From afd38f0c617d6f89b2b4532c6c44f116617e2b6f Mon Sep 17 00:00:00 2001 From: Felix Thomasmathibalan Date: Wed, 27 Sep 2023 17:46:17 +0100 Subject: Apply clang-format on repository Code is formatted as per a revised clang format configuration file(not part of this delivery). Version 14.0.6 is used. Exclusion List: - files with .cl extension - files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...) And the following directories - compute_kernel_writer/validation/ - tests/ - include/ - src/core/NEON/kernels/convolution/ - src/core/NEON/kernels/arm_gemm/ - src/core/NEON/kernels/arm_conv/ - data/ There will be a follow up for formatting of .cl files and the files under tests/ and compute_kernel_writer/validation/. Signed-off-by: Felix Thomasmathibalan Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir --- .../NEON/kernels/NENormalizationLayerKernel.cpp | 144 ++++++++++++--------- 1 file changed, 82 insertions(+), 62 deletions(-) (limited to 'src/core/NEON/kernels/NENormalizationLayerKernel.cpp') diff --git a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp index 49a045382d..2c61bda147 100644 --- a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp @@ -29,19 +29,23 @@ #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" + #include "src/core/CPP/Validate.h" -#include "src/core/NEON/NEFixedPoint.h" -#include "src/core/NEON/NEMath.h" -#include "src/core/NEON/wrapper/wrapper.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/NormalizationHelpers.h" #include "src/core/helpers/WindowHelpers.h" +#include "src/core/NEON/NEFixedPoint.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" namespace arm_compute { namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *input_squared, const ITensorInfo *output, const NormalizationLayerInfo &norm_info) +Status validate_arguments(const ITensorInfo *input, + const ITensorInfo *input_squared, + const ITensorInfo *output, + const NormalizationLayerInfo &norm_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_squared, output); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); @@ -52,7 +56,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *input_squ ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd"); // Checks performed when output is configured - if(output->total_size() != 0) + if (output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); @@ -69,7 +73,10 @@ NENormalizationLayerKernel::NENormalizationLayerKernel() { } -void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor *input_squared, ITensor *output, NormalizationLayerInfo norm_info) +void NENormalizationLayerKernel::configure(const ITensor *input, + const ITensor *input_squared, + ITensor *output, + NormalizationLayerInfo norm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_squared, output); // Output tensor auto initialization if not yet initialized @@ -85,15 +92,15 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * _output = output; _norm_info = norm_info; - switch(_input->info()->data_type()) + switch (_input->info()->data_type()) { case DataType::F32: { - switch(norm_idx) + switch (norm_idx) { case 0: { - if(norm_info.type() == NormType::IN_MAP_2D) + if (norm_info.type() == NormType::IN_MAP_2D) { _func = &NENormalizationLayerKernel::normalize_float; } @@ -104,7 +111,7 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * break; } case 1: - if(norm_info.type() == NormType::IN_MAP_2D) + if (norm_info.type() == NormType::IN_MAP_2D) { _func = &NENormalizationLayerKernel::normalize_float; } @@ -124,11 +131,11 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC case DataType::F16: { - switch(norm_idx) + switch (norm_idx) { case 0: { - if(norm_info.type() == NormType::IN_MAP_2D) + if (norm_info.type() == NormType::IN_MAP_2D) { _func = &NENormalizationLayerKernel::normalize_float; } @@ -139,7 +146,7 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * break; } case 1: - if(norm_info.type() == NormType::IN_MAP_2D) + if (norm_info.type() == NormType::IN_MAP_2D) { _func = &NENormalizationLayerKernel::normalize_float; } @@ -196,8 +203,9 @@ void NENormalizationLayerKernel::normalize_float(const Window &window) const auto beta_vec = wrapper::vdup_n(static_cast(_norm_info.beta()), ExactTagType{}); const auto kappa_vec = wrapper::vdup_n(static_cast(_norm_info.kappa()), ExactTagType{}); - auto sequential_normalization = [&](const int x, const Coordinates & id, const int current_row, const int first_row, const int last_row, const T * input_ptr, const uint8_t *input_squared_start_ptr, - T * output_ptr) + auto sequential_normalization = [&](const int x, const Coordinates &id, const int current_row, const int first_row, + const int last_row, const T *input_ptr, const uint8_t *input_squared_start_ptr, + T *output_ptr) { const int current_slice = dim == 0 ? x : id[dim]; const int first_slice = std::max(current_slice - radius, 0); @@ -206,75 +214,87 @@ void NENormalizationLayerKernel::normalize_float(const Window &window) const uint8_t *const input_squared_x_ptr = input_squared_start_ptr + x * input_squared_stride_x; // Accumulate 2D In-Map values auto accu = static_cast(0.f); - for(int j = first_row; j <= last_row; ++j) + for (int j = first_row; j <= last_row; ++j) { // Compute row displacement const uint8_t *const input_squared_ptr = input_squared_x_ptr + (j - current_row) * input_squared_stride_row; - for(int i = first_slice; i <= last_slice; ++i) + for (int i = first_slice; i <= last_slice; ++i) { - accu += *reinterpret_cast(input_squared_ptr + (i - current_slice) * input_squared_stride_slice); + accu += + *reinterpret_cast(input_squared_ptr + (i - current_slice) * input_squared_stride_slice); } } // Normalize - const auto normalized = std::pow(accu * static_cast(_norm_info.scale_coeff()) + static_cast(_norm_info.kappa()), _norm_info.beta()); + const auto normalized = std::pow( + accu * static_cast(_norm_info.scale_coeff()) + static_cast(_norm_info.kappa()), _norm_info.beta()); const auto normalized_pixel = (*(input_ptr + x)) / normalized; *(output_ptr + x) = normalized_pixel; }; - execute_window_loop(win, [&](const Coordinates & id) - { - const auto input_ptr = reinterpret_cast(input.ptr()); - auto output_ptr = reinterpret_cast(output.ptr()); + execute_window_loop( + win, + [&](const Coordinates &id) + { + const auto input_ptr = reinterpret_cast(input.ptr()); + auto output_ptr = reinterpret_cast(output.ptr()); - // Get range to normalize - const int current_row = do_2D_norm ? id[dim_y] : 0; - const int first_row = do_2D_norm ? std::max(current_row - radius, 0) : 0; - const int last_row = do_2D_norm ? std::min(current_row + radius, max_bottom) : 0; + // Get range to normalize + const int current_row = do_2D_norm ? id[dim_y] : 0; + const int first_row = do_2D_norm ? std::max(current_row - radius, 0) : 0; + const int last_row = do_2D_norm ? std::min(current_row + radius, max_bottom) : 0; - int x = window_start_x; - // Compute serially starting elements for the case x dimension is width - for(; x < radius && x < window_end_x && dim == 0; ++x) - { - sequential_normalization(x, id, current_row, first_row, last_row, input_ptr, input_squared.ptr(), output_ptr); - } + int x = window_start_x; + // Compute serially starting elements for the case x dimension is width + for (; x < radius && x < window_end_x && dim == 0; ++x) + { + sequential_normalization(x, id, current_row, first_row, last_row, input_ptr, input_squared.ptr(), + output_ptr); + } - // Compute vectorized - for(; x <= window_end_x - window_step_x - radius; x += window_step_x) - { - const int current_slice = dim == 0 ? x : id[dim]; - const int first_slice = std::max(current_slice - radius, 0); - const int last_slice = std::min(current_slice + radius, max_right); - - const uint8_t *const input_squared_x_ptr = input_squared.ptr() + x * input_squared_stride_x; - // Accumulate 2D In-Map values - auto accu = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); - for(int j = first_row; j <= last_row; ++j) + // Compute vectorized + for (; x <= window_end_x - window_step_x - radius; x += window_step_x) { - // Compute row displacement - const uint8_t *const input_squared_ptr = input_squared_x_ptr + (j - current_row) * input_squared_stride_row; - for(int i = first_slice; i <= last_slice; ++i) + const int current_slice = dim == 0 ? x : id[dim]; + const int first_slice = std::max(current_slice - radius, 0); + const int last_slice = std::min(current_slice + radius, max_right); + + const uint8_t *const input_squared_x_ptr = input_squared.ptr() + x * input_squared_stride_x; + // Accumulate 2D In-Map values + auto accu = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); + for (int j = first_row; j <= last_row; ++j) { - accu = wrapper::vadd(accu, wrapper::vloadq(reinterpret_cast(input_squared_ptr + (i - current_slice) * input_squared_stride_slice))); + // Compute row displacement + const uint8_t *const input_squared_ptr = + input_squared_x_ptr + (j - current_row) * input_squared_stride_row; + for (int i = first_slice; i <= last_slice; ++i) + { + accu = wrapper::vadd( + accu, wrapper::vloadq(reinterpret_cast( + input_squared_ptr + (i - current_slice) * input_squared_stride_slice))); + } } - } - // Normalize - const auto normalized = wrapper::vpow(wrapper::vmla(kappa_vec, coeff_vec, accu), beta_vec); - const auto normalized_pixel = wrapper::vmul(wrapper::vloadq(input_ptr + x), wrapper::vinv(normalized)); - wrapper::vstore(reinterpret_cast(output_ptr + x), normalized_pixel); - } + // Normalize + const auto normalized = wrapper::vpow(wrapper::vmla(kappa_vec, coeff_vec, accu), beta_vec); + const auto normalized_pixel = wrapper::vmul(wrapper::vloadq(input_ptr + x), wrapper::vinv(normalized)); + wrapper::vstore(reinterpret_cast(output_ptr + x), normalized_pixel); + } - // Compute left-over elements - for(; x < window_end_x; ++x) - { - sequential_normalization(x, id, current_row, first_row, last_row, input_ptr, input_squared.ptr(), output_ptr); - } - }, - input, input_squared, output); + // Compute left-over elements + for (; x < window_end_x; ++x) + { + sequential_normalization(x, id, current_row, first_row, last_row, input_ptr, input_squared.ptr(), + output_ptr); + } + }, + input, input_squared, output); } -Status NENormalizationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *input_squared, const ITensorInfo *output, const NormalizationLayerInfo norm_info) +Status NENormalizationLayerKernel::validate(const ITensorInfo *input, + const ITensorInfo *input_squared, + const ITensorInfo *output, + const NormalizationLayerInfo norm_info) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, input_squared, output, norm_info)); -- cgit v1.2.1