From 3b162e53267d13d18891baf3372f971f1d4213d3 Mon Sep 17 00:00:00 2001 From: David Svantesson Date: Tue, 28 Mar 2023 14:13:32 +0000 Subject: Reorder added Adds Reorder kernel exposing blocking reorders from arm_gemm Resolves ONCPUML-1232 Change-Id: I42bf4166311fe1771565134d3ed7039fc8e30230 Signed-off-by: David Svantesson Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9500 Comments-Addressed: Arm Jenkins Reviewed-by: SiCong Li Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .bazelrc | 1 + Android.bp | 2 + BUILD.bazel | 23 ++- CMakeLists.txt | 12 ++ arm_compute/runtime/NEON/NEFunctions.h | 1 + .../runtime/NEON/functions/NEReorderLayer.h | 84 ++++++++ cmake/Options.cmake | 12 +- filelist.json | 8 + src/BUILD.bazel | 2 + src/CMakeLists.txt | 2 + src/core/NEON/NEKernels.h | 3 +- src/core/NEON/kernels/NEReorderKernel.cpp | 223 +++++++++++++++++++++ src/core/NEON/kernels/NEReorderKernel.h | 93 +++++++++ src/runtime/NEON/functions/NEReorder.cpp | 53 +++++ tests/BUILD.bazel | 7 + tests/CMakeLists.txt | 1 + tests/datasets/ReorderLayerDataset.h | 158 +++++++++++++++ tests/validation/CMakeLists.txt | 1 + tests/validation/NEON/ReorderLayer.cpp | 66 ++++++ tests/validation/fixtures/ReorderFixture.h | 109 ++++++++++ tests/validation/reference/Reorder.cpp | 156 ++++++++++++++ tests/validation/reference/Reorder.h | 44 ++++ 22 files changed, 1046 insertions(+), 15 deletions(-) create mode 100644 arm_compute/runtime/NEON/functions/NEReorderLayer.h create mode 100644 src/core/NEON/kernels/NEReorderKernel.cpp create mode 100644 src/core/NEON/kernels/NEReorderKernel.h create mode 100644 src/runtime/NEON/functions/NEReorder.cpp create mode 100644 tests/datasets/ReorderLayerDataset.h create mode 100644 tests/validation/NEON/ReorderLayer.cpp create mode 100644 tests/validation/fixtures/ReorderFixture.h create mode 100644 tests/validation/reference/Reorder.cpp create mode 100644 tests/validation/reference/Reorder.h diff --git a/.bazelrc b/.bazelrc index 8611db3d4a..d6a5e62871 100644 --- a/.bazelrc +++ b/.bazelrc @@ -37,3 +37,4 @@ build --flag_alias=logging=//:logging build --flag_alias=openmp=//:openmp build --flag_alias=cppthreads=//:cppthreads build --flag_alias=enable_bf16_validation=//:enable_bf16_validation +build --flag_alias=enable_sve_validation=//:enable_sve_validation diff --git a/Android.bp b/Android.bp index bfc747405f..83b971bd0e 100644 --- a/Android.bp +++ b/Android.bp @@ -296,6 +296,7 @@ cc_library_static { "src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp", "src/core/NEON/kernels/NERangeKernel.cpp", "src/core/NEON/kernels/NEReductionOperationKernel.cpp", + "src/core/NEON/kernels/NEReorderKernel.cpp", "src/core/NEON/kernels/NEReorgLayerKernel.cpp", "src/core/NEON/kernels/NEReverseKernel.cpp", "src/core/NEON/kernels/NESelectKernel.cpp", @@ -963,6 +964,7 @@ cc_library_static { "src/runtime/NEON/functions/NERange.cpp", "src/runtime/NEON/functions/NEReduceMean.cpp", "src/runtime/NEON/functions/NEReductionOperation.cpp", + "src/runtime/NEON/functions/NEReorder.cpp", "src/runtime/NEON/functions/NEReorgLayer.cpp", "src/runtime/NEON/functions/NEReshapeLayer.cpp", "src/runtime/NEON/functions/NEReverse.cpp", diff --git a/BUILD.bazel b/BUILD.bazel index e3ad75abd8..d246336c29 100644 --- a/BUILD.bazel +++ b/BUILD.bazel @@ -72,6 +72,12 @@ bool_flag( visibility = ["//visibility:public"], ) +bool_flag( + name = "enable_sve_validation", + build_setting_default = False, + visibility = ["//visibility:public"], +) + #--------------------------------------------------------------------- # Flag variables config_setting( @@ -116,6 +122,13 @@ config_setting( }, ) +config_setting( + name = "sve_validation_flag", + flag_values = { + ":enable_sve_validation": "true", + }, +) + #--------------------------------------------------------------------- # Common defines used for all targets @@ -138,8 +151,6 @@ cc_library( "DARM_COMPUTE_GRAPH_ENABLED", "ARM_COMPUTE_ENABLE_SVEF32MM", "ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS", - "ENABLE_SVE", - "ARM_COMPUTE_ENABLE_SVE", "_GLIBCXX_USE_NANOSLEEP" ] + select({ "//:debug_flag": [ @@ -243,8 +254,10 @@ cc_library( "//conditions:default": [], }), local_defines = [ + "ENABLE_SVE", + "ARM_COMPUTE_ENABLE_SVE", "ARM_COMPUTE_ENABLE_SVE2", - "ARM_COMPUTE_ENABLE_BF16" + "ARM_COMPUTE_ENABLE_BF16", ], deps = [ "//:common_defines", @@ -292,6 +305,8 @@ cc_library( "//conditions:default": [], }), local_defines = [ + "ENABLE_SVE", + "ARM_COMPUTE_ENABLE_SVE", "ARM_COMPUTE_ENABLE_BF16", ], deps = [ @@ -348,6 +363,8 @@ cc_library( "//conditions:default": [], }), local_defines = [ + "ENABLE_SVE", + "ARM_COMPUTE_ENABLE_SVE", "ARM_COMPUTE_ENABLE_BF16", ], visibility = ["//visibility:public"], diff --git a/CMakeLists.txt b/CMakeLists.txt index 72992ed038..060370d33a 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -149,6 +149,8 @@ add_library(arm_compute_sve "") target_compile_options(arm_compute_sve PRIVATE "-march=armv8.2-a+sve+fp16+dotprod") target_compile_definitions(arm_compute_sve PRIVATE ARM_COMPUTE_ENABLE_BF16) +target_compile_definitions(arm_compute_sve PRIVATE ENABLE_SVE) +target_compile_definitions(arm_compute_sve PRIVATE ARM_COMPUTE_ENABLE_SVE) target_include_directories( arm_compute_sve PUBLIC $ @@ -170,6 +172,8 @@ target_compile_options(arm_compute_sve2 PRIVATE "-march=armv8.6-a+sve2+fp16+dotprod") target_compile_definitions(arm_compute_sve2 PRIVATE ARM_COMPUTE_ENABLE_SVE2) target_compile_definitions(arm_compute_sve2 PRIVATE ARM_COMPUTE_ENABLE_BF16) +target_compile_definitions(arm_compute_sve2 PRIVATE ENABLE_SVE) +target_compile_definitions(arm_compute_sve2 PRIVATE ARM_COMPUTE_ENABLE_SVE) target_include_directories( arm_compute_sve2 PUBLIC $ @@ -189,6 +193,8 @@ target_include_directories( add_library(arm_compute_core "") target_compile_options(arm_compute_core PRIVATE "-march=armv8.2-a+fp16") target_compile_definitions(arm_compute_core PRIVATE ARM_COMPUTE_ENABLE_BF16) +target_compile_definitions(arm_compute_core PRIVATE ENABLE_SVE) +target_compile_definitions(arm_compute_core PRIVATE ARM_COMPUTE_ENABLE_SVE) target_include_directories( arm_compute_core PUBLIC $ @@ -211,6 +217,8 @@ target_link_libraries( add_library(arm_compute_graph "") target_compile_options(arm_compute_graph PRIVATE "-march=armv8.2-a+fp16") +target_compile_definitions(arm_compute_graph PRIVATE ENABLE_SVE) +target_compile_definitions(arm_compute_graph PRIVATE ARM_COMPUTE_ENABLE_SVE) # add_subdirectory(src/graph) target_include_directories( @@ -262,6 +270,10 @@ if(BUILD_TESTING) if(ENABLE_BF16_VALIDATION) target_compile_definitions(arm_compute_validation PRIVATE ARM_COMPUTE_ENABLE_BF16) endif() + if(ENABLE_SVE_VALIDATION) + target_compile_definitions(arm_compute_validation PRIVATE ENABLE_SVE) + target_compile_definitions(arm_compute_validation PRIVATE ARM_COMPUTE_ENABLE_SVE) + endif() add_subdirectory(tests/validation) target_compile_options(arm_compute_validation PUBLIC ${COMMON_CXX_FLAGS}) set_target_properties( diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h index 836cba7699..3a10310452 100644 --- a/arm_compute/runtime/NEON/NEFunctions.h +++ b/arm_compute/runtime/NEON/NEFunctions.h @@ -93,6 +93,7 @@ #include "arm_compute/runtime/NEON/functions/NERange.h" #include "arm_compute/runtime/NEON/functions/NEReduceMean.h" #include "arm_compute/runtime/NEON/functions/NEReductionOperation.h" +#include "arm_compute/runtime/NEON/functions/NEReorderLayer.h" #include "arm_compute/runtime/NEON/functions/NEReorgLayer.h" #include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h" #include "arm_compute/runtime/NEON/functions/NEReverse.h" diff --git a/arm_compute/runtime/NEON/functions/NEReorderLayer.h b/arm_compute/runtime/NEON/functions/NEReorderLayer.h new file mode 100644 index 0000000000..4d5e3fa850 --- /dev/null +++ b/arm_compute/runtime/NEON/functions/NEReorderLayer.h @@ -0,0 +1,84 @@ +/* + * Copyright (c) 2023 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 ACL_ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEREORDERLAYER +#define ACL_ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEREORDERLAYER + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/IFunction.h" +#include "src/core/NEON/kernels/NEReorderKernel.h" + +namespace arm_compute +{ +class ITensor; +class ITensorInfo; +/** Function to compute blocked reorder. */ +class NEReorderLayer : public IFunction +{ +public: + /** Default constructor */ + NEReorderLayer(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEReorderLayer(const NEReorderLayer &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEReorderLayer &operator=(const NEReorderLayer &) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + NEReorderLayer(NEReorderLayer &&) = delete; + /** Prevent instances of this class from being moved (As this class contains non movable objects) */ + NEReorderLayer &operator=(NEReorderLayer &&) = delete; + /** Default destructor */ + ~NEReorderLayer() = default; + /** Set the input and output tensors. + * + * Valid data layouts: + * - NCHW + * + * Valid data type configurations: + * |src |dst | + * |:--------|:---------| + * |F32 |F32 | + * + * @param[in] input Source tensor. Data type supported: F32. Data layouts supported: NCHW. + * @param[out] output Destination with the same dimensions, data type, data layout as @p input + * except last dimension of data layout which needs to be multiple of blocking parameter ksize + * @param[in] input_wf WeightFormat of input. + * @param[in] output_wf WeightFormat of output. + */ + void configure(const ITensor *input, ITensor *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf); + + /** Static function to check if given info will lead to a valid configuration of @ref NEReorderLayer + * + * Similar to @ref NEReorderLayer::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf); + + // Inherited methods overridden: + void run() override; + +private: + std::unique_ptr _reorder_kernel; /**< Reorder layer kernel */ +}; +} // namespace arm_compute +#endif /* ACL_ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEREORDERLAYER */ diff --git a/cmake/Options.cmake b/cmake/Options.cmake index 170dad9f69..e2aa2bfbff 100644 --- a/cmake/Options.cmake +++ b/cmake/Options.cmake @@ -66,9 +66,8 @@ option(ENABLE_NCHW_KERNELS "" ON) option(ARM_COMPUTE_GRAPH_ENABLED "" ON) option(ARM_COMPUTE_ENABLE_SVEF32MM "" ON) option(ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS "" ON) -option(ENABLE_SVE "" ON) -option(ARM_COMPUTE_ENABLE_SVE "" ON) option(ENABLE_BF16_VALIDATION "" OFF) +option(ENABLE_SVE_VALIDATION "" OFF) if(ENABLE_NEON) add_definitions(-DENABLE_NEON) @@ -112,19 +111,10 @@ endif() if(ARM_COMPUTE_GRAPH_ENABLED) add_definitions(-DARM_COMPUTE_GRAPH_ENABLED) endif() -if(ARM_COMPUTE_ENABLE_BF16) - add_definitions(-DARM_COMPUTE_ENABLE_BF16) -endif() if(ARM_COMPUTE_ENABLE_SVEF32MM) add_definitions(-DARM_COMPUTE_ENABLE_SVEF32MM) endif() if(ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS) add_definitions(-DARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS) endif() -if(ENABLE_SVE) - add_definitions(-DENABLE_SVE) -endif() -if(ARM_COMPUTE_ENABLE_SVE) - add_definitions(-DARM_COMPUTE_ENABLE_SVE) -endif() add_definitions(-D_GLIBCXX_USE_NANOSLEEP) \ No newline at end of file diff --git a/filelist.json b/filelist.json index e62520dc46..52e29ef9bb 100644 --- a/filelist.json +++ b/filelist.json @@ -2077,6 +2077,14 @@ ] } }, + "Reorder": { + "files": { + "common": [ + "src/runtime/NEON/functions/NEReorder.cpp", + "src/core/NEON/kernels/NEReorderKernel.cpp" + ] + } + }, "Reverse": { "files": { "common": [ diff --git a/src/BUILD.bazel b/src/BUILD.bazel index e0140addb7..6e73f2eba6 100644 --- a/src/BUILD.bazel +++ b/src/BUILD.bazel @@ -416,6 +416,7 @@ filegroup( "core/NEON/kernels/NEROIPoolingLayerKernel.cpp", "core/NEON/kernels/NERangeKernel.cpp", "core/NEON/kernels/NEReductionOperationKernel.cpp", + "core/NEON/kernels/NEReorderKernel.cpp", "core/NEON/kernels/NEReorgLayerKernel.cpp", "core/NEON/kernels/NEReverseKernel.cpp", "core/NEON/kernels/NESelectKernel.cpp", @@ -953,6 +954,7 @@ filegroup( "runtime/NEON/functions/NERange.cpp", "runtime/NEON/functions/NEReduceMean.cpp", "runtime/NEON/functions/NEReductionOperation.cpp", + "runtime/NEON/functions/NEReorder.cpp", "runtime/NEON/functions/NEReorgLayer.cpp", "runtime/NEON/functions/NEReshapeLayer.cpp", "runtime/NEON/functions/NEReverse.cpp", diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 28028e85ef..35d1d9966e 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -408,6 +408,7 @@ target_sources( core/NEON/kernels/NEROIPoolingLayerKernel.cpp core/NEON/kernels/NERangeKernel.cpp core/NEON/kernels/NEReductionOperationKernel.cpp + core/NEON/kernels/NEReorderKernel.cpp core/NEON/kernels/NEReorgLayerKernel.cpp core/NEON/kernels/NEReverseKernel.cpp core/NEON/kernels/NESelectKernel.cpp @@ -945,6 +946,7 @@ target_sources( runtime/NEON/functions/NERange.cpp runtime/NEON/functions/NEReduceMean.cpp runtime/NEON/functions/NEReductionOperation.cpp + runtime/NEON/functions/NEReorder.cpp runtime/NEON/functions/NEReorgLayer.cpp runtime/NEON/functions/NEReshapeLayer.cpp runtime/NEON/functions/NEReverse.cpp diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h index 12a1bb9a11..75016b0db1 100644 --- a/src/core/NEON/NEKernels.h +++ b/src/core/NEON/NEKernels.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2022 Arm Limited. + * Copyright (c) 2016-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -54,6 +54,7 @@ #include "src/core/NEON/kernels/NEROIPoolingLayerKernel.h" #include "src/core/NEON/kernels/NERangeKernel.h" #include "src/core/NEON/kernels/NEReductionOperationKernel.h" +#include "src/core/NEON/kernels/NEReorderKernel.h" #include "src/core/NEON/kernels/NEReorgLayerKernel.h" #include "src/core/NEON/kernels/NEReverseKernel.h" #include "src/core/NEON/kernels/NESelectKernel.h" diff --git a/src/core/NEON/kernels/NEReorderKernel.cpp b/src/core/NEON/kernels/NEReorderKernel.cpp new file mode 100644 index 0000000000..c09745604f --- /dev/null +++ b/src/core/NEON/kernels/NEReorderKernel.cpp @@ -0,0 +1,223 @@ +/* + * Copyright (c) 2023 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/core/NEON/kernels/NEReorderKernel.h" +#include "src/common/utils/Log.h" +#include "src/core/NEON/kernels/arm_gemm/transform.hpp" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Validate.h" + +namespace arm_compute +{ + +void NEReorderKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + switch(_input->info()->data_type()) + { + case DataType::F32: + { + const int ksize_rows_elements = _xmax * _ksize; + const int jump_rows = ksize_rows_elements * window.x().start(); + const int k_start = window.x().start() * _ksize; + const int k_end = std::min(window.x().end() * _ksize, _kmax); + const int stride = _kmax; + if(k_start < k_end) + { + + switch(_output_wf) + { + case WeightFormat::OHWIo4: + { + arm_gemm::Transform<4, 1, true, arm_gemm::VLType::None>(reinterpret_cast(_output->buffer()) + jump_rows, reinterpret_cast(_input->buffer()), stride, k_start, k_end, 0, _xmax); + break; + } +#if defined(ARM_COMPUTE_ENABLE_SVE) + case WeightFormat::OHWIo8: + { + arm_gemm::Transform<1, 1, true, arm_gemm::VLType::SVE>(reinterpret_cast(_output->buffer()) + jump_rows, reinterpret_cast(_input->buffer()), stride, k_start, k_end, 0, _xmax); + break; + } +#endif /* ARM_COMPUTE_ENABLE_SVE */ + default: + { + ARM_COMPUTE_ERROR("Unsupported data type!"); + break; + } + } + } + break; + } + default: + ARM_COMPUTE_ERROR("Unsupported data type!"); + } +} + +NEReorderKernel::NEReorderKernel() + : _input(nullptr), _output(nullptr), _ksize(0), _kmax(0), _xmax(0), _input_wf(WeightFormat::ANY), _output_wf(WeightFormat::ANY) +{ +} + +void NEReorderKernel::configure(const ITensor *input, ITensor *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf) +{ + ARM_COMPUTE_LOG_PARAMS(input, output, input_wf, output_wf); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), input_wf, output_wf)); + + // Set variables + _input = input; + _output = output; + _input_wf = input_wf; + _output_wf = output_wf; + + // Setting parameters for transform + auto dims = input->info()->num_dimensions(); + switch(dims) + { + case 2: + { + _xmax = input->info()->dimension(0); // Number of columns in input matrix + _kmax = input->info()->dimension(1); // Number of rows in input matrix + break; + } + case 4: + { + _xmax = input->info()->dimension(2); // Number of columns in input matrix + _kmax = input->info()->dimension(3); // Number of rows in input matrix + break; + } + default: + { + ARM_COMPUTE_ERROR("Only 2 or 4 dimensions supported."); + } + } + + // Configure kernel window + // Window size is set by rows / _ksize + Window win; + int window_size = 0; + switch(_output_wf) + { +#if defined(ARM_COMPUTE_ENABLE_SVE) + case WeightFormat::OHWIo8: + { + _ksize = 8; + window_size = _kmax / _ksize; + break; + } +#endif /* ARM_COMPUTE_ENABLE_SVE */ + case WeightFormat::OHWIo4: + { + _ksize = 4; + window_size = _kmax / _ksize; + break; + } + default: + { + ARM_COMPUTE_ERROR("Unsupported weight format."); + break; + } + } + if(_kmax % _ksize != 0) + { + window_size += 1; + } + + win.set(Window::DimX, Window::Dimension(0, window_size, 1)); + + INEKernel::configure(win); +} + +Status NEReorderKernel::validate(const ITensorInfo *input, const ITensorInfo *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); + if(output->tensor_shape().total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); + // Only input WeightFormat OHWI supported + ARM_COMPUTE_RETURN_ERROR_ON(input_wf != arm_compute::WeightFormat::OHWI); + int input_x_dim; + int input_k_dim; + int output_x_dim; + int output_k_dim; + auto dims = output->num_dimensions(); + switch(dims) + { + case 2: + { + input_x_dim = input->dimension(0); // Number of columns in input matrix + input_k_dim = input->dimension(1); // Number of rows in input matrix + output_x_dim = output->dimension(0); // Number of columns in output matrix + output_k_dim = output->dimension(1); // Number of rows in output matrix + break; + } + case 4: + { + input_x_dim = input->dimension(2); // Number of columns in input matrix + input_k_dim = input->dimension(3); // Number of rows in input matrix + output_x_dim = output->dimension(2); // Number of columns in output matrix + output_k_dim = output->dimension(3); // Number of rows in output matrix + break; + } + default: + { + ARM_COMPUTE_RETURN_ERROR_MSG("Only 2 or 4 dimensions supported."); + } + } + + int ksize; + switch(output_wf) + { + case WeightFormat::OHWIo8: + { + ksize = 8; + break; + } + case WeightFormat::OHWIo4: + { + ksize = 4; + break; + } + default: + { + ARM_COMPUTE_RETURN_ERROR_MSG("Unsupported weight format."); + break; + } + } + + // output k_dim needs to be same as input but multiple of ksize + int32_t rnd_up_input_kdim = arm_compute::ceil_to_multiple(input_k_dim, ksize); + ARM_COMPUTE_RETURN_ERROR_ON(rnd_up_input_kdim != output_k_dim); + // output x_dim needs to be same as input + ARM_COMPUTE_RETURN_ERROR_ON(input_x_dim != output_x_dim); + + } + return Status{}; +} + +} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEReorderKernel.h b/src/core/NEON/kernels/NEReorderKernel.h new file mode 100644 index 0000000000..e744ff9e52 --- /dev/null +++ b/src/core/NEON/kernels/NEReorderKernel.h @@ -0,0 +1,93 @@ +/* + * Copyright (c) 2023 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 ACL_SRC_CORE_NEON_KERNELS_NEREORDERKERNEL +#define ACL_SRC_CORE_NEON_KERNELS_NEREORDERKERNEL + +#include "src/core/NEON/INEKernel.h" +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ + +/** Interface kernel to reorder tensor into blocked format. */ +class NEReorderKernel : public INEKernel +{ +public: + + const char *name() const override + { + return "NEReorderKernel"; + } + + /** Default constructor */ + NEReorderKernel(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEReorderKernel(const NEReorderKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEReorderKernel &operator=(const NEReorderKernel &) = delete; + /** Allow instances of this class to be moved */ + NEReorderKernel(NEReorderKernel &&) = default; + /** Allow instances of this class to be moved */ + NEReorderKernel &operator=(NEReorderKernel &&) = default; + /** Default destructor */ + ~NEReorderKernel() = default; + + /** Initialise the kernel's input and outputs. + * + * @param[in] input Source tensor with 2 or 4 dimensions. Data types supported: F32. + * @param[out] output Destination tensor. Data type supported: same as @p input. Shape same as @p input expect last dimension which needs to be multiple of blocking parameter _ksize. + * @param[in] input_wf WeightFormat of input. + * @param[in] output_wf WeightFormat of output. + */ + void configure(const ITensor *input, ITensor *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf); + + /** Static function to check if given info will lead to a valid configuration of @ref NEReorderKernel + * + * @param[in] input Source tensor with 2 or 4 dimensions. Data types supported: F32. + * @param[in] output Destination tensor. Data type supported: same as @p input. Shape same as @p input expect last dimension which needs to be multiple of blocking parameter _ksize. + * @param[in] input_wf WeightFormat of input. + * @param[in] output_wf WeightFormat of output. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf); + + // Inherited methods overridden: + void run(const Window &window, const ThreadInfo &info) override; + + +/*****************************************************************************/ + +private: + const ITensor *_input{nullptr}; // Input tensor + ITensor *_output{nullptr}; // Output tensor + int32_t _ksize{0}; // Blocking parameter, how many rows kernel reorders on each call + int32_t _kmax{0}; // Rows in input tensor + int32_t _xmax{0}; // Columns in input tensor + WeightFormat _input_wf{WeightFormat::UNSPECIFIED}; // WeightFormat of input tensor + WeightFormat _output_wf{WeightFormat::UNSPECIFIED}; // WeightFormat of output tensor +}; + +} // namespace arm_compute +#endif /* ACL_SRC_CORE_NEON_KERNELS_NEREORDERKERNEL */ diff --git a/src/runtime/NEON/functions/NEReorder.cpp b/src/runtime/NEON/functions/NEReorder.cpp new file mode 100644 index 0000000000..f9bbeba90c --- /dev/null +++ b/src/runtime/NEON/functions/NEReorder.cpp @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2023 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/runtime/NEON/NEScheduler.h" +#include "arm_compute/runtime/NEON/functions/NEReorderLayer.h" + +namespace arm_compute +{ + +NEReorderLayer::NEReorderLayer() + : _reorder_kernel(std::make_unique()) +{ +} + +void NEReorderLayer::configure(const ITensor *input, ITensor *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf) +{ + auto k = std::make_unique(); + k->configure(input, output, input_wf, output_wf); + _reorder_kernel = std::move(k); +} + +void NEReorderLayer::run() +{ + // Run Reorder + NEScheduler::get().schedule(_reorder_kernel.get(), Window::DimX); +} + +Status NEReorderLayer::validate(const ITensorInfo *input, const ITensorInfo *output, arm_compute::WeightFormat input_wf, arm_compute::WeightFormat output_wf) +{ + return NEReorderKernel::validate(input, output, input_wf, output_wf); +} + +} // namespace arm_compute \ No newline at end of file diff --git a/tests/BUILD.bazel b/tests/BUILD.bazel index 8122652991..606f403f38 100644 --- a/tests/BUILD.bazel +++ b/tests/BUILD.bazel @@ -106,6 +106,13 @@ cc_binary( "ARM_COMPUTE_ENABLE_BF16", ], "//conditions:default": [], + }) + + select({ + "//:sve_validation_flag": [ + "ENABLE_SVE", + "ARM_COMPUTE_ENABLE_SVE", + ], + "//conditions:default": [], }) ) diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 163ab82e91..c4b12e770a 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -107,6 +107,7 @@ target_sources( validation/reference/DepthwiseConvolutionLayer.cpp validation/reference/FullyConnectedLayer.cpp validation/reference/ConvolutionLayer.cpp + validation/reference/Reorder.cpp framework/Framework.cpp framework/Utils.cpp framework/Exceptions.cpp diff --git a/tests/datasets/ReorderLayerDataset.h b/tests/datasets/ReorderLayerDataset.h new file mode 100644 index 0000000000..8e1a8422b2 --- /dev/null +++ b/tests/datasets/ReorderLayerDataset.h @@ -0,0 +1,158 @@ +/* + * Copyright (c) 2023 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 ACL_TESTS_DATASETS_REORDERLAYERDATASET +#define ACL_TESTS_DATASETS_REORDERLAYERDATASET + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "utils/TypePrinter.h" + +namespace arm_compute +{ +namespace test +{ +namespace datasets +{ +/** [ReorderLayer datasets] **/ +class ReorderLayerDataset +{ +public: + using type = std::tuple; + + struct iterator + { + iterator(std::vector::const_iterator in_it, + std::vector::const_iterator out_it, + std::vector::const_iterator _wf_in_it, + std::vector::const_iterator _wf_out_it) + : _in_it{ std::move(in_it) }, + _out_it{ std::move(out_it) }, + _wf_in_it{ std::move(_wf_in_it) }, + _wf_out_it{ std::move(_wf_out_it) } + { + } + + std::string description() const + { + std::stringstream description; + description << "In=" << *_in_it << ":"; + description << "Out=" << *_out_it << ":"; + description << "Wf_In=" << *_wf_in_it << ":"; + description << "Wf_Out=" << *_wf_out_it; + return description.str(); + } + + ReorderLayerDataset::type operator*() const + { + return std::make_tuple(*_in_it, *_out_it, *_wf_in_it, *_wf_out_it); + } + + iterator &operator++() + { + ++_in_it; + ++_out_it; + ++_wf_in_it; + ++_wf_out_it; + + return *this; + } + + private: + std::vector::const_iterator _in_it; + std::vector::const_iterator _out_it; + std::vector::const_iterator _wf_in_it; + std::vector::const_iterator _wf_out_it; + }; + + iterator begin() const + { + return iterator(_in_shapes.begin(), _out_shapes.begin(), _in_wfs.begin(), _out_wfs.begin()); + } + + int size() const + { + return std::min(_in_shapes.size(), std::min(_out_shapes.size(), std::min(_in_wfs.size(), _out_wfs.size()))); + } + + void add_config(TensorShape in, TensorShape out, WeightFormat in_wf, WeightFormat out_wf) + { + _in_shapes.emplace_back(std::move(in)); + _out_shapes.emplace_back(std::move(out)); + _in_wfs.emplace_back(std::move(in_wf)); + _out_wfs.emplace_back(std::move(out_wf)); + } + + // protected: + ReorderLayerDataset() = default; + ReorderLayerDataset(ReorderLayerDataset &&) = default; + + private: + std::vector _in_shapes{}; + std::vector _out_shapes{}; + std::vector _in_wfs{}; + std::vector _out_wfs{}; +}; + +/** [ReorderLayer datasets] **/ + +class ReorderLayerDatasetBlock4 final : public ReorderLayerDataset +{ + public: + ReorderLayerDatasetBlock4() + { + add_config(TensorShape(10U, 9U), TensorShape(10U, 12U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(16U, 16U), TensorShape(16U, 16U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(10U, 511U), TensorShape(10U, 512U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(234U, 301U), TensorShape(234U, 304U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(1024U, 1024U), TensorShape(1024U, 1024U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(10U, 9U, 1U, 1U), TensorShape(10U, 12U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(16U, 16U, 1U, 1U), TensorShape(16U, 16U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(10U, 511U, 1U, 1U), TensorShape(10U, 512U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(234U, 301U, 1U, 1U), TensorShape(234U, 304U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo4); + add_config(TensorShape(1024U, 1024U, 1U, 1U), TensorShape(1024U, 1024U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo4); + } +}; + +class ReorderLayerDatasetBlock8 final : public ReorderLayerDataset +{ + public: + ReorderLayerDatasetBlock8() + { + add_config(TensorShape(10U, 9U), TensorShape(10U, 16U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(16U, 16U), TensorShape(16U, 16U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(10U, 511U), TensorShape(10U, 512U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(234U, 301U), TensorShape(234U, 304U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(1024U, 1024U), TensorShape(1024U, 1024U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(10U, 9U, 1U, 1U), TensorShape(10U, 16U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(16U, 16U, 1U, 1U), TensorShape(16U, 16U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(10U, 511U, 1U, 1U), TensorShape(10U, 512U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(234U, 301U, 1U, 1U), TensorShape(234U, 304U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo8); + add_config(TensorShape(1024U, 1024U, 1U, 1U), TensorShape(1024U, 1024U, 1U, 1U), WeightFormat::OHWI, WeightFormat::OHWIo8); + } +}; + +} // namespace datasets +} // namespace test +} // namespace arm_compute +#endif /* ACL_TESTS_DATASETS_REORDERLAYERDATASET */ diff --git a/tests/validation/CMakeLists.txt b/tests/validation/CMakeLists.txt index eef062b5f9..448e96c4f9 100644 --- a/tests/validation/CMakeLists.txt +++ b/tests/validation/CMakeLists.txt @@ -138,6 +138,7 @@ if(ENABLE_NEON) NEON/FullyConnectedLayer.cpp NEON/ConvolutionLayer.cpp NEON/StridedSlice.cpp + NEON/ReorderLayer.cpp NEON/UNIT/DynamicTensor.cpp NEON/UNIT/TensorAllocator.cpp NEON/UNIT/MemoryManager.cpp diff --git a/tests/validation/NEON/ReorderLayer.cpp b/tests/validation/NEON/ReorderLayer.cpp new file mode 100644 index 0000000000..89d860fa4f --- /dev/null +++ b/tests/validation/NEON/ReorderLayer.cpp @@ -0,0 +1,66 @@ +/* + * Copyright (c) 2023 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/runtime/NEON/functions/NEReorderLayer.h" +#include "arm_compute/runtime/Tensor.h" +#include "tests/NEON/Accessor.h" +#include "tests/datasets/ReorderLayerDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/ReorderFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(NEON) +TEST_SUITE(ReorderLayer) + +template +using NEReorderLayerAlias = ReorderValidationFixture; + +TEST_SUITE(FP32) +#if defined(ARM_COMPUTE_ENABLE_SVE) +FIXTURE_DATA_TEST_CASE(RunBlock8, NEReorderLayerAlias, framework::DatasetMode::ALL, combine(datasets::ReorderLayerDatasetBlock8(), framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +#endif // ARM_COMPUTE_ENABLE_SVE +FIXTURE_DATA_TEST_CASE(RunBlock4, NEReorderLayerAlias, framework::DatasetMode::ALL, combine(datasets::ReorderLayerDatasetBlock4(), framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +TEST_SUITE_END() // FP32 + +TEST_SUITE_END() // ReorderLayer +TEST_SUITE_END() // NEON +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/ReorderFixture.h b/tests/validation/fixtures/ReorderFixture.h new file mode 100644 index 0000000000..bb3e3730df --- /dev/null +++ b/tests/validation/fixtures/ReorderFixture.h @@ -0,0 +1,109 @@ +/* + * Copyright (c) 2023 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 ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE +#define ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/Reorder.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/** [ReorderLayer fixture] **/ +template +class ReorderValidationFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape input_shape, TensorShape output_shape, WeightFormat input_wf, WeightFormat output_wf, DataType data_type) + { + _target = compute_target(input_shape, output_shape, input_wf, output_wf, data_type); + _reference = compute_reference(input_shape, output_shape, output_wf, data_type); + } + + protected: + template + void fill(U &&tensor) + { + library->fill_tensor_uniform(tensor, 0); + } + + TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, WeightFormat input_wf, WeightFormat output_wf, DataType data_type) + { + // Create tensors + TensorType src = create_tensor(input_shape, data_type); + TensorType dst = create_tensor(output_shape, data_type); + + // Create and configure function + FunctionType reorder; + + reorder.configure(&src, &dst, input_wf, output_wf); + + ARM_COMPUTE_ASSERT(src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + // Allocate tensors + src.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(src)); + + // Compute function + reorder.run(); + + return dst; + } + + SimpleTensor compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, WeightFormat output_wf, DataType data_type) + { + // Create reference + SimpleTensor src{ input_shape, data_type }; + + // Fill reference + fill(src); + + return reference::reorder_layer(src, output_shape, output_wf); + } + + TensorType _target{}; + SimpleTensor _reference{}; +}; +/** [ReorderLayer fixture] **/ +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE */ diff --git a/tests/validation/reference/Reorder.cpp b/tests/validation/reference/Reorder.cpp new file mode 100644 index 0000000000..8abb372596 --- /dev/null +++ b/tests/validation/reference/Reorder.cpp @@ -0,0 +1,156 @@ +/* + * Copyright (c) 2023 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 "Reorder.h" +#include "src/core/NEON/kernels/arm_gemm/utils.hpp" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ + +/* + * Generic transform. + * + * Assuming the untransposed case, this works by first reading + * consecutive values from the first input row. This same number of values + * are then read from the next rows. Now return to the first + * input row and repeat. + * + * Need to cope with the work requested in either dimension not actually + * being a multiple of the block sizes. + */ +template +struct Transform_ref +{ + template + static void Transform(TOut &out, const TIn in, const int stride, + const int y0, const int ymax, const int x0, const int xmax) + { + // NOTE: This code is disabled to avoid the call to get_vector_length(), so templated transforms will not be + // correct for SVE. This is not an issue as we have specializations for all SVE cases. + // For SVE cases we multiply the interleave factor by the vector length. + // const unsigned int IntBy = tIntBy * (vlt == VLType::SVE ? get_vector_length() / BlockBy : 1); + const unsigned int IntBy = tIntBy; + int out_index = 0; + + const int n_whole_y_blocks = (ymax - y0) / IntBy; + const int y_remainders = (ymax - y0) % IntBy; + const int n_y_blocks = n_whole_y_blocks + (y_remainders ? 1 : 0); + + const int n_whole_x_blocks = (xmax - x0) / BlockBy; + const int x_remainders = (xmax - x0) % BlockBy; + const int n_x_blocks = n_whole_x_blocks + (x_remainders ? 1 : 0); + + // "Y" loop: advance down the rows of the source IntBy rows at a time. + // Set up fill_rows to show the number rows to copy from, and blank_rows + // for the number of blank rows to add. + for(int y_block = 0; y_block < n_y_blocks; y_block++) + { + const int fill_rows = (y_block < n_whole_y_blocks) ? IntBy : y_remainders; + const int blank_rows = IntBy - fill_rows; + + const int y_base = y0 + (y_block * IntBy); + + // So now advance along this block of rows, BlockBy columns at a time. + for(int x_block = 0; x_block < n_x_blocks; x_block++) + { + const int fill_cols = (x_block < n_whole_x_blocks) ? BlockBy : x_remainders; + const int blank_cols = BlockBy - fill_cols; + + const int x_base = x0 + (x_block * BlockBy); + + for(int row = 0; row < fill_rows; row++) + { + for(int col = 0; col < fill_cols; col++) + { + // In-range copy. If it's transposed, we reverse the sense of rows and columns here. + if(Transposed) + { + out[out_index] = in[(x_base + col) * stride + y_base + row]; + out_index++; + } + else + { + out[out_index] = in[(y_base + row) * stride + x_base + col]; + out_index++; + } + } + // "col" tail - row is in range but column is out of range. + for(int col = 0; col < blank_cols; col++) + { + out[out_index] = 0; + out_index++; + } + } + // "row" tail - row is out of range so fill with zeros always. + const d_type zeroval = 0; + const int pads = blank_rows * (fill_cols + blank_cols); + + for(int i = 0; i < pads; i++) + { + out[out_index] = zeroval; + } + + out_index += pads; + } + } + } +}; + +template +SimpleTensor reorder_layer(const SimpleTensor &src, const TensorShape &output_shape, WeightFormat output_wf) +{ + SimpleTensor dst{ output_shape, src.data_type() }; + const int cols = src.shape()[0]; + const int rows = src.shape()[1]; + + switch(output_wf) + { + case WeightFormat::OHWIo4: + { + Transform_ref<4, 1, true, sizeof(float), sizeof(float), float, arm_gemm::VLType::None>::Transform &, SimpleTensor>(dst, src, rows, 0, rows, 0, cols); + break; + } + case WeightFormat::OHWIo8: + { + Transform_ref<8, 1, true, sizeof(float), sizeof(float), float, arm_gemm::VLType::None>::Transform &, SimpleTensor>(dst, src, rows, 0, rows, 0, cols); + break; + } + default: + break; + } + + return dst; +} + +template SimpleTensor reorder_layer(const SimpleTensor &src, const TensorShape &output_shape, WeightFormat output_wf); + +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/Reorder.h b/tests/validation/reference/Reorder.h new file mode 100644 index 0000000000..94ee5078f8 --- /dev/null +++ b/tests/validation/reference/Reorder.h @@ -0,0 +1,44 @@ +/* + * Copyright (c) 2023 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 ACL_TESTS_VALIDATION_REFERENCE_REORDER +#define ACL_TESTS_VALIDATION_REFERENCE_REORDER + +#include "tests/SimpleTensor.h" +#include "tests/Types.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template +SimpleTensor reorder_layer(const SimpleTensor &src, const TensorShape &output_shape, WeightFormat output_wf); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ACL_TESTS_VALIDATION_REFERENCE_REORDER */ -- cgit v1.2.1