diff options
Diffstat (limited to 'src/cpu/kernels/elementwise_binary/generic/sve')
-rw-r--r-- | src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp | 107 | ||||
-rw-r--r-- | src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp | 102 | ||||
-rw-r--r-- | src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp | 297 | ||||
-rw-r--r-- | src/cpu/kernels/elementwise_binary/generic/sve/impl.h | 167 | ||||
-rw-r--r-- | src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp | 199 |
5 files changed, 872 insertions, 0 deletions
diff --git a/src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp b/src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp new file mode 100644 index 0000000000..600c7f1c05 --- /dev/null +++ b/src/cpu/kernels/elementwise_binary/generic/sve/fp16.cpp @@ -0,0 +1,107 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) + +#include "arm_compute/core/Helpers.h" + +#include "src/cpu/kernels/elementwise_binary/generic/sve/impl.h" +namespace arm_compute +{ +namespace cpu +{ +template <ArithmeticOperation op> +void sve_fp16_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_arithmetic_op<float16_t>(in1, in2, out, op, window); +} + +template void sve_fp16_elementwise_binary<ArithmeticOperation::ADD>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::SUB>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::DIV>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::MIN>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::MAX>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::SQUARED_DIFF>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::POWER>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_elementwise_binary<ArithmeticOperation::PRELU>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +template <ComparisonOperation op> +void sve_fp16_comparison_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_comparison_op<float16_t>(in1, in2, out, op, window); +} + +template void sve_fp16_comparison_elementwise_binary<ComparisonOperation::Equal>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_comparison_elementwise_binary<ComparisonOperation::NotEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_comparison_elementwise_binary<ComparisonOperation::Greater>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_comparison_elementwise_binary<ComparisonOperation::GreaterEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_comparison_elementwise_binary<ComparisonOperation::Less>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp16_comparison_elementwise_binary<ComparisonOperation::LessEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +} // namespace cpu +} // namespace arm_compute + +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp b/src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp new file mode 100644 index 0000000000..832a966883 --- /dev/null +++ b/src/cpu/kernels/elementwise_binary/generic/sve/fp32.cpp @@ -0,0 +1,102 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ + +#include "arm_compute/core/Helpers.h" + +#include "src/cpu/kernels/elementwise_binary/generic/sve/impl.h" +namespace arm_compute +{ +namespace cpu +{ +template <ArithmeticOperation op> +void sve_fp32_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_arithmetic_op<float32_t>(in1, in2, out, op, window); +} + +template void sve_fp32_elementwise_binary<ArithmeticOperation::ADD>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::SUB>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::DIV>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::MIN>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::MAX>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::SQUARED_DIFF>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::POWER>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_elementwise_binary<ArithmeticOperation::PRELU>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +template <ComparisonOperation op> +void sve_fp32_comparison_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_comparison_op<float>(in1, in2, out, op, window); +} +template void sve_fp32_comparison_elementwise_binary<ComparisonOperation::Equal>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_comparison_elementwise_binary<ComparisonOperation::NotEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_comparison_elementwise_binary<ComparisonOperation::Greater>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_comparison_elementwise_binary<ComparisonOperation::GreaterEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_comparison_elementwise_binary<ComparisonOperation::Less>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_fp32_comparison_elementwise_binary<ComparisonOperation::LessEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp b/src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp new file mode 100644 index 0000000000..fa48407e9b --- /dev/null +++ b/src/cpu/kernels/elementwise_binary/generic/sve/impl.cpp @@ -0,0 +1,297 @@ +/* + * Copyright (c) 2021-2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ + +#include "src/cpu/kernels/elementwise_binary/generic/sve/impl.h" + +#include "src/core/NEON/SVEMath.h" + +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +using namespace arm_compute::wrapper; + +template <typename ScalarType> +void elementwise_arithmetic_op( + const ITensor *in1, const ITensor *in2, ITensor *out, ArithmeticOperation op, const Window &window) +{ + using VectorType = typename sve_vector<ScalarType>::type; + + const auto all_true_pg = svptrue<ScalarType>(); + + // Create input windows + Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()); + + // Clear X