From f2c714b244d0cdd8c38816bcc6b9e7eb3be7ee66 Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Mon, 15 Jun 2020 16:50:35 +0100 Subject: COMPMID-3153: Remove padding from NENormalizationLayerKernel Change-Id: Ib84308ea18bfa31ffbc3269a1f005d7d302139f7 Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3350 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Reviewed-by: Michele Di Giorgio --- .../core/NEON/kernels/NENormalizationLayerKernel.h | 4 +- .../runtime/NEON/functions/NENormalizationLayer.h | 3 +- .../NEON/kernels/NENormalizationLayerKernel.cpp | 171 +++++++++++---------- .../NEON/functions/NENormalizationLayer.cpp | 11 +- tests/validation/NEON/NormalizationLayer.cpp | 25 ++- 5 files changed, 118 insertions(+), 96 deletions(-) diff --git a/arm_compute/core/NEON/kernels/NENormalizationLayerKernel.h b/arm_compute/core/NEON/kernels/NENormalizationLayerKernel.h index 4727164d00..ad898d960c 100644 --- a/arm_compute/core/NEON/kernels/NENormalizationLayerKernel.h +++ b/arm_compute/core/NEON/kernels/NENormalizationLayerKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -76,7 +76,6 @@ public: // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; - BorderSize border_size() const override; private: /** Function to perform normalization depending on the given template @@ -104,7 +103,6 @@ private: const ITensor *_input_squared; ITensor *_output; NormalizationLayerInfo _norm_info; - BorderSize _border_size; }; } // namespace arm_compute #endif /*ARM_COMPUTE_NENORMALIZATIONLAYERKERNEL_H */ diff --git a/arm_compute/runtime/NEON/functions/NENormalizationLayer.h b/arm_compute/runtime/NEON/functions/NENormalizationLayer.h index af34147bfe..8683e44d3c 100644 --- a/arm_compute/runtime/NEON/functions/NENormalizationLayer.h +++ b/arm_compute/runtime/NEON/functions/NENormalizationLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -78,7 +78,6 @@ private: MemoryGroup _memory_group; /**< Function memory group */ NENormalizationLayerKernel _norm_kernel; /**< Normalization layer kernel */ NEPixelWiseMultiplicationKernel _multiply_kernel; /**< Pixel multiplication kernel */ - NEFillBorderKernel _border_handