); diff --git a/third_party/flatbuffers b/third_party/flatbuffers -Subproject 697147a2e686486424b9d15fc3e1612586a60f9 +Subproject 0100f6a5779831fa7a651e4b67ef389a8752bd9 diff --git a/third_party/half/ChangeLog.txt b/third_party/half/ChangeLog.txt new file mode 100644 index 0000000..37f3dbf --- /dev/null +++ b/third_party/half/ChangeLog.txt @@ -0,0 +1,213 @@ +Release Notes {#changelog}
+=============
+
+2.2.0 release (2021-06-12):
+---------------------------
+
+- Added `rsqrt` function for inverse square root.
+- Improved performance of `pow` function.
+- Fixed bug that forgot to include `<immintrin.h>` for F16C intrinsics.
+
+
+2.1.0 release (2019-08-05):
+---------------------------
+
+- Added detection of IEEE floating-point exceptions to operators and functions.
+- Added configuration options for automatic exception handling.
+- Added functions for explicitly managing floating-point exception flags.
+- Improved accuracy of `pow` and `atan2` functions.
+
+
+2.0.0 release (2019-07-23):
+---------------------------
+
+- Made internal implementation independent from built-in floating point
+ facilities for increased reliability and IEEE-conformance.
+- Changed default rounding mode to rounding to nearest.
+- Always round ties to even when rounding to nearest.
+- Extended `constexpr` support to comparison and classification functions.
+- Added support for F16C compiler intrinsics for conversions.
+- Enabled C++11 feature detection for Intel compilers.
+
+
+1.12.0 release (2017-03-06):
+----------------------------
+
+- Changed behaviour of `half_cast` to perform conversions to/from `double`
+ and `long double` directly according to specified rounding mode, without an
+ intermediate `float` conversion.
+- Added `noexcept` specifiers to constructors.
+- Fixed minor portability problem with `logb` and `ilogb`.
+- Tested for *VC++ 2015*.
+
+
+1.11.0 release (2013-11-16):
+----------------------------
+
+- Made tie-breaking behaviour in round to nearest configurable by
+ `HALF_ROUND_TIES_TO_EVEN` macro.
+- Completed support for all C++11 mathematical functions even if single-
+ precision versions from `<cmath>` are unsupported.
+- Fixed inability to disable support for C++11 mathematical functions on
+ *VC++ 2013*.
+
+
+1.10.0 release (2013-11-09):
+----------------------------
+
+- Made default rounding mode configurable by `HALF_ROUND_STYLE` macro.
+- Added support for non-IEEE single-precision implementations.
+- Added `HALF_ENABLE_CPP11_TYPE_TRAITS` preprocessor flag for checking
+ support for C++11 type traits and TMP features.
+- Restricted `half_cast` to support built-in arithmetic types only.
+- Changed behaviour of `half_cast` to respect rounding mode when casting
+ to/from integer types.
+
+
+1.9.2 release (2013-11-01):
+---------------------------
+
+- Tested for *gcc 4.8*.
+- Tested and fixed for *VC++ 2013*.
+- Removed unnecessary warnings in *MSVC*.
+
+
+1.9.1 release (2013-08-08):
+---------------------------
+
+- Fixed problems with older gcc and MSVC versions.
+- Small fix to non-C++11 implementations of `remainder` and `remquo`.
+
+
+1.9.0 release (2013-08-07):
+---------------------------
+
+- Changed behaviour of `nearbyint`, `rint`, `lrint` and `llrint` to use
+ rounding mode of half-precision implementation (which is
+ truncating/indeterminate) instead of single-precision rounding mode.
+- Added support for more C++11 mathematical functions even if single-
+ precision versions from `<cmath>` are unsupported, in particular
+ `remainder`, `remquo` and `cbrt`.
+- Minor implementation changes.
+
+
+1.8.1 release (2013-01-22):
+---------------------------
+
+- Fixed bug resulting in multiple definitions of the `nanh` function due to
+ a missing `inline` specification.
+
+
+1.8.0 release (2013-01-19):
+---------------------------
+
+- Added support for more C++11 mathematical functions even if single-
+ precision versions from `<cmath>` are unsupported, in particular
+ exponential and logarithm functions, hyperbolic area functions and the
+ hypotenuse function.
+- Made `fma` function use default implementation if single-precision version
+ from `<cmath>` is not faster and thus `FP_FAST_FMAH` to be defined always.
+- Fixed overload resolution issues when invoking certain mathematical
+ functions by unqualified calls.
+
+
+1.7.0 release (2012-10-26):
+---------------------------
+
+- Added support for C++11 `noexcept` specifiers.
+- Changed C++11 `long long` to be supported on *VC++ 2003* and up.
+
+
+1.6.1 release (2012-09-13):
+---------------------------
+
+- Made `fma` and `fdim` functions available even if corresponding
+ single-precision functions are not.
+
+
+1.6.0 release (2012-09-12):
+---------------------------
+
+- Added `HALF_ENABLE_CPP11_LONG_LONG` to control support for `long long`
+ integers and corresponding mathematical functions.
+- Fixed C++98 compatibility on non-VC compilers.
+
+
+1.5.1 release (2012-08-17):
+---------------------------
+
+- Recorrected `std::numeric_limits::round_style` to always return
+ `std::round_indeterminate`, due to overflow-handling deviating from
+ correct round-toward-zero behaviour.
+
+
+1.5.0 release (2012-08-16):
+---------------------------
+
+- Added `half_cast` for explicitly casting between half and any type
+ convertible to/from `float` and allowing the explicit specification of
+ the rounding mode to use.
+
+
+1.4.0 release (2012-08-12):
+---------------------------
+
+- Added support for C++11 generalized constant expressions (`constexpr`).
+
+
+1.3.1 release (2012-08-11):
+---------------------------
+
+- Fixed requirement for `std::signbit` and `std::isnan` (even if C++11
+ `<cmath>` functions disabled) on non-VC compilers.
+
+
+1.3.0 release (2012-08-10):
+---------------------------
+
+- Made requirement for `<cstdint>` and `static_assert` optional and thus
+ made the library C++98-compatible.
+- Made support for C++11 features user-overridable through explicit
+ definition of corresponding preprocessor symbols to either 0 or 1.
+- Renamed `HALF_ENABLE_HASH` to `HALF_ENABLE_CPP11_HASH` in correspondence
+ with other C++11 preprocessor symbols.
+
+
+1.2.0 release (2012-08-07):
+---------------------------
+
+- Added proper preprocessor definitions for `HUGE_VALH` and `FP_FAST_FMAH`
+ in correspondence with their single-precision counterparts from `<cmath>`.
+- Fixed internal preprocessor macros to be properly undefined after use.
+
+
+1.1.2 release (2012-08-07):
+---------------------------
+
+- Revised `std::numeric_limits::round_style` to return
+ `std::round_toward_zero` if the `float` version also does and
+ `std::round_indeterminate` otherwise.
+- Fixed `std::numeric_limits::round_error` to reflect worst-case round
+ toward zero behaviour.
+
+
+1.1.1 release (2012-08-06):
+---------------------------
+
+- Fixed `std::numeric_limits::min` to return smallest positive normal
+ number, instead of subnormal number.
+- Fixed `std::numeric_limits::round_style` to return
+ `std::round_indeterminate` due to mixture of separately rounded
+ single-precision arithmetics with truncating single-to-half conversions.
+
+
+1.1.0 release (2012-08-06):
+---------------------------
+
+- Added half-precision literals.
+
+
+1.0.0 release (2012-08-05):
+---------------------------
+
+- First release.
diff --git a/third_party/half/LICENSE.txt b/third_party/half/LICENSE.txt new file mode 100644 index 0000000..45f55db --- /dev/null +++ b/third_party/half/LICENSE.txt @@ -0,0 +1,21 @@ +The MIT License
+
+Copyright (c) 2012-2021 Christian Rau
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
diff --git a/third_party/half/README.txt b/third_party/half/README.txt new file mode 100644 index 0000000..3dd0d1c --- /dev/null +++ b/third_party/half/README.txt @@ -0,0 +1,317 @@ +HALF-PRECISION FLOATING-POINT LIBRARY (Version 2.2.0)
+-----------------------------------------------------
+
+This is a C++ header-only library to provide an IEEE 754 conformant 16-bit
+half-precision floating-point type along with corresponding arithmetic
+operators, type conversions and common mathematical functions. It aims for both
+efficiency and ease of use, trying to accurately mimic the behaviour of the
+built-in floating-point types at the best performance possible.
+
+
+INSTALLATION AND REQUIREMENTS
+-----------------------------
+
+Conveniently, the library consists of just a single header file containing all
+the functionality, which can be directly included by your projects, without the
+neccessity to build anything or link to anything.
+
+Whereas this library is fully C++98-compatible, it can profit from certain
+C++11 features. Support for those features is checked automatically at compile
+(or rather preprocessing) time, but can be explicitly enabled or disabled by
+predefining the corresponding preprocessor symbols to either 1 or 0 yourself
+before including half.hpp. This is useful when the automatic detection fails
+(for more exotic implementations) or when a feature should be explicitly
+disabled:
+
+ - 'long long' integer type for mathematical functions returning 'long long'
+ results (enabled for VC++ 2003 and icc 11.1 and newer, gcc and clang,
+ overridable with 'HALF_ENABLE_CPP11_LONG_LONG').
+
+ - Static assertions for extended compile-time checks (enabled for VC++ 2010,
+ gcc 4.3, clang 2.9, icc 11.1 and newer, overridable with
+ 'HALF_ENABLE_CPP11_STATIC_ASSERT').
+
+ - Generalized constant expressions (enabled for VC++ 2015, gcc 4.6, clang 3.1,
+ icc 14.0 and newer, overridable with 'HALF_ENABLE_CPP11_CONSTEXPR').
+
+ - noexcept exception specifications (enabled for VC++ 2015, gcc 4.6,
+ clang 3.0, icc 14.0 and newer, overridable with 'HALF_ENABLE_CPP11_NOEXCEPT').
+
+ - User-defined literals for half-precision literals to work (enabled for
+ VC++ 2015, gcc 4.7, clang 3.1, icc 15.0 and newer, overridable with
+ 'HALF_ENABLE_CPP11_USER_LITERALS').
+
+ - Thread-local storage for per-thread floating-point exception flags (enabled
+ for VC++ 2015, gcc 4.8, clang 3.3, icc 15.0 and newer, overridable with
+ 'HALF_ENABLE_CPP11_THREAD_LOCAL').
+
+ - Type traits and template meta-programming features from <type_traits>
+ (enabled for VC++ 2010, libstdc++ 4.3, libc++ and newer, overridable with
+ 'HALF_ENABLE_CPP11_TYPE_TRAITS').
+
+ - Special integer types from <cstdint> (enabled for VC++ 2010, libstdc++ 4.3,
+ libc++ and newer, overridable with 'HALF_ENABLE_CPP11_CSTDINT').
+
+ - Certain C++11 single-precision mathematical functions from <cmath> for
+ floating-point classification during conversions from higher precision types
+ (enabled for VC++ 2013, libstdc++ 4.3, libc++ and newer, overridable with
+ 'HALF_ENABLE_CPP11_CMATH').
+
+ - Floating-point environment control from <cfenv> for possible exception
+ propagation to the built-in floating-point platform (enabled for VC++ 2013,
+ libstdc++ 4.3, libc++ and newer, overridable with 'HALF_ENABLE_CPP11_CFENV').
+
+ - Hash functor 'std::hash' from <functional> (enabled for VC++ 2010,
+ libstdc++ 4.3, libc++ and newer, overridable with 'HALF_ENABLE_CPP11_HASH').
+
+The library has been tested successfully with Visual C++ 2005-2015, gcc 4-8
+and clang 3-8 on 32- and 64-bit x86 systems. Please contact me if you have any
+problems, suggestions or even just success testing it on other platforms.
+
+
+DOCUMENTATION
+-------------
+
+What follows are some general words about the usage of the library and its
+implementation. For a complete documentation of its interface consult the
+corresponding website http://half.sourceforge.net. You may also generate the
+complete developer documentation from the library's only include file's doxygen
+comments, but this is more relevant to developers rather than mere users.
+
+BASIC USAGE
+
+To make use of the library just include its only header file half.hpp, which
+defines all half-precision functionality inside the 'half_float' namespace. The
+actual 16-bit half-precision data type is represented by the 'half' type, which
+uses the standard IEEE representation with 1 sign bit, 5 exponent bits and 11
+mantissa bits (including the hidden bit) and supports all types of special
+values, like subnormal values, infinity and NaNs. This type behaves like the
+built-in floating-point types as much as possible, supporting the usual
+arithmetic, comparison and streaming operators, which makes its use pretty
+straight-forward:
+
+ using half_float::half;
+ half a(3.4), b(5);
+ half c = a * b;
+ c += 3;
+ if(c > a)
+ std::cout << c << std::endl;
+
+Additionally the 'half_float' namespace also defines half-precision versions
+for all mathematical functions of the C++ standard library, which can be used
+directly through ADL:
+
+ half a(-3.14159);
+ half s = sin(abs(a));
+ long l = lround(s);
+
+You may also specify explicit half-precision literals, since the library
+provides a user-defined literal inside the 'half_float::literal' namespace,
+which you just need to import (assuming support for C++11 user-defined literals):
+
+ using namespace half_float::literal;
+ half x = 1.0_h;
+
+Furthermore the library provides proper specializations for
+'std::numeric_limits', defining various implementation properties, and
+'std::hash' for hashing half-precision numbers (assuming support for C++11
+'std::hash'). Similar to the corresponding preprocessor symbols from <cmath>
+the library also defines the 'HUGE_VALH' constant and maybe the 'FP_FAST_FMAH'
+symbol.
+
+CONVERSIONS AND ROUNDING
+
+The half is explicitly constructible/convertible from a single-precision float
+argument. Thus it is also explicitly constructible/convertible from any type
+implicitly convertible to float, but constructing it from types like double or
+int will involve the usual warnings arising when implicitly converting those to
+float because of the lost precision. On the one hand those warnings are
+intentional, because converting those types to half neccessarily also reduces
+precision. But on the other hand they are raised for explicit conversions from
+those types, when the user knows what he is doing. So if those warnings keep
+bugging you, then you won't get around first explicitly converting to float
+before converting to half, or use the 'half_cast' described below. In addition
+you can also directly assign float values to halfs.
+
+In contrast to the float-to-half conversion, which reduces precision, the
+conversion from half to float (and thus to any other type implicitly
+convertible from float) is implicit, because all values represetable with
+half-precision are also representable with single-precision. This way the
+half-to-float conversion behaves similar to the builtin float-to-double
+conversion and all arithmetic expressions involving both half-precision and
+single-precision arguments will be of single-precision type. This way you can
+also directly use the mathematical functions of the C++ standard library,
+though in this case you will invoke the single-precision versions which will
+also return single-precision values, which is (even if maybe performing the
+exact same computation, see below) not as conceptually clean when working in a
+half-precision environment.
+
+The default rounding mode for conversions between half and more precise types
+as well as for rounding results of arithmetic operations and mathematical
+functions rounds to the nearest representable value. But by predefining the
+'HALF_ROUND_STYLE' preprocessor symbol this default can be overridden with one
+of the other standard rounding modes using their respective constants or the
+equivalent values of 'std::float_round_style' (it can even be synchronized with
+the built-in single-precision implementation by defining it to
+'std::numeric_limits<float>::round_style'):
+
+ - 'std::round_indeterminate' (-1) for the fastest rounding.
+
+ - 'std::round_toward_zero' (0) for rounding toward zero.
+
+ - 'std::round_to_nearest' (1) for rounding to the nearest value (default).
+
+ - 'std::round_toward_infinity' (2) for rounding toward positive infinity.
+
+ - 'std::round_toward_neg_infinity' (3) for rounding toward negative infinity.
+
+In addition to changing the overall default rounding mode one can also use the
+'half_cast'. This converts between half and any built-in arithmetic type using
+a configurable rounding mode (or the default rounding mode if none is
+specified). In addition to a configurable rounding mode, 'half_cast' has
+another big difference to a mere 'static_cast': Any conversions are performed
+directly using the given rounding mode, without any intermediate conversion
+to/from 'float'. This is especially relevant for conversions to integer types,
+which don't necessarily truncate anymore. But also for conversions from
+'double' or 'long double' this may produce more precise results than a
+pre-conversion to 'float' using the single-precision implementation's current
+rounding mode would.
+
+ half a = half_cast<half>(4.2);
+ half b = half_cast<half,std::numeric_limits<float>::round_style>(4.2f);
+ assert( half_cast<int, std::round_to_nearest>( 0.7_h ) == 1 );
+ assert( half_cast<half,std::round_toward_zero>( 4097 ) == 4096.0_h );
+ assert( half_cast<half,std::round_toward_infinity>( 4097 ) == 4100.0_h );
+ assert( half_cast<half,std::round_toward_infinity>( std::numeric_limits<double>::min() ) > 0.0_h );
+
+ACCURACY AND PERFORMANCE
+
+From version 2.0 onward the library is implemented without employing the
+underlying floating-point implementation of the system (except for conversions,
+of course), providing an entirely self-contained half-precision implementation
+with results independent from the system's existing single- or double-precision
+implementation and its rounding behaviour.
+
+As to accuracy, many of the operators and functions provided by this library
+are exact to rounding for all rounding modes, i.e. the error to the exact
+result is at most 0.5 ULP (unit in the last place) for rounding to nearest and
+less than 1 ULP for all other rounding modes. This holds for all the operations
+required by the IEEE 754 standard and many more. Specifically the following
+functions might exhibit a deviation from the correctly rounded exact result by
+1 ULP for a select few input values: 'expm1', 'log1p', 'pow', 'atan2', 'erf',
+'erfc', 'lgamma', 'tgamma' (for more details see the documentation of the
+individual functions). All other functions and operators are always exact to
+rounding or independent of the rounding mode altogether.
+
+The increased IEEE-conformance and cleanliness of this implementation comes
+with a certain performance cost compared to doing computations and mathematical
+functions in hardware-accelerated single-precision. On average and depending on
+the platform, the arithemtic operators are about 75% as fast and the
+mathematical functions about 33-50% as fast as performing the corresponding
+operations in single-precision and converting between the inputs and outputs.
+However, directly computing with half-precision values is a rather rare
+use-case and usually using actual 'float' values for all computations and
+temproraries and using 'half's only for storage is the recommended way. But
+nevertheless the goal of this library was to provide a complete and
+conceptually clean IEEE-confromant half-precision implementation and in the few
+cases when you do need to compute directly in half-precision you do so for a
+reason and want accurate results.
+
+If necessary, this internal implementation can be overridden by predefining the
+'HALF_ARITHMETIC_TYPE' preprocessor symbol to one of the built-in
+floating-point types ('float', 'double' or 'long double'), which will cause the
+library to use this type for computing arithmetic operations and mathematical
+functions (if available). However, due to using the platform's floating-point
+implementation (and its rounding behaviour) internally, this might cause
+results to deviate from the specified half-precision rounding mode. It will of
+course also inhibit the automatic exception detection described below.
+
+The conversion operations between half-precision and single-precision types can
+also make use of the F16C extension for x86 processors by using the
+corresponding compiler intrinsics from <immintrin.h>. Support for this is
+checked at compile-time by looking for the '__F16C__' macro which at least gcc
+and clang define based on the target platform. It can also be enabled manually
+by predefining the 'HALF_ENABLE_F16C_INTRINSICS' preprocessor symbol to 1, or 0
+for explicitly disabling it. However, this will directly use the corresponding
+intrinsics for conversion without checking if they are available at runtime
+(possibly crashing if they are not), so make sure they are supported on the
+target platform before enabling this.
+
+EXCEPTION HANDLING
+
+The half-precision implementation supports all 5 required floating-point
+exceptions from the IEEE standard to indicate erroneous inputs or inexact
+results during operations. These are represented by exception flags which
+actually use the same values as the corresponding 'FE_...' flags defined in
+C++11's <cfenv> header if supported, specifically:
+
+ - 'FE_INVALID' for invalid inputs to an operation.
+ - 'FE_DIVBYZERO' for finite inputs producing infinite results.
+ - 'FE_OVERFLOW' if a result is too large to represent finitely.
+ - 'FE_UNDERFLOW' for a subnormal or zero result after rounding.
+ - 'FE_INEXACT' if a result needed rounding to be representable.
+ - 'FE_ALL_EXCEPT' as a convenient OR of all possible exception flags.
+
+The internal exception flag state will start with all flags cleared and is
+maintained per thread if C++11 thread-local storage is supported, otherwise it
+will be maintained globally and will theoretically NOT be thread-safe (while
+practically being as thread-safe as a simple integer variable can be). These
+flags can be managed explicitly using the library's error handling functions,
+which again try to mimic the built-in functions for handling floating-point
+exceptions from <cfenv>. You can clear them with 'feclearexcept' (which is the
+only way a flag can be cleared), test them with 'fetestexcept', explicitly
+raise errors with 'feraiseexcept' and save and restore their state using
+'fegetexceptflag' and 'fesetexceptflag'. You can also throw corresponding C++
+exceptions based on the current flag state using 'fethrowexcept'.
+
+However, any automatic exception detection and handling during half-precision
+operations and functions is DISABLED by default, since it comes with a minor
+performance overhead due to runtime checks, and reacting to IEEE floating-point
+exceptions is rarely ever needed in application code. But the library fully
+supports IEEE-conformant detection of floating-point exceptions and various
+ways for handling them, which can be enabled by pre-defining the corresponding
+preprocessor symbols to 1. They can be enabled individually or all at once and
+they will be processed in the order they are listed here:
+
+ - 'HALF_ERRHANDLING_FLAGS' sets the internal exception flags described above
+ whenever the corresponding exception occurs.
+ - 'HALF_ERRHANDLING_ERRNO' sets the value of 'errno' from <cerrno> similar to
+ the behaviour of the built-in floating-point types when 'MATH_ERRNO' is used.
+ - 'HALF_ERRHANDLING_FENV' will propagate exceptions to the built-in
+ floating-point implementation using 'std::feraiseexcept' if support for
+ C++11 floating-point control is enabled. However, this does not synchronize
+ exceptions: neither will clearing propagate nor will it work in reverse.
+ - 'HALF_ERRHANDLING_THROW_...' can be defined to a string literal which will
+ be used as description message for a C++ exception that is thrown whenever
+ a 'FE_...' exception occurs, similar to the behaviour of 'fethrowexcept'.
+
+If any of the above error handling is activated, non-quiet operations on
+half-precision values will also raise a 'FE_INVALID' exception whenever
+they encounter a signaling NaN value, in addition to transforming the value
+into a quiet NaN. If error handling is disabled, signaling NaNs will be
+treated like quiet NaNs (while still getting explicitly quieted if propagated
+to the result). There can also be additional treatment of overflow and
+underflow errors after they have been processed as above, which is ENABLED by
+default (but of course only takes effect if any other exception handling is
+activated) unless overridden by pre-defining the corresponding preprocessor
+symbol to 0:
+
+ - 'HALF_ERRHANDLING_OVERFLOW_TO_INEXACT' will cause overflow errors to also
+ raise a 'FE_INEXACT' exception.
+ - 'HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT' will cause underflow errors to also
+ raise a 'FE_INEXACT' exception. This will also slightly change the
+ behaviour of the underflow exception, which will ONLY be raised if the
+ result is actually inexact due to underflow. If this is disabled, underflow
+ exceptions will be raised for ANY (possibly exact) subnormal result.
+
+
+CREDITS AND CONTACT
+-------------------
+
+This library is developed by CHRISTIAN RAU and released under the MIT License
+(see LICENSE.txt). If you have any questions or problems with it, feel free to
+contact me at rauy@users.sourceforge.net.
+
+Additional credit goes to JEROEN VAN DER ZIJP for his paper on "Fast Half Float
+Conversions", whose algorithms have been used in the library for converting
+between half-precision and single-precision values.
diff --git a/third_party/half/include/half.hpp b/third_party/half/include/half.hpp new file mode 100644 index 0000000..ee8819a --- /dev/null +++ b/third_party/half/include/half.hpp @@ -0,0 +1,4605 @@ +// half - IEEE 754-based half-precision floating-point library.
+//
+// Copyright (c) 2012-2021 Christian Rau <rauy@users.sourceforge.net>
+// Copyright (c) 2023, ARM Limited.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation
+// files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy,
+// modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE
+// WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+// Version 2.2.0
+
+/// \file
+/// Main header file for half-precision functionality.
+
+#ifndef HALF_HALF_HPP
+#define HALF_HALF_HPP
+
+#define HALF_GCC_VERSION (__GNUC__*100+__GNUC_MINOR__)
+
+#if defined(__INTEL_COMPILER)
+ #define HALF_ICC_VERSION __INTEL_COMPILER
+#elif defined(__ICC)
+ #define HALF_ICC_VERSION __ICC
+#elif defined(__ICL)
+ #define HALF_ICC_VERSION __ICL
+#else
+ #define HALF_ICC_VERSION 0
+#endif
+
+// check C++11 language features
+#if defined(__clang__) // clang
+ #if __has_feature(cxx_static_assert) && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
+ #endif
+ #if __has_feature(cxx_constexpr) && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
+ #endif
+ #if __has_feature(cxx_noexcept) && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
+ #endif
+ #if __has_feature(cxx_user_literals) && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
+ #endif
+ #if __has_feature(cxx_thread_local) && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
+ #endif
+ #if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L) && !defined(HALF_ENABLE_CPP11_LONG_LONG)
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
+ #endif
+#elif HALF_ICC_VERSION && defined(__INTEL_CXX11_MODE__) // Intel C++
+ #if HALF_ICC_VERSION >= 1500 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
+ #endif
+ #if HALF_ICC_VERSION >= 1500 && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
+ #endif
+ #if HALF_ICC_VERSION >= 1400 && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
+ #endif
+ #if HALF_ICC_VERSION >= 1400 && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
+ #endif
+ #if HALF_ICC_VERSION >= 1110 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
+ #endif
+ #if HALF_ICC_VERSION >= 1110 && !defined(HALF_ENABLE_CPP11_LONG_LONG)
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
+ #endif
+#elif defined(__GNUC__) // gcc
+ #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L
+ #if HALF_GCC_VERSION >= 408 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
+ #endif
+ #if HALF_GCC_VERSION >= 407 && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
+ #endif
+ #if HALF_GCC_VERSION >= 406 && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
+ #endif
+ #if HALF_GCC_VERSION >= 406 && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
+ #endif
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
+ #endif
+ #if !defined(HALF_ENABLE_CPP11_LONG_LONG)
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
+ #endif
+ #endif
+ #define HALF_TWOS_COMPLEMENT_INT 1
+#elif defined(_MSC_VER) // Visual C++
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL)
+ #define HALF_ENABLE_CPP11_THREAD_LOCAL 1
+ #endif
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_USER_LITERALS)
+ #define HALF_ENABLE_CPP11_USER_LITERALS 1
+ #endif
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_CONSTEXPR)
+ #define HALF_ENABLE_CPP11_CONSTEXPR 1
+ #endif
+ #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_NOEXCEPT)
+ #define HALF_ENABLE_CPP11_NOEXCEPT 1
+ #endif
+ #if _MSC_VER >= 1600 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT)
+ #define HALF_ENABLE_CPP11_STATIC_ASSERT 1
+ #endif
+ #if _MSC_VER >= 1310 && !defined(HALF_ENABLE_CPP11_LONG_LONG)
+ #define HALF_ENABLE_CPP11_LONG_LONG 1
+ #endif
+ #define HALF_TWOS_COMPLEMENT_INT 1
+ #define HALF_POP_WARNINGS 1
+ #pragma warning(push)
+ #pragma warning(disable : 4099 4127 4146) //struct vs class, constant in if, negative unsigned
+#endif
+
+// check C++11 library features
+#include <utility>
+#if defined(_LIBCPP_VERSION) // libc++
+ #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103
+ #ifndef HALF_ENABLE_CPP11_TYPE_TRAITS
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
+ #endif
+ #ifndef HALF_ENABLE_CPP11_CSTDINT
+ #define HALF_ENABLE_CPP11_CSTDINT 1
+ #endif
+ #ifndef HALF_ENABLE_CPP11_CMATH
+ #define HALF_ENABLE_CPP11_CMATH 1
+ #endif
+ #ifndef HALF_ENABLE_CPP11_HASH
+ #define HALF_ENABLE_CPP11_HASH 1
+ #endif
+ #ifndef HALF_ENABLE_CPP11_CFENV
+ #define HALF_ENABLE_CPP11_CFENV 1
+ #endif
+ #endif
+#elif defined(__GLIBCXX__) // libstdc++
+ #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103
+ #ifdef __clang__
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS)
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
+ #endif
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CSTDINT)
+ #define HALF_ENABLE_CPP11_CSTDINT 1
+ #endif
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CMATH)
+ #define HALF_ENABLE_CPP11_CMATH 1
+ #endif
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_HASH)
+ #define HALF_ENABLE_CPP11_HASH 1
+ #endif
+ #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CFENV)
+ #define HALF_ENABLE_CPP11_CFENV 1
+ #endif
+ #else
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS)
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
+ #endif
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CSTDINT)
+ #define HALF_ENABLE_CPP11_CSTDINT 1
+ #endif
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CMATH)
+ #define HALF_ENABLE_CPP11_CMATH 1
+ #endif
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_HASH)
+ #define HALF_ENABLE_CPP11_HASH 1
+ #endif
+ #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CFENV)
+ #define HALF_ENABLE_CPP11_CFENV 1
+ #endif
+ #endif
+ #endif
+#elif defined(_CPPLIB_VER) // Dinkumware/Visual C++
+ #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS)
+ #define HALF_ENABLE_CPP11_TYPE_TRAITS 1
+ #endif
+ #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_CSTDINT)
+ #define HALF_ENABLE_CPP11_CSTDINT 1
+ #endif
+ #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_HASH)
+ #define HALF_ENABLE_CPP11_HASH 1
+ #endif
+ #if _CPPLIB_VER >= 610 && !defined(HALF_ENABLE_CPP11_CMATH)
+ #define HALF_ENABLE_CPP11_CMATH 1
+ #endif
+ #if _CPPLIB_VER >= 610 && !defined(HALF_ENABLE_CPP11_CFENV)
+ #define HALF_ENABLE_CPP11_CFENV 1
+ #endif
+#endif
+#undef HALF_GCC_VERSION
+#undef HALF_ICC_VERSION
+
+// any error throwing C++ exceptions?
+#if defined(HALF_ERRHANDLING_THROW_INVALID) || defined(HALF_ERRHANDLING_THROW_DIVBYZERO) || defined(HALF_ERRHANDLING_THROW_OVERFLOW) || defined(HALF_ERRHANDLING_THROW_UNDERFLOW) || defined(HALF_ERRHANDLING_THROW_INEXACT)
+#define HALF_ERRHANDLING_THROWS 1
+#endif
+
+// any error handling enabled?
+#define HALF_ERRHANDLING (HALF_ERRHANDLING_FLAGS||HALF_ERRHANDLING_ERRNO||HALF_ERRHANDLING_FENV||HALF_ERRHANDLING_THROWS)
+
+#if HALF_ERRHANDLING
+ #define HALF_UNUSED_NOERR(name) name
+#else
+ #define HALF_UNUSED_NOERR(name)
+#endif
+
+// support constexpr
+#if HALF_ENABLE_CPP11_CONSTEXPR
+ #define HALF_CONSTEXPR constexpr
+ #define HALF_CONSTEXPR_CONST constexpr
+ #if HALF_ERRHANDLING
+ #define HALF_CONSTEXPR_NOERR
+ #else
+ #define HALF_CONSTEXPR_NOERR constexpr
+ #endif
+#else
+ #define HALF_CONSTEXPR
+ #define HALF_CONSTEXPR_CONST const
+ #define HALF_CONSTEXPR_NOERR
+#endif
+
+// support noexcept
+#if HALF_ENABLE_CPP11_NOEXCEPT
+ #define HALF_NOEXCEPT noexcept
+ #define HALF_NOTHROW noexcept
+#else
+ #define HALF_NOEXCEPT
+ #define HALF_NOTHROW throw()
+#endif
+
+// support thread storage
+#if HALF_ENABLE_CPP11_THREAD_LOCAL
+ #define HALF_THREAD_LOCAL thread_local
+#else
+ #define HALF_THREAD_LOCAL static
+#endif
+
+#include <utility>
+#include <algorithm>
+#include <istream>
+#include <ostream>
+#include <limits>
+#include <stdexcept>
+#include <climits>
+#include <cmath>
+#include <cstring>
+#include <cstdlib>
+#if HALF_ENABLE_CPP11_TYPE_TRAITS
+ #include <type_traits>
+#endif
+#if HALF_ENABLE_CPP11_CSTDINT
+ #include <cstdint>
+#endif
+#if HALF_ERRHANDLING_ERRNO
+ #include <cerrno>
+#endif
+#if HALF_ENABLE_CPP11_CFENV
+ #include <cfenv>
+#endif
+#if HALF_ENABLE_CPP11_HASH
+ #include <functional>
+#endif
+
+
+#ifndef HALF_ENABLE_F16C_INTRINSICS
+ /// Enable F16C intruction set intrinsics.
+ /// Defining this to 1 enables the use of [F16C compiler intrinsics](https://en.wikipedia.org/wiki/F16C) for converting between
+ /// half-precision and single-precision values which may result in improved performance. This will not perform additional checks
+ /// for support of the F16C instruction set, so an appropriate target platform is required when enabling this feature.
+ ///
+ /// Unless predefined it will be enabled automatically when the `__F16C__` symbol is defined, which some compilers do on supporting platforms.
+ #define HALF_ENABLE_F16C_INTRINSICS __F16C__
+#endif
+#if HALF_ENABLE_F16C_INTRINSICS
+ #include <immintrin.h>
+#endif
+
+#ifdef HALF_DOXYGEN_ONLY
+/// Type for internal floating-point computations.
+/// This can be predefined to a built-in floating-point type (`float`, `double` or `long double`) to override the internal
+/// half-precision implementation to use this type for computing arithmetic operations and mathematical function (if available).
+/// This can result in improved performance for arithmetic operators and mathematical functions but might cause results to
+/// deviate from the specified half-precision rounding mode and inhibits proper detection of half-precision exceptions.
+#define HALF_ARITHMETIC_TYPE (undefined)
+
+/// Enable internal exception flags.
+/// Defining this to 1 causes operations on half-precision values to raise internal floating-point exception flags according to
+/// the IEEE 754 standard. These can then be cleared and checked with clearexcept(), testexcept().
+#define HALF_ERRHANDLING_FLAGS 0
+
+/// Enable exception propagation to `errno`.
+/// Defining this to 1 causes operations on half-precision values to propagate floating-point exceptions to
+/// [errno](https://en.cppreference.com/w/cpp/error/errno) from `<cerrno>`. Specifically this will propagate domain errors as
+/// [EDOM](https://en.cppreference.com/w/cpp/error/errno_macros) and pole, overflow and underflow errors as
+/// [ERANGE](https://en.cppreference.com/w/cpp/error/errno_macros). Inexact errors won't be propagated.
+#define HALF_ERRHANDLING_ERRNO 0
+
+/// Enable exception propagation to built-in floating-point platform.
+/// Defining this to 1 causes operations on half-precision values to propagate floating-point exceptions to the built-in
+/// single- and double-precision implementation's exception flags using the
+/// [C++11 floating-point environment control](https://en.cppreference.com/w/cpp/numeric/fenv) from `<cfenv>`. However, this
+/// does not work in reverse and single- or double-precision exceptions will not raise the corresponding half-precision
+/// exception flags, nor will explicitly clearing flags clear the corresponding built-in flags.
+#define HALF_ERRHANDLING_FENV 0
+
+/// Throw C++ exception on domain errors.
+/// Defining this to a string literal causes operations on half-precision values to throw a
+/// [std::domain_error](https://en.cppreference.com/w/cpp/error/domain_error) with the specified message on domain errors.
+#define HALF_ERRHANDLING_THROW_INVALID (undefined)
+
+/// Throw C++ exception on pole errors.
+/// Defining this to a string literal causes operations on half-precision values to throw a
+/// [std::domain_error](https://en.cppreference.com/w/cpp/error/domain_error) with the specified message on pole errors.
+#define HALF_ERRHANDLING_THROW_DIVBYZERO (undefined)
+
+/// Throw C++ exception on overflow errors.
+/// Defining this to a string literal causes operations on half-precision values to throw a
+/// [std::overflow_error](https://en.cppreference.com/w/cpp/error/overflow_error) with the specified message on overflows.
+#define HALF_ERRHANDLING_THROW_OVERFLOW (undefined)
+
+/// Throw C++ exception on underflow errors.
+/// Defining this to a string literal causes operations on half-precision values to throw a
+/// [std::underflow_error](https://en.cppreference.com/w/cpp/error/underflow_error) with the specified message on underflows.
+#define HALF_ERRHANDLING_THROW_UNDERFLOW (undefined)
+
+/// Throw C++ exception on rounding errors.
+/// Defining this to 1 causes operations on half-precision values to throw a
+/// [std::range_error](https://en.cppreference.com/w/cpp/error/range_error) with the specified message on general rounding errors.
+#define HALF_ERRHANDLING_THROW_INEXACT (undefined)
+#endif
+
+#ifndef HALF_ERRHANDLING_OVERFLOW_TO_INEXACT
+/// Raise INEXACT exception on overflow.
+/// Defining this to 1 (default) causes overflow errors to automatically raise inexact exceptions in addition.
+/// These will be raised after any possible handling of the underflow exception.
+#define HALF_ERRHANDLING_OVERFLOW_TO_INEXACT 1
+#endif
+
+#ifndef HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT
+/// Raise INEXACT exception on underflow.
+/// Defining this to 1 (default) causes underflow errors to automatically raise inexact exceptions in addition.
+/// These will be raised after any possible handling of the underflow exception.
+///
+/// **Note:** This will actually cause underflow (and the accompanying inexact) exceptions to be raised *only* when the result
+/// is inexact, while if disabled bare underflow errors will be raised for *any* (possibly exact) subnormal result.
+#define HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT 1
+#endif
+
+/// Default rounding mode.
+/// This specifies the rounding mode used for all conversions between [half](\ref half_float::half)s and more precise types
+/// (unless using half_cast() and specifying the rounding mode directly) as well as in arithmetic operations and mathematical
+/// functions. It can be redefined (before including half.hpp) to one of the standard rounding modes using their respective
+/// constants or the equivalent values of
+/// [std::float_round_style](https://en.cppreference.com/w/cpp/types/numeric_limits/float_round_style):
+///
+/// `std::float_round_style` | value | rounding
+/// ---------------------------------|-------|-------------------------
+/// `std::round_indeterminate` | -1 | fastest
+/// `std::round_toward_zero` | 0 | toward zero
+/// `std::round_to_nearest` | 1 | to nearest (default)
+/// `std::round_toward_infinity` | 2 | toward positive infinity
+/// `std::round_toward_neg_infinity` | 3 | toward negative infinity
+///
+/// By default this is set to `1` (`std::round_to_nearest`), which rounds results to the nearest representable value. It can even
+/// be set to [std::numeric_limits<float>::round_style](https://en.cppreference.com/w/cpp/types/numeric_limits/round_style) to synchronize
+/// the rounding mode with that of the built-in single-precision implementation (which is likely `std::round_to_nearest`, though).
+#ifndef HALF_ROUND_STYLE
+ #define HALF_ROUND_STYLE 1 // = std::round_to_nearest
+#endif
+
+/// Value signaling overflow.
+/// In correspondence with `HUGE_VAL[F|L]` from `<cmath>` this symbol expands to a positive value signaling the overflow of an
+/// operation, in particular it just evaluates to positive infinity.
+///
+/// **See also:** Documentation for [HUGE_VAL](https://en.cppreference.com/w/cpp/numeric/math/HUGE_VAL)
+#define HUGE_VALH std::numeric_limits<half_float::half>::infinity()
+
+/// Fast half-precision fma function.
+/// This symbol is defined if the fma() function generally executes as fast as, or faster than, a separate
+/// half-precision multiplication followed by an addition, which is always the case.
+///
+/// **See also:** Documentation for [FP_FAST_FMA](https://en.cppreference.com/w/cpp/numeric/math/fma)
+#define FP_FAST_FMAH 1
+
+/// Half rounding mode.
+/// In correspondence with `FLT_ROUNDS` from `<cfloat>` this symbol expands to the rounding mode used for
+/// half-precision operations. It is an alias for [HALF_ROUND_STYLE](\ref HALF_ROUND_STYLE).
+///
+/// **See also:** Documentation for [FLT_ROUNDS](https://en.cppreference.com/w/cpp/types/climits/FLT_ROUNDS)
+#define HLF_ROUNDS HALF_ROUND_STYLE
+
+#ifndef FP_ILOGB0
+ #define FP_ILOGB0 INT_MIN
+#endif
+#ifndef FP_ILOGBNAN
+ #define FP_ILOGBNAN INT_MAX
+#endif
+#ifndef FP_SUBNORMAL
+ #define FP_SUBNORMAL 0
+#endif
+#ifndef FP_ZERO
+ #define FP_ZERO 1
+#endif
+#ifndef FP_NAN
+ #define FP_NAN 2
+#endif
+#ifndef FP_INFINITE
+ #define FP_INFINITE 3
+#endif
+#ifndef FP_NORMAL
+ #define FP_NORMAL 4
+#endif
+
+#if !HALF_ENABLE_CPP11_CFENV && !defined(FE_ALL_EXCEPT)
+ #define FE_INVALID 0x10
+ #define FE_DIVBYZERO 0x08
+ #define FE_OVERFLOW 0x04
+ #define FE_UNDERFLOW 0x02
+ #define FE_INEXACT 0x01
+ #define FE_ALL_EXCEPT (FE_INVALID|FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW|FE_INEXACT)
+#endif
+
+
+/// Main namespace for half-precision functionality.
+/// This namespace contains all the functionality provided by the library.
+namespace half_float
+{
+ class half;
+
+#if HALF_ENABLE_CPP11_USER_LITERALS
+ /// Library-defined half-precision literals.
+ /// Import this namespace to enable half-precision floating-point literals:
+ /// ~~~~{.cpp}
+ /// using namespace half_float::literal;
+ /// half_float::half = 4.2_h;
+ /// ~~~~
+ namespace literal
+ {
+ half operator "" _h(long double);
+ }
+#endif
+
+ /// \internal
+ /// \brief Implementation details.
+ namespace detail
+ {
+ #if HALF_ENABLE_CPP11_TYPE_TRAITS
+ /// Conditional type.
+ template<bool B,typename T,typename F> struct conditional : std::conditional<B,T,F> {};
+
+ /// Helper for tag dispatching.
+ template<bool B> struct bool_type : std::integral_constant<bool,B> {};
+ using std::true_type;
+ using std::false_type;
+
+ /// Type traits for floating-point types.
+ template<typename T> struct is_float : std::is_floating_point<T> {};
+ #else
+ /// Conditional type.
+ template<bool,typename T,typename> struct conditional { typedef T type; };
+ template<typename T,typename F> struct conditional<false,T,F> { typedef F type; };
+
+ /// Helper for tag dispatching.
+ template<bool> struct bool_type {};
+ typedef bool_type<true> true_type;
+ typedef bool_type<false> false_type;
+
+ /// Type traits for floating-point types.
+ template<typename> struct is_float : false_type {};
+ template<typename T> struct is_float<const T> : is_float<T> {};
+ template<typename T> struct is_float<volatile T> : is_float<T> {};
+ template<typename T> struct is_float<const volatile T> : is_float<T> {};
+ template<> struct is_float<float> : true_type {};
+ template<> struct is_float<double> : true_type {};
+ template<> struct is_float<long double> : true_type {};
+ #endif
+
+ /// Type traits for floating-point bits.
+ template<typename T> struct bits { typedef unsigned char type; };
+ template<typename T> struct bits<const T> : bits<T> {};
+ template<typename T> struct bits<volatile T> : bits<T> {};
+ template<typename T> struct bits<const volatile T> : bits<T> {};
+
+ #if HALF_ENABLE_CPP11_CSTDINT
+ /// Unsigned integer of (at least) 16 bits width.
+ typedef std::uint_least16_t uint16;
+
+ /// Fastest unsigned integer of (at least) 32 bits width.
+ typedef std::uint_fast32_t uint32;
+
+ /// Fastest signed integer of (at least) 32 bits width.
+ typedef std::int_fast32_t int32;
+
+ /// Unsigned integer of (at least) 32 bits width.
+ template<> struct bits<float> { typedef std::uint_least32_t type; };
+
+ /// Unsigned integer of (at least) 64 bits width.
+ template<> struct bits<double> { typedef std::uint_least64_t type; };
+ #else
+ /// Unsigned integer of (at least) 16 bits width.
+ typedef unsigned short uint16;
+
+ /// Fastest unsigned integer of (at least) 32 bits width.
+ typedef unsigned long uint32;
+
+ /// Fastest unsigned integer of (at least) 32 bits width.
+ typedef long int32;
+
+ /// Unsigned integer of (at least) 32 bits width.
+ template<> struct bits<float> : conditional<std::numeric_limits<unsigned int>::digits>=32,unsigned int,unsigned long> {};
+
+ #if HALF_ENABLE_CPP11_LONG_LONG
+ /// Unsigned integer of (at least) 64 bits width.
+ template<> struct bits<double> : conditional<std::numeric_limits<unsigned long>::digits>=64,unsigned long,unsigned long long> {};
+ #else
+ /// Unsigned integer of (at least) 64 bits width.
+ template<> struct bits<double> { typedef unsigned long type; };
+ #endif
+ #endif
+
+ #ifdef HALF_ARITHMETIC_TYPE
+ /// Type to use for arithmetic computations and mathematic functions internally.
+ typedef HALF_ARITHMETIC_TYPE internal_t;
+ #endif
+
+ /// Tag type for binary construction.
+ struct binary_t {};
+
+ /// Tag for binary construction.
+ HALF_CONSTEXPR_CONST binary_t binary = binary_t();
+
+ /// \name Implementation defined classification and arithmetic
+ /// \{
+
+ /// Check for infinity.
+ /// \tparam T argument type (builtin floating-point type)
+ /// \param arg value to query
+ /// \retval true if infinity
+ /// \retval false else
+ template<typename T> bool builtin_isinf(T arg)
+ {
+ #if HALF_ENABLE_CPP11_CMATH
+ return std::isinf(arg);
+ #elif defined(_MSC_VER)
+ return !::_finite(static_cast<double>(arg)) && !::_isnan(static_cast<double>(arg));
+ #else
+ return arg == std::numeric_limits<T>::infinity() || arg == -std::numeric_limits<T>::infinity();
+ #endif
+ }
+
+ /// Check for NaN.
+ /// \tparam T argument type (builtin floating-point type)
+ /// \param arg value to query
+ /// \retval true if not a number
+ /// \retval false else
+ template<typename T> bool builtin_isnan(T arg)
+ {
+ #if HALF_ENABLE_CPP11_CMATH
+ return std::isnan(arg);
+ #elif defined(_MSC_VER)
+ return ::_isnan(static_cast<double>(arg)) != 0;
+ #else
+ return arg != arg;
+ #endif
+ }
+
+ /// Check sign.
+ /// \tparam T argument type (builtin floating-point type)
+ /// \param arg value to query
+ /// \retval true if signbit set
+ /// \retval false else
+ template<typename T> bool builtin_signbit(T arg)
+ {
+ #if HALF_ENABLE_CPP11_CMATH
+ return std::signbit(arg);
+ #else
+ return arg < T() || (arg == T() && T(1)/arg < T());
+ #endif
+ }
+
+ /// Platform-independent sign mask.
+ /// \param arg integer value in two's complement
+ /// \retval -1 if \a arg negative
+ /// \retval 0 if \a arg positive
+ inline uint32 sign_mask(uint32 arg)
+ {
+ static const int N = std::numeric_limits<uint32>::digits - 1;
+ #if HALF_TWOS_COMPLEMENT_INT
+ return static_cast<int32>(arg) >> N;
+ #else
+ return -((arg>>N)&1);
+ #endif
+ }
+
+ /// Platform-independent arithmetic right shift.
+ /// \param arg integer value in two's complement
+ /// \param i shift amount (at most 31)
+ /// \return \a arg right shifted for \a i bits with possible sign extension
+ inline uint32 arithmetic_shift(uint32 arg, int i)
+ {
+ #if HALF_TWOS_COMPLEMENT_INT
+ return static_cast<int32>(arg) >> i;
+ #else
+ return static_cast<int32>(arg)/(static_cast<int32>(1)<<i) - ((arg>>(std::numeric_limits<uint32>::digits-1))&1);
+ #endif
+ }
+
+ /// \}
+ /// \name Error handling
+ /// \{
+
+ /// Internal exception flags.
+ /// \return reference to global exception flags
+ inline int& errflags() { HALF_THREAD_LOCAL int flags = 0; return flags; }
+
+ /// Raise floating-point exception.
+ /// \param flags exceptions to raise
+ /// \param cond condition to raise exceptions for
+ inline void raise(int HALF_UNUSED_NOERR(flags), bool HALF_UNUSED_NOERR(cond) = true)
+ {
+ #if HALF_ERRHANDLING
+ if(!cond)
+ return;
+ #if HALF_ERRHANDLING_FLAGS
+ errflags() |= flags;
+ #endif
+ #if HALF_ERRHANDLING_ERRNO
+ if(flags & FE_INVALID)
+ errno = EDOM;
+ else if(flags & (FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW))
+ errno = ERANGE;
+ #endif
+ #if HALF_ERRHANDLING_FENV && HALF_ENABLE_CPP11_CFENV
+ std::feraiseexcept(flags);
+ #endif
+ #ifdef HALF_ERRHANDLING_THROW_INVALID
+ if(flags & FE_INVALID)
+ throw std::domain_error(HALF_ERRHANDLING_THROW_INVALID);
+ #endif
+ #ifdef HALF_ERRHANDLING_THROW_DIVBYZERO
+ if(flags & FE_DIVBYZERO)
+ throw std::domain_error(HALF_ERRHANDLING_THROW_DIVBYZERO);
+ #endif
+ #ifdef HALF_ERRHANDLING_THROW_OVERFLOW
+ if(flags & FE_OVERFLOW)
+ throw std::overflow_error(HALF_ERRHANDLING_THROW_OVERFLOW);
+ #endif
+ #ifdef HALF_ERRHANDLING_THROW_UNDERFLOW
+ if(flags & FE_UNDERFLOW)
+ throw std::underflow_error(HALF_ERRHANDLING_THROW_UNDERFLOW);
+ #endif
+ #ifdef HALF_ERRHANDLING_THROW_INEXACT
+ if(flags & FE_INEXACT)
+ throw std::range_error(HALF_ERRHANDLING_THROW_INEXACT);
+ #endif
+ #if HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT
+ if((flags & FE_UNDERFLOW) && !(flags & FE_INEXACT))
+ raise(FE_INEXACT);
+ #endif
+ #if HALF_ERRHANDLING_OVERFLOW_TO_INEXACT
+ if((flags & FE_OVERFLOW) && !(flags & FE_INEXACT))
+ raise(FE_INEXACT);
+ #endif
+ #endif
+ }
+
+ /// Check and signal for any NaN.
+ /// \param x first half-precision value to check
+ /// \param y second half-precision value to check
+ /// \retval true if either \a x or \a y is NaN
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool compsignal(unsigned int x, unsigned int y)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_INVALID, (x&0x7FFF)>0x7C00 || (y&0x7FFF)>0x7C00);
+ #endif
+ return (x&0x7FFF) > 0x7C00 || (y&0x7FFF) > 0x7C00;
+ }
+
+ /// Signal and silence signaling NaN.
+ /// \param nan half-precision NaN value
+ /// \return quiet NaN
+ /// \exception FE_INVALID if \a nan is signaling NaN
+ inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int nan)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_INVALID, !(nan&0x200));
+ #endif
+ return nan | 0x200;
+ }
+
+ /// Signal and silence signaling NaNs.
+ /// \param x first half-precision value to check
+ /// \param y second half-precision value to check
+ /// \return quiet NaN
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
+ inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int x, unsigned int y)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_INVALID, ((x&0x7FFF)>0x7C00 && !(x&0x200)) || ((y&0x7FFF)>0x7C00 && !(y&0x200)));
+ #endif
+ return ((x&0x7FFF)>0x7C00) ? (x|0x200) : (y|0x200);
+ }
+
+ /// Signal and silence signaling NaNs.
+ /// \param x first half-precision value to check
+ /// \param y second half-precision value to check
+ /// \param z third half-precision value to check
+ /// \return quiet NaN
+ /// \exception FE_INVALID if \a x, \a y or \a z is signaling NaN
+ inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int x, unsigned int y, unsigned int z)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_INVALID, ((x&0x7FFF)>0x7C00 && !(x&0x200)) || ((y&0x7FFF)>0x7C00 && !(y&0x200)) || ((z&0x7FFF)>0x7C00 && !(z&0x200)));
+ #endif
+ return ((x&0x7FFF)>0x7C00) ? (x|0x200) : ((y&0x7FFF)>0x7C00) ? (y|0x200) : (z|0x200);
+ }
+
+ /// Select value or signaling NaN.
+ /// \param x preferred half-precision value
+ /// \param y ignored half-precision value except for signaling NaN
+ /// \return \a y if signaling NaN, \a x otherwise
+ /// \exception FE_INVALID if \a y is signaling NaN
+ inline HALF_CONSTEXPR_NOERR unsigned int select(unsigned int x, unsigned int HALF_UNUSED_NOERR(y))
+ {
+ #if HALF_ERRHANDLING
+ return (((y&0x7FFF)>0x7C00) && !(y&0x200)) ? signal(y) : x;
+ #else
+ return x;
+ #endif
+ }
+
+ /// Raise domain error and return NaN.
+ /// return quiet NaN
+ /// \exception FE_INVALID
+ inline HALF_CONSTEXPR_NOERR unsigned int invalid()
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_INVALID);
+ #endif
+ return 0x7FFF;
+ }
+
+ /// Raise pole error and return infinity.
+ /// \param sign half-precision value with sign bit only
+ /// \return half-precision infinity with sign of \a sign
+ /// \exception FE_DIVBYZERO
+ inline HALF_CONSTEXPR_NOERR unsigned int pole(unsigned int sign = 0)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_DIVBYZERO);
+ #endif
+ return sign | 0x7C00;
+ }
+
+ /// Check value for underflow.
+ /// \param arg non-zero half-precision value to check
+ /// \return \a arg
+ /// \exception FE_UNDERFLOW if arg is subnormal
+ inline HALF_CONSTEXPR_NOERR unsigned int check_underflow(unsigned int arg)
+ {
+ #if HALF_ERRHANDLING && !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT
+ raise(FE_UNDERFLOW, !(arg&0x7C00));
+ #endif
+ return arg;
+ }
+
+ /// \}
+ /// \name Conversion and rounding
+ /// \{
+
+ /// Half-precision overflow.
+ /// \tparam R rounding mode to use
+ /// \param sign half-precision value with sign bit only
+ /// \return rounded overflowing half-precision value
+ /// \exception FE_OVERFLOW
+ template<std::float_round_style R> HALF_CONSTEXPR_NOERR unsigned int overflow(unsigned int sign = 0)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_OVERFLOW);
+ #endif
+ return (R==std::round_toward_infinity) ? (sign+0x7C00-(sign>>15)) :
+ (R==std::round_toward_neg_infinity) ? (sign+0x7BFF+(sign>>15)) :
+ (R==std::round_toward_zero) ? (sign|0x7BFF) :
+ (sign|0x7C00);
+ }
+
+ /// Half-precision underflow.
+ /// \tparam R rounding mode to use
+ /// \param sign half-precision value with sign bit only
+ /// \return rounded underflowing half-precision value
+ /// \exception FE_UNDERFLOW
+ template<std::float_round_style R> HALF_CONSTEXPR_NOERR unsigned int underflow(unsigned int sign = 0)
+ {
+ #if HALF_ERRHANDLING
+ raise(FE_UNDERFLOW);
+ #endif
+ return (R==std::round_toward_infinity) ? (sign+1-(sign>>15)) :
+ (R==std::round_toward_neg_infinity) ? (sign+(sign>>15)) :
+ sign;
+ }
+
+ /// Round half-precision number.
+ /// \tparam R rounding mode to use
+ /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results
+ /// \param value finite half-precision number to round
+ /// \param g guard bit (most significant discarded bit)
+ /// \param s sticky bit (or of all but the most significant discarded bits)
+ /// \return rounded half-precision value
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded or \a I is `true`
+ template<std::float_round_style R,bool I> HALF_CONSTEXPR_NOERR unsigned int rounded(unsigned int value, int g, int s)
+ {
+ #if HALF_ERRHANDLING
+ value += (R==std::round_to_nearest) ? (g&(s|value)) :
+ (R==std::round_toward_infinity) ? (~(value>>15)&(g|s)) :
+ (R==std::round_toward_neg_infinity) ? ((value>>15)&(g|s)) : 0;
+ if((value&0x7C00) == 0x7C00)
+ raise(FE_OVERFLOW);
+ else if(value & 0x7C00)
+ raise(FE_INEXACT, I || (g|s)!=0);
+ else
+ raise(FE_UNDERFLOW, !(HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT) || I || (g|s)!=0);
+ return value;
+ #else
+ return (R==std::round_to_nearest) ? (value+(g&(s|value))) :
+ (R==std::round_toward_infinity) ? (value+(~(value>>15)&(g|s))) :
+ (R==std::round_toward_neg_infinity) ? (value+((value>>15)&(g|s))) :
+ value;
+ #endif
+ }
+
+ /// Round half-precision number to nearest integer value.
+ /// \tparam R rounding mode to use
+ /// \tparam E `true` for round to even, `false` for round away from zero
+ /// \tparam I `true` to raise INEXACT exception (if inexact), `false` to never raise it
+ /// \param value half-precision value to round
+ /// \return half-precision bits for nearest integral value
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT if value had to be rounded and \a I is `true`
+ template<std::float_round_style R,bool E,bool I> unsigned int integral(unsigned int value)
+ {
+ unsigned int abs = value & 0x7FFF;
+ if(abs < 0x3C00)
+ {
+ raise(FE_INEXACT, I);
+ return ((R==std::round_to_nearest) ? (0x3C00&-static_cast<unsigned>(abs>=(0x3800+E))) :
+ (R==std::round_toward_infinity) ? (0x3C00&-(~(value>>15)&(abs!=0))) :
+ (R==std::round_toward_neg_infinity) ? (0x3C00&-static_cast<unsigned>(value>0x8000)) :
+ 0) | (value&0x8000);
+ }
+ if(abs >= 0x6400)
+ return (abs>0x7C00) ? signal(value) : value;
+ unsigned int exp = 25 - (abs>>10), mask = (1<<exp) - 1;
+ raise(FE_INEXACT, I && (value&mask));
+ return (( (R==std::round_to_nearest) ? ((1<<(exp-1))-(~(value>>exp)&E)) :
+ (R==std::round_toward_infinity) ? (mask&((value>>15)-1)) :
+ (R==std::round_toward_neg_infinity) ? (mask&-(value>>15)) :
+ 0) + value) & ~mask;
+ }
+
+ /// Convert fixed point to half-precision floating-point.
+ /// \tparam R rounding mode to use
+ /// \tparam F number of fractional bits in [11,31]
+ /// \tparam S `true` for signed, `false` for unsigned
+ /// \tparam N `true` for additional normalization step, `false` if already normalized to 1.F
+ /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results
+ /// \param m mantissa in Q1.F fixed point format
+ /// \param exp biased exponent - 1
+ /// \param sign half-precision value with sign bit only
+ /// \param s sticky bit (or of all but the most significant already discarded bits)
+ /// \return value converted to half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded or \a I is `true`
+ template<std::float_round_style R,unsigned int F,bool S,bool N,bool I> unsigned int fixed2half(uint32 m, int exp = 14, unsigned int sign = 0, int s = 0)
+ {
+ if(S)
+ {
+ uint32 msign = sign_mask(m);
+ m = (m^msign) - msign;
+ sign = msign & 0x8000;
+ }
+ if(N)
+ for(; m<(static_cast<uint32>(1)<<F) && exp; m<<=1,--exp) ;
+ else if(exp < 0)
+ return rounded<R,I>(sign+(m>>(F-10-exp)), (m>>(F-11-exp))&1, s|((m&((static_cast<uint32>(1)<<(F-11-exp))-1))!=0));
+ return rounded<R,I>(sign+(exp<<10)+(m>>(F-10)), (m>>(F-11))&1, s|((m&((static_cast<uint32>(1)<<(F-11))-1))!=0));
+ }
+
+ /// Convert IEEE single-precision to half-precision.
+ /// Credit for this goes to [Jeroen van der Zijp](ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf).
+ /// \tparam R rounding mode to use
+ /// \param value single-precision value to convert
+ /// \return rounded half-precision value
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded
+ template<std::float_round_style R> unsigned int float2half_impl(float value, true_type)
+ {
+ #if HALF_ENABLE_F16C_INTRINSICS
+ return _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(value),
+ (R==std::round_to_nearest) ? _MM_FROUND_TO_NEAREST_INT :
+ (R==std::round_toward_zero) ? _MM_FROUND_TO_ZERO :
+ (R==std::round_toward_infinity) ? _MM_FROUND_TO_POS_INF :
+ (R==std::round_toward_neg_infinity) ? _MM_FROUND_TO_NEG_INF :
+ _MM_FROUND_CUR_DIRECTION));
+ #else
+ bits<float>::type fbits;
+ std::memcpy(&fbits, &value, sizeof(float));
+ #if 1
+ unsigned int sign = (fbits>>16) & 0x8000;
+ fbits &= 0x7FFFFFFF;
+ if(fbits >= 0x7F800000)
+ return sign | 0x7C00 | ((fbits>0x7F800000) ? (0x200|((fbits>>13)&0x3FF)) : 0);
+ if(fbits >= 0x47800000)
+ return overflow<R>(sign);
+ if(fbits >= 0x38800000)
+ return rounded<R,false>(sign|(((fbits>>23)-112)<<10)|((fbits>>13)&0x3FF), (fbits>>12)&1, (fbits&0xFFF)!=0);
+ if(fbits >= 0x33000000)
+ {
+ int i = 125 - (fbits>>23);
+ fbits = (fbits&0x7FFFFF) | 0x800000;
+ return rounded<R,false>(sign|(fbits>>(i+1)), (fbits>>i)&1, (fbits&((static_cast<uint32>(1)<<i)-1))!=0);
+ }
+ if(fbits != 0)
+ return underflow<R>(sign);
+ return sign;
+ #else
+ static const uint16 base_table[512] = {
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+ 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080, 0x0100,
+ 0x0200, 0x0400, 0x0800, 0x0C00, 0x1000, 0x1400, 0x1800, 0x1C00, 0x2000, 0x2400, 0x2800, 0x2C00, 0x3000, 0x3400, 0x3800, 0x3C00,
+ 0x4000, 0x4400, 0x4800, 0x4C00, 0x5000, 0x5400, 0x5800, 0x5C00, 0x6000, 0x6400, 0x6800, 0x6C00, 0x7000, 0x7400, 0x7800, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF,
+ 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7C00,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+ 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008, 0x8010, 0x8020, 0x8040, 0x8080, 0x8100,
+ 0x8200, 0x8400, 0x8800, 0x8C00, 0x9000, 0x9400, 0x9800, 0x9C00, 0xA000, 0xA400, 0xA800, 0xAC00, 0xB000, 0xB400, 0xB800, 0xBC00,
+ 0xC000, 0xC400, 0xC800, 0xCC00, 0xD000, 0xD400, 0xD800, 0xDC00, 0xE000, 0xE400, 0xE800, 0xEC00, 0xF000, 0xF400, 0xF800, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF,
+ 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFC00 };
+ static const unsigned char shift_table[256] = {
+ 24, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+ 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+ 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+ 25, 25, 25, 25, 25, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
+ 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+ 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+ 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+ 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 13 };
+ int sexp = fbits >> 23, exp = sexp & 0xFF, i = shift_table[exp];
+ fbits &= 0x7FFFFF;
+ uint32 m = (fbits|((exp!=0)<<23)) & -static_cast<uint32>(exp!=0xFF);
+ return rounded<R,false>(base_table[sexp]+(fbits>>i), (m>>(i-1))&1, (((static_cast<uint32>(1)<<(i-1))-1)&m)!=0);
+ #endif
+ #endif
+ }
+
+ /// Convert IEEE double-precision to half-precision.
+ /// \tparam R rounding mode to use
+ /// \param value double-precision value to convert
+ /// \return rounded half-precision value
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded
+ template<std::float_round_style R> unsigned int float2half_impl(double value, true_type)
+ {
+ #if HALF_ENABLE_F16C_INTRINSICS
+ if(R == std::round_indeterminate)
+ return _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_cvtpd_ps(_mm_set_sd(value)), _MM_FROUND_CUR_DIRECTION));
+ #endif
+ bits<double>::type dbits;
+ std::memcpy(&dbits, &value, sizeof(double));
+ uint32 hi = dbits >> 32, lo = dbits & 0xFFFFFFFF;
+ unsigned int sign = (hi>>16) & 0x8000;
+ hi &= 0x7FFFFFFF;
+ if(hi >= 0x7FF00000)
+ return sign | 0x7C00 | ((dbits&0xFFFFFFFFFFFFF) ? (0x200|((hi>>10)&0x3FF)) : 0);
+ if(hi >= 0x40F00000)
+ return overflow<R>(sign);
+ if(hi >= 0x3F100000)
+ return rounded<R,false>(sign|(((hi>>20)-1008)<<10)|((hi>>10)&0x3FF), (hi>>9)&1, ((hi&0x1FF)|lo)!=0);
+ if(hi >= 0x3E600000)
+ {
+ int i = 1018 - (hi>>20);
+ hi = (hi&0xFFFFF) | 0x100000;
+ return rounded<R,false>(sign|(hi>>(i+1)), (hi>>i)&1, ((hi&((static_cast<uint32>(1)<<i)-1))|lo)!=0);
+ }
+ if((hi|lo) != 0)
+ return underflow<R>(sign);
+ return sign;
+ }
+
+ /// Convert non-IEEE floating-point to half-precision.
+ /// \tparam R rounding mode to use
+ /// \tparam T source type (builtin floating-point type)
+ /// \param value floating-point value to convert
+ /// \return rounded half-precision value
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded
+ template<std::float_round_style R,typename T> unsigned int float2half_impl(T value, ...)
+ {
+ unsigned int hbits = static_cast<unsigned>(builtin_signbit(value)) << 15;
+ if(value == T())
+ return hbits;
+ if(builtin_isnan(value))
+ return hbits | 0x7FFF;
+ if(builtin_isinf(value))
+ return hbits | 0x7C00;
+ int exp;
+ std::frexp(value, &exp);
+ if(exp > 16)
+ return overflow<R>(hbits);
+ if(exp < -13)
+ value = std::ldexp(value, 25);
+ else
+ {
+ value = std::ldexp(value, 12-exp);
+ hbits |= ((exp+13)<<10);
+ }
+ T ival, frac = std::modf(value, &ival);
+ int m = std::abs(static_cast<int>(ival));
+ return rounded<R,false>(hbits+(m>>1), m&1, frac!=T());
+ }
+
+ /// Convert floating-point to half-precision.
+ /// \tparam R rounding mode to use
+ /// \tparam T source type (builtin floating-point type)
+ /// \param value floating-point value to convert
+ /// \return rounded half-precision value
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded
+ template<std::float_round_style R,typename T> unsigned int float2half(T value)
+ {
+ return float2half_impl<R>(value, bool_type<std::numeric_limits<T>::is_iec559&&sizeof(typename bits<T>::type)==sizeof(T)>());
+ }
+
+ /// Convert integer to half-precision floating-point.
+ /// \tparam R rounding mode to use
+ /// \tparam T type to convert (builtin integer type)
+ /// \param value integral value to convert
+ /// \return rounded half-precision value
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_INEXACT if value had to be rounded
+ template<std::float_round_style R,typename T> unsigned int int2half(T value)
+ {
+ unsigned int bits = static_cast<unsigned>(value<0) << 15;
+ if(!value)
+ return bits;
+ if(bits)
+ value = -value;
+ if(value > 0xFFFF)
+ return overflow<R>(bits);
+ unsigned int m = static_cast<unsigned int>(value), exp = 24;
+ for(; m<0x400; m<<=1,--exp) ;
+ for(; m>0x7FF; m>>=1,++exp) ;
+ bits |= (exp<<10) + m;
+ return (exp>24) ? rounded<R,false>(bits, (value>>(exp-25))&1, (((1<<(exp-25))-1)&value)!=0) : bits;
+ }
+
+ /// Convert half-precision to IEEE single-precision.
+ /// Credit for this goes to [Jeroen van der Zijp](ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf).
+ /// \param value half-precision value to convert
+ /// \return single-precision value
+ inline float half2float_impl(unsigned int value, float, true_type)
+ {
+ #if HALF_ENABLE_F16C_INTRINSICS
+ return _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(value)));
+ #else
+ #if 0
+ bits<float>::type fbits = static_cast<bits<float>::type>(value&0x8000) << 16;
+ int abs = value & 0x7FFF;
+ if(abs)
+ {
+ fbits |= 0x38000000 << static_cast<unsigned>(abs>=0x7C00);
+ for(; abs<0x400; abs<<=1,fbits-=0x800000) ;
+ fbits += static_cast<bits<float>::type>(abs) << 13;
+ }
+ #else
+ static const bits<float>::type mantissa_table[2048] = {
+ 0x00000000, 0x33800000, 0x34000000, 0x34400000, 0x34800000, 0x34A00000, 0x34C00000, 0x34E00000, 0x35000000, 0x35100000, 0x35200000, 0x35300000, 0x35400000, 0x35500000, 0x35600000, 0x35700000,
+ 0x35800000, 0x35880000, 0x35900000, 0x35980000, 0x35A00000, 0x35A80000, 0x35B00000, 0x35B80000, 0x35C00000, 0x35C80000, 0x35D00000, 0x35D80000, 0x35E00000, 0x35E80000, 0x35F00000, 0x35F80000,
+ 0x36000000, 0x36040000, 0x36080000, 0x360C0000, 0x36100000, 0x36140000, 0x36180000, 0x361C0000, 0x36200000, 0x36240000, 0x36280000, 0x362C0000, 0x36300000, 0x36340000, 0x36380000, 0x363C0000,
+ 0x36400000, 0x36440000, 0x36480000, 0x364C0000, 0x36500000, 0x36540000, 0x36580000, 0x365C0000, 0x36600000, 0x36640000, 0x36680000, 0x366C0000, 0x36700000, 0x36740000, 0x36780000, 0x367C0000,
+ 0x36800000, 0x36820000, 0x36840000, 0x36860000, 0x36880000, 0x368A0000, 0x368C0000, 0x368E0000, 0x36900000, 0x36920000, 0x36940000, 0x36960000, 0x36980000, 0x369A0000, 0x369C0000, 0x369E0000,
+ 0x36A00000, 0x36A20000, 0x36A40000, 0x36A60000, 0x36A80000, 0x36AA0000, 0x36AC0000, 0x36AE0000, 0x36B00000, 0x36B20000, 0x36B40000, 0x36B60000, 0x36B80000, 0x36BA0000, 0x36BC0000, 0x36BE0000,
+ 0x36C00000, 0x36C20000, 0x36C40000, 0x36C60000, 0x36C80000, 0x36CA0000, 0x36CC0000, 0x36CE0000, 0x36D00000, 0x36D20000, 0x36D40000, 0x36D60000, 0x36D80000, 0x36DA0000, 0x36DC0000, 0x36DE0000,
+ 0x36E00000, 0x36E20000, 0x36E40000, 0x36E60000, 0x36E80000, 0x36EA0000, 0x36EC0000, 0x36EE0000, 0x36F00000, 0x36F20000, 0x36F40000, 0x36F60000, 0x36F80000, 0x36FA0000, 0x36FC0000, 0x36FE0000,
+ 0x37000000, 0x37010000, 0x37020000, 0x37030000, 0x37040000, 0x37050000, 0x37060000, 0x37070000, 0x37080000, 0x37090000, 0x370A0000, 0x370B0000, 0x370C0000, 0x370D0000, 0x370E0000, 0x370F0000,
+ 0x37100000, 0x37110000, 0x37120000, 0x37130000, 0x37140000, 0x37150000, 0x37160000, 0x37170000, 0x37180000, 0x37190000, 0x371A0000, 0x371B0000, 0x371C0000, 0x371D0000, 0x371E0000, 0x371F0000,
+ 0x37200000, 0x37210000, 0x37220000, 0x37230000, 0x37240000, 0x37250000, 0x37260000, 0x37270000, 0x37280000, 0x37290000, 0x372A0000, 0x372B0000, 0x372C0000, 0x372D0000, 0x372E0000, 0x372F0000,
+ 0x37300000, 0x37310000, 0x37320000, 0x37330000, 0x37340000, 0x37350000, 0x37360000, 0x37370000, 0x37380000, 0x37390000, 0x373A0000, 0x373B0000, 0x373C0000, 0x373D0000, 0x373E0000, 0x373F0000,
+ 0x37400000, 0x37410000, 0x37420000, 0x37430000, 0x37440000, 0x37450000, 0x37460000, 0x37470000, 0x37480000, 0x37490000, 0x374A0000, 0x374B0000, 0x374C0000, 0x374D0000, 0x374E0000, 0x374F0000,
+ 0x37500000, 0x37510000, 0x37520000, 0x37530000, 0x37540000, 0x37550000, 0x37560000, 0x37570000, 0x37580000, 0x37590000, 0x375A0000, 0x375B0000, 0x375C0000, 0x375D0000, 0x375E0000, 0x375F0000,
+ 0x37600000, 0x37610000, 0x37620000, 0x37630000, 0x37640000, 0x37650000, 0x37660000, 0x37670000, 0x37680000, 0x37690000, 0x376A0000, 0x376B0000, 0x376C0000, 0x376D0000, 0x376E0000, 0x376F0000,
+ 0x37700000, 0x37710000, 0x37720000, 0x37730000, 0x37740000, 0x37750000, 0x37760000, 0x37770000, 0x37780000, 0x37790000, 0x377A0000, 0x377B0000, 0x377C0000, 0x377D0000, 0x377E0000, 0x377F0000,
+ 0x37800000, 0x37808000, 0x37810000, 0x37818000, 0x37820000, 0x37828000, 0x37830000, 0x37838000, 0x37840000, 0x37848000, 0x37850000, 0x37858000, 0x37860000, 0x37868000, 0x37870000, 0x37878000,
+ 0x37880000, 0x37888000, 0x37890000, 0x37898000, 0x378A0000, 0x378A8000, 0x378B0000, 0x378B8000, 0x378C0000, 0x378C8000, 0x378D0000, 0x378D8000, 0x378E0000, 0x378E8000, 0x378F0000, 0x378F8000,
+ 0x37900000, 0x37908000, 0x37910000, 0x37918000, 0x37920000, 0x37928000, 0x37930000, 0x37938000, 0x37940000, 0x37948000, 0x37950000, 0x37958000, 0x37960000, 0x37968000, 0x37970000, 0x37978000,
+ 0x37980000, 0x37988000, 0x37990000, 0x37998000, 0x379A0000, 0x379A8000, 0x379B0000, 0x379B8000, 0x379C0000, 0x379C8000, 0x379D0000, 0x379D8000, 0x379E0000, 0x379E8000, 0x379F0000, 0x379F8000,
+ 0x37A00000, 0x37A08000, 0x37A10000, 0x37A18000, 0x37A20000, 0x37A28000, 0x37A30000, 0x37A38000, 0x37A40000, 0x37A48000, 0x37A50000, 0x37A58000, 0x37A60000, 0x37A68000, 0x37A70000, 0x37A78000,
+ 0x37A80000, 0x37A88000, 0x37A90000, 0x37A98000, 0x37AA0000, 0x37AA8000, 0x37AB0000, 0x37AB8000, 0x37AC0000, 0x37AC8000, 0x37AD0000, 0x37AD8000, 0x37AE0000, 0x37AE8000, 0x37AF0000, 0x37AF8000,
+ 0x37B00000, 0x37B08000, 0x37B10000, 0x37B18000, 0x37B20000, 0x37B28000, 0x37B30000, 0x37B38000, 0x37B40000, 0x37B48000, 0x37B50000, 0x37B58000, 0x37B60000, 0x37B68000, 0x37B70000, 0x37B78000,
+ 0x37B80000, 0x37B88000, 0x37B90000, 0x37B98000, 0x37BA0000, 0x37BA8000, 0x37BB0000, 0x37BB8000, 0x37BC0000, 0x37BC8000, 0x37BD0000, 0x37BD8000, 0x37BE0000, 0x37BE8000, 0x37BF0000, 0x37BF8000,
+ 0x37C00000, 0x37C08000, 0x37C10000, 0x37C18000, 0x37C20000, 0x37C28000, 0x37C30000, 0x37C38000, 0x37C40000, 0x37C48000, 0x37C50000, 0x37C58000, 0x37C60000, 0x37C68000, 0x37C70000, 0x37C78000,
+ 0x37C80000, 0x37C88000, 0x37C90000, 0x37C98000, 0x37CA0000, 0x37CA8000, 0x37CB0000, 0x37CB8000, 0x37CC0000, 0x37CC8000, 0x37CD0000, 0x37CD8000, 0x37CE0000, 0x37CE8000, 0x37CF0000, 0x37CF8000,
+ 0x37D00000, 0x37D08000, 0x37D10000, 0x37D18000, 0x37D20000, 0x37D28000, 0x37D30000, 0x37D38000, 0x37D40000, 0x37D48000, 0x37D50000, 0x37D58000, 0x37D60000, 0x37D68000, 0x37D70000, 0x37D78000,
+ 0x37D80000, 0x37D88000, 0x37D90000, 0x37D98000, 0x37DA0000, 0x37DA8000, 0x37DB0000, 0x37DB8000, 0x37DC0000, 0x37DC8000, 0x37DD0000, 0x37DD8000, 0x37DE0000, 0x37DE8000, 0x37DF0000, 0x37DF8000,
+ 0x37E00000, 0x37E08000, 0x37E10000, 0x37E18000, 0x37E20000, 0x37E28000, 0x37E30000, 0x37E38000, 0x37E40000, 0x37E48000, 0x37E50000, 0x37E58000, 0x37E60000, 0x37E68000, 0x37E70000, 0x37E78000,
+ 0x37E80000, 0x37E88000, 0x37E90000, 0x37E98000, 0x37EA0000, 0x37EA8000, 0x37EB0000, 0x37EB8000, 0x37EC0000, 0x37EC8000, 0x37ED0000, 0x37ED8000, 0x37EE0000, 0x37EE8000, 0x37EF0000, 0x37EF8000,
+ 0x37F00000, 0x37F08000, 0x37F10000, 0x37F18000, 0x37F20000, 0x37F28000, 0x37F30000, 0x37F38000, 0x37F40000, 0x37F48000, 0x37F50000, 0x37F58000, 0x37F60000, 0x37F68000, 0x37F70000, 0x37F78000,
+ 0x37F80000, 0x37F88000, 0x37F90000, 0x37F98000, 0x37FA0000, 0x37FA8000, 0x37FB0000, 0x37FB8000, 0x37FC0000, 0x37FC8000, 0x37FD0000, 0x37FD8000, 0x37FE0000, 0x37FE8000, 0x37FF0000, 0x37FF8000,
+ 0x38000000, 0x38004000, 0x38008000, 0x3800C000, 0x38010000, 0x38014000, 0x38018000, 0x3801C000, 0x38020000, 0x38024000, 0x38028000, 0x3802C000, 0x38030000, 0x38034000, 0x38038000, 0x3803C000,
+ 0x38040000, 0x38044000, 0x38048000, 0x3804C000, 0x38050000, 0x38054000, 0x38058000, 0x3805C000, 0x38060000, 0x38064000, 0x38068000, 0x3806C000, 0x38070000, 0x38074000, 0x38078000, 0x3807C000,
+ 0x38080000, 0x38084000, 0x38088000, 0x3808C000, 0x38090000, 0x38094000, 0x38098000, 0x3809C000, 0x380A0000, 0x380A4000, 0x380A8000, 0x380AC000, 0x380B0000, 0x380B4000, 0x380B8000, 0x380BC000,
+ 0x380C0000, 0x380C4000, 0x380C8000, 0x380CC000, 0x380D0000, 0x380D4000, 0x380D8000, 0x380DC000, 0x380E0000, 0x380E4000, 0x380E8000, 0x380EC000, 0x380F0000, 0x380F4000, 0x380F8000, 0x380FC000,
+ 0x38100000, 0x38104000, 0x38108000, 0x3810C000, 0x38110000, 0x38114000, 0x38118000, 0x3811C000, 0x38120000, 0x38124000, 0x38128000, 0x3812C000, 0x38130000, 0x38134000, 0x38138000, 0x3813C000,
+ 0x38140000, 0x38144000, 0x38148000, 0x3814C000, 0x38150000, 0x38154000, 0x38158000, 0x3815C000, 0x38160000, 0x38164000, 0x38168000, 0x3816C000, 0x38170000, 0x38174000, 0x38178000, 0x3817C000,
+ 0x38180000, 0x38184000, 0x38188000, 0x3818C000, 0x38190000, 0x38194000, 0x38198000, 0x3819C000, 0x381A0000, 0x381A4000, 0x381A8000, 0x381AC000, 0x381B0000, 0x381B4000, 0x381B8000, 0x381BC000,
+ 0x381C0000, 0x381C4000, 0x381C8000, 0x381CC000, 0x381D0000, 0x381D4000, 0x381D8000, 0x381DC000, 0x381E0000, 0x381E4000, 0x381E8000, 0x381EC000, 0x381F0000, 0x381F4000, 0x381F8000, 0x381FC000,
+ 0x38200000, 0x38204000, 0x38208000, 0x3820C000, 0x38210000, 0x38214000, 0x38218000, 0x3821C000, 0x38220000, 0x38224000, 0x38228000, 0x3822C000, 0x38230000, 0x38234000, 0x38238000, 0x3823C000,
+ 0x38240000, 0x38244000, 0x38248000, 0x3824C000, 0x38250000, 0x38254000, 0x38258000, 0x3825C000, 0x38260000, 0x38264000, 0x38268000, 0x3826C000, 0x38270000, 0x38274000, 0x38278000, 0x3827C000,
+ 0x38280000, 0x38284000, 0x38288000, 0x3828C000, 0x38290000, 0x38294000, 0x38298000, 0x3829C000, 0x382A0000, 0x382A4000, 0x382A8000, 0x382AC000, 0x382B0000, 0x382B4000, 0x382B8000, 0x382BC000,
+ 0x382C0000, 0x382C4000, 0x382C8000, 0x382CC000, 0x382D0000, 0x382D4000, 0x382D8000, 0x382DC000, 0x382E0000, 0x382E4000, 0x382E8000, 0x382EC000, 0x382F0000, 0x382F4000, 0x382F8000, 0x382FC000,
+ 0x38300000, 0x38304000, 0x38308000, 0x3830C000, 0x38310000, 0x38314000, 0x38318000, 0x3831C000, 0x38320000, 0x38324000, 0x38328000, 0x3832C000, 0x38330000, 0x38334000, 0x38338000, 0x3833C000,
+ 0x38340000, 0x38344000, 0x38348000, 0x3834C000, 0x38350000, 0x38354000, 0x38358000, 0x3835C000, 0x38360000, 0x38364000, 0x38368000, 0x3836C000, 0x38370000, 0x38374000, 0x38378000, 0x3837C000,
+ 0x38380000, 0x38384000, 0x38388000, 0x3838C000, 0x38390000, 0x38394000, 0x38398000, 0x3839C000, 0x383A0000, 0x383A4000, 0x383A8000, 0x383AC000, 0x383B0000, 0x383B4000, 0x383B8000, 0x383BC000,
+ 0x383C0000, 0x383C4000, 0x383C8000, 0x383CC000, 0x383D0000, 0x383D4000, 0x383D8000, 0x383DC000, 0x383E0000, 0x383E4000, 0x383E8000, 0x383EC000, 0x383F0000, 0x383F4000, 0x383F8000, 0x383FC000,
+ 0x38400000, 0x38404000, 0x38408000, 0x3840C000, 0x38410000, 0x38414000, 0x38418000, 0x3841C000, 0x38420000, 0x38424000, 0x38428000, 0x3842C000, 0x38430000, 0x38434000, 0x38438000, 0x3843C000,
+ 0x38440000, 0x38444000, 0x38448000, 0x3844C000, 0x38450000, 0x38454000, 0x38458000, 0x3845C000, 0x38460000, 0x38464000, 0x38468000, 0x3846C000, 0x38470000, 0x38474000, 0x38478000, 0x3847C000,
+ 0x38480000, 0x38484000, 0x38488000, 0x3848C000, 0x38490000, 0x38494000, 0x38498000, 0x3849C000, 0x384A0000, 0x384A4000, 0x384A8000, 0x384AC000, 0x384B0000, 0x384B4000, 0x384B8000, 0x384BC000,
+ 0x384C0000, 0x384C4000, 0x384C8000, 0x384CC000, 0x384D0000, 0x384D4000, 0x384D8000, 0x384DC000, 0x384E0000, 0x384E4000, 0x384E8000, 0x384EC000, 0x384F0000, 0x384F4000, 0x384F8000, 0x384FC000,
+ 0x38500000, 0x38504000, 0x38508000, 0x3850C000, 0x38510000, 0x38514000, 0x38518000, 0x3851C000, 0x38520000, 0x38524000, 0x38528000, 0x3852C000, 0x38530000, 0x38534000, 0x38538000, 0x3853C000,
+ 0x38540000, 0x38544000, 0x38548000, 0x3854C000, 0x38550000, 0x38554000, 0x38558000, 0x3855C000, 0x38560000, 0x38564000, 0x38568000, 0x3856C000, 0x38570000, 0x38574000, 0x38578000, 0x3857C000,
+ 0x38580000, 0x38584000, 0x38588000, 0x3858C000, 0x38590000, 0x38594000, 0x38598000, 0x3859C000, 0x385A0000, 0x385A4000, 0x385A8000, 0x385AC000, 0x385B0000, 0x385B4000, 0x385B8000, 0x385BC000,
+ 0x385C0000, 0x385C4000, 0x385C8000, 0x385CC000, 0x385D0000, 0x385D4000, 0x385D8000, 0x385DC000, 0x385E0000, 0x385E4000, 0x385E8000, 0x385EC000, 0x385F0000, 0x385F4000, 0x385F8000, 0x385FC000,
+ 0x38600000, 0x38604000, 0x38608000, 0x3860C000, 0x38610000, 0x38614000, 0x38618000, 0x3861C000, 0x38620000, 0x38624000, 0x38628000, 0x3862C000, 0x38630000, 0x38634000, 0x38638000, 0x3863C000,
+ 0x38640000, 0x38644000, 0x38648000, 0x3864C000, 0x38650000, 0x38654000, 0x38658000, 0x3865C000, 0x38660000, 0x38664000, 0x38668000, 0x3866C000, 0x38670000, 0x38674000, 0x38678000, 0x3867C000,
+ 0x38680000, 0x38684000, 0x38688000, 0x3868C000, 0x38690000, 0x38694000, 0x38698000, 0x3869C000, 0x386A0000, 0x386A4000, 0x386A8000, 0x386AC000, 0x386B0000, 0x386B4000, 0x386B8000, 0x386BC000,
+ 0x386C0000, 0x386C4000, 0x386C8000, 0x386CC000, 0x386D0000, 0x386D4000, 0x386D8000, 0x386DC000, 0x386E0000, 0x386E4000, 0x386E8000, 0x386EC000, 0x386F0000, 0x386F4000, 0x386F8000, 0x386FC000,
+ 0x38700000, 0x38704000, 0x38708000, 0x3870C000, 0x38710000, 0x38714000, 0x38718000, 0x3871C000, 0x38720000, 0x38724000, 0x38728000, 0x3872C000, 0x38730000, 0x38734000, 0x38738000, 0x3873C000,
+ 0x38740000, 0x38744000, 0x38748000, 0x3874C000, 0x38750000, 0x38754000, 0x38758000, 0x3875C000, 0x38760000, 0x38764000, 0x38768000, 0x3876C000, 0x38770000, 0x38774000, 0x38778000, 0x3877C000,
+ 0x38780000, 0x38784000, 0x38788000, 0x3878C000, 0x38790000, 0x38794000, 0x38798000, 0x3879C000, 0x387A0000, 0x387A4000, 0x387A8000, 0x387AC000, 0x387B0000, 0x387B4000, 0x387B8000, 0x387BC000,
+ 0x387C0000, 0x387C4000, 0x387C8000, 0x387CC000, 0x387D0000, 0x387D4000, 0x387D8000, 0x387DC000, 0x387E0000, 0x387E4000, 0x387E8000, 0x387EC000, 0x387F0000, 0x387F4000, 0x387F8000, 0x387FC000,
+ 0x38000000, 0x38002000, 0x38004000, 0x38006000, 0x38008000, 0x3800A000, 0x3800C000, 0x3800E000, 0x38010000, 0x38012000, 0x38014000, 0x38016000, 0x38018000, 0x3801A000, 0x3801C000, 0x3801E000,
+ 0x38020000, 0x38022000, 0x38024000, 0x38026000, 0x38028000, 0x3802A000, 0x3802C000, 0x3802E000, 0x38030000, 0x38032000, 0x38034000, 0x38036000, 0x38038000, 0x3803A000, 0x3803C000, 0x3803E000,
+ 0x38040000, 0x38042000, 0x38044000, 0x38046000, 0x38048000, 0x3804A000, 0x3804C000, 0x3804E000, 0x38050000, 0x38052000, 0x38054000, 0x38056000, 0x38058000, 0x3805A000, 0x3805C000, 0x3805E000,
+ 0x38060000, 0x38062000, 0x38064000, 0x38066000, 0x38068000, 0x3806A000, 0x3806C000, 0x3806E000, 0x38070000, 0x38072000, 0x38074000, 0x38076000, 0x38078000, 0x3807A000, 0x3807C000, 0x3807E000,
+ 0x38080000, 0x38082000, 0x38084000, 0x38086000, 0x38088000, 0x3808A000, 0x3808C000, 0x3808E000, 0x38090000, 0x38092000, 0x38094000, 0x38096000, 0x38098000, 0x3809A000, 0x3809C000, 0x3809E000,
+ 0x380A0000, 0x380A2000, 0x380A4000, 0x380A6000, 0x380A8000, 0x380AA000, 0x380AC000, 0x380AE000, 0x380B0000, 0x380B2000, 0x380B4000, 0x380B6000, 0x380B8000, 0x380BA000, 0x380BC000, 0x380BE000,
+ 0x380C0000, 0x380C2000, 0x380C4000, 0x380C6000, 0x380C8000, 0x380CA000, 0x380CC000, 0x380CE000, 0x380D0000, 0x380D2000, 0x380D4000, 0x380D6000, 0x380D8000, 0x380DA000, 0x380DC000, 0x380DE000,
+ 0x380E0000, 0x380E2000, 0x380E4000, 0x380E6000, 0x380E8000, 0x380EA000, 0x380EC000, 0x380EE000, 0x380F0000, 0x380F2000, 0x380F4000, 0x380F6000, 0x380F8000, 0x380FA000, 0x380FC000, 0x380FE000,
+ 0x38100000, 0x38102000, 0x38104000, 0x38106000, 0x38108000, 0x3810A000, 0x3810C000, 0x3810E000, 0x38110000, 0x38112000, 0x38114000, 0x38116000, 0x38118000, 0x3811A000, 0x3811C000, 0x3811E000,
+ 0x38120000, 0x38122000, 0x38124000, 0x38126000, 0x38128000, 0x3812A000, 0x3812C000, 0x3812E000, 0x38130000, 0x38132000, 0x38134000, 0x38136000, 0x38138000, 0x3813A000, 0x3813C000, 0x3813E000,
+ 0x38140000, 0x38142000, 0x38144000, 0x38146000, 0x38148000, 0x3814A000, 0x3814C000, 0x3814E000, 0x38150000, 0x38152000, 0x38154000, 0x38156000, 0x38158000, 0x3815A000, 0x3815C000, 0x3815E000,
+ 0x38160000, 0x38162000, 0x38164000, 0x38166000, 0x38168000, 0x3816A000, 0x3816C000, 0x3816E000, 0x38170000, 0x38172000, 0x38174000, 0x38176000, 0x38178000, 0x3817A000, 0x3817C000, 0x3817E000,
+ 0x38180000, 0x38182000, 0x38184000, 0x38186000, 0x38188000, 0x3818A000, 0x3818C000, 0x3818E000, 0x38190000, 0x38192000, 0x38194000, 0x38196000, 0x38198000, 0x3819A000, 0x3819C000, 0x3819E000,
+ 0x381A0000, 0x381A2000, 0x381A4000, 0x381A6000, 0x381A8000, 0x381AA000, 0x381AC000, 0x381AE000, 0x381B0000, 0x381B2000, 0x381B4000, 0x381B6000, 0x381B8000, 0x381BA000, 0x381BC000, 0x381BE000,
+ 0x381C0000, 0x381C2000, 0x381C4000, 0x381C6000, 0x381C8000, 0x381CA000, 0x381CC000, 0x381CE000, 0x381D0000, 0x381D2000, 0x381D4000, 0x381D6000, 0x381D8000, 0x381DA000, 0x381DC000, 0x381DE000,
+ 0x381E0000, 0x381E2000, 0x381E4000, 0x381E6000, 0x381E8000, 0x381EA000, 0x381EC000, 0x381EE000, 0x381F0000, 0x381F2000, 0x381F4000, 0x381F6000, 0x381F8000, 0x381FA000, 0x381FC000, 0x381FE000,
+ 0x38200000, 0x38202000, 0x38204000, 0x38206000, 0x38208000, 0x3820A000, 0x3820C000, 0x3820E000, 0x38210000, 0x38212000, 0x38214000, 0x38216000, 0x38218000, 0x3821A000, 0x3821C000, 0x3821E000,
+ 0x38220000, 0x38222000, 0x38224000, 0x38226000, 0x38228000, 0x3822A000, 0x3822C000, 0x3822E000, 0x38230000, 0x38232000, 0x38234000, 0x38236000, 0x38238000, 0x3823A000, 0x3823C000, 0x3823E000,
+ 0x38240000, 0x38242000, 0x38244000, 0x38246000, 0x38248000, 0x3824A000, 0x3824C000, 0x3824E000, 0x38250000, 0x38252000, 0x38254000, 0x38256000, 0x38258000, 0x3825A000, 0x3825C000, 0x3825E000,
+ 0x38260000, 0x38262000, 0x38264000, 0x38266000, 0x38268000, 0x3826A000, 0x3826C000, 0x3826E000, 0x38270000, 0x38272000, 0x38274000, 0x38276000, 0x38278000, 0x3827A000, 0x3827C000, 0x3827E000,
+ 0x38280000, 0x38282000, 0x38284000, 0x38286000, 0x38288000, 0x3828A000, 0x3828C000, 0x3828E000, 0x38290000, 0x38292000, 0x38294000, 0x38296000, 0x38298000, 0x3829A000, 0x3829C000, 0x3829E000,
+ 0x382A0000, 0x382A2000, 0x382A4000, 0x382A6000, 0x382A8000, 0x382AA000, 0x382AC000, 0x382AE000, 0x382B0000, 0x382B2000, 0x382B4000, 0x382B6000, 0x382B8000, 0x382BA000, 0x382BC000, 0x382BE000,
+ 0x382C0000, 0x382C2000, 0x382C4000, 0x382C6000, 0x382C8000, 0x382CA000, 0x382CC000, 0x382CE000, 0x382D0000, 0x382D2000, 0x382D4000, 0x382D6000, 0x382D8000, 0x382DA000, 0x382DC000, 0x382DE000,
+ 0x382E0000, 0x382E2000, 0x382E4000, 0x382E6000, 0x382E8000, 0x382EA000, 0x382EC000, 0x382EE000, 0x382F0000, 0x382F2000, 0x382F4000, 0x382F6000, 0x382F8000, 0x382FA000, 0x382FC000, 0x382FE000,
+ 0x38300000, 0x38302000, 0x38304000, 0x38306000, 0x38308000, 0x3830A000, 0x3830C000, 0x3830E000, 0x38310000, 0x38312000, 0x38314000, 0x38316000, 0x38318000, 0x3831A000, 0x3831C000, 0x3831E000,
+ 0x38320000, 0x38322000, 0x38324000, 0x38326000, 0x38328000, 0x3832A000, 0x3832C000, 0x3832E000, 0x38330000, 0x38332000, 0x38334000, 0x38336000, 0x38338000, 0x3833A000, 0x3833C000, 0x3833E000,
+ 0x38340000, 0x38342000, 0x38344000, 0x38346000, 0x38348000, 0x3834A000, 0x3834C000, 0x3834E000, 0x38350000, 0x38352000, 0x38354000, 0x38356000, 0x38358000, 0x3835A000, 0x3835C000, 0x3835E000,
+ 0x38360000, 0x38362000, 0x38364000, 0x38366000, 0x38368000, 0x3836A000, 0x3836C000, 0x3836E000, 0x38370000, 0x38372000, 0x38374000, 0x38376000, 0x38378000, 0x3837A000, 0x3837C000, 0x3837E000,
+ 0x38380000, 0x38382000, 0x38384000, 0x38386000, 0x38388000, 0x3838A000, 0x3838C000, 0x3838E000, 0x38390000, 0x38392000, 0x38394000, 0x38396000, 0x38398000, 0x3839A000, 0x3839C000, 0x3839E000,
+ 0x383A0000, 0x383A2000, 0x383A4000, 0x383A6000, 0x383A8000, 0x383AA000, 0x383AC000, 0x383AE000, 0x383B0000, 0x383B2000, 0x383B4000, 0x383B6000, 0x383B8000, 0x383BA000, 0x383BC000, 0x383BE000,
+ 0x383C0000, 0x383C2000, 0x383C4000, 0x383C6000, 0x383C8000, 0x383CA000, 0x383CC000, 0x383CE000, 0x383D0000, 0x383D2000, 0x383D4000, 0x383D6000, 0x383D8000, 0x383DA000, 0x383DC000, 0x383DE000,
+ 0x383E0000, 0x383E2000, 0x383E4000, 0x383E6000, 0x383E8000, 0x383EA000, 0x383EC000, 0x383EE000, 0x383F0000, 0x383F2000, 0x383F4000, 0x383F6000, 0x383F8000, 0x383FA000, 0x383FC000, 0x383FE000,
+ 0x38400000, 0x38402000, 0x38404000, 0x38406000, 0x38408000, 0x3840A000, 0x3840C000, 0x3840E000, 0x38410000, 0x38412000, 0x38414000, 0x38416000, 0x38418000, 0x3841A000, 0x3841C000, 0x3841E000,
+ 0x38420000, 0x38422000, 0x38424000, 0x38426000, 0x38428000, 0x3842A000, 0x3842C000, 0x3842E000, 0x38430000, 0x38432000, 0x38434000, 0x38436000, 0x38438000, 0x3843A000, 0x3843C000, 0x3843E000,
+ 0x38440000, 0x38442000, 0x38444000, 0x38446000, 0x38448000, 0x3844A000, 0x3844C000, 0x3844E000, 0x38450000, 0x38452000, 0x38454000, 0x38456000, 0x38458000, 0x3845A000, 0x3845C000, 0x3845E000,
+ 0x38460000, 0x38462000, 0x38464000, 0x38466000, 0x38468000, 0x3846A000, 0x3846C000, 0x3846E000, 0x38470000, 0x38472000, 0x38474000, 0x38476000, 0x38478000, 0x3847A000, 0x3847C000, 0x3847E000,
+ 0x38480000, 0x38482000, 0x38484000, 0x38486000, 0x38488000, 0x3848A000, 0x3848C000, 0x3848E000, 0x38490000, 0x38492000, 0x38494000, 0x38496000, 0x38498000, 0x3849A000, 0x3849C000, 0x3849E000,
+ 0x384A0000, 0x384A2000, 0x384A4000, 0x384A6000, 0x384A8000, 0x384AA000, 0x384AC000, 0x384AE000, 0x384B0000, 0x384B2000, 0x384B4000, 0x384B6000, 0x384B8000, 0x384BA000, 0x384BC000, 0x384BE000,
+ 0x384C0000, 0x384C2000, 0x384C4000, 0x384C6000, 0x384C8000, 0x384CA000, 0x384CC000, 0x384CE000, 0x384D0000, 0x384D2000, 0x384D4000, 0x384D6000, 0x384D8000, 0x384DA000, 0x384DC000, 0x384DE000,
+ 0x384E0000, 0x384E2000, 0x384E4000, 0x384E6000, 0x384E8000, 0x384EA000, 0x384EC000, 0x384EE000, 0x384F0000, 0x384F2000, 0x384F4000, 0x384F6000, 0x384F8000, 0x384FA000, 0x384FC000, 0x384FE000,
+ 0x38500000, 0x38502000, 0x38504000, 0x38506000, 0x38508000, 0x3850A000, 0x3850C000, 0x3850E000, 0x38510000, 0x38512000, 0x38514000, 0x38516000, 0x38518000, 0x3851A000, 0x3851C000, 0x3851E000,
+ 0x38520000, 0x38522000, 0x38524000, 0x38526000, 0x38528000, 0x3852A000, 0x3852C000, 0x3852E000, 0x38530000, 0x38532000, 0x38534000, 0x38536000, 0x38538000, 0x3853A000, 0x3853C000, 0x3853E000,
+ 0x38540000, 0x38542000, 0x38544000, 0x38546000, 0x38548000, 0x3854A000, 0x3854C000, 0x3854E000, 0x38550000, 0x38552000, 0x38554000, 0x38556000, 0x38558000, 0x3855A000, 0x3855C000, 0x3855E000,
+ 0x38560000, 0x38562000, 0x38564000, 0x38566000, 0x38568000, 0x3856A000, 0x3856C000, 0x3856E000, 0x38570000, 0x38572000, 0x38574000, 0x38576000, 0x38578000, 0x3857A000, 0x3857C000, 0x3857E000,
+ 0x38580000, 0x38582000, 0x38584000, 0x38586000, 0x38588000, 0x3858A000, 0x3858C000, 0x3858E000, 0x38590000, 0x38592000, 0x38594000, 0x38596000, 0x38598000, 0x3859A000, 0x3859C000, 0x3859E000,
+ 0x385A0000, 0x385A2000, 0x385A4000, 0x385A6000, 0x385A8000, 0x385AA000, 0x385AC000, 0x385AE000, 0x385B0000, 0x385B2000, 0x385B4000, 0x385B6000, 0x385B8000, 0x385BA000, 0x385BC000, 0x385BE000,
+ 0x385C0000, 0x385C2000, 0x385C4000, 0x385C6000, 0x385C8000, 0x385CA000, 0x385CC000, 0x385CE000, 0x385D0000, 0x385D2000, 0x385D4000, 0x385D6000, 0x385D8000, 0x385DA000, 0x385DC000, 0x385DE000,
+ 0x385E0000, 0x385E2000, 0x385E4000, 0x385E6000, 0x385E8000, 0x385EA000, 0x385EC000, 0x385EE000, 0x385F0000, 0x385F2000, 0x385F4000, 0x385F6000, 0x385F8000, 0x385FA000, 0x385FC000, 0x385FE000,
+ 0x38600000, 0x38602000, 0x38604000, 0x38606000, 0x38608000, 0x3860A000, 0x3860C000, 0x3860E000, 0x38610000, 0x38612000, 0x38614000, 0x38616000, 0x38618000, 0x3861A000, 0x3861C000, 0x3861E000,
+ 0x38620000, 0x38622000, 0x38624000, 0x38626000, 0x38628000, 0x3862A000, 0x3862C000, 0x3862E000, 0x38630000, 0x38632000, 0x38634000, 0x38636000, 0x38638000, 0x3863A000, 0x3863C000, 0x3863E000,
+ 0x38640000, 0x38642000, 0x38644000, 0x38646000, 0x38648000, 0x3864A000, 0x3864C000, 0x3864E000, 0x38650000, 0x38652000, 0x38654000, 0x38656000, 0x38658000, 0x3865A000, 0x3865C000, 0x3865E000,
+ 0x38660000, 0x38662000, 0x38664000, 0x38666000, 0x38668000, 0x3866A000, 0x3866C000, 0x3866E000, 0x38670000, 0x38672000, 0x38674000, 0x38676000, 0x38678000, 0x3867A000, 0x3867C000, 0x3867E000,
+ 0x38680000, 0x38682000, 0x38684000, 0x38686000, 0x38688000, 0x3868A000, 0x3868C000, 0x3868E000, 0x38690000, 0x38692000, 0x38694000, 0x38696000, 0x38698000, 0x3869A000, 0x3869C000, 0x3869E000,
+ 0x386A0000, 0x386A2000, 0x386A4000, 0x386A6000, 0x386A8000, 0x386AA000, 0x386AC000, 0x386AE000, 0x386B0000, 0x386B2000, 0x386B4000, 0x386B6000, 0x386B8000, 0x386BA000, 0x386BC000, 0x386BE000,
+ 0x386C0000, 0x386C2000, 0x386C4000, 0x386C6000, 0x386C8000, 0x386CA000, 0x386CC000, 0x386CE000, 0x386D0000, 0x386D2000, 0x386D4000, 0x386D6000, 0x386D8000, 0x386DA000, 0x386DC000, 0x386DE000,
+ 0x386E0000, 0x386E2000, 0x386E4000, 0x386E6000, 0x386E8000, 0x386EA000, 0x386EC000, 0x386EE000, 0x386F0000, 0x386F2000, 0x386F4000, 0x386F6000, 0x386F8000, 0x386FA000, 0x386FC000, 0x386FE000,
+ 0x38700000, 0x38702000, 0x38704000, 0x38706000, 0x38708000, 0x3870A000, 0x3870C000, 0x3870E000, 0x38710000, 0x38712000, 0x38714000, 0x38716000, 0x38718000, 0x3871A000, 0x3871C000, 0x3871E000,
+ 0x38720000, 0x38722000, 0x38724000, 0x38726000, 0x38728000, 0x3872A000, 0x3872C000, 0x3872E000, 0x38730000, 0x38732000, 0x38734000, 0x38736000, 0x38738000, 0x3873A000, 0x3873C000, 0x3873E000,
+ 0x38740000, 0x38742000, 0x38744000, 0x38746000, 0x38748000, 0x3874A000, 0x3874C000, 0x3874E000, 0x38750000, 0x38752000, 0x38754000, 0x38756000, 0x38758000, 0x3875A000, 0x3875C000, 0x3875E000,
+ 0x38760000, 0x38762000, 0x38764000, 0x38766000, 0x38768000, 0x3876A000, 0x3876C000, 0x3876E000, 0x38770000, 0x38772000, 0x38774000, 0x38776000, 0x38778000, 0x3877A000, 0x3877C000, 0x3877E000,
+ 0x38780000, 0x38782000, 0x38784000, 0x38786000, 0x38788000, 0x3878A000, 0x3878C000, 0x3878E000, 0x38790000, 0x38792000, 0x38794000, 0x38796000, 0x38798000, 0x3879A000, 0x3879C000, 0x3879E000,
+ 0x387A0000, 0x387A2000, 0x387A4000, 0x387A6000, 0x387A8000, 0x387AA000, 0x387AC000, 0x387AE000, 0x387B0000, 0x387B2000, 0x387B4000, 0x387B6000, 0x387B8000, 0x387BA000, 0x387BC000, 0x387BE000,
+ 0x387C0000, 0x387C2000, 0x387C4000, 0x387C6000, 0x387C8000, 0x387CA000, 0x387CC000, 0x387CE000, 0x387D0000, 0x387D2000, 0x387D4000, 0x387D6000, 0x387D8000, 0x387DA000, 0x387DC000, 0x387DE000,
+ 0x387E0000, 0x387E2000, 0x387E4000, 0x387E6000, 0x387E8000, 0x387EA000, 0x387EC000, 0x387EE000, 0x387F0000, 0x387F2000, 0x387F4000, 0x387F6000, 0x387F8000, 0x387FA000, 0x387FC000, 0x387FE000 };
+ static const bits<float>::type exponent_table[64] = {
+ 0x00000000, 0x00800000, 0x01000000, 0x01800000, 0x02000000, 0x02800000, 0x03000000, 0x03800000, 0x04000000, 0x04800000, 0x05000000, 0x05800000, 0x06000000, 0x06800000, 0x07000000, 0x07800000,
+ 0x08000000, 0x08800000, 0x09000000, 0x09800000, 0x0A000000, 0x0A800000, 0x0B000000, 0x0B800000, 0x0C000000, 0x0C800000, 0x0D000000, 0x0D800000, 0x0E000000, 0x0E800000, 0x0F000000, 0x47800000,
+ 0x80000000, 0x80800000, 0x81000000, 0x81800000, 0x82000000, 0x82800000, 0x83000000, 0x83800000, 0x84000000, 0x84800000, 0x85000000, 0x85800000, 0x86000000, 0x86800000, 0x87000000, 0x87800000,
+ 0x88000000, 0x88800000, 0x89000000, 0x89800000, 0x8A000000, 0x8A800000, 0x8B000000, 0x8B800000, 0x8C000000, 0x8C800000, 0x8D000000, 0x8D800000, 0x8E000000, 0x8E800000, 0x8F000000, 0xC7800000 };
+ static const unsigned short offset_table[64] = {
+ 0, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024,
+ 0, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024 };
+ bits<float>::type fbits = mantissa_table[offset_table[value>>10]+(value&0x3FF)] + exponent_table[value>>10];
+ #endif
+ float out;
+ std::memcpy(&out, &fbits, sizeof(float));
+ return out;
+ #endif
+ }
+
+ /// Convert half-precision to IEEE double-precision.
+ /// \param value half-precision value to convert
+ /// \return double-precision value
+ inline double half2float_impl(unsigned int value, double, true_type)
+ {
+ #if HALF_ENABLE_F16C_INTRINSICS
+ return _mm_cvtsd_f64(_mm_cvtps_pd(_mm_cvtph_ps(_mm_cvtsi32_si128(value))));
+ #else
+ uint32 hi = static_cast<uint32>(value&0x8000) << 16;
+ unsigned int abs = value & 0x7FFF;
+ if(abs)
+ {
+ hi |= 0x3F000000 << static_cast<unsigned>(abs>=0x7C00);
+ for(; abs<0x400; abs<<=1,hi-=0x100000) ;
+ hi += static_cast<uint32>(abs) << 10;
+ }
+ bits<double>::type dbits = static_cast<bits<double>::type>(hi) << 32;
+ double out;
+ std::memcpy(&out, &dbits, sizeof(double));
+ return out;
+ #endif
+ }
+
+ /// Convert half-precision to non-IEEE floating-point.
+ /// \tparam T type to convert to (builtin integer type)
+ /// \param value half-precision value to convert
+ /// \return floating-point value
+ template<typename T> T half2float_impl(unsigned int value, T, ...)
+ {
+ T out;
+ unsigned int abs = value & 0x7FFF;
+ if(abs > 0x7C00)
+ out = (std::numeric_limits<T>::has_signaling_NaN && !(abs&0x200)) ? std::numeric_limits<T>::signaling_NaN() :
+ std::numeric_limits<T>::has_quiet_NaN ? std::numeric_limits<T>::quiet_NaN() : T();
+ else if(abs == 0x7C00)
+ out = std::numeric_limits<T>::has_infinity ? std::numeric_limits<T>::infinity() : std::numeric_limits<T>::max();
+ else if(abs > 0x3FF)
+ out = std::ldexp(static_cast<T>((abs&0x3FF)|0x400), (abs>>10)-25);
+ else
+ out = std::ldexp(static_cast<T>(abs), -24);
+ return (value&0x8000) ? -out : out;
+ }
+
+ /// Convert half-precision to floating-point.
+ /// \tparam T type to convert to (builtin integer type)
+ /// \param value half-precision value to convert
+ /// \return floating-point value
+ template<typename T> T half2float(unsigned int value)
+ {
+ return half2float_impl(value, T(), bool_type<std::numeric_limits<T>::is_iec559&&sizeof(typename bits<T>::type)==sizeof(T)>());
+ }
+
+ /// Convert half-precision floating-point to integer.
+ /// \tparam R rounding mode to use
+ /// \tparam E `true` for round to even, `false` for round away from zero
+ /// \tparam I `true` to raise INEXACT exception (if inexact), `false` to never raise it
+ /// \tparam T type to convert to (buitlin integer type with at least 16 bits precision, excluding any implicit sign bits)
+ /// \param value half-precision value to convert
+ /// \return rounded integer value
+ /// \exception FE_INVALID if value is not representable in type \a T
+ /// \exception FE_INEXACT if value had to be rounded and \a I is `true`
+ template<std::float_round_style R,bool E,bool I,typename T> T half2int(unsigned int value)
+ {
+ unsigned int abs = value & 0x7FFF;
+ if(abs >= 0x7C00)
+ {
+ raise(FE_INVALID);
+ return (value&0x8000) ? std::numeric_limits<T>::min() : std::numeric_limits<T>::max();
+ }
+ if(abs < 0x3800)
+ {
+ raise(FE_INEXACT, I);
+ return (R==std::round_toward_infinity) ? T(~(value>>15)&(abs!=0)) :
+ (R==std::round_toward_neg_infinity) ? -T(value>0x8000) :
+ T();
+ }
+ int exp = 25 - (abs>>10);
+ unsigned int m = (value&0x3FF) | 0x400;
+ int32 i = static_cast<int32>((exp<=0) ? (m<<-exp) : ((m+(
+ (R==std::round_to_nearest) ? ((1<<(exp-1))-(~(m>>exp)&E)) :
+ (R==std::round_toward_infinity) ? (((1<<exp)-1)&((value>>15)-1)) :
+ (R==std::round_toward_neg_infinity) ? (((1<<exp)-1)&-(value>>15)) : 0))>>exp));
+ if((!std::numeric_limits<T>::is_signed && (value&0x8000)) || (std::numeric_limits<T>::digits<16 &&
+ ((value&0x8000) ? (-i<std::numeric_limits<T>::min()) : (i>std::numeric_limits<T>::max()))))
+ raise(FE_INVALID);
+ else if(I && exp > 0 && (m&((1<<exp)-1)))
+ raise(FE_INEXACT);
+ return static_cast<T>((value&0x8000) ? -i : i);
+ }
+
+ /// \}
+ /// \name Mathematics
+ /// \{
+
+ /// upper part of 64-bit multiplication.
+ /// \tparam R rounding mode to use
+ /// \param x first factor
+ /// \param y second factor
+ /// \return upper 32 bit of \a x * \a y
+ template<std::float_round_style R> uint32 mulhi(uint32 x, uint32 y)
+ {
+ uint32 xy = (x>>16) * (y&0xFFFF), yx = (x&0xFFFF) * (y>>16), c = (xy&0xFFFF) + (yx&0xFFFF) + (((x&0xFFFF)*(y&0xFFFF))>>16);
+ return (x>>16)*(y>>16) + (xy>>16) + (yx>>16) + (c>>16) +
+ ((R==std::round_to_nearest) ? ((c>>15)&1) : (R==std::round_toward_infinity) ? ((c&0xFFFF)!=0) : 0);
+ }
+
+ /// 64-bit multiplication.
+ /// \param x first factor
+ /// \param y second factor
+ /// \return upper 32 bit of \a x * \a y rounded to nearest
+ inline uint32 multiply64(uint32 x, uint32 y)
+ {
+ #if HALF_ENABLE_CPP11_LONG_LONG
+ return static_cast<uint32>((static_cast<unsigned long long>(x)*static_cast<unsigned long long>(y)+0x80000000)>>32);
+ #else
+ return mulhi<std::round_to_nearest>(x, y);
+ #endif
+ }
+
+ /// 64-bit division.
+ /// \param x upper 32 bit of dividend
+ /// \param y divisor
+ /// \param s variable to store sticky bit for rounding
+ /// \return (\a x << 32) / \a y
+ inline uint32 divide64(uint32 x, uint32 y, int &s)
+ {
+ #if HALF_ENABLE_CPP11_LONG_LONG
+ unsigned long long xx = static_cast<unsigned long long>(x) << 32;
+ return s = (xx%y!=0), static_cast<uint32>(xx/y);
+ #else
+ y >>= 1;
+ uint32 rem = x, div = 0;
+ for(unsigned int i=0; i<32; ++i)
+ {
+ div <<= 1;
+ if(rem >= y)
+ {
+ rem -= y;
+ div |= 1;
+ }
+ rem <<= 1;
+ }
+ return s = rem > 1, div;
+ #endif
+ }
+
+ /// Half precision positive modulus.
+ /// \tparam Q `true` to compute full quotient, `false` else
+ /// \tparam R `true` to compute signed remainder, `false` for positive remainder
+ /// \param x first operand as positive finite half-precision value
+ /// \param y second operand as positive finite half-precision value
+ /// \param quo adress to store quotient at, `nullptr` if \a Q `false`
+ /// \return modulus of \a x / \a y
+ template<bool Q,bool R> unsigned int mod(unsigned int x, unsigned int y, int *quo = NULL)
+ {
+ unsigned int q = 0;
+ if(x > y)
+ {
+ int absx = x, absy = y, expx = 0, expy = 0;
+ for(; absx<0x400; absx<<=1,--expx) ;
+ for(; absy<0x400; absy<<=1,--expy) ;
+ expx += absx >> 10;
+ expy += absy >> 10;
+ int mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400;
+ for(int d=expx-expy; d; --d)
+ {
+ if(!Q && mx == my)
+ return 0;
+ if(mx >= my)
+ {
+ mx -= my;
+ q += Q;
+ }
+ mx <<= 1;
+ q <<= static_cast<int>(Q);
+ }
+ if(!Q && mx == my)
+ return 0;
+ if(mx >= my)
+ {
+ mx -= my;
+ ++q;
+ }
+ if(Q)
+ {
+ q &= (1<<(std::numeric_limits<int>::digits-1)) - 1;
+ if(!mx)
+ return *quo = q, 0;
+ }
+ for(; mx<0x400; mx<<=1,--expy) ;
+ x = (expy>0) ? ((expy<<10)|(mx&0x3FF)) : (mx>>(1-expy));
+ }
+ if(R)
+ {
+ unsigned int a, b;
+ if(y < 0x800)
+ {
+ a = (x<0x400) ? (x<<1) : (x+0x400);
+ b = y;
+ }
+ else
+ {
+ a = x;
+ b = y - 0x400;
+ }
+ if(a > b || (a == b && (q&1)))
+ {
+ int exp = (y>>10) + (y<=0x3FF), d = exp - (x>>10) - (x<=0x3FF);
+ int m = (((y&0x3FF)|((y>0x3FF)<<10))<<1) - (((x&0x3FF)|((x>0x3FF)<<10))<<(1-d));
+ for(; m<0x800 && exp>1; m<<=1,--exp) ;
+ x = 0x8000 + ((exp-1)<<10) + (m>>1);
+ q += Q;
+ }
+ }
+ if(Q)
+ *quo = q;
+ return x;
+ }
+
+ /// Fixed point square root.
+ /// \tparam F number of fractional bits
+ /// \param r radicand in Q1.F fixed point format
+ /// \param exp exponent
+ /// \return square root as Q1.F/2
+ template<unsigned int F> uint32 sqrt(uint32 &r, int &exp)
+ {
+ int i = exp & 1;
+ r <<= i;
+ exp = (exp-i) / 2;
+ uint32 m = 0;
+ for(uint32 bit=static_cast<uint32>(1)<<F; bit; bit>>=2)
+ {
+ if(r < m+bit)
+ m >>= 1;
+ else
+ {
+ r -= m + bit;
+ m = (m>>1) + bit;
+ }
+ }
+ return m;
+ }
+
+ /// Fixed point binary exponential.
+ /// This uses the BKM algorithm in E-mode.
+ /// \param m exponent in [0,1) as Q0.31
+ /// \param n number of iterations (at most 32)
+ /// \return 2 ^ \a m as Q1.31
+ inline uint32 exp2(uint32 m, unsigned int n = 32)
+ {
+ static const uint32 logs[] = {
+ 0x80000000, 0x4AE00D1D, 0x2934F098, 0x15C01A3A, 0x0B31FB7D, 0x05AEB4DD, 0x02DCF2D1, 0x016FE50B,
+ 0x00B84E23, 0x005C3E10, 0x002E24CA, 0x001713D6, 0x000B8A47, 0x0005C53B, 0x0002E2A3, 0x00017153,
+ 0x0000B8AA, 0x00005C55, 0x00002E2B, 0x00001715, 0x00000B8B, 0x000005C5, 0x000002E3, 0x00000171,
+ 0x000000B9, 0x0000005C, 0x0000002E, 0x00000017, 0x0000000C, 0x00000006, 0x00000003, 0x00000001 };
+ if(!m)
+ return 0x80000000;
+ uint32 mx = 0x80000000, my = 0;
+ for(unsigned int i=1; i<n; ++i)
+ {
+ uint32 mz = my + logs[i];
+ if(mz <= m)
+ {
+ my = mz;
+ mx += mx >> i;
+ }
+ }
+ return mx;
+ }
+
+ /// Fixed point binary logarithm.
+ /// This uses the BKM algorithm in L-mode.
+ /// \param m mantissa in [1,2) as Q1.30
+ /// \param n number of iterations (at most 32)
+ /// \return log2(\a m) as Q0.31
+ inline uint32 log2(uint32 m, unsigned int n = 32)
+ {
+ static const uint32 logs[] = {
+ 0x80000000, 0x4AE00D1D, 0x2934F098, 0x15C01A3A, 0x0B31FB7D, 0x05AEB4DD, 0x02DCF2D1, 0x016FE50B,
+ 0x00B84E23, 0x005C3E10, 0x002E24CA, 0x001713D6, 0x000B8A47, 0x0005C53B, 0x0002E2A3, 0x00017153,
+ 0x0000B8AA, 0x00005C55, 0x00002E2B, 0x00001715, 0x00000B8B, 0x000005C5, 0x000002E3, 0x00000171,
+ 0x000000B9, 0x0000005C, 0x0000002E, 0x00000017, 0x0000000C, 0x00000006, 0x00000003, 0x00000001 };
+ if(m == 0x40000000)
+ return 0;
+ uint32 mx = 0x40000000, my = 0;
+ for(unsigned int i=1; i<n; ++i)
+ {
+ uint32 mz = mx + (mx>>i);
+ if(mz <= m)
+ {
+ mx = mz;
+ my += logs[i];
+ }
+ }
+ return my;
+ }
+
+ /// Fixed point sine and cosine.
+ /// This uses the CORDIC algorithm in rotation mode.
+ /// \param mz angle in [-pi/2,pi/2] as Q1.30
+ /// \param n number of iterations (at most 31)
+ /// \return sine and cosine of \a mz as Q1.30
+ inline std::pair<uint32,uint32> sincos(uint32 mz, unsigned int n = 31)
+ {
+ static const uint32 angles[] = {
+ 0x3243F6A9, 0x1DAC6705, 0x0FADBAFD, 0x07F56EA7, 0x03FEAB77, 0x01FFD55C, 0x00FFFAAB, 0x007FFF55,
+ 0x003FFFEB, 0x001FFFFD, 0x00100000, 0x00080000, 0x00040000, 0x00020000, 0x00010000, 0x00008000,
+ 0x00004000, 0x00002000, 0x00001000, 0x00000800, 0x00000400, 0x00000200, 0x00000100, 0x00000080,
+ 0x00000040, 0x00000020, 0x00000010, 0x00000008, 0x00000004, 0x00000002, 0x00000001 };
+ uint32 mx = 0x26DD3B6A, my = 0;
+ for(unsigned int i=0; i<n; ++i)
+ {
+ uint32 sign = sign_mask(mz);
+ uint32 tx = mx - (arithmetic_shift(my, i)^sign) + sign;
+ uint32 ty = my + (arithmetic_shift(mx, i)^sign) - sign;
+ mx = tx; my = ty; mz -= (angles[i]^sign) - sign;
+ }
+ return std::make_pair(my, mx);
+ }
+
+ /// Fixed point arc tangent.
+ /// This uses the CORDIC algorithm in vectoring mode.
+ /// \param my y coordinate as Q0.30
+ /// \param mx x coordinate as Q0.30
+ /// \param n number of iterations (at most 31)
+ /// \return arc tangent of \a my / \a mx as Q1.30
+ inline uint32 atan2(uint32 my, uint32 mx, unsigned int n = 31)
+ {
+ static const uint32 angles[] = {
+ 0x3243F6A9, 0x1DAC6705, 0x0FADBAFD, 0x07F56EA7, 0x03FEAB77, 0x01FFD55C, 0x00FFFAAB, 0x007FFF55,
+ 0x003FFFEB, 0x001FFFFD, 0x00100000, 0x00080000, 0x00040000, 0x00020000, 0x00010000, 0x00008000,
+ 0x00004000, 0x00002000, 0x00001000, 0x00000800, 0x00000400, 0x00000200, 0x00000100, 0x00000080,
+ 0x00000040, 0x00000020, 0x00000010, 0x00000008, 0x00000004, 0x00000002, 0x00000001 };
+ uint32 mz = 0;
+ for(unsigned int i=0; i<n; ++i)
+ {
+ uint32 sign = sign_mask(my);
+ uint32 tx = mx + (arithmetic_shift(my, i)^sign) - sign;
+ uint32 ty = my - (arithmetic_shift(mx, i)^sign) + sign;
+ mx = tx; my = ty; mz += (angles[i]^sign) - sign;
+ }
+ return mz;
+ }
+
+ /// Reduce argument for trigonometric functions.
+ /// \param abs half-precision floating-point value
+ /// \param k value to take quarter period
+ /// \return \a abs reduced to [-pi/4,pi/4] as Q0.30
+ inline uint32 angle_arg(unsigned int abs, int &k)
+ {
+ uint32 m = (abs&0x3FF) | ((abs>0x3FF)<<10);
+ int exp = (abs>>10) + (abs<=0x3FF) - 15;
+ if(abs < 0x3A48)
+ return k = 0, m << (exp+20);
+ #if HALF_ENABLE_CPP11_LONG_LONG
+ unsigned long long y = m * 0xA2F9836E4E442, mask = (1ULL<<(62-exp)) - 1, yi = (y+(mask>>1)) & ~mask, f = y - yi;
+ uint32 sign = -static_cast<uint32>(f>>63);
+ k = static_cast<int>(yi>>(62-exp));
+ return (multiply64(static_cast<uint32>((sign ? -f : f)>>(31-exp)), 0xC90FDAA2)^sign) - sign;
+ #else
+ uint32 yh = m*0xA2F98 + mulhi<std::round_toward_zero>(m, 0x36E4E442), yl = (m*0x36E4E442) & 0xFFFFFFFF;
+ uint32 mask = (static_cast<uint32>(1)<<(30-exp)) - 1, yi = (yh+(mask>>1)) & ~mask, sign = -static_cast<uint32>(yi>yh);
+ k = static_cast<int>(yi>>(30-exp));
+ uint32 fh = (yh^sign) + (yi^~sign) - ~sign, fl = (yl^sign) - sign;
+ return (multiply64((exp>-1) ? (((fh<<(1+exp))&0xFFFFFFFF)|((fl&0xFFFFFFFF)>>(31-exp))) : fh, 0xC90FDAA2)^sign) - sign;
+ #endif
+ }
+
+ /// Get arguments for atan2 function.
+ /// \param abs half-precision floating-point value
+ /// \return \a abs and sqrt(1 - \a abs^2) as Q0.30
+ inline std::pair<uint32,uint32> atan2_args(unsigned int abs)
+ {
+ int exp = -15;
+ for(; abs<0x400; abs<<=1,--exp) ;
+ exp += abs >> 10;
+ uint32 my = ((abs&0x3FF)|0x400) << 5, r = my * my;
+ int rexp = 2 * exp;
+ r = 0x40000000 - ((rexp>-31) ? ((r>>-rexp)|((r&((static_cast<uint32>(1)<<-rexp)-1))!=0)) : 1);
+ for(rexp=0; r<0x40000000; r<<=1,--rexp) ;
+ uint32 mx = sqrt<30>(r, rexp);
+ int d = exp - rexp;
+ if(d < 0)
+ return std::make_pair((d<-14) ? ((my>>(-d-14))+((my>>(-d-15))&1)) : (my<<(14+d)), (mx<<14)+(r<<13)/mx);
+ if(d > 0)
+ return std::make_pair(my<<14, (d>14) ? ((mx>>(d-14))+((mx>>(d-15))&1)) : ((d==14) ? mx : ((mx<<(14-d))+(r<<(13-d))/mx)));
+ return std::make_pair(my<<13, (mx<<13)+(r<<12)/mx);
+ }
+
+ /// Get exponentials for hyperbolic computation
+ /// \param abs half-precision floating-point value
+ /// \param exp variable to take unbiased exponent of larger result
+ /// \param n number of BKM iterations (at most 32)
+ /// \return exp(abs) and exp(-\a abs) as Q1.31 with same exponent
+ inline std::pair<uint32,uint32> hyperbolic_args(unsigned int abs, int &exp, unsigned int n = 32)
+ {
+ uint32 mx = detail::multiply64(static_cast<uint32>((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29), my;
+ int e = (abs>>10) + (abs<=0x3FF);
+ if(e < 14)
+ {
+ exp = 0;
+ mx >>= 14 - e;
+ }
+ else
+ {
+ exp = mx >> (45-e);
+ mx = (mx<<(e-14)) & 0x7FFFFFFF;
+ }
+ mx = exp2(mx, n);
+ int d = exp << 1, s;
+ if(mx > 0x80000000)
+ {
+ my = divide64(0x80000000, mx, s);
+ my |= s;
+ ++d;
+ }
+ else
+ my = mx;
+ return std::make_pair(mx, (d<31) ? ((my>>d)|((my&((static_cast<uint32>(1)<<d)-1))!=0)) : 1);
+ }
+
+ /// Postprocessing for binary exponential.
+ /// \tparam R rounding mode to use
+ /// \param m fractional part of as Q0.31
+ /// \param exp absolute value of unbiased exponent
+ /// \param esign sign of actual exponent
+ /// \param sign sign bit of result
+ /// \param n number of BKM iterations (at most 32)
+ /// \return value converted to half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded or \a I is `true`
+ template<std::float_round_style R> unsigned int exp2_post(uint32 m, int exp, bool esign, unsigned int sign = 0, unsigned int n = 32)
+ {
+ if(esign)
+ {
+ exp = -exp - (m!=0);
+ if(exp < -25)
+ return underflow<R>(sign);
+ else if(exp == -25)
+ return rounded<R,false>(sign, 1, m!=0);
+ }
+ else if(exp > 15)
+ return overflow<R>(sign);
+ if(!m)
+ return sign | (((exp+=15)>0) ? (exp<<10) : check_underflow(0x200>>-exp));
+ m = exp2(m, n);
+ int s = 0;
+ if(esign)
+ m = divide64(0x80000000, m, s);
+ return fixed2half<R,31,false,false,true>(m, exp+14, sign, s);
+ }
+
+ /// Postprocessing for binary logarithm.
+ /// \tparam R rounding mode to use
+ /// \tparam L logarithm for base transformation as Q1.31
+ /// \param m fractional part of logarithm as Q0.31
+ /// \param ilog signed integer part of logarithm
+ /// \param exp biased exponent of result
+ /// \param sign sign bit of result
+ /// \return value base-transformed and converted to half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if no other exception occurred
+ template<std::float_round_style R,uint32 L> unsigned int log2_post(uint32 m, int ilog, int exp, unsigned int sign = 0)
+ {
+ uint32 msign = sign_mask(ilog);
+ m = (((static_cast<uint32>(ilog)<<27)+(m>>4))^msign) - msign;
+ if(!m)
+ return 0;
+ for(; m<0x80000000; m<<=1,--exp) ;
+ int i = m >= L, s;
+ exp += i;
+ m >>= 1 + i;
+ sign ^= msign & 0x8000;
+ if(exp < -11)
+ return underflow<R>(sign);
+ m = divide64(m, L, s);
+ return fixed2half<R,30,false,false,true>(m, exp, sign, 1);
+ }
+
+ /// Hypotenuse square root and postprocessing.
+ /// \tparam R rounding mode to use
+ /// \param r mantissa as Q2.30
+ /// \param exp biased exponent
+ /// \return square root converted to half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if value had to be rounded
+ template<std::float_round_style R> unsigned int hypot_post(uint32 r, int exp)
+ {
+ int i = r >> 31;
+ if((exp+=i) > 46)
+ return overflow<R>();
+ if(exp < -34)
+ return underflow<R>();
+ r = (r>>i) | (r&i);
+ uint32 m = sqrt<30>(r, exp+=15);
+ return fixed2half<R,15,false,false,false>(m, exp-1, 0, r!=0);
+ }
+
+ /// Division and postprocessing for tangents.
+ /// \tparam R rounding mode to use
+ /// \param my dividend as Q1.31
+ /// \param mx divisor as Q1.31
+ /// \param exp biased exponent of result
+ /// \param sign sign bit of result
+ /// \return quotient converted to half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if no other exception occurred
+ template<std::float_round_style R> unsigned int tangent_post(uint32 my, uint32 mx, int exp, unsigned int sign = 0)
+ {
+ int i = my >= mx, s;
+ exp += i;
+ if(exp > 29)
+ return overflow<R>(sign);
+ if(exp < -11)
+ return underflow<R>(sign);
+ uint32 m = divide64(my>>(i+1), mx, s);
+ return fixed2half<R,30,false,false,true>(m, exp, sign, s);
+ }
+
+ /// Area function and postprocessing.
+ /// This computes the value directly in Q2.30 using the representation `asinh|acosh(x) = log(x+sqrt(x^2+|-1))`.
+ /// \tparam R rounding mode to use
+ /// \tparam S `true` for asinh, `false` for acosh
+ /// \param arg half-precision argument
+ /// \return asinh|acosh(\a arg) converted to half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if no other exception occurred
+ template<std::float_round_style R,bool S> unsigned int area(unsigned int arg)
+ {
+ int abs = arg & 0x7FFF, expx = (abs>>10) + (abs<=0x3FF) - 15, expy = -15, ilog, i;
+ uint32 mx = static_cast<uint32>((abs&0x3FF)|((abs>0x3FF)<<10)) << 20, my, r;
+ for(; abs<0x400; abs<<=1,--expy) ;
+ expy += abs >> 10;
+ r = ((abs&0x3FF)|0x400) << 5;
+ r *= r;
+ i = r >> 31;
+ expy = 2*expy + i;
+ r >>= i;
+ if(S)
+ {
+ if(expy < 0)
+ {
+ r = 0x40000000 + ((expy>-30) ? ((r>>-expy)|((r&((static_cast<uint32>(1)<<-expy)-1))!=0)) : 1);
+ expy = 0;
+ }
+ else
+ {
+ r += 0x40000000 >> expy;
+ i = r >> 31;
+ r = (r>>i) | (r&i);
+ expy += i;
+ }
+ }
+ else
+ {
+ r -= 0x40000000 >> expy;
+ for(; r<0x40000000; r<<=1,--expy) ;
+ }
+ my = sqrt<30>(r, expy);
+ my = (my<<15) + (r<<14)/my;
+ if(S)
+ {
+ mx >>= expy - expx;
+ ilog = expy;
+ }
+ else
+ {
+ my >>= expx - expy;
+ ilog = expx;
+ }
+ my += mx;
+ i = my >> 31;
+ static const int G = S && (R==std::round_to_nearest);
+ return log2_post<R,0xB8AA3B2A>(log2(my>>i, 26+S+G)+(G<<3), ilog+i, 17, arg&(static_cast<unsigned>(S)<<15));
+ }
+
+ /// Class for 1.31 unsigned floating-point computation
+ struct f31
+ {
+ /// Constructor.
+ /// \param mant mantissa as 1.31
+ /// \param e exponent
+ HALF_CONSTEXPR f31(uint32 mant, int e) : m(mant), exp(e) {}
+
+ /// Constructor.
+ /// \param abs unsigned half-precision value
+ f31(unsigned int abs) : exp(-15)
+ {
+ for(; abs<0x400; abs<<=1,--exp) ;
+ m = static_cast<uint32>((abs&0x3FF)|0x400) << 21;
+ exp += (abs>>10);
+ }
+
+ /// Addition operator.
+ /// \param a first operand
+ /// \param b second operand
+ /// \return \a a + \a b
+ friend f31 operator+(f31 a, f31 b)
+ {
+ if(b.exp > a.exp)
+ std::swap(a, b);
+ int d = a.exp - b.exp;
+ uint32 m = a.m + ((d<32) ? (b.m>>d) : 0);
+ int i = (m&0xFFFFFFFF) < a.m;
+ return f31(((m+i)>>i)|0x80000000, a.exp+i);
+ }
+
+ /// Subtraction operator.
+ /// \param a first operand
+ /// \param b second operand
+ /// \return \a a - \a b
+ friend f31 operator-(f31 a, f31 b)
+ {
+ int d = a.exp - b.exp, exp = a.exp;
+ uint32 m = a.m - ((d<32) ? (b.m>>d) : 0);
+ if(!m)
+ return f31(0, -32);
+ for(; m<0x80000000; m<<=1,--exp) ;
+ return f31(m, exp);
+ }
+
+ /// Multiplication operator.
+ /// \param a first operand
+ /// \param b second operand
+ /// \return \a a * \a b
+ friend f31 operator*(f31 a, f31 b)
+ {
+ uint32 m = multiply64(a.m, b.m);
+ int i = m >> 31;
+ return f31(m<<(1-i), a.exp + b.exp + i);
+ }
+
+ /// Division operator.
+ /// \param a first operand
+ /// \param b second operand
+ /// \return \a a / \a b
+ friend f31 operator/(f31 a, f31 b)
+ {
+ int i = a.m >= b.m, s;
+ uint32 m = divide64((a.m+i)>>i, b.m, s);
+ return f31(m, a.exp - b.exp + i - 1);
+ }
+
+ uint32 m; ///< mantissa as 1.31.
+ int exp; ///< exponent.
+ };
+
+ /// Error function and postprocessing.
+ /// This computes the value directly in Q1.31 using the approximations given
+ /// [here](https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions).
+ /// \tparam R rounding mode to use
+ /// \tparam C `true` for comlementary error function, `false` else
+ /// \param arg half-precision function argument
+ /// \return approximated value of error function in half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if no other exception occurred
+ template<std::float_round_style R,bool C> unsigned int erf(unsigned int arg)
+ {
+ unsigned int abs = arg & 0x7FFF, sign = arg & 0x8000;
+ f31 x(abs), x2 = x * x * f31(0xB8AA3B29, 0), t = f31(0x80000000, 0) / (f31(0x80000000, 0)+f31(0xA7BA054A, -2)*x), t2 = t * t;
+ f31 e = ((f31(0x87DC2213, 0)*t2+f31(0xB5F0E2AE, 0))*t2+f31(0x82790637, -2)-(f31(0xBA00E2B8, 0)*t2+f31(0x91A98E62, -2))*t) * t /
+ ((x2.exp<0) ? f31(exp2((x2.exp>-32) ? (x2.m>>-x2.exp) : 0, 30), 0) : f31(exp2((x2.m<<x2.exp)&0x7FFFFFFF, 22), x2.m>>(31-x2.exp)));
+ return (!C || sign) ? fixed2half<R,31,false,true,true>(0x80000000-(e.m>>(C-e.exp)), 14+C, sign&(C-1U)) :
+ (e.exp<-25) ? underflow<R>() : fixed2half<R,30,false,false,true>(e.m>>1, e.exp+14, 0, e.m&1);
+ }
+
+ /// Gamma function and postprocessing.
+ /// This approximates the value of either the gamma function or its logarithm directly in Q1.31.
+ /// \tparam R rounding mode to use
+ /// \tparam L `true` for lograithm of gamma function, `false` for gamma function
+ /// \param arg half-precision floating-point value
+ /// \return lgamma/tgamma(\a arg) in half-precision
+ /// \exception FE_OVERFLOW on overflows
+ /// \exception FE_UNDERFLOW on underflows
+ /// \exception FE_INEXACT if \a arg is not a positive integer
+ template<std::float_round_style R,bool L> unsigned int gamma(unsigned int arg)
+ {
+/* static const double p[] ={ 2.50662827563479526904, 225.525584619175212544, -268.295973841304927459, 80.9030806934622512966, -5.00757863970517583837, 0.0114684895434781459556 };
+ double t = arg + 4.65, s = p[0];
+ for(unsigned int i=0; i<5; ++i)
+ s += p[i+1] / (arg+i);
+ return std::log(s) + (arg-0.5)*std::log(t) - t;
+*/ static const f31 pi(0xC90FDAA2, 1), lbe(0xB8AA3B29, 0);
+ unsigned int abs = arg & 0x7FFF, sign = arg & 0x8000;
+ bool bsign = sign != 0;
+ f31 z(abs), x = sign ? (z+f31(0x80000000, 0)) : z, t = x + f31(0x94CCCCCD, 2), s =
+ f31(0xA06C9901, 1) + f31(0xBBE654E2, -7)/(x+f31(0x80000000, 2)) + f31(0xA1CE6098, 6)/(x+f31(0x80000000, 1))
+ + f31(0xE1868CB7, 7)/x - f31(0x8625E279, 8)/(x+f31(0x80000000, 0)) - f31(0xA03E158F, 2)/(x+f31(0xC0000000, 1));
+ int i = (s.exp>=2) + (s.exp>=4) + (s.exp>=8) + (s.exp>=16);
+ s = f31((static_cast<uint32>(s.exp)<<(31-i))+(log2(s.m>>1, 28)>>i), i) / lbe;
+ if(x.exp != -1 || x.m != 0x80000000)
+ {
+ i = (t.exp>=2) + (t.exp>=4) + (t.exp>=8);
+ f31 l = f31((static_cast<uint32>(t.exp)<<(31-i))+(log2(t.m>>1, 30)>>i), i) / lbe;
+ s = (x.exp<-1) ? (s-(f31(0x80000000, -1)-x)*l) : (s+(x-f31(0x80000000, -1))*l);
+ }
+ s = x.exp ? (s-t) : (t-s);
+ if(bsign)
+ {
+ if(z.exp >= 0)
+ {
+ sign &= (L|((z.m>>(31-z.exp))&1)) - 1;
+ for(z=f31((z.m<<(1+z.exp))&0xFFFFFFFF, -1); z.m<0x80000000; z.m<<=1,--z.exp) ;
+ }
+ if(z.exp == -1)
+ z = f31(0x80000000, 0) - z;
+ if(z.exp < -1)
+ {
+ z = z * pi;
+ z.m = sincos(z.m>>(1-z.exp), 30).first;
+ for(z.exp=1; z.m<0x80000000; z.m<<=1,--z.exp) ;
+ }
+ else
+ z = f31(0x80000000, 0);
+ }
+ if(L)
+ {
+ if(bsign)
+ {
+ f31 l(0x92868247, 0);
+ if(z.exp < 0)
+ {
+ uint32 m = log2((z.m+1)>>1, 27);
+ z = f31(-((static_cast<uint32>(z.exp)<<26)+(m>>5)), 5);
+ for(; z.m<0x80000000; z.m<<=1,--z.exp) ;
+ l = l + z / lbe;
+ }
+ sign = static_cast<unsigned>(x.exp&&(l.exp<s.exp||(l.exp==s.exp&&l.m<s.m))) << 15;
+ s = sign ? (s-l) : x.exp ? (l-s) : (l+s);
+ }
+ else
+ {
+ sign = static_cast<unsigned>(x.exp==0) << 15;
+ if(s.exp < -24)
+ return underflow<R>(sign);
+ if(s.exp > 15)
+ return overflow<R>(sign);
+ }
+ }
+ else
+ {
+ s = s * lbe;
+ uint32 m;
+ if(s.exp < 0)
+ {
+ m = s.m >> -s.exp;
+ s.exp = 0;
+ }
+ else
+ {
+ m = (s.m<<s.exp) & 0x7FFFFFFF;
+ s.exp = (s.m>>(31-s.exp));
+ }
+ s.m = exp2(m, 27);
+ if(!x.exp)
+ s = f31(0x80000000, 0) / s;
+ if(bsign)
+ {
+ if(z.exp < 0)
+ s = s * z;
+ s = pi / s;
+ if(s.exp < -24)
+ return underflow<R>(sign);
+ }
+ else if(z.exp > 0 && !(z.m&((1<<(31-z.exp))-1)))
+ return ((s.exp+14)<<10) + (s.m>>21);
+ if(s.exp > 15)
+ return overflow<R>(sign);
+ }
+ return fixed2half<R,31,false,false,true>(s.m, s.exp+14, sign);
+ }
+ /// \}
+
+ template<typename,typename,std::float_round_style> struct half_caster;
+ }
+
+ /// Half-precision floating-point type.
+ /// This class implements an IEEE-conformant half-precision floating-point type with the usual arithmetic
+ /// operators and conversions. It is implicitly convertible to single-precision floating-point, which makes artihmetic
+ /// expressions and functions with mixed-type operands to be of the most precise operand type.
+ ///
+ /// According to the C++98/03 definition, the half type is not a POD type. But according to C++11's less strict and
+ /// extended definitions it is both a standard layout type and a trivially copyable type (even if not a POD type), which
+ /// means it can be standard-conformantly copied using raw binary copies. But in this context some more words about the
+ /// actual size of the type. Although the half is representing an IEEE 16-bit type, it does not neccessarily have to be of
+ /// exactly 16-bits size. But on any reasonable implementation the actual binary representation of this type will most
+ /// probably not ivolve any additional "magic" or padding beyond the simple binary representation of the underlying 16-bit
+ /// IEEE number, even if not strictly guaranteed by the standard. But even then it only has an actual size of 16 bits if
+ /// your C++ implementation supports an unsigned integer type of exactly 16 bits width. But this should be the case on
+ /// nearly any reasonable platform.
+ ///
+ /// So if your C++ implementation is not totally exotic or imposes special alignment requirements, it is a reasonable
+ /// assumption that the data of a half is just comprised of the 2 bytes of the underlying IEEE representation.
+ class half
+ {
+ public:
+ /// \name Construction and assignment
+ /// \{
+
+ /// Default constructor.
+ /// This initializes the half to 0. Although this does not match the builtin types' default-initialization semantics
+ /// and may be less efficient than no initialization, it is needed to provide proper value-initialization semantics.
+ HALF_CONSTEXPR half() HALF_NOEXCEPT : data_() {}
+
+ /// Conversion constructor.
+ /// \param rhs float to convert
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ explicit half(float rhs) : data_(static_cast<detail::uint16>(detail::float2half<round_style>(rhs))) {}
+
+ /// Conversion to single-precision.
+ /// \return single precision value representing expression value
+ operator float() const { return detail::half2float<float>(data_); }
+
+ /// Assignment operator.
+ /// \param rhs single-precision value to copy from
+ /// \return reference to this half
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ half& operator=(float rhs) { data_ = static_cast<detail::uint16>(detail::float2half<round_style>(rhs)); return *this; }
+
+ /// \}
+ /// \name Arithmetic updates
+ /// \{
+
+ /// Arithmetic assignment.
+ /// \tparam T type of concrete half expression
+ /// \param rhs half expression to add
+ /// \return reference to this half
+ /// \exception FE_... according to operator+(half,half)
+ half& operator+=(half rhs) { return *this = *this + rhs; }
+
+ /// Arithmetic assignment.
+ /// \tparam T type of concrete half expression
+ /// \param rhs half expression to subtract
+ /// \return reference to this half
+ /// \exception FE_... according to operator-(half,half)
+ half& operator-=(half rhs) { return *this = *this - rhs; }
+
+ /// Arithmetic assignment.
+ /// \tparam T type of concrete half expression
+ /// \param rhs half expression to multiply with
+ /// \return reference to this half
+ /// \exception FE_... according to operator*(half,half)
+ half& operator*=(half rhs) { return *this = *this * rhs; }
+
+ /// Arithmetic assignment.
+ /// \tparam T type of concrete half expression
+ /// \param rhs half expression to divide by
+ /// \return reference to this half
+ /// \exception FE_... according to operator/(half,half)
+ half& operator/=(half rhs) { return *this = *this / rhs; }
+
+ /// Arithmetic assignment.
+ /// \param rhs single-precision value to add
+ /// \return reference to this half
+ /// \exception FE_... according to operator=()
+ half& operator+=(float rhs) { return *this = *this + rhs; }
+
+ /// Arithmetic assignment.
+ /// \param rhs single-precision value to subtract
+ /// \return reference to this half
+ /// \exception FE_... according to operator=()
+ half& operator-=(float rhs) { return *this = *this - rhs; }
+
+ /// Arithmetic assignment.
+ /// \param rhs single-precision value to multiply with
+ /// \return reference to this half
+ /// \exception FE_... according to operator=()
+ half& operator*=(float rhs) { return *this = *this * rhs; }
+
+ /// Arithmetic assignment.
+ /// \param rhs single-precision value to divide by
+ /// \return reference to this half
+ /// \exception FE_... according to operator=()
+ half& operator/=(float rhs) { return *this = *this / rhs; }
+
+ /// \}
+ /// \name Increment and decrement
+ /// \{
+
+ /// Prefix increment.
+ /// \return incremented half value
+ /// \exception FE_... according to operator+(half,half)
+ half& operator++() { return *this = *this + half(detail::binary, 0x3C00); }
+
+ /// Prefix decrement.
+ /// \return decremented half value
+ /// \exception FE_... according to operator-(half,half)
+ half& operator--() { return *this = *this + half(detail::binary, 0xBC00); }
+
+ /// Postfix increment.
+ /// \return non-incremented half value
+ /// \exception FE_... according to operator+(half,half)
+ half operator++(int) { half out(*this); ++*this; return out; }
+
+ /// Postfix decrement.
+ /// \return non-decremented half value
+ /// \exception FE_... according to operator-(half,half)
+ half operator--(int) { half out(*this); --*this; return out; }
+ /// \}
+
+ private:
+ /// Rounding mode to use
+ static const std::float_round_style round_style = (std::float_round_style)(HALF_ROUND_STYLE);
+
+ /// Constructor.
+ /// \param bits binary representation to set half to
+ HALF_CONSTEXPR half(detail::binary_t, unsigned int bits) HALF_NOEXCEPT : data_(static_cast<detail::uint16>(bits)) {}
+
+ /// Internal binary representation
+ detail::uint16 data_;
+
+ #ifndef HALF_DOXYGEN_ONLY
+ friend HALF_CONSTEXPR_NOERR bool operator==(half, half);
+ friend HALF_CONSTEXPR_NOERR bool operator!=(half, half);
+ friend HALF_CONSTEXPR_NOERR bool operator<(half, half);
+ friend HALF_CONSTEXPR_NOERR bool operator>(half, half);
+ friend HALF_CONSTEXPR_NOERR bool operator<=(half, half);
+ friend HALF_CONSTEXPR_NOERR bool operator>=(half, half);
+ friend HALF_CONSTEXPR half operator-(half);
+ friend half operator+(half, half);
+ friend half operator-(half, half);
+ friend half operator*(half, half);
+ friend half operator/(half, half);
+ template<typename charT,typename traits> friend std::basic_ostream<charT,traits>& operator<<(std::basic_ostream<charT,traits>&, half);
+ template<typename charT,typename traits> friend std::basic_istream<charT,traits>& operator>>(std::basic_istream<charT,traits>&, half&);
+ friend HALF_CONSTEXPR half fabs(half);
+ friend half fmod(half, half);
+ friend half remainder(half, half);
+ friend half remquo(half, half, int*);
+ friend half fma(half, half, half);
+ friend HALF_CONSTEXPR_NOERR half fmax(half, half);
+ friend HALF_CONSTEXPR_NOERR half fmin(half, half);
+ friend half fdim(half, half);
+ friend half nanh(const char*);
+ friend half exp(half);
+ friend half exp2(half);
+ friend half expm1(half);
+ friend half log(half);
+ friend half log10(half);
+ friend half log2(half);
+ friend half log1p(half);
+ friend half sqrt(half);
+ friend half rsqrt(half);
+ friend half cbrt(half);
+ friend half hypot(half, half);
+ friend half hypot(half, half, half);
+ friend half pow(half, half);
+ friend void sincos(half, half*, half*);
+ friend half sin(half);
+ friend half cos(half);
+ friend half tan(half);
+ friend half asin(half);
+ friend half acos(half);
+ friend half atan(half);
+ friend half atan2(half, half);
+ friend half sinh(half);
+ friend half cosh(half);
+ friend half tanh(half);
+ friend half asinh(half);
+ friend half acosh(half);
+ friend half atanh(half);
+ friend half erf(half);
+ friend half erfc(half);
+ friend half lgamma(half);
+ friend half tgamma(half);
+ friend half ceil(half);
+ friend half floor(half);
+ friend half trunc(half);
+ friend half round(half);
+ friend long lround(half);
+ friend half rint(half);
+ friend long lrint(half);
+ friend half nearbyint(half);
+ #ifdef HALF_ENABLE_CPP11_LONG_LONG
+ friend long long llround(half);
+ friend long long llrint(half);
+ #endif
+ friend half frexp(half, int*);
+ friend half scalbln(half, long);
+ friend half modf(half, half*);
+ friend int ilogb(half);
+ friend half logb(half);
+ friend half nextafter(half, half);
+ friend half nexttoward(half, long double);
+ friend HALF_CONSTEXPR half copysign(half, half);
+ friend HALF_CONSTEXPR int fpclassify(half);
+ friend HALF_CONSTEXPR bool isfinite(half);
+ friend HALF_CONSTEXPR bool isinf(half);
+ friend HALF_CONSTEXPR bool isnan(half);
+ friend HALF_CONSTEXPR bool isnormal(half);
+ friend HALF_CONSTEXPR bool signbit(half);
+ friend HALF_CONSTEXPR bool isgreater(half, half);
+ friend HALF_CONSTEXPR bool isgreaterequal(half, half);
+ friend HALF_CONSTEXPR bool isless(half, half);
+ friend HALF_CONSTEXPR bool islessequal(half, half);
+ friend HALF_CONSTEXPR bool islessgreater(half, half);
+ template<typename,typename,std::float_round_style> friend struct detail::half_caster;
+ friend class std::numeric_limits<half>;
+ #if HALF_ENABLE_CPP11_HASH
+ friend struct std::hash<half>;
+ #endif
+ #if HALF_ENABLE_CPP11_USER_LITERALS
+ friend half literal::operator "" _h(long double);
+ #endif
+ #endif
+ };
+
+#if HALF_ENABLE_CPP11_USER_LITERALS
+ namespace literal
+ {
+ /// Half literal.
+ /// While this returns a properly rounded half-precision value, half literals can unfortunately not be constant
+ /// expressions due to rather involved conversions. So don't expect this to be a literal literal without involving
+ /// conversion operations at runtime. It is a convenience feature, not a performance optimization.
+ /// \param value literal value
+ /// \return half with of given value (possibly rounded)
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half operator "" _h(long double value) { return half(detail::binary, detail::float2half<half::round_style>(value)); }
+ }
+#endif
+
+ namespace detail
+ {
+ /// Helper class for half casts.
+ /// This class template has to be specialized for all valid cast arguments to define an appropriate static
+ /// `cast` member function and a corresponding `type` member denoting its return type.
+ /// \tparam T destination type
+ /// \tparam U source type
+ /// \tparam R rounding mode to use
+ template<typename T,typename U,std::float_round_style R=(std::float_round_style)(HALF_ROUND_STYLE)> struct half_caster {};
+ template<typename U,std::float_round_style R> struct half_caster<half,U,R>
+ {
+ #if HALF_ENABLE_CPP11_STATIC_ASSERT && HALF_ENABLE_CPP11_TYPE_TRAITS
+ static_assert(std::is_arithmetic<U>::value, "half_cast from non-arithmetic type unsupported");
+ #endif
+
+ static half cast(U arg) { return cast_impl(arg, is_float<U>()); };
+
+ private:
+ static half cast_impl(U arg, true_type) { return half(binary, float2half<R>(arg)); }
+ static half cast_impl(U arg, false_type) { return half(binary, int2half<R>(arg)); }
+ };
+ template<typename T,std::float_round_style R> struct half_caster<T,half,R>
+ {
+ #if HALF_ENABLE_CPP11_STATIC_ASSERT && HALF_ENABLE_CPP11_TYPE_TRAITS
+ static_assert(std::is_arithmetic<T>::value, "half_cast to non-arithmetic type unsupported");
+ #endif
+
+ static T cast(half arg) { return cast_impl(arg, is_float<T>()); }
+
+ private:
+ static T cast_impl(half arg, true_type) { return half2float<T>(arg.data_); }
+ static T cast_impl(half arg, false_type) { return half2int<R,true,true,T>(arg.data_); }
+ };
+ template<std::float_round_style R> struct half_caster<half,half,R>
+ {
+ static half cast(half arg) { return arg; }
+ };
+ }
+}
+
+/// Extensions to the C++ standard library.
+namespace std
+{
+ /// Numeric limits for half-precision floats.
+ /// **See also:** Documentation for [std::numeric_limits](https://en.cppreference.com/w/cpp/types/numeric_limits)
+ template<> class numeric_limits<half_float::half>
+ {
+ public:
+ /// Is template specialization.
+ static HALF_CONSTEXPR_CONST bool is_specialized = true;
+
+ /// Supports signed values.
+ static HALF_CONSTEXPR_CONST bool is_signed = true;
+
+ /// Is not an integer type.
+ static HALF_CONSTEXPR_CONST bool is_integer = false;
+
+ /// Is not exact.
+ static HALF_CONSTEXPR_CONST bool is_exact = false;
+
+ /// Doesn't provide modulo arithmetic.
+ static HALF_CONSTEXPR_CONST bool is_modulo = false;
+
+ /// Has a finite set of values.
+ static HALF_CONSTEXPR_CONST bool is_bounded = true;
+
+ /// IEEE conformant.
+ static HALF_CONSTEXPR_CONST bool is_iec559 = true;
+
+ /// Supports infinity.
+ static HALF_CONSTEXPR_CONST bool has_infinity = true;
+
+ /// Supports quiet NaNs.
+ static HALF_CONSTEXPR_CONST bool has_quiet_NaN = true;
+
+ /// Supports signaling NaNs.
+ static HALF_CONSTEXPR_CONST bool has_signaling_NaN = true;
+
+ /// Supports subnormal values.
+ static HALF_CONSTEXPR_CONST float_denorm_style has_denorm = denorm_present;
+
+ /// Supports no denormalization detection.
+ static HALF_CONSTEXPR_CONST bool has_denorm_loss = false;
+
+ #if HALF_ERRHANDLING_THROWS
+ static HALF_CONSTEXPR_CONST bool traps = true;
+ #else
+ /// Traps only if [HALF_ERRHANDLING_THROW_...](\ref HALF_ERRHANDLING_THROW_INVALID) is acitvated.
+ static HALF_CONSTEXPR_CONST bool traps = false;
+ #endif
+
+ /// Does not support no pre-rounding underflow detection.
+ static HALF_CONSTEXPR_CONST bool tinyness_before = false;
+
+ /// Rounding mode.
+ static HALF_CONSTEXPR_CONST float_round_style round_style = half_float::half::round_style;
+
+ /// Significant digits.
+ static HALF_CONSTEXPR_CONST int digits = 11;
+
+ /// Significant decimal digits.
+ static HALF_CONSTEXPR_CONST int digits10 = 3;
+
+ /// Required decimal digits to represent all possible values.
+ static HALF_CONSTEXPR_CONST int max_digits10 = 5;
+
+ /// Number base.
+ static HALF_CONSTEXPR_CONST int radix = 2;
+
+ /// One more than smallest exponent.
+ static HALF_CONSTEXPR_CONST int min_exponent = -13;
+
+ /// Smallest normalized representable power of 10.
+ static HALF_CONSTEXPR_CONST int min_exponent10 = -4;
+
+ /// One more than largest exponent
+ static HALF_CONSTEXPR_CONST int max_exponent = 16;
+
+ /// Largest finitely representable power of 10.
+ static HALF_CONSTEXPR_CONST int max_exponent10 = 4;
+
+ /// Smallest positive normal value.
+ static HALF_CONSTEXPR half_float::half min() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x0400); }
+
+ /// Smallest finite value.
+ static HALF_CONSTEXPR half_float::half lowest() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0xFBFF); }
+
+ /// Largest finite value.
+ static HALF_CONSTEXPR half_float::half max() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7BFF); }
+
+ /// Difference between 1 and next representable value.
+ static HALF_CONSTEXPR half_float::half epsilon() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x1400); }
+
+ /// Maximum rounding error in ULP (units in the last place).
+ static HALF_CONSTEXPR half_float::half round_error() HALF_NOTHROW
+ { return half_float::half(half_float::detail::binary, (round_style==std::round_to_nearest) ? 0x3800 : 0x3C00); }
+
+ /// Positive infinity.
+ static HALF_CONSTEXPR half_float::half infinity() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7C00); }
+
+ /// Quiet NaN.
+ static HALF_CONSTEXPR half_float::half quiet_NaN() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7FFF); }
+
+ /// Signaling NaN.
+ static HALF_CONSTEXPR half_float::half signaling_NaN() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7DFF); }
+
+ /// Smallest positive subnormal value.
+ static HALF_CONSTEXPR half_float::half denorm_min() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x0001); }
+ };
+
+#if HALF_ENABLE_CPP11_HASH
+ /// Hash function for half-precision floats.
+ /// This is only defined if C++11 `std::hash` is supported and enabled.
+ ///
+ /// **See also:** Documentation for [std::hash](https://en.cppreference.com/w/cpp/utility/hash)
+ template<> struct hash<half_float::half>
+ {
+ /// Type of function argument.
+ typedef half_float::half argument_type;
+
+ /// Function return type.
+ typedef size_t result_type;
+
+ /// Compute hash function.
+ /// \param arg half to hash
+ /// \return hash value
+ result_type operator()(argument_type arg) const { return hash<half_float::detail::uint16>()(arg.data_&-static_cast<unsigned>(arg.data_!=0x8000)); }
+ };
+#endif
+}
+
+namespace half_float
+{
+ /// \anchor compop
+ /// \name Comparison operators
+ /// \{
+
+ /// Comparison for equality.
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if operands equal
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool operator==(half x, half y)
+ {
+ return !detail::compsignal(x.data_, y.data_) && (x.data_==y.data_ || !((x.data_|y.data_)&0x7FFF));
+ }
+
+ /// Comparison for inequality.
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if operands not equal
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool operator!=(half x, half y)
+ {
+ return detail::compsignal(x.data_, y.data_) || (x.data_!=y.data_ && ((x.data_|y.data_)&0x7FFF));
+ }
+
+ /// Comparison for less than.
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x less than \a y
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool operator<(half x, half y)
+ {
+ return !detail::compsignal(x.data_, y.data_) &&
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) < ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
+ }
+
+ /// Comparison for greater than.
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x greater than \a y
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool operator>(half x, half y)
+ {
+ return !detail::compsignal(x.data_, y.data_) &&
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) > ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
+ }
+
+ /// Comparison for less equal.
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x less equal \a y
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool operator<=(half x, half y)
+ {
+ return !detail::compsignal(x.data_, y.data_) &&
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) <= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
+ }
+
+ /// Comparison for greater equal.
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x greater equal \a y
+ /// \retval false else
+ /// \exception FE_INVALID if \a x or \a y is NaN
+ inline HALF_CONSTEXPR_NOERR bool operator>=(half x, half y)
+ {
+ return !detail::compsignal(x.data_, y.data_) &&
+ ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) >= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15));
+ }
+
+ /// \}
+ /// \anchor arithmetics
+ /// \name Arithmetic operators
+ /// \{
+
+ /// Identity.
+ /// \param arg operand
+ /// \return unchanged operand
+ inline HALF_CONSTEXPR half operator+(half arg) { return arg; }
+
+ /// Negation.
+ /// \param arg operand
+ /// \return negated operand
+ inline HALF_CONSTEXPR half operator-(half arg) { return half(detail::binary, arg.data_^0x8000); }
+
+ /// Addition.
+ /// This operation is exact to rounding for all rounding modes.
+ /// \param x left operand
+ /// \param y right operand
+ /// \return sum of half expressions
+ /// \exception FE_INVALID if \a x and \a y are infinities with different signs or signaling NaNs
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half operator+(half x, half y)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)+detail::half2float<detail::internal_t>(y.data_)));
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF;
+ bool sub = ((x.data_^y.data_)&0x8000) != 0;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : (absy!=0x7C00) ? x.data_ :
+ (sub && absx==0x7C00) ? detail::invalid() : y.data_);
+ if(!absx)
+ return absy ? y : half(detail::binary, (half::round_style==std::round_toward_neg_infinity) ? (x.data_|y.data_) : (x.data_&y.data_));
+ if(!absy)
+ return x;
+ unsigned int sign = ((sub && absy>absx) ? y.data_ : x.data_) & 0x8000;
+ if(absy > absx)
+ std::swap(absx, absy);
+ int exp = (absx>>10) + (absx<=0x3FF), d = exp - (absy>>10) - (absy<=0x3FF), mx = ((absx&0x3FF)|((absx>0x3FF)<<10)) << 3, my;
+ if(d < 13)
+ {
+ my = ((absy&0x3FF)|((absy>0x3FF)<<10)) << 3;
+ my = (my>>d) | ((my&((1<<d)-1))!=0);
+ }
+ else
+ my = 1;
+ if(sub)
+ {
+ if(!(mx-=my))
+ return half(detail::binary, static_cast<unsigned>(half::round_style==std::round_toward_neg_infinity)<<15);
+ for(; mx<0x2000 && exp>1; mx<<=1,--exp) ;
+ }
+ else
+ {
+ mx += my;
+ int i = mx >> 14;
+ if((exp+=i) > 30)
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
+ mx = (mx>>i) | (mx&i);
+ }
+ return half(detail::binary, detail::rounded<half::round_style,false>(sign+((exp-1)<<10)+(mx>>3), (mx>>2)&1, (mx&0x3)!=0));
+ #endif
+ }
+
+ /// Subtraction.
+ /// This operation is exact to rounding for all rounding modes.
+ /// \param x left operand
+ /// \param y right operand
+ /// \return difference of half expressions
+ /// \exception FE_INVALID if \a x and \a y are infinities with equal signs or signaling NaNs
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half operator-(half x, half y)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)-detail::half2float<detail::internal_t>(y.data_)));
+ #else
+ return x + -y;
+ #endif
+ }
+
+ /// Multiplication.
+ /// This operation is exact to rounding for all rounding modes.
+ /// \param x left operand
+ /// \param y right operand
+ /// \return product of half expressions
+ /// \exception FE_INVALID if multiplying 0 with infinity or if \a x or \a y is signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half operator*(half x, half y)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)*detail::half2float<detail::internal_t>(y.data_)));
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = -16;
+ unsigned int sign = (x.data_^y.data_) & 0x8000;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
+ ((absx==0x7C00 && !absy)||(absy==0x7C00 && !absx)) ? detail::invalid() : (sign|0x7C00));
+ if(!absx || !absy)
+ return half(detail::binary, sign);
+ for(; absx<0x400; absx<<=1,--exp) ;
+ for(; absy<0x400; absy<<=1,--exp) ;
+ detail::uint32 m = static_cast<detail::uint32>((absx&0x3FF)|0x400) * static_cast<detail::uint32>((absy&0x3FF)|0x400);
+ int i = m >> 21, s = m & i;
+ exp += (absx>>10) + (absy>>10) + i;
+ if(exp > 29)
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
+ else if(exp < -11)
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
+ return half(detail::binary, detail::fixed2half<half::round_style,20,false,false,false>(m>>i, exp, sign, s));
+ #endif
+ }
+
+ /// Division.
+ /// This operation is exact to rounding for all rounding modes.
+ /// \param x left operand
+ /// \param y right operand
+ /// \return quotient of half expressions
+ /// \exception FE_INVALID if dividing 0s or infinities with each other or if \a x or \a y is signaling NaN
+ /// \exception FE_DIVBYZERO if dividing finite value by 0
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half operator/(half x, half y)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(detail::half2float<detail::internal_t>(x.data_)/detail::half2float<detail::internal_t>(y.data_)));
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = 14;
+ unsigned int sign = (x.data_^y.data_) & 0x8000;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
+ (absx==absy) ? detail::invalid() : (sign|((absx==0x7C00) ? 0x7C00 : 0)));
+ if(!absx)
+ return half(detail::binary, absy ? sign : detail::invalid());
+ if(!absy)
+ return half(detail::binary, detail::pole(sign));
+ for(; absx<0x400; absx<<=1,--exp) ;
+ for(; absy<0x400; absy<<=1,++exp) ;
+ detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400;
+ int i = mx < my;
+ exp += (absx>>10) - (absy>>10) - i;
+ if(exp > 29)
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
+ else if(exp < -11)
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
+ mx <<= 12 + i;
+ my <<= 1;
+ return half(detail::binary, detail::fixed2half<half::round_style,11,false,false,false>(mx/my, exp, sign, mx%my!=0));
+ #endif
+ }
+
+ /// \}
+ /// \anchor streaming
+ /// \name Input and output
+ /// \{
+
+ /// Output operator.
+ /// This uses the built-in functionality for streaming out floating-point numbers.
+ /// \param out output stream to write into
+ /// \param arg half expression to write
+ /// \return reference to output stream
+ template<typename charT,typename traits> std::basic_ostream<charT,traits>& operator<<(std::basic_ostream<charT,traits> &out, half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return out << detail::half2float<detail::internal_t>(arg.data_);
+ #else
+ return out << detail::half2float<float>(arg.data_);
+ #endif
+ }
+
+ /// Input operator.
+ /// This uses the built-in functionality for streaming in floating-point numbers, specifically double precision floating
+ /// point numbers (unless overridden with [HALF_ARITHMETIC_TYPE](\ref HALF_ARITHMETIC_TYPE)). So the input string is first
+ /// rounded to double precision using the underlying platform's current floating-point rounding mode before being rounded
+ /// to half-precision using the library's half-precision rounding mode.
+ /// \param in input stream to read from
+ /// \param arg half to read into
+ /// \return reference to input stream
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ template<typename charT,typename traits> std::basic_istream<charT,traits>& operator>>(std::basic_istream<charT,traits> &in, half &arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ detail::internal_t f;
+ #else
+ double f;
+ #endif
+ if(in >> f)
+ arg.data_ = detail::float2half<half::round_style>(f);
+ return in;
+ }
+
+ /// \}
+ /// \anchor basic
+ /// \name Basic mathematical operations
+ /// \{
+
+ /// Absolute value.
+ /// **See also:** Documentation for [std::fabs](https://en.cppreference.com/w/cpp/numeric/math/fabs).
+ /// \param arg operand
+ /// \return absolute value of \a arg
+ inline HALF_CONSTEXPR half fabs(half arg) { return half(detail::binary, arg.data_&0x7FFF); }
+
+ /// Absolute value.
+ /// **See also:** Documentation for [std::abs](https://en.cppreference.com/w/cpp/numeric/math/fabs).
+ /// \param arg operand
+ /// \return absolute value of \a arg
+ inline HALF_CONSTEXPR half abs(half arg) { return fabs(arg); }
+
+ /// Remainder of division.
+ /// **See also:** Documentation for [std::fmod](https://en.cppreference.com/w/cpp/numeric/math/fmod).
+ /// \param x first operand
+ /// \param y second operand
+ /// \return remainder of floating-point division.
+ /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN
+ inline half fmod(half x, half y)
+ {
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, sign = x.data_ & 0x8000;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
+ (absx==0x7C00) ? detail::invalid() : x.data_);
+ if(!absy)
+ return half(detail::binary, detail::invalid());
+ if(!absx)
+ return x;
+ if(absx == absy)
+ return half(detail::binary, sign);
+ return half(detail::binary, sign|detail::mod<false,false>(absx, absy));
+ }
+
+ /// Remainder of division.
+ /// **See also:** Documentation for [std::remainder](https://en.cppreference.com/w/cpp/numeric/math/remainder).
+ /// \param x first operand
+ /// \param y second operand
+ /// \return remainder of floating-point division.
+ /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN
+ inline half remainder(half x, half y)
+ {
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, sign = x.data_ & 0x8000;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
+ (absx==0x7C00) ? detail::invalid() : x.data_);
+ if(!absy)
+ return half(detail::binary, detail::invalid());
+ if(absx == absy)
+ return half(detail::binary, sign);
+ return half(detail::binary, sign^detail::mod<false,true>(absx, absy));
+ }
+
+ /// Remainder of division.
+ /// **See also:** Documentation for [std::remquo](https://en.cppreference.com/w/cpp/numeric/math/remquo).
+ /// \param x first operand
+ /// \param y second operand
+ /// \param quo address to store some bits of quotient at
+ /// \return remainder of floating-point division.
+ /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN
+ inline half remquo(half x, half y, int *quo)
+ {
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, value = x.data_ & 0x8000;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
+ (absx==0x7C00) ? detail::invalid() : (*quo = 0, x.data_));
+ if(!absy)
+ return half(detail::binary, detail::invalid());
+ bool qsign = ((value^y.data_)&0x8000) != 0;
+ int q = 1;
+ if(absx != absy)
+ value ^= detail::mod<true, true>(absx, absy, &q);
+ return *quo = qsign ? -q : q, half(detail::binary, value);
+ }
+
+ /// Fused multiply add.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::fma](https://en.cppreference.com/w/cpp/numeric/math/fma).
+ /// \param x first operand
+ /// \param y second operand
+ /// \param z third operand
+ /// \return ( \a x * \a y ) + \a z rounded as one operation.
+ /// \exception FE_INVALID according to operator*() and operator+() unless any argument is a quiet NaN and no argument is a signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding the final addition
+ inline half fma(half x, half y, half z)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ detail::internal_t fx = detail::half2float<detail::internal_t>(x.data_), fy = detail::half2float<detail::internal_t>(y.data_), fz = detail::half2float<detail::internal_t>(z.data_);
+ #if HALF_ENABLE_CPP11_CMATH && FP_FAST_FMA
+ return half(detail::binary, detail::float2half<half::round_style>(std::fma(fx, fy, fz)));
+ #else
+ return half(detail::binary, detail::float2half<half::round_style>(fx*fy+fz));
+ #endif
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, absz = z.data_ & 0x7FFF, exp = -15;
+ unsigned int sign = (x.data_^y.data_) & 0x8000;
+ bool sub = ((sign^z.data_)&0x8000) != 0;
+ if(absx >= 0x7C00 || absy >= 0x7C00 || absz >= 0x7C00)
+ return (absx>0x7C00 || absy>0x7C00 || absz>0x7C00) ? half(detail::binary, detail::signal(x.data_, y.data_, z.data_)) :
+ (absx==0x7C00) ? half(detail::binary, (!absy || (sub && absz==0x7C00)) ? detail::invalid() : (sign|0x7C00)) :
+ (absy==0x7C00) ? half(detail::binary, (!absx || (sub && absz==0x7C00)) ? detail::invalid() : (sign|0x7C00)) : z;
+ if(!absx || !absy)
+ return absz ? z : half(detail::binary, (half::round_style==std::round_toward_neg_infinity) ? (z.data_|sign) : (z.data_&sign));
+ for(; absx<0x400; absx<<=1,--exp) ;
+ for(; absy<0x400; absy<<=1,--exp) ;
+ detail::uint32 m = static_cast<detail::uint32>((absx&0x3FF)|0x400) * static_cast<detail::uint32>((absy&0x3FF)|0x400);
+ int i = m >> 21;
+ exp += (absx>>10) + (absy>>10) + i;
+ m <<= 3 - i;
+ if(absz)
+ {
+ int expz = 0;
+ for(; absz<0x400; absz<<=1,--expz) ;
+ expz += absz >> 10;
+ detail::uint32 mz = static_cast<detail::uint32>((absz&0x3FF)|0x400) << 13;
+ if(expz > exp || (expz == exp && mz > m))
+ {
+ std::swap(m, mz);
+ std::swap(exp, expz);
+ if(sub)
+ sign = z.data_ & 0x8000;
+ }
+ int d = exp - expz;
+ mz = (d<23) ? ((mz>>d)|((mz&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
+ if(sub)
+ {
+ m = m - mz;
+ if(!m)
+ return half(detail::binary, static_cast<unsigned>(half::round_style==std::round_toward_neg_infinity)<<15);
+ for(; m<0x800000; m<<=1,--exp) ;
+ }
+ else
+ {
+ m += mz;
+ i = m >> 24;
+ m = (m>>i) | (m&i);
+ exp += i;
+ }
+ }
+ if(exp > 30)
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
+ else if(exp < -10)
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
+ return half(detail::binary, detail::fixed2half<half::round_style,23,false,false,false>(m, exp-1, sign));
+ #endif
+ }
+
+ /// Maximum of half expressions.
+ /// **See also:** Documentation for [std::fmax](https://en.cppreference.com/w/cpp/numeric/math/fmax).
+ /// \param x first operand
+ /// \param y second operand
+ /// \return maximum of operands, ignoring quiet NaNs
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
+ inline HALF_CONSTEXPR_NOERR half fmax(half x, half y)
+ {
+ return half(detail::binary, (!isnan(y) && (isnan(x) || (x.data_^(0x8000|(0x8000-(x.data_>>15)))) <
+ (y.data_^(0x8000|(0x8000-(y.data_>>15)))))) ? detail::select(y.data_, x.data_) : detail::select(x.data_, y.data_));
+ }
+
+ /// Minimum of half expressions.
+ /// **See also:** Documentation for [std::fmin](https://en.cppreference.com/w/cpp/numeric/math/fmin).
+ /// \param x first operand
+ /// \param y second operand
+ /// \return minimum of operands, ignoring quiet NaNs
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
+ inline HALF_CONSTEXPR_NOERR half fmin(half x, half y)
+ {
+ return half(detail::binary, (!isnan(y) && (isnan(x) || (x.data_^(0x8000|(0x8000-(x.data_>>15)))) >
+ (y.data_^(0x8000|(0x8000-(y.data_>>15)))))) ? detail::select(y.data_, x.data_) : detail::select(x.data_, y.data_));
+ }
+
+ /// Positive difference.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::fdim](https://en.cppreference.com/w/cpp/numeric/math/fdim).
+ /// \param x first operand
+ /// \param y second operand
+ /// \return \a x - \a y or 0 if difference negative
+ /// \exception FE_... according to operator-(half,half)
+ inline half fdim(half x, half y)
+ {
+ if(isnan(x) || isnan(y))
+ return half(detail::binary, detail::signal(x.data_, y.data_));
+ return (x.data_^(0x8000|(0x8000-(x.data_>>15)))) <= (y.data_^(0x8000|(0x8000-(y.data_>>15)))) ? half(detail::binary, 0) : (x-y);
+ }
+
+ /// Get NaN value.
+ /// **See also:** Documentation for [std::nan](https://en.cppreference.com/w/cpp/numeric/math/nan).
+ /// \param arg string code
+ /// \return quiet NaN
+ inline half nanh(const char *arg)
+ {
+ unsigned int value = 0x7FFF;
+ while(*arg)
+ value ^= static_cast<unsigned>(*arg++) & 0xFF;
+ return half(detail::binary, value);
+ }
+
+ /// \}
+ /// \anchor exponential
+ /// \name Exponential functions
+ /// \{
+
+ /// Exponential function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::exp](https://en.cppreference.com/w/cpp/numeric/math/exp).
+ /// \param arg function argument
+ /// \return e raised to \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half exp(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::exp(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, e = (abs>>10) + (abs<=0x3FF), exp;
+ if(!abs)
+ return half(detail::binary, 0x3C00);
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? (0x7C00&((arg.data_>>15)-1U)) : detail::signal(arg.data_));
+ if(abs >= 0x4C80)
+ return half(detail::binary, (arg.data_&0x8000) ? detail::underflow<half::round_style>() : detail::overflow<half::round_style>());
+ detail::uint32 m = detail::multiply64(static_cast<detail::uint32>((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29);
+ if(e < 14)
+ {
+ exp = 0;
+ m >>= 14 - e;
+ }
+ else
+ {
+ exp = m >> (45-e);
+ m = (m<<(e-14)) & 0x7FFFFFFF;
+ }
+ return half(detail::binary, detail::exp2_post<half::round_style>(m, exp, (arg.data_&0x8000)!=0, 0, 26));
+ #endif
+ }
+
+ /// Binary exponential.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::exp2](https://en.cppreference.com/w/cpp/numeric/math/exp2).
+ /// \param arg function argument
+ /// \return 2 raised to \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half exp2(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::exp2(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, e = (abs>>10) + (abs<=0x3FF), exp = (abs&0x3FF) + ((abs>0x3FF)<<10);
+ if(!abs)
+ return half(detail::binary, 0x3C00);
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? (0x7C00&((arg.data_>>15)-1U)) : detail::signal(arg.data_));
+ if(abs >= 0x4E40)
+ return half(detail::binary, (arg.data_&0x8000) ? detail::underflow<half::round_style>() : detail::overflow<half::round_style>());
+ return half(detail::binary, detail::exp2_post<half::round_style>(
+ (static_cast<detail::uint32>(exp)<<(6+e))&0x7FFFFFFF, exp>>(25-e), (arg.data_&0x8000)!=0, 0, 28));
+ #endif
+ }
+
+ /// Exponential minus one.
+ /// This function may be 1 ULP off the correctly rounded exact result in <0.05% of inputs for `std::round_to_nearest`
+ /// and in <1% of inputs for any other rounding mode.
+ ///
+ /// **See also:** Documentation for [std::expm1](https://en.cppreference.com/w/cpp/numeric/math/expm1).
+ /// \param arg function argument
+ /// \return e raised to \a arg and subtracted by 1
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half expm1(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::expm1(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000, e = (abs>>10) + (abs<=0x3FF), exp;
+ if(!abs)
+ return arg;
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? (0x7C00+(sign>>1)) : detail::signal(arg.data_));
+ if(abs >= 0x4A00)
+ return half(detail::binary, (arg.data_&0x8000) ? detail::rounded<half::round_style,true>(0xBBFF, 1, 1) : detail::overflow<half::round_style>());
+ detail::uint32 m = detail::multiply64(static_cast<detail::uint32>((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29);
+ if(e < 14)
+ {
+ exp = 0;
+ m >>= 14 - e;
+ }
+ else
+ {
+ exp = m >> (45-e);
+ m = (m<<(e-14)) & 0x7FFFFFFF;
+ }
+ m = detail::exp2(m);
+ if(sign)
+ {
+ int s = 0;
+ if(m > 0x80000000)
+ {
+ ++exp;
+ m = detail::divide64(0x80000000, m, s);
+ }
+ m = 0x80000000 - ((m>>exp)|((m&((static_cast<detail::uint32>(1)<<exp)-1))!=0)|s);
+ exp = 0;
+ }
+ else
+ m -= (exp<31) ? (0x80000000>>exp) : 1;
+ for(exp+=14; m<0x80000000 && exp; m<<=1,--exp) ;
+ if(exp > 29)
+ return half(detail::binary, detail::overflow<half::round_style>());
+ return half(detail::binary, detail::rounded<half::round_style,true>(sign+(exp<<10)+(m>>21), (m>>20)&1, (m&0xFFFFF)!=0));
+ #endif
+ }
+
+ /// Natural logarithm.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::log](https://en.cppreference.com/w/cpp/numeric/math/log).
+ /// \param arg function argument
+ /// \return logarithm of \a arg to base e
+ /// \exception FE_INVALID for signaling NaN or negative argument
+ /// \exception FE_DIVBYZERO for 0
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half log(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::log(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = -15;
+ if(!abs)
+ return half(detail::binary, detail::pole(0x8000));
+ if(arg.data_ & 0x8000)
+ return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs >= 0x7C00)
+ return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
+ for(; abs<0x400; abs<<=1,--exp) ;
+ exp += abs >> 10;
+ return half(detail::binary, detail::log2_post<half::round_style,0xB8AA3B2A>(
+ detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 27)+8, exp, 17));
+ #endif
+ }
+
+ /// Common logarithm.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::log10](https://en.cppreference.com/w/cpp/numeric/math/log10).
+ /// \param arg function argument
+ /// \return logarithm of \a arg to base 10
+ /// \exception FE_INVALID for signaling NaN or negative argument
+ /// \exception FE_DIVBYZERO for 0
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half log10(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::log10(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = -15;
+ if(!abs)
+ return half(detail::binary, detail::pole(0x8000));
+ if(arg.data_ & 0x8000)
+ return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs >= 0x7C00)
+ return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
+ switch(abs)
+ {
+ case 0x4900: return half(detail::binary, 0x3C00);
+ case 0x5640: return half(detail::binary, 0x4000);
+ case 0x63D0: return half(detail::binary, 0x4200);
+ case 0x70E2: return half(detail::binary, 0x4400);
+ }
+ for(; abs<0x400; abs<<=1,--exp) ;
+ exp += abs >> 10;
+ return half(detail::binary, detail::log2_post<half::round_style,0xD49A784C>(
+ detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 27)+8, exp, 16));
+ #endif
+ }
+
+ /// Binary logarithm.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::log2](https://en.cppreference.com/w/cpp/numeric/math/log2).
+ /// \param arg function argument
+ /// \return logarithm of \a arg to base 2
+ /// \exception FE_INVALID for signaling NaN or negative argument
+ /// \exception FE_DIVBYZERO for 0
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half log2(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::log2(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = -15, s = 0;
+ if(!abs)
+ return half(detail::binary, detail::pole(0x8000));
+ if(arg.data_ & 0x8000)
+ return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs >= 0x7C00)
+ return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
+ if(abs == 0x3C00)
+ return half(detail::binary, 0);
+ for(; abs<0x400; abs<<=1,--exp) ;
+ exp += (abs>>10);
+ if(!(abs&0x3FF))
+ {
+ unsigned int value = static_cast<unsigned>(exp<0) << 15, m = std::abs(exp) << 6;
+ for(exp=18; m<0x400; m<<=1,--exp) ;
+ return half(detail::binary, value+(exp<<10)+m);
+ }
+ detail::uint32 ilog = exp, sign = detail::sign_mask(ilog), m =
+ (((ilog<<27)+(detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 28)>>4))^sign) - sign;
+ if(!m)
+ return half(detail::binary, 0);
+ for(exp=14; m<0x8000000 && exp; m<<=1,--exp) ;
+ for(; m>0xFFFFFFF; m>>=1,++exp)
+ s |= m & 1;
+ return half(detail::binary, detail::fixed2half<half::round_style,27,false,false,true>(m, exp, sign&0x8000, s));
+ #endif
+ }
+
+ /// Natural logarithm plus one.
+ /// This function may be 1 ULP off the correctly rounded exact result in <0.05% of inputs for `std::round_to_nearest`
+ /// and in ~1% of inputs for any other rounding mode.
+ ///
+ /// **See also:** Documentation for [std::log1p](https://en.cppreference.com/w/cpp/numeric/math/log1p).
+ /// \param arg function argument
+ /// \return logarithm of \a arg plus 1 to base e
+ /// \exception FE_INVALID for signaling NaN or argument <-1
+ /// \exception FE_DIVBYZERO for -1
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half log1p(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::log1p(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ if(arg.data_ >= 0xBC00)
+ return half(detail::binary, (arg.data_==0xBC00) ? detail::pole(0x8000) : (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_));
+ int abs = arg.data_ & 0x7FFF, exp = -15;
+ if(!abs || abs >= 0x7C00)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ for(; abs<0x400; abs<<=1,--exp) ;
+ exp += abs >> 10;
+ detail::uint32 m = static_cast<detail::uint32>((abs&0x3FF)|0x400) << 20;
+ if(arg.data_ & 0x8000)
+ {
+ m = 0x40000000 - (m>>-exp);
+ for(exp=0; m<0x40000000; m<<=1,--exp) ;
+ }
+ else
+ {
+ if(exp < 0)
+ {
+ m = 0x40000000 + (m>>-exp);
+ exp = 0;
+ }
+ else
+ {
+ m += 0x40000000 >> exp;
+ int i = m >> 31;
+ m >>= i;
+ exp += i;
+ }
+ }
+ return half(detail::binary, detail::log2_post<half::round_style,0xB8AA3B2A>(detail::log2(m), exp, 17));
+ #endif
+ }
+
+ /// \}
+ /// \anchor power
+ /// \name Power functions
+ /// \{
+
+ /// Square root.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::sqrt](https://en.cppreference.com/w/cpp/numeric/math/sqrt).
+ /// \param arg function argument
+ /// \return square root of \a arg
+ /// \exception FE_INVALID for signaling NaN and negative arguments
+ /// \exception FE_INEXACT according to rounding
+ inline half sqrt(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::sqrt(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = 15;
+ if(!abs || arg.data_ >= 0x7C00)
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_>0x8000) ? detail::invalid() : arg.data_);
+ for(; abs<0x400; abs<<=1,--exp) ;
+ detail::uint32 r = static_cast<detail::uint32>((abs&0x3FF)|0x400) << 10, m = detail::sqrt<20>(r, exp+=abs>>10);
+ return half(detail::binary, detail::rounded<half::round_style,false>((exp<<10)+(m&0x3FF), r>m, r!=0));
+ #endif
+ }
+
+ /// Inverse square root.
+ /// This function is exact to rounding for all rounding modes and thus generally more accurate than directly computing
+ /// 1 / sqrt(\a arg) in half-precision, in addition to also being faster.
+ /// \param arg function argument
+ /// \return reciprocal of square root of \a arg
+ /// \exception FE_INVALID for signaling NaN and negative arguments
+ /// \exception FE_INEXACT according to rounding
+ inline half rsqrt(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(detail::internal_t(1)/std::sqrt(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF, bias = 0x4000;
+ if(!abs || arg.data_ >= 0x7C00)
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_>0x8000) ?
+ detail::invalid() : !abs ? detail::pole(arg.data_&0x8000) : 0);
+ for(; abs<0x400; abs<<=1,bias-=0x400) ;
+ unsigned int frac = (abs+=bias) & 0x7FF;
+ if(frac == 0x400)
+ return half(detail::binary, 0x7A00-(abs>>1));
+ if((half::round_style == std::round_to_nearest && (frac == 0x3FE || frac == 0x76C)) ||
+ (half::round_style != std::round_to_nearest && (frac == 0x15A || frac == 0x3FC || frac == 0x401 || frac == 0x402 || frac == 0x67B)))
+ return pow(arg, half(detail::binary, 0xB800));
+ detail::uint32 f = 0x17376 - abs, mx = (abs&0x3FF) | 0x400, my = ((f>>1)&0x3FF) | 0x400, mz = my * my;
+ int expy = (f>>11) - 31, expx = 32 - (abs>>10), i = mz >> 21;
+ for(mz=0x60000000-(((mz>>i)*mx)>>(expx-2*expy-i)); mz<0x40000000; mz<<=1,--expy) ;
+ i = (my*=mz>>10) >> 31;
+ expy += i;
+ my = (my>>(20+i)) + 1;
+ i = (mz=my*my) >> 21;
+ for(mz=0x60000000-(((mz>>i)*mx)>>(expx-2*expy-i)); mz<0x40000000; mz<<=1,--expy) ;
+ i = (my*=(mz>>10)+1) >> 31;
+ return half(detail::binary, detail::fixed2half<half::round_style,30,false,false,true>(my>>i, expy+i+14));
+ #endif
+ }
+
+ /// Cubic root.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::cbrt](https://en.cppreference.com/w/cpp/numeric/math/cbrt).
+ /// \param arg function argument
+ /// \return cubic root of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT according to rounding
+ inline half cbrt(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::cbrt(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = -15;
+ if(!abs || abs == 0x3C00 || abs >= 0x7C00)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ for(; abs<0x400; abs<<=1, --exp);
+ detail::uint32 ilog = exp + (abs>>10), sign = detail::sign_mask(ilog), f, m =
+ (((ilog<<27)+(detail::log2(static_cast<detail::uint32>((abs&0x3FF)|0x400)<<20, 24)>>4))^sign) - sign;
+ for(exp=2; m<0x80000000; m<<=1,--exp) ;
+ m = detail::multiply64(m, 0xAAAAAAAB);
+ int i = m >> 31, s;
+ exp += i;
+ m <<= 1 - i;
+ if(exp < 0)
+ {
+ f = m >> -exp;
+ exp = 0;
+ }
+ else
+ {
+ f = (m<<exp) & 0x7FFFFFFF;
+ exp = m >> (31-exp);
+ }
+ m = detail::exp2(f, (half::round_style==std::round_to_nearest) ? 29 : 26);
+ if(sign)
+ {
+ if(m > 0x80000000)
+ {
+ m = detail::divide64(0x80000000, m, s);
+ ++exp;
+ }
+ exp = -exp;
+ }
+ return half(detail::binary, (half::round_style==std::round_to_nearest) ?
+ detail::fixed2half<half::round_style,31,false,false,false>(m, exp+14, arg.data_&0x8000) :
+ detail::fixed2half<half::round_style,23,false,false,false>((m+0x80)>>8, exp+14, arg.data_&0x8000));
+ #endif
+ }
+
+ /// Hypotenuse function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::hypot](https://en.cppreference.com/w/cpp/numeric/math/hypot).
+ /// \param x first argument
+ /// \param y second argument
+ /// \return square root of sum of squares without internal over- or underflows
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding of the final square root
+ inline half hypot(half x, half y)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ detail::internal_t fx = detail::half2float<detail::internal_t>(x.data_), fy = detail::half2float<detail::internal_t>(y.data_);
+ #if HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::hypot(fx, fy)));
+ #else
+ return half(detail::binary, detail::float2half<half::round_style>(std::sqrt(fx*fx+fy*fy)));
+ #endif
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, expx = 0, expy = 0;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx==0x7C00) ? detail::select(0x7C00, y.data_) :
+ (absy==0x7C00) ? detail::select(0x7C00, x.data_) : detail::signal(x.data_, y.data_));
+ if(!absx)
+ return half(detail::binary, absy ? detail::check_underflow(absy) : 0);
+ if(!absy)
+ return half(detail::binary, detail::check_underflow(absx));
+ if(absy > absx)
+ std::swap(absx, absy);
+ for(; absx<0x400; absx<<=1,--expx) ;
+ for(; absy<0x400; absy<<=1,--expy) ;
+ detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400;
+ mx *= mx;
+ my *= my;
+ int ix = mx >> 21, iy = my >> 21;
+ expx = 2*(expx+(absx>>10)) - 15 + ix;
+ expy = 2*(expy+(absy>>10)) - 15 + iy;
+ mx <<= 10 - ix;
+ my <<= 10 - iy;
+ int d = expx - expy;
+ my = (d<30) ? ((my>>d)|((my&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
+ return half(detail::binary, detail::hypot_post<half::round_style>(mx+my, expx));
+ #endif
+ }
+
+ /// Hypotenuse function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::hypot](https://en.cppreference.com/w/cpp/numeric/math/hypot).
+ /// \param x first argument
+ /// \param y second argument
+ /// \param z third argument
+ /// \return square root of sum of squares without internal over- or underflows
+ /// \exception FE_INVALID if \a x, \a y or \a z is signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding of the final square root
+ inline half hypot(half x, half y, half z)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ detail::internal_t fx = detail::half2float<detail::internal_t>(x.data_), fy = detail::half2float<detail::internal_t>(y.data_), fz = detail::half2float<detail::internal_t>(z.data_);
+ return half(detail::binary, detail::float2half<half::round_style>(std::sqrt(fx*fx+fy*fy+fz*fz)));
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, absz = z.data_ & 0x7FFF, expx = 0, expy = 0, expz = 0;
+ if(!absx)
+ return hypot(y, z);
+ if(!absy)
+ return hypot(x, z);
+ if(!absz)
+ return hypot(x, y);
+ if(absx >= 0x7C00 || absy >= 0x7C00 || absz >= 0x7C00)
+ return half(detail::binary, (absx==0x7C00) ? detail::select(0x7C00, detail::select(y.data_, z.data_)) :
+ (absy==0x7C00) ? detail::select(0x7C00, detail::select(x.data_, z.data_)) :
+ (absz==0x7C00) ? detail::select(0x7C00, detail::select(x.data_, y.data_)) :
+ detail::signal(x.data_, y.data_, z.data_));
+ if(absz > absy)
+ std::swap(absy, absz);
+ if(absy > absx)
+ std::swap(absx, absy);
+ if(absz > absy)
+ std::swap(absy, absz);
+ for(; absx<0x400; absx<<=1,--expx) ;
+ for(; absy<0x400; absy<<=1,--expy) ;
+ for(; absz<0x400; absz<<=1,--expz) ;
+ detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400, mz = (absz&0x3FF) | 0x400;
+ mx *= mx;
+ my *= my;
+ mz *= mz;
+ int ix = mx >> 21, iy = my >> 21, iz = mz >> 21;
+ expx = 2*(expx+(absx>>10)) - 15 + ix;
+ expy = 2*(expy+(absy>>10)) - 15 + iy;
+ expz = 2*(expz+(absz>>10)) - 15 + iz;
+ mx <<= 10 - ix;
+ my <<= 10 - iy;
+ mz <<= 10 - iz;
+ int d = expy - expz;
+ mz = (d<30) ? ((mz>>d)|((mz&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
+ my += mz;
+ if(my & 0x80000000)
+ {
+ my = (my>>1) | (my&1);
+ if(++expy > expx)
+ {
+ std::swap(mx, my);
+ std::swap(expx, expy);
+ }
+ }
+ d = expx - expy;
+ my = (d<30) ? ((my>>d)|((my&((static_cast<detail::uint32>(1)<<d)-1))!=0)) : 1;
+ return half(detail::binary, detail::hypot_post<half::round_style>(mx+my, expx));
+ #endif
+ }
+
+ /// Power function.
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.00025% of inputs.
+ ///
+ /// **See also:** Documentation for [std::pow](https://en.cppreference.com/w/cpp/numeric/math/pow).
+ /// \param x base
+ /// \param y exponent
+ /// \return \a x raised to \a y
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN or if \a x is finite an negative and \a y is finite and not integral
+ /// \exception FE_DIVBYZERO if \a x is 0 and \a y is negative
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half pow(half x, half y)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::pow(detail::half2float<detail::internal_t>(x.data_), detail::half2float<detail::internal_t>(y.data_))));
+ #else
+ int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = -15;
+ if(!absy || x.data_ == 0x3C00)
+ return half(detail::binary, detail::select(0x3C00, (x.data_==0x3C00) ? y.data_ : x.data_));
+ bool is_int = absy >= 0x6400 || (absy>=0x3C00 && !(absy&((1<<(25-(absy>>10)))-1)));
+ unsigned int sign = x.data_ & (static_cast<unsigned>((absy<0x6800)&&is_int&&((absy>>(25-(absy>>10)))&1))<<15);
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) :
+ (absy==0x7C00) ? ((absx==0x3C00) ? 0x3C00 : (!absx && y.data_==0xFC00) ? detail::pole() :
+ (0x7C00&-((y.data_>>15)^(absx>0x3C00)))) : (sign|(0x7C00&((y.data_>>15)-1U))));
+ if(!absx)
+ return half(detail::binary, (y.data_&0x8000) ? detail::pole(sign) : sign);
+ if((x.data_&0x8000) && !is_int)
+ return half(detail::binary, detail::invalid());
+ if(x.data_ == 0xBC00)
+ return half(detail::binary, sign|0x3C00);
+ switch(y.data_)
+ {
+ case 0x3800: return sqrt(x);
+ case 0x3C00: return half(detail::binary, detail::check_underflow(x.data_));
+ case 0x4000: return x * x;
+ case 0xBC00: return half(detail::binary, 0x3C00) / x;
+ }
+ for(; absx<0x400; absx<<=1,--exp) ;
+ detail::uint32 ilog = exp + (absx>>10), msign = detail::sign_mask(ilog), f, m =
+ (((ilog<<27)+((detail::log2(static_cast<detail::uint32>((absx&0x3FF)|0x400)<<20)+8)>>4))^msign) - msign;
+ for(exp=-11; m<0x80000000; m<<=1,--exp) ;
+ for(; absy<0x400; absy<<=1,--exp) ;
+ m = detail::multiply64(m, static_cast<detail::uint32>((absy&0x3FF)|0x400)<<21);
+ int i = m >> 31;
+ exp += (absy>>10) + i;
+ m <<= 1 - i;
+ if(exp < 0)
+ {
+ f = m >> -exp;
+ exp = 0;
+ }
+ else
+ {
+ f = (m<<exp) & 0x7FFFFFFF;
+ exp = m >> (31-exp);
+ }
+ return half(detail::binary, detail::exp2_post<half::round_style>(f, exp, ((msign&1)^(y.data_>>15))!=0, sign));
+ #endif
+ }
+
+ /// \}
+ /// \anchor trigonometric
+ /// \name Trigonometric functions
+ /// \{
+
+ /// Compute sine and cosine simultaneously.
+ /// This returns the same results as sin() and cos() but is faster than calling each function individually.
+ ///
+ /// This function is exact to rounding for all rounding modes.
+ /// \param arg function argument
+ /// \param sin variable to take sine of \a arg
+ /// \param cos variable to take cosine of \a arg
+ /// \exception FE_INVALID for signaling NaN or infinity
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline void sincos(half arg, half *sin, half *cos)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ detail::internal_t f = detail::half2float<detail::internal_t>(arg.data_);
+ *sin = half(detail::binary, detail::float2half<half::round_style>(std::sin(f)));
+ *cos = half(detail::binary, detail::float2half<half::round_style>(std::cos(f)));
+ #else
+ int abs = arg.data_ & 0x7FFF, sign = arg.data_ >> 15, k;
+ if(abs >= 0x7C00)
+ *sin = *cos = half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
+ else if(!abs)
+ {
+ *sin = arg;
+ *cos = half(detail::binary, 0x3C00);
+ }
+ else if(abs < 0x2500)
+ {
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x3BFF, 1, 1));
+ }
+ else
+ {
+ if(half::round_style != std::round_to_nearest)
+ {
+ switch(abs)
+ {
+ case 0x48B7:
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x1D07, 1, 1));
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0xBBFF, 1, 1));
+ return;
+ case 0x598C:
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x3BFF, 1, 1));
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x80FC, 1, 1));
+ return;
+ case 0x6A64:
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x3BFE, 1, 1));
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x27FF, 1, 1));
+ return;
+ case 0x6D8C:
+ *sin = half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x0FE6, 1, 1));
+ *cos = half(detail::binary, detail::rounded<half::round_style,true>(0x3BFF, 1, 1));
+ return;
+ }
+ }
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 28);
+ switch(k & 3)
+ {
+ case 1: sc = std::make_pair(sc.second, -sc.first); break;
+ case 2: sc = std::make_pair(-sc.first, -sc.second); break;
+ case 3: sc = std::make_pair(-sc.second, sc.first); break;
+ }
+ *sin = half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>((sc.first^-static_cast<detail::uint32>(sign))+sign));
+ *cos = half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>(sc.second));
+ }
+ #endif
+ }
+
+ /// Sine function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::sin](https://en.cppreference.com/w/cpp/numeric/math/sin).
+ /// \param arg function argument
+ /// \return sine value of \a arg
+ /// \exception FE_INVALID for signaling NaN or infinity
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half sin(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::sin(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, k;
+ if(!abs)
+ return arg;
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs < 0x2900)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
+ if(half::round_style != std::round_to_nearest)
+ switch(abs)
+ {
+ case 0x48B7: return half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x1D07, 1, 1));
+ case 0x6A64: return half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x3BFE, 1, 1));
+ case 0x6D8C: return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x0FE6, 1, 1));
+ }
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 28);
+ detail::uint32 sign = -static_cast<detail::uint32>(((k>>1)&1)^(arg.data_>>15));
+ return half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>((((k&1) ? sc.second : sc.first)^sign) - sign));
+ #endif
+ }
+
+ /// Cosine function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::cos](https://en.cppreference.com/w/cpp/numeric/math/cos).
+ /// \param arg function argument
+ /// \return cosine value of \a arg
+ /// \exception FE_INVALID for signaling NaN or infinity
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half cos(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::cos(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, k;
+ if(!abs)
+ return half(detail::binary, 0x3C00);
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs < 0x2500)
+ return half(detail::binary, detail::rounded<half::round_style,true>(0x3BFF, 1, 1));
+ if(half::round_style != std::round_to_nearest && abs == 0x598C)
+ return half(detail::binary, detail::rounded<half::round_style,true>(0x80FC, 1, 1));
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 28);
+ detail::uint32 sign = -static_cast<detail::uint32>(((k>>1)^k)&1);
+ return half(detail::binary, detail::fixed2half<half::round_style,30,true,true,true>((((k&1) ? sc.first : sc.second)^sign) - sign));
+ #endif
+ }
+
+ /// Tangent function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::tan](https://en.cppreference.com/w/cpp/numeric/math/tan).
+ /// \param arg function argument
+ /// \return tangent value of \a arg
+ /// \exception FE_INVALID for signaling NaN or infinity
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half tan(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::tan(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = 13, k;
+ if(!abs)
+ return arg;
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs < 0x2700)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
+ if(half::round_style != std::round_to_nearest)
+ switch(abs)
+ {
+ case 0x658C: return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x07E6, 1, 1));
+ case 0x7330: return half(detail::binary, detail::rounded<half::round_style,true>((~arg.data_&0x8000)|0x4B62, 1, 1));
+ }
+ std::pair<detail::uint32,detail::uint32> sc = detail::sincos(detail::angle_arg(abs, k), 30);
+ if(k & 1)
+ sc = std::make_pair(-sc.second, sc.first);
+ detail::uint32 signy = detail::sign_mask(sc.first), signx = detail::sign_mask(sc.second);
+ detail::uint32 my = (sc.first^signy) - signy, mx = (sc.second^signx) - signx;
+ for(; my<0x80000000; my<<=1,--exp) ;
+ for(; mx<0x80000000; mx<<=1,++exp) ;
+ return half(detail::binary, detail::tangent_post<half::round_style>(my, mx, exp, (signy^signx^arg.data_)&0x8000));
+ #endif
+ }
+
+ /// Arc sine.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::asin](https://en.cppreference.com/w/cpp/numeric/math/asin).
+ /// \param arg function argument
+ /// \return arc sine value of \a arg
+ /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half asin(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::asin(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
+ if(!abs)
+ return arg;
+ if(abs >= 0x3C00)
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (abs>0x3C00) ? detail::invalid() :
+ detail::rounded<half::round_style,true>(sign|0x3E48, 0, 1));
+ if(abs < 0x2900)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
+ if(half::round_style != std::round_to_nearest && (abs == 0x2B44 || abs == 0x2DC3))
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_+1, 1, 1));
+ std::pair<detail::uint32,detail::uint32> sc = detail::atan2_args(abs);
+ detail::uint32 m = detail::atan2(sc.first, sc.second, (half::round_style==std::round_to_nearest) ? 27 : 26);
+ return half(detail::binary, detail::fixed2half<half::round_style,30,false,true,true>(m, 14, sign));
+ #endif
+ }
+
+ /// Arc cosine function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::acos](https://en.cppreference.com/w/cpp/numeric/math/acos).
+ /// \param arg function argument
+ /// \return arc cosine value of \a arg
+ /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half acos(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::acos(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ >> 15;
+ if(!abs)
+ return half(detail::binary, detail::rounded<half::round_style,true>(0x3E48, 0, 1));
+ if(abs >= 0x3C00)
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (abs>0x3C00) ? detail::invalid() :
+ sign ? detail::rounded<half::round_style,true>(0x4248, 0, 1) : 0);
+ std::pair<detail::uint32,detail::uint32> cs = detail::atan2_args(abs);
+ detail::uint32 m = detail::atan2(cs.second, cs.first, 28);
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,true,true>(sign ? (0xC90FDAA2-m) : m, 15, 0, sign));
+ #endif
+ }
+
+ /// Arc tangent function.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::atan](https://en.cppreference.com/w/cpp/numeric/math/atan).
+ /// \param arg function argument
+ /// \return arc tangent value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half atan(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::atan(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
+ if(!abs)
+ return arg;
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? detail::rounded<half::round_style,true>(sign|0x3E48, 0, 1) : detail::signal(arg.data_));
+ if(abs <= 0x2700)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
+ int exp = (abs>>10) + (abs<=0x3FF);
+ detail::uint32 my = (abs&0x3FF) | ((abs>0x3FF)<<10);
+ detail::uint32 m = (exp>15) ? detail::atan2(my<<19, 0x20000000>>(exp-15), (half::round_style==std::round_to_nearest) ? 26 : 24) :
+ detail::atan2(my<<(exp+4), 0x20000000, (half::round_style==std::round_to_nearest) ? 30 : 28);
+ return half(detail::binary, detail::fixed2half<half::round_style,30,false,true,true>(m, 14, sign));
+ #endif
+ }
+
+ /// Arc tangent function.
+ /// This function may be 1 ULP off the correctly rounded exact result in ~0.005% of inputs for `std::round_to_nearest`,
+ /// in ~0.1% of inputs for `std::round_toward_zero` and in ~0.02% of inputs for any other rounding mode.
+ ///
+ /// **See also:** Documentation for [std::atan2](https://en.cppreference.com/w/cpp/numeric/math/atan2).
+ /// \param y numerator
+ /// \param x denominator
+ /// \return arc tangent value
+ /// \exception FE_INVALID if \a x or \a y is signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half atan2(half y, half x)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::atan2(detail::half2float<detail::internal_t>(y.data_), detail::half2float<detail::internal_t>(x.data_))));
+ #else
+ unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, signx = x.data_ >> 15, signy = y.data_ & 0x8000;
+ if(absx >= 0x7C00 || absy >= 0x7C00)
+ {
+ if(absx > 0x7C00 || absy > 0x7C00)
+ return half(detail::binary, detail::signal(x.data_, y.data_));
+ if(absy == 0x7C00)
+ return half(detail::binary, (absx<0x7C00) ? detail::rounded<half::round_style,true>(signy|0x3E48, 0, 1) :
+ signx ? detail::rounded<half::round_style,true>(signy|0x40B6, 0, 1) :
+ detail::rounded<half::round_style,true>(signy|0x3A48, 0, 1));
+ return (x.data_==0x7C00) ? half(detail::binary, signy) : half(detail::binary, detail::rounded<half::round_style,true>(signy|0x4248, 0, 1));
+ }
+ if(!absy)
+ return signx ? half(detail::binary, detail::rounded<half::round_style,true>(signy|0x4248, 0, 1)) : y;
+ if(!absx)
+ return half(detail::binary, detail::rounded<half::round_style,true>(signy|0x3E48, 0, 1));
+ int d = (absy>>10) + (absy<=0x3FF) - (absx>>10) - (absx<=0x3FF);
+ if(d > (signx ? 18 : 12))
+ return half(detail::binary, detail::rounded<half::round_style,true>(signy|0x3E48, 0, 1));
+ if(signx && d < -11)
+ return half(detail::binary, detail::rounded<half::round_style,true>(signy|0x4248, 0, 1));
+ if(!signx && d < ((half::round_style==std::round_toward_zero) ? -15 : -9))
+ {
+ for(; absy<0x400; absy<<=1,--d) ;
+ detail::uint32 mx = ((absx<<1)&0x7FF) | 0x800, my = ((absy<<1)&0x7FF) | 0x800;
+ int i = my < mx;
+ d -= i;
+ if(d < -25)
+ return half(detail::binary, detail::underflow<half::round_style>(signy));
+ my <<= 11 + i;
+ return half(detail::binary, detail::fixed2half<half::round_style,11,false,false,true>(my/mx, d+14, signy, my%mx!=0));
+ }
+ detail::uint32 m = detail::atan2( ((absy&0x3FF)|((absy>0x3FF)<<10))<<(19+((d<0) ? d : (d>0) ? 0 : -1)),
+ ((absx&0x3FF)|((absx>0x3FF)<<10))<<(19-((d>0) ? d : (d<0) ? 0 : 1)));
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,true,true>(signx ? (0xC90FDAA2-m) : m, 15, signy, signx));
+ #endif
+ }
+
+ /// \}
+ /// \anchor hyperbolic
+ /// \name Hyperbolic functions
+ /// \{
+
+ /// Hyperbolic sine.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::sinh](https://en.cppreference.com/w/cpp/numeric/math/sinh).
+ /// \param arg function argument
+ /// \return hyperbolic sine value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half sinh(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::sinh(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp;
+ if(!abs || abs >= 0x7C00)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ if(abs <= 0x2900)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
+ std::pair<detail::uint32,detail::uint32> mm = detail::hyperbolic_args(abs, exp, (half::round_style==std::round_to_nearest) ? 29 : 27);
+ detail::uint32 m = mm.first - mm.second;
+ for(exp+=13; m<0x80000000 && exp; m<<=1,--exp) ;
+ unsigned int sign = arg.data_ & 0x8000;
+ if(exp > 29)
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,false,true>(m, exp, sign));
+ #endif
+ }
+
+ /// Hyperbolic cosine.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::cosh](https://en.cppreference.com/w/cpp/numeric/math/cosh).
+ /// \param arg function argument
+ /// \return hyperbolic cosine value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half cosh(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::cosh(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp;
+ if(!abs)
+ return half(detail::binary, 0x3C00);
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : 0x7C00);
+ std::pair<detail::uint32,detail::uint32> mm = detail::hyperbolic_args(abs, exp, (half::round_style==std::round_to_nearest) ? 23 : 26);
+ detail::uint32 m = mm.first + mm.second, i = (~m&0xFFFFFFFF) >> 31;
+ m = (m>>i) | (m&i) | 0x80000000;
+ if((exp+=13+i) > 29)
+ return half(detail::binary, detail::overflow<half::round_style>());
+ return half(detail::binary, detail::fixed2half<half::round_style,31,false,false,true>(m, exp));
+ #endif
+ }
+
+ /// Hyperbolic tangent.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::tanh](https://en.cppreference.com/w/cpp/numeric/math/tanh).
+ /// \param arg function argument
+ /// \return hyperbolic tangent value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half tanh(half arg)
+ {
+ #ifdef HALF_ARITHMETIC_TYPE
+ return half(detail::binary, detail::float2half<half::round_style>(std::tanh(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp;
+ if(!abs)
+ return arg;
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_-0x4000));
+ if(abs >= 0x4500)
+ return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x3BFF, 1, 1));
+ if(abs < 0x2700)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
+ if(half::round_style != std::round_to_nearest && abs == 0x2D3F)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-3, 0, 1));
+ std::pair<detail::uint32,detail::uint32> mm = detail::hyperbolic_args(abs, exp, 27);
+ detail::uint32 my = mm.first - mm.second - (half::round_style!=std::round_to_nearest), mx = mm.first + mm.second, i = (~mx&0xFFFFFFFF) >> 31;
+ for(exp=13; my<0x80000000; my<<=1,--exp) ;
+ mx = (mx>>i) | 0x80000000;
+ return half(detail::binary, detail::tangent_post<half::round_style>(my, mx, exp-i, arg.data_&0x8000));
+ #endif
+ }
+
+ /// Hyperbolic area sine.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::asinh](https://en.cppreference.com/w/cpp/numeric/math/asinh).
+ /// \param arg function argument
+ /// \return area sine value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half asinh(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::asinh(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF;
+ if(!abs || abs >= 0x7C00)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ if(abs <= 0x2900)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-1, 1, 1));
+ if(half::round_style != std::round_to_nearest)
+ switch(abs)
+ {
+ case 0x32D4: return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-13, 1, 1));
+ case 0x3B5B: return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_-197, 1, 1));
+ }
+ return half(detail::binary, detail::area<half::round_style,true>(arg.data_));
+ #endif
+ }
+
+ /// Hyperbolic area cosine.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::acosh](https://en.cppreference.com/w/cpp/numeric/math/acosh).
+ /// \param arg function argument
+ /// \return area cosine value of \a arg
+ /// \exception FE_INVALID for signaling NaN or arguments <1
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half acosh(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::acosh(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF;
+ if((arg.data_&0x8000) || abs < 0x3C00)
+ return half(detail::binary, (abs<=0x7C00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs == 0x3C00)
+ return half(detail::binary, 0);
+ if(arg.data_ >= 0x7C00)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ return half(detail::binary, detail::area<half::round_style,false>(arg.data_));
+ #endif
+ }
+
+ /// Hyperbolic area tangent.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::atanh](https://en.cppreference.com/w/cpp/numeric/math/atanh).
+ /// \param arg function argument
+ /// \return area tangent value of \a arg
+ /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1
+ /// \exception FE_DIVBYZERO for +/-1
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half atanh(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::atanh(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF, exp = 0;
+ if(!abs)
+ return arg;
+ if(abs >= 0x3C00)
+ return half(detail::binary, (abs==0x3C00) ? detail::pole(arg.data_&0x8000) : (abs<=0x7C00) ? detail::invalid() : detail::signal(arg.data_));
+ if(abs < 0x2700)
+ return half(detail::binary, detail::rounded<half::round_style,true>(arg.data_, 0, 1));
+ detail::uint32 m = static_cast<detail::uint32>((abs&0x3FF)|((abs>0x3FF)<<10)) << ((abs>>10)+(abs<=0x3FF)+6), my = 0x80000000 + m, mx = 0x80000000 - m;
+ for(; mx<0x80000000; mx<<=1,++exp) ;
+ int i = my >= mx, s;
+ return half(detail::binary, detail::log2_post<half::round_style,0xB8AA3B2A>(detail::log2(
+ (detail::divide64(my>>i, mx, s)+1)>>1, 27)+0x10, exp+i-1, 16, arg.data_&0x8000));
+ #endif
+ }
+
+ /// \}
+ /// \anchor special
+ /// \name Error and gamma functions
+ /// \{
+
+ /// Error function.
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.5% of inputs.
+ ///
+ /// **See also:** Documentation for [std::erf](https://en.cppreference.com/w/cpp/numeric/math/erf).
+ /// \param arg function argument
+ /// \return error function value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half erf(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::erf(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF;
+ if(!abs || abs >= 0x7C00)
+ return (abs>=0x7C00) ? half(detail::binary, (abs==0x7C00) ? (arg.data_-0x4000) : detail::signal(arg.data_)) : arg;
+ if(abs >= 0x4200)
+ return half(detail::binary, detail::rounded<half::round_style,true>((arg.data_&0x8000)|0x3BFF, 1, 1));
+ return half(detail::binary, detail::erf<half::round_style,false>(arg.data_));
+ #endif
+ }
+
+ /// Complementary error function.
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.5% of inputs.
+ ///
+ /// **See also:** Documentation for [std::erfc](https://en.cppreference.com/w/cpp/numeric/math/erfc).
+ /// \param arg function argument
+ /// \return 1 minus error function value of \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half erfc(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::erfc(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
+ if(abs >= 0x7C00)
+ return (abs>=0x7C00) ? half(detail::binary, (abs==0x7C00) ? (sign>>1) : detail::signal(arg.data_)) : arg;
+ if(!abs)
+ return half(detail::binary, 0x3C00);
+ if(abs >= 0x4400)
+ return half(detail::binary, detail::rounded<half::round_style,true>((sign>>1)-(sign>>15), sign>>15, 1));
+ return half(detail::binary, detail::erf<half::round_style,true>(arg.data_));
+ #endif
+ }
+
+ /// Natural logarithm of gamma function.
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.025% of inputs.
+ ///
+ /// **See also:** Documentation for [std::lgamma](https://en.cppreference.com/w/cpp/numeric/math/lgamma).
+ /// \param arg function argument
+ /// \return natural logarith of gamma function for \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_DIVBYZERO for 0 or negative integer arguments
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half lgamma(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::lgamma(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ int abs = arg.data_ & 0x7FFF;
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? 0x7C00 : detail::signal(arg.data_));
+ if(!abs || arg.data_ >= 0xE400 || (arg.data_ >= 0xBC00 && !(abs&((1<<(25-(abs>>10)))-1))))
+ return half(detail::binary, detail::pole());
+ if(arg.data_ == 0x3C00 || arg.data_ == 0x4000)
+ return half(detail::binary, 0);
+ return half(detail::binary, detail::gamma<half::round_style,true>(arg.data_));
+ #endif
+ }
+
+ /// Gamma function.
+ /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.25% of inputs.
+ ///
+ /// **See also:** Documentation for [std::tgamma](https://en.cppreference.com/w/cpp/numeric/math/tgamma).
+ /// \param arg function argument
+ /// \return gamma function value of \a arg
+ /// \exception FE_INVALID for signaling NaN, negative infinity or negative integer arguments
+ /// \exception FE_DIVBYZERO for 0
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half tgamma(half arg)
+ {
+ #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH
+ return half(detail::binary, detail::float2half<half::round_style>(std::tgamma(detail::half2float<detail::internal_t>(arg.data_))));
+ #else
+ unsigned int abs = arg.data_ & 0x7FFF;
+ if(!abs)
+ return half(detail::binary, detail::pole(arg.data_));
+ if(abs >= 0x7C00)
+ return (arg.data_==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_));
+ if(arg.data_ >= 0xE400 || (arg.data_ >= 0xBC00 && !(abs&((1<<(25-(abs>>10)))-1))))
+ return half(detail::binary, detail::invalid());
+ if(arg.data_ >= 0xCA80)
+ return half(detail::binary, detail::underflow<half::round_style>((1-((abs>>(25-(abs>>10)))&1))<<15));
+ if(arg.data_ <= 0x100 || (arg.data_ >= 0x4900 && arg.data_ < 0x8000))
+ return half(detail::binary, detail::overflow<half::round_style>());
+ if(arg.data_ == 0x3C00)
+ return arg;
+ return half(detail::binary, detail::gamma<half::round_style,false>(arg.data_));
+ #endif
+ }
+
+ /// \}
+ /// \anchor rounding
+ /// \name Rounding
+ /// \{
+
+ /// Nearest integer not less than half value.
+ /// **See also:** Documentation for [std::ceil](https://en.cppreference.com/w/cpp/numeric/math/ceil).
+ /// \param arg half to round
+ /// \return nearest integer not less than \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT if value had to be rounded
+ inline half ceil(half arg) { return half(detail::binary, detail::integral<std::round_toward_infinity,true,true>(arg.data_)); }
+
+ /// Nearest integer not greater than half value.
+ /// **See also:** Documentation for [std::floor](https://en.cppreference.com/w/cpp/numeric/math/floor).
+ /// \param arg half to round
+ /// \return nearest integer not greater than \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT if value had to be rounded
+ inline half floor(half arg) { return half(detail::binary, detail::integral<std::round_toward_neg_infinity,true,true>(arg.data_)); }
+
+ /// Nearest integer not greater in magnitude than half value.
+ /// **See also:** Documentation for [std::trunc](https://en.cppreference.com/w/cpp/numeric/math/trunc).
+ /// \param arg half to round
+ /// \return nearest integer not greater in magnitude than \a arg
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT if value had to be rounded
+ inline half trunc(half arg) { return half(detail::binary, detail::integral<std::round_toward_zero,true,true>(arg.data_)); }
+
+ /// Nearest integer.
+ /// **See also:** Documentation for [std::round](https://en.cppreference.com/w/cpp/numeric/math/round).
+ /// \param arg half to round
+ /// \return nearest integer, rounded away from zero in half-way cases
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT if value had to be rounded
+ inline half round(half arg) { return half(detail::binary, detail::integral<std::round_to_nearest,false,true>(arg.data_)); }
+
+ /// Nearest integer.
+ /// **See also:** Documentation for [std::lround](https://en.cppreference.com/w/cpp/numeric/math/round).
+ /// \param arg half to round
+ /// \return nearest integer, rounded away from zero in half-way cases
+ /// \exception FE_INVALID if value is not representable as `long`
+ inline long lround(half arg) { return detail::half2int<std::round_to_nearest,false,false,long>(arg.data_); }
+
+ /// Nearest integer using half's internal rounding mode.
+ /// **See also:** Documentation for [std::rint](https://en.cppreference.com/w/cpp/numeric/math/rint).
+ /// \param arg half expression to round
+ /// \return nearest integer using default rounding mode
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_INEXACT if value had to be rounded
+ inline half rint(half arg) { return half(detail::binary, detail::integral<half::round_style,true,true>(arg.data_)); }
+
+ /// Nearest integer using half's internal rounding mode.
+ /// **See also:** Documentation for [std::lrint](https://en.cppreference.com/w/cpp/numeric/math/rint).
+ /// \param arg half expression to round
+ /// \return nearest integer using default rounding mode
+ /// \exception FE_INVALID if value is not representable as `long`
+ /// \exception FE_INEXACT if value had to be rounded
+ inline long lrint(half arg) { return detail::half2int<half::round_style,true,true,long>(arg.data_); }
+
+ /// Nearest integer using half's internal rounding mode.
+ /// **See also:** Documentation for [std::nearbyint](https://en.cppreference.com/w/cpp/numeric/math/nearbyint).
+ /// \param arg half expression to round
+ /// \return nearest integer using default rounding mode
+ /// \exception FE_INVALID for signaling NaN
+ inline half nearbyint(half arg) { return half(detail::binary, detail::integral<half::round_style,true,false>(arg.data_)); }
+#if HALF_ENABLE_CPP11_LONG_LONG
+ /// Nearest integer.
+ /// **See also:** Documentation for [std::llround](https://en.cppreference.com/w/cpp/numeric/math/round).
+ /// \param arg half to round
+ /// \return nearest integer, rounded away from zero in half-way cases
+ /// \exception FE_INVALID if value is not representable as `long long`
+ inline long long llround(half arg) { return detail::half2int<std::round_to_nearest,false,false,long long>(arg.data_); }
+
+ /// Nearest integer using half's internal rounding mode.
+ /// **See also:** Documentation for [std::llrint](https://en.cppreference.com/w/cpp/numeric/math/rint).
+ /// \param arg half expression to round
+ /// \return nearest integer using default rounding mode
+ /// \exception FE_INVALID if value is not representable as `long long`
+ /// \exception FE_INEXACT if value had to be rounded
+ inline long long llrint(half arg) { return detail::half2int<half::round_style,true,true,long long>(arg.data_); }
+#endif
+
+ /// \}
+ /// \anchor float
+ /// \name Floating point manipulation
+ /// \{
+
+ /// Decompress floating-point number.
+ /// **See also:** Documentation for [std::frexp](https://en.cppreference.com/w/cpp/numeric/math/frexp).
+ /// \param arg number to decompress
+ /// \param exp address to store exponent at
+ /// \return significant in range [0.5, 1)
+ /// \exception FE_INVALID for signaling NaN
+ inline half frexp(half arg, int *exp)
+ {
+ *exp = 0;
+ unsigned int abs = arg.data_ & 0x7FFF;
+ if(abs >= 0x7C00 || !abs)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ for(; abs<0x400; abs<<=1,--*exp) ;
+ *exp += (abs>>10) - 14;
+ return half(detail::binary, (arg.data_&0x8000)|0x3800|(abs&0x3FF));
+ }
+
+ /// Multiply by power of two.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::scalbln](https://en.cppreference.com/w/cpp/numeric/math/scalbn).
+ /// \param arg number to modify
+ /// \param exp power of two to multiply with
+ /// \return \a arg multplied by 2 raised to \a exp
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half scalbln(half arg, long exp)
+ {
+ unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000;
+ if(abs >= 0x7C00 || !abs)
+ return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg;
+ for(; abs<0x400; abs<<=1,--exp) ;
+ exp += abs >> 10;
+ if(exp > 30)
+ return half(detail::binary, detail::overflow<half::round_style>(sign));
+ else if(exp < -10)
+ return half(detail::binary, detail::underflow<half::round_style>(sign));
+ else if(exp > 0)
+ return half(detail::binary, sign|(exp<<10)|(abs&0x3FF));
+ unsigned int m = (abs&0x3FF) | 0x400;
+ return half(detail::binary, detail::rounded<half::round_style,false>(sign|(m>>(1-exp)), (m>>-exp)&1, (m&((1<<-exp)-1))!=0));
+ }
+
+ /// Multiply by power of two.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::scalbn](https://en.cppreference.com/w/cpp/numeric/math/scalbn).
+ /// \param arg number to modify
+ /// \param exp power of two to multiply with
+ /// \return \a arg multplied by 2 raised to \a exp
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half scalbn(half arg, int exp) { return scalbln(arg, exp); }
+
+ /// Multiply by power of two.
+ /// This function is exact to rounding for all rounding modes.
+ ///
+ /// **See also:** Documentation for [std::ldexp](https://en.cppreference.com/w/cpp/numeric/math/ldexp).
+ /// \param arg number to modify
+ /// \param exp power of two to multiply with
+ /// \return \a arg multplied by 2 raised to \a exp
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ inline half ldexp(half arg, int exp) { return scalbln(arg, exp); }
+
+ /// Extract integer and fractional parts.
+ /// **See also:** Documentation for [std::modf](https://en.cppreference.com/w/cpp/numeric/math/modf).
+ /// \param arg number to decompress
+ /// \param iptr address to store integer part at
+ /// \return fractional part
+ /// \exception FE_INVALID for signaling NaN
+ inline half modf(half arg, half *iptr)
+ {
+ unsigned int abs = arg.data_ & 0x7FFF;
+ if(abs > 0x7C00)
+ {
+ arg = half(detail::binary, detail::signal(arg.data_));
+ return *iptr = arg, arg;
+ }
+ if(abs >= 0x6400)
+ return *iptr = arg, half(detail::binary, arg.data_&0x8000);
+ if(abs < 0x3C00)
+ return iptr->data_ = arg.data_ & 0x8000, arg;
+ unsigned int exp = abs >> 10, mask = (1<<(25-exp)) - 1, m = arg.data_ & mask;
+ iptr->data_ = arg.data_ & ~mask;
+ if(!m)
+ return half(detail::binary, arg.data_&0x8000);
+ for(; m<0x400; m<<=1,--exp) ;
+ return half(detail::binary, (arg.data_&0x8000)|(exp<<10)|(m&0x3FF));
+ }
+
+ /// Extract exponent.
+ /// **See also:** Documentation for [std::ilogb](https://en.cppreference.com/w/cpp/numeric/math/ilogb).
+ /// \param arg number to query
+ /// \return floating-point exponent
+ /// \retval FP_ILOGB0 for zero
+ /// \retval FP_ILOGBNAN for NaN
+ /// \retval INT_MAX for infinity
+ /// \exception FE_INVALID for 0 or infinite values
+ inline int ilogb(half arg)
+ {
+ int abs = arg.data_ & 0x7FFF, exp;
+ if(!abs || abs >= 0x7C00)
+ {
+ detail::raise(FE_INVALID);
+ return !abs ? FP_ILOGB0 : (abs==0x7C00) ? INT_MAX : FP_ILOGBNAN;
+ }
+ for(exp=(abs>>10)-15; abs<0x200; abs<<=1,--exp) ;
+ return exp;
+ }
+
+ /// Extract exponent.
+ /// **See also:** Documentation for [std::logb](https://en.cppreference.com/w/cpp/numeric/math/logb).
+ /// \param arg number to query
+ /// \return floating-point exponent
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_DIVBYZERO for 0
+ inline half logb(half arg)
+ {
+ int abs = arg.data_ & 0x7FFF, exp;
+ if(!abs)
+ return half(detail::binary, detail::pole(0x8000));
+ if(abs >= 0x7C00)
+ return half(detail::binary, (abs==0x7C00) ? 0x7C00 : detail::signal(arg.data_));
+ for(exp=(abs>>10)-15; abs<0x200; abs<<=1,--exp) ;
+ unsigned int value = static_cast<unsigned>(exp<0) << 15;
+ if(exp)
+ {
+ unsigned int m = std::abs(exp) << 6;
+ for(exp=18; m<0x400; m<<=1,--exp) ;
+ value |= (exp<<10) + m;
+ }
+ return half(detail::binary, value);
+ }
+
+ /// Next representable value.
+ /// **See also:** Documentation for [std::nextafter](https://en.cppreference.com/w/cpp/numeric/math/nextafter).
+ /// \param from value to compute next representable value for
+ /// \param to direction towards which to compute next value
+ /// \return next representable value after \a from in direction towards \a to
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW for infinite result from finite argument
+ /// \exception FE_UNDERFLOW for subnormal result
+ inline half nextafter(half from, half to)
+ {
+ int fabs = from.data_ & 0x7FFF, tabs = to.data_ & 0x7FFF;
+ if(fabs > 0x7C00 || tabs > 0x7C00)
+ return half(detail::binary, detail::signal(from.data_, to.data_));
+ if(from.data_ == to.data_ || !(fabs|tabs))
+ return to;
+ if(!fabs)
+ {
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT);
+ return half(detail::binary, (to.data_&0x8000)+1);
+ }
+ unsigned int out = from.data_ + (((from.data_>>15)^static_cast<unsigned>(
+ (from.data_^(0x8000|(0x8000-(from.data_>>15))))<(to.data_^(0x8000|(0x8000-(to.data_>>15))))))<<1) - 1;
+ detail::raise(FE_OVERFLOW, fabs<0x7C00 && (out&0x7C00)==0x7C00);
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT && (out&0x7C00)<0x400);
+ return half(detail::binary, out);
+ }
+
+ /// Next representable value.
+ /// **See also:** Documentation for [std::nexttoward](https://en.cppreference.com/w/cpp/numeric/math/nexttoward).
+ /// \param from value to compute next representable value for
+ /// \param to direction towards which to compute next value
+ /// \return next representable value after \a from in direction towards \a to
+ /// \exception FE_INVALID for signaling NaN
+ /// \exception FE_OVERFLOW for infinite result from finite argument
+ /// \exception FE_UNDERFLOW for subnormal result
+ inline half nexttoward(half from, long double to)
+ {
+ int fabs = from.data_ & 0x7FFF;
+ if(fabs > 0x7C00)
+ return half(detail::binary, detail::signal(from.data_));
+ long double lfrom = static_cast<long double>(from);
+ if(detail::builtin_isnan(to) || lfrom == to)
+ return half(static_cast<float>(to));
+ if(!fabs)
+ {
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT);
+ return half(detail::binary, (static_cast<unsigned>(detail::builtin_signbit(to))<<15)+1);
+ }
+ unsigned int out = from.data_ + (((from.data_>>15)^static_cast<unsigned>(lfrom<to))<<1) - 1;
+ detail::raise(FE_OVERFLOW, (out&0x7FFF)==0x7C00);
+ detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT && (out&0x7FFF)<0x400);
+ return half(detail::binary, out);
+ }
+
+ /// Take sign.
+ /// **See also:** Documentation for [std::copysign](https://en.cppreference.com/w/cpp/numeric/math/copysign).
+ /// \param x value to change sign for
+ /// \param y value to take sign from
+ /// \return value equal to \a x in magnitude and to \a y in sign
+ inline HALF_CONSTEXPR half copysign(half x, half y) { return half(detail::binary, x.data_^((x.data_^y.data_)&0x8000)); }
+
+ /// \}
+ /// \anchor classification
+ /// \name Floating point classification
+ /// \{
+
+ /// Classify floating-point value.
+ /// **See also:** Documentation for [std::fpclassify](https://en.cppreference.com/w/cpp/numeric/math/fpclassify).
+ /// \param arg number to classify
+ /// \retval FP_ZERO for positive and negative zero
+ /// \retval FP_SUBNORMAL for subnormal numbers
+ /// \retval FP_INFINITY for positive and negative infinity
+ /// \retval FP_NAN for NaNs
+ /// \retval FP_NORMAL for all other (normal) values
+ inline HALF_CONSTEXPR int fpclassify(half arg)
+ {
+ return !(arg.data_&0x7FFF) ? FP_ZERO :
+ ((arg.data_&0x7FFF)<0x400) ? FP_SUBNORMAL :
+ ((arg.data_&0x7FFF)<0x7C00) ? FP_NORMAL :
+ ((arg.data_&0x7FFF)==0x7C00) ? FP_INFINITE :
+ FP_NAN;
+ }
+
+ /// Check if finite number.
+ /// **See also:** Documentation for [std::isfinite](https://en.cppreference.com/w/cpp/numeric/math/isfinite).
+ /// \param arg number to check
+ /// \retval true if neither infinity nor NaN
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isfinite(half arg) { return (arg.data_&0x7C00) != 0x7C00; }
+
+ /// Check for infinity.
+ /// **See also:** Documentation for [std::isinf](https://en.cppreference.com/w/cpp/numeric/math/isinf).
+ /// \param arg number to check
+ /// \retval true for positive or negative infinity
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isinf(half arg) { return (arg.data_&0x7FFF) == 0x7C00; }
+
+ /// Check for NaN.
+ /// **See also:** Documentation for [std::isnan](https://en.cppreference.com/w/cpp/numeric/math/isnan).
+ /// \param arg number to check
+ /// \retval true for NaNs
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isnan(half arg) { return (arg.data_&0x7FFF) > 0x7C00; }
+
+ /// Check if normal number.
+ /// **See also:** Documentation for [std::isnormal](https://en.cppreference.com/w/cpp/numeric/math/isnormal).
+ /// \param arg number to check
+ /// \retval true if normal number
+ /// \retval false if either subnormal, zero, infinity or NaN
+ inline HALF_CONSTEXPR bool isnormal(half arg) { return ((arg.data_&0x7C00)!=0) & ((arg.data_&0x7C00)!=0x7C00); }
+
+ /// Check sign.
+ /// **See also:** Documentation for [std::signbit](https://en.cppreference.com/w/cpp/numeric/math/signbit).
+ /// \param arg number to check
+ /// \retval true for negative number
+ /// \retval false for positive number
+ inline HALF_CONSTEXPR bool signbit(half arg) { return (arg.data_&0x8000) != 0; }
+
+ /// \}
+ /// \anchor compfunc
+ /// \name Comparison
+ /// \{
+
+ /// Quiet comparison for greater than.
+ /// **See also:** Documentation for [std::isgreater](https://en.cppreference.com/w/cpp/numeric/math/isgreater).
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x greater than \a y
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isgreater(half x, half y)
+ {
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) > ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
+ }
+
+ /// Quiet comparison for greater equal.
+ /// **See also:** Documentation for [std::isgreaterequal](https://en.cppreference.com/w/cpp/numeric/math/isgreaterequal).
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x greater equal \a y
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isgreaterequal(half x, half y)
+ {
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) >= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
+ }
+
+ /// Quiet comparison for less than.
+ /// **See also:** Documentation for [std::isless](https://en.cppreference.com/w/cpp/numeric/math/isless).
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x less than \a y
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isless(half x, half y)
+ {
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) < ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
+ }
+
+ /// Quiet comparison for less equal.
+ /// **See also:** Documentation for [std::islessequal](https://en.cppreference.com/w/cpp/numeric/math/islessequal).
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if \a x less equal \a y
+ /// \retval false else
+ inline HALF_CONSTEXPR bool islessequal(half x, half y)
+ {
+ return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) <= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y);
+ }
+
+ /// Quiet comarison for less or greater.
+ /// **See also:** Documentation for [std::islessgreater](https://en.cppreference.com/w/cpp/numeric/math/islessgreater).
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if either less or greater
+ /// \retval false else
+ inline HALF_CONSTEXPR bool islessgreater(half x, half y)
+ {
+ return x.data_!=y.data_ && ((x.data_|y.data_)&0x7FFF) && !isnan(x) && !isnan(y);
+ }
+
+ /// Quiet check if unordered.
+ /// **See also:** Documentation for [std::isunordered](https://en.cppreference.com/w/cpp/numeric/math/isunordered).
+ /// \param x first operand
+ /// \param y second operand
+ /// \retval true if unordered (one or two NaN operands)
+ /// \retval false else
+ inline HALF_CONSTEXPR bool isunordered(half x, half y) { return isnan(x) || isnan(y); }
+
+ /// \}
+ /// \anchor casting
+ /// \name Casting
+ /// \{
+
+ /// Cast to or from half-precision floating-point number.
+ /// This casts between [half](\ref half_float::half) and any built-in arithmetic type. The values are converted
+ /// directly using the default rounding mode, without any roundtrip over `float` that a `static_cast` would otherwise do.
+ ///
+ /// Using this cast with neither of the two types being a [half](\ref half_float::half) or with any of the two types
+ /// not being a built-in arithmetic type (apart from [half](\ref half_float::half), of course) results in a compiler
+ /// error and casting between [half](\ref half_float::half)s returns the argument unmodified.
+ /// \tparam T destination type (half or built-in arithmetic type)
+ /// \tparam U source type (half or built-in arithmetic type)
+ /// \param arg value to cast
+ /// \return \a arg converted to destination type
+ /// \exception FE_INVALID if \a T is integer type and result is not representable as \a T
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ template<typename T,typename U> T half_cast(U arg) { return detail::half_caster<T,U>::cast(arg); }
+
+ /// Cast to or from half-precision floating-point number.
+ /// This casts between [half](\ref half_float::half) and any built-in arithmetic type. The values are converted
+ /// directly using the specified rounding mode, without any roundtrip over `float` that a `static_cast` would otherwise do.
+ ///
+ /// Using this cast with neither of the two types being a [half](\ref half_float::half) or with any of the two types
+ /// not being a built-in arithmetic type (apart from [half](\ref half_float::half), of course) results in a compiler
+ /// error and casting between [half](\ref half_float::half)s returns the argument unmodified.
+ /// \tparam T destination type (half or built-in arithmetic type)
+ /// \tparam R rounding mode to use.
+ /// \tparam U source type (half or built-in arithmetic type)
+ /// \param arg value to cast
+ /// \return \a arg converted to destination type
+ /// \exception FE_INVALID if \a T is integer type and result is not representable as \a T
+ /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding
+ template<typename T,std::float_round_style R,typename U> T half_cast(U arg) { return detail::half_caster<T,U,R>::cast(arg); }
+ /// \}
+
+ /// \}
+ /// \anchor errors
+ /// \name Error handling
+ /// \{
+
+ /// Clear exception flags.
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
+ /// but in that case manual flag management is the only way to raise flags.
+ ///
+ /// **See also:** Documentation for [std::feclearexcept](https://en.cppreference.com/w/cpp/numeric/fenv/feclearexcept).
+ /// \param excepts OR of exceptions to clear
+ /// \retval 0 all selected flags cleared successfully
+ inline int feclearexcept(int excepts) { detail::errflags() &= ~excepts; return 0; }
+
+ /// Test exception flags.
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
+ /// but in that case manual flag management is the only way to raise flags.
+ ///
+ /// **See also:** Documentation for [std::fetestexcept](https://en.cppreference.com/w/cpp/numeric/fenv/fetestexcept).
+ /// \param excepts OR of exceptions to test
+ /// \return OR of selected exceptions if raised
+ inline int fetestexcept(int excepts) { return detail::errflags() & excepts; }
+
+ /// Raise exception flags.
+ /// This raises the specified floating point exceptions and also invokes any additional automatic exception handling as
+ /// configured with the [HALF_ERRHANDLIG_...](\ref HALF_ERRHANDLING_ERRNO) preprocessor symbols.
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
+ /// but in that case manual flag management is the only way to raise flags.
+ ///
+ /// **See also:** Documentation for [std::feraiseexcept](https://en.cppreference.com/w/cpp/numeric/fenv/feraiseexcept).
+ /// \param excepts OR of exceptions to raise
+ /// \retval 0 all selected exceptions raised successfully
+ inline int feraiseexcept(int excepts) { detail::errflags() |= excepts; detail::raise(excepts); return 0; }
+
+ /// Save exception flags.
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
+ /// but in that case manual flag management is the only way to raise flags.
+ ///
+ /// **See also:** Documentation for [std::fegetexceptflag](https://en.cppreference.com/w/cpp/numeric/fenv/feexceptflag).
+ /// \param flagp adress to store flag state at
+ /// \param excepts OR of flags to save
+ /// \retval 0 for success
+ inline int fegetexceptflag(int *flagp, int excepts) { *flagp = detail::errflags() & excepts; return 0; }
+
+ /// Restore exception flags.
+ /// This only copies the specified exception state (including unset flags) without incurring any additional exception handling.
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
+ /// but in that case manual flag management is the only way to raise flags.
+ ///
+ /// **See also:** Documentation for [std::fesetexceptflag](https://en.cppreference.com/w/cpp/numeric/fenv/feexceptflag).
+ /// \param flagp adress to take flag state from
+ /// \param excepts OR of flags to restore
+ /// \retval 0 for success
+ inline int fesetexceptflag(const int *flagp, int excepts) { detail::errflags() = (detail::errflags()|(*flagp&excepts)) & (*flagp|~excepts); return 0; }
+
+ /// Throw C++ exceptions based on set exception flags.
+ /// This function manually throws a corresponding C++ exception if one of the specified flags is set,
+ /// no matter if automatic throwing (via [HALF_ERRHANDLING_THROW_...](\ref HALF_ERRHANDLING_THROW_INVALID)) is enabled or not.
+ /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled,
+ /// but in that case manual flag management is the only way to raise flags.
+ /// \param excepts OR of exceptions to test
+ /// \param msg error message to use for exception description
+ /// \throw std::domain_error if `FE_INVALID` or `FE_DIVBYZERO` is selected and set
+ /// \throw std::overflow_error if `FE_OVERFLOW` is selected and set
+ /// \throw std::underflow_error if `FE_UNDERFLOW` is selected and set
+ /// \throw std::range_error if `FE_INEXACT` is selected and set
+
+ #if not defined HALF_ENABLE_CPP11_NOEXCEPT
+ inline void fethrowexcept(int excepts, const char *msg = "")
+ {
+ excepts &= detail::errflags();
+ if(excepts & (FE_INVALID|FE_DIVBYZERO))
+ throw std::domain_error(msg);
+ if(excepts & FE_OVERFLOW)
+ throw std::overflow_error(msg);
+ if(excepts & FE_UNDERFLOW)
+ throw std::underflow_error(msg);
+ if(excepts & FE_INEXACT)
+ throw std::range_error(msg);
+ }
+ #endif //HALF_ENABLE_CPP11_NOEXCEPT
+ /// \}
+}
+
+
+#undef HALF_UNUSED_NOERR
+#undef HALF_CONSTEXPR
+#undef HALF_CONSTEXPR_CONST
+#undef HALF_CONSTEXPR_NOERR
+#undef HALF_NOEXCEPT
+#undef HALF_NOTHROW
+#undef HALF_THREAD_LOCAL
+#undef HALF_TWOS_COMPLEMENT_INT
+#ifdef HALF_POP_WARNINGS
+ #pragma warning(pop)
+ #undef HALF_POP_WARNINGS
+#endif
+
+#endif
|