Dimension on execution window as we handle manually + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x(); + + if (is_broadcast_across_x) + { + const bool is_broadcast_input_2 = input2_win.x().step() == 0; + Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; + Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; + const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; + + // Clear X Dimension on execution window as we handle manually + non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator broadcast_input(broadcast_tensor, broadcast_win); + Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); + Iterator output(out, win); + + execute_window_loop( + win, + [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr()); + const ScalarType broadcast_value = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr()); + const auto broadcast_vector = svdup_n(broadcast_value); + + int x = window_start_x; + + svbool_t pg = svwhilelt<ScalarType>(x, window_end_x); + do + { + const auto non_broadcast_vector = svld1(pg, non_broadcast_input_ptr + x); + VectorType res{}; + + if (is_broadcast_input_2) + { + res = elementwise_arithmetic_op<typename sve_vector<ScalarType>::type>(pg, non_broadcast_vector, + broadcast_vector, op); + } + else + { + res = elementwise_arithmetic_op<typename sve_vector<ScalarType>::type>( + pg, broadcast_vector, non_broadcast_vector, op); + } + svst1(pg, output_ptr + x, res); + + x += svcnt<ScalarType>(); + pg = svwhilelt<ScalarType>(x, window_end_x); + } while (svptest_any(all_true_pg, pg)); + }, + broadcast_input, non_broadcast_input, output); + } + else + { + // Clear X Dimension on execution window as we handle manually + input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input1(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + execute_window_loop( + win, + [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr()); + + int x = window_start_x; + + svbool_t pg = svwhilelt<ScalarType>(x, window_end_x); + do + { + const auto in1 = svld1(pg, input1_ptr + x); + const auto in2 = svld1(pg, input2_ptr + x); + const auto res = elementwise_arithmetic_op<typename sve_vector<ScalarType>::type>(pg, in1, in2, op); + svst1(pg, output_ptr + x, res); + + x += svcnt<ScalarType>(); + pg = svwhilelt<ScalarType>(x, window_end_x); + } while (svptest_any(all_true_pg, pg)); + }, + input1, input2, output); + } +} +template void elementwise_arithmetic_op<float32_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ArithmeticOperation op, const Window &window); +template void elementwise_arithmetic_op<float16_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ArithmeticOperation op, const Window &window); +template void elementwise_arithmetic_op<int16_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ArithmeticOperation op, const Window &window); +template void elementwise_arithmetic_op<int32_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ArithmeticOperation op, const Window &window); + +template <typename InputScalarType, typename OutputScalarType> +void elementwise_comparison_op( + const ITensor *in1, const ITensor *in2, ITensor *out, ComparisonOperation op, const Window &window) +{ + static_assert(sizeof(InputScalarType) >= sizeof(OutputScalarType), + "input data type's width should be equal to or greater than output data type's width"); + + using OutputVectorType = typename sve_vector<OutputScalarType>::type; + const auto all_true_pg = svptrue<InputScalarType>(); + + // Create input windows + Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()); + + // Clear X Dimension on execution window as we handle manually + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x(); + + if (is_broadcast_across_x) + { + const bool is_broadcast_input_2 = input2_win.x().step() == 0; + Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; + Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; + const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; + + // Clear X Dimension on execution window as we handle manually + non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator broadcast_input(broadcast_tensor, broadcast_win); + Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); + Iterator output(out, win); + + execute_window_loop( + win, + [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr()); + const auto non_broadcast_input_ptr = + reinterpret_cast<const InputScalarType *>(non_broadcast_input.ptr()); + const InputScalarType broadcast_value = + *reinterpret_cast<const InputScalarType *>(broadcast_input.ptr()); + const auto broadcast_vector = svdup_n(broadcast_value); + + int x = window_start_x; + + svbool_t pg = svwhilelt<InputScalarType>(x, window_end_x); + do + { + const auto non_broadcast_vector = svld1(pg, non_broadcast_input_ptr + x); + const svbool_t output_pg = narrow_to_byte_predicate<sizeof(InputScalarType)>(pg); + OutputVectorType res{}; + if (is_broadcast_input_2) + { + res = elementwise_comparison_op<typename sve_vector<InputScalarType>::type, + typename sve_vector<OutputScalarType>::type>( + pg, non_broadcast_vector, broadcast_vector, op); + } + else + { + res = elementwise_comparison_op<typename sve_vector<InputScalarType>::type, + typename sve_vector<OutputScalarType>::type>( + pg, broadcast_vector, non_broadcast_vector, op); + } + svst1(output_pg, output_ptr + x, res); + + x += svcnt<InputScalarType>(); + pg = svwhilelt<InputScalarType>(x, window_end_x); + } while (svptest_any(all_true_pg, pg)); + }, + broadcast_input, non_broadcast_input, output); + } + else + { + // Clear X Dimension on execution window as we handle manually + input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input1(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + execute_window_loop( + win, + [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast<OutputScalarType *>(output.ptr()); + const auto input1_ptr = reinterpret_cast<const InputScalarType *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const InputScalarType *>(input2.ptr()); + + int x = window_start_x; + + svbool_t pg = svwhilelt<InputScalarType>(x, window_end_x); + do + { + const auto in1 = svld1(pg, input1_ptr + x); + const auto in2 = svld1(pg, input2_ptr + x); + const auto res = + elementwise_comparison_op<typename sve_vector<InputScalarType>::type, + typename sve_vector<OutputScalarType>::type>(pg, in1, in2, op); + const svbool_t output_pg = narrow_to_byte_predicate<sizeof(InputScalarType)>(pg); + svst1(output_pg, output_ptr + x, res); + + x += svcnt<InputScalarType>(); + pg = svwhilelt<InputScalarType>(x, window_end_x); + } while (svptest_any(all_true_pg, pg)); + }, + input1, input2, output); + } +} + +template void elementwise_comparison_op<float32_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ComparisonOperation op, const Window &window); +template void elementwise_comparison_op<float16_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ComparisonOperation op, const Window &window); +template void elementwise_comparison_op<uint8_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ComparisonOperation op, const Window &window); +template void elementwise_comparison_op<int16_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ComparisonOperation op, const Window &window); +template void elementwise_comparison_op<int32_t>( + const ITensor *in1, const ITensor *in2, ITensor *out, const ComparisonOperation op, const Window &window); + +template <> +svint32_t elementwise_pow<svint32_t>(svbool_t &pg, const svint32_t &a, const svint32_t &b) +{ + return svcvt_s32_z(pg, svpow_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b))); +} + +template <> +svint32_t elementwise_div<svint32_t>(svbool_t &pg, const svint32_t &a, const svint32_t &b) +{ + return svcvt_s32_z(pg, svdiv_z(pg, svcvt_f32_z(pg, a), svcvt_f32_z(pg, b))); +} + +template <> +svint16_t elementwise_div<svint16_t>(svbool_t &pg, const svint16_t &a, const svint16_t &b) +{ + ARM_COMPUTE_UNUSED(pg, a, b); + ARM_COMPUTE_ERROR("Not supported"); +} + +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/elementwise_binary/generic/sve/impl.h b/src/cpu/kernels/elementwise_binary/generic/sve/impl.h new file mode 100644 index 0000000000..4c61b9f315 --- /dev/null +++ b/src/cpu/kernels/elementwise_binary/generic/sve/impl.h @@ -0,0 +1,167 @@ +/* + * Copyright (c) 2021-2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ +#ifndef SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H +#define SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H + +#include "arm_compute/core/Helpers.h" + +#include "src/core/NEON/wrapper/intrinsics/intrinsics.h" +#include "src/core/NEON/wrapper/svtraits.h" + +namespace arm_compute +{ +namespace cpu +{ +using namespace arm_compute::wrapper; + +template <typename VectorType> +VectorType elementwise_pow(svbool_t &pg, const VectorType &a, const VectorType &b) +{ + return svpow_z(pg, a, b); +} + +template <typename VectorType> +VectorType elementwise_div(svbool_t &pg, const VectorType &a, const VectorType &b) +{ + return svdiv_z(pg, a, b); +} + +template <uint32_t bytewidth> +svbool_t narrow_to_byte_predicate(svbool_t pg) +{ + const auto all_false = svpfalse(); + + switch (bytewidth) + { + case 8: + pg = svuzp1_b32(pg, all_false); + /* fall through */ + case 4: + pg = svuzp1_b16(pg, all_false); + /* fall through */ + case 2: + pg = svuzp1_b8(pg, all_false); + /* fall through */ + default: + break; + } + return pg; +} + +template <typename VectorType> +VectorType elementwise_arithmetic_op(svbool_t &pg, const VectorType &a, const VectorType &b, ArithmeticOperation op) +{ + using ScalarType = typename wrapper::sve_scalar<VectorType>::type; + VectorType res{}; + + switch (op) + { + case ArithmeticOperation::MAX: + res = svmax_z(pg, a, b); + break; + case ArithmeticOperation::MIN: + res = svmin_z(pg, a, b); + break; + case ArithmeticOperation::SQUARED_DIFF: + { + const auto tmp = svsub_z(pg, a, b); + res = svmul_z(pg, tmp, tmp); + break; + } + case ArithmeticOperation::PRELU: + { + const auto zero = svdup_n(ScalarType(0)); + const auto tmp = svmul_z(pg, a, b); + const auto gt = svcmpgt(pg, a, zero); + res = svsel(gt, a, tmp); + break; + } + case ArithmeticOperation::DIV: + { + res = elementwise_div(pg, a, b); + break; + } + case ArithmeticOperation::POWER: + { + res = elementwise_pow(pg, a, b); + break; + } + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + + return res; +} + +template <typename InputVectorType, typename OutputVectorType> +OutputVectorType +elementwise_comparison_op(svbool_t &pg, const InputVectorType &a, const InputVectorType &b, ComparisonOperation op) +{ + svbool_t selection_vector{}; + + switch (op) + { + case ComparisonOperation::Equal: + selection_vector = svcmpeq(pg, a, b); + break; + case ComparisonOperation::NotEqual: + selection_vector = svcmpne(pg, a, b); + break; + case ComparisonOperation::Greater: + selection_vector = svcmpgt(pg, a, b); + break; + case ComparisonOperation::GreaterEqual: + selection_vector = svcmpge(pg, a, b); + break; + case ComparisonOperation::Less: + selection_vector = svcmplt(pg, a, b); + break; + case ComparisonOperation::LessEqual: + selection_vector = svcmple(pg, a, b); + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + + using InputScalarType = typename wrapper::sve_scalar<InputVectorType>::type; + selection_vector = narrow_to_byte_predicate<sizeof(InputScalarType)>(selection_vector); + + using OutputScalarType = typename wrapper::sve_scalar<OutputVectorType>::type; + const auto false_vector = svdup_n(static_cast<OutputScalarType>((uint32_t)0)); + const auto true_vector = svdup_n(static_cast<OutputScalarType>(~(uint32_t)0)); + auto ret = svsel(selection_vector, true_vector, false_vector); + + return ret; +} + +template <typename ScalarType> +void elementwise_arithmetic_op( + const ITensor *in1, const ITensor *in2, ITensor *out, ArithmeticOperation op, const Window &window); + +template <typename ScalarType, typename OutputScalarType = uint8_t> +void elementwise_comparison_op( + const ITensor *in1, const ITensor *in2, ITensor *out, ComparisonOperation op, const Window &window); +} // namespace cpu +} // namespace arm_compute +#endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_LIST_H */ diff --git a/src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp b/src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp new file mode 100644 index 0000000000..f7714ff7e9 --- /dev/null +++ b/src/cpu/kernels/elementwise_binary/generic/sve/integer.cpp @@ -0,0 +1,199 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * 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. + */ + +#include "arm_compute/core/Helpers.h" + +#include "src/cpu/kernels/elementwise_binary/generic/sve/impl.h" +namespace arm_compute +{ +namespace cpu +{ +template <ArithmeticOperation op> +void sve_s32_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_arithmetic_op<int32_t>(in1, in2, out, op, window); +} +template void sve_s32_elementwise_binary<ArithmeticOperation::ADD>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::SUB>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::DIV>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::MIN>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::MAX>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::SQUARED_DIFF>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::POWER>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_elementwise_binary<ArithmeticOperation::PRELU>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +template <ArithmeticOperation op> +void sve_s16_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_arithmetic_op<int16_t>(in1, in2, out, op, window); +} +template void sve_s16_elementwise_binary<ArithmeticOperation::ADD>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::SUB>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::DIV>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::MIN>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::MAX>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::SQUARED_DIFF>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::POWER>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_elementwise_binary<ArithmeticOperation::PRELU>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +template <ComparisonOperation op> +void sve_u8_comparison_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_comparison_op<uint8_t>(in1, in2, out, op, window); +} +template void sve_u8_comparison_elementwise_binary<ComparisonOperation::Equal>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_u8_comparison_elementwise_binary<ComparisonOperation::NotEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_u8_comparison_elementwise_binary<ComparisonOperation::Greater>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_u8_comparison_elementwise_binary<ComparisonOperation::GreaterEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_u8_comparison_elementwise_binary<ComparisonOperation::Less>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_u8_comparison_elementwise_binary<ComparisonOperation::LessEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +template <ComparisonOperation op> +void sve_s16_comparison_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_comparison_op<int16_t>(in1, in2, out, op, window); +} +template void sve_s16_comparison_elementwise_binary<ComparisonOperation::Equal>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_comparison_elementwise_binary<ComparisonOperation::NotEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_comparison_elementwise_binary<ComparisonOperation::Greater>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_comparison_elementwise_binary<ComparisonOperation::GreaterEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_comparison_elementwise_binary<ComparisonOperation::Less>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s16_comparison_elementwise_binary<ComparisonOperation::LessEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +template <ComparisonOperation op> +void sve_s32_comparison_elementwise_binary(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + return elementwise_comparison_op<int32_t>(in1, in2, out, op, window); +} +template void sve_s32_comparison_elementwise_binary<ComparisonOperation::Equal>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_comparison_elementwise_binary<ComparisonOperation::NotEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_comparison_elementwise_binary<ComparisonOperation::Greater>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_comparison_elementwise_binary<ComparisonOperation::GreaterEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_comparison_elementwise_binary<ComparisonOperation::Less>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); +template void sve_s32_comparison_elementwise_binary<ComparisonOperation::LessEqual>(const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window); + +} // namespace cpu +} // namespace arm_compute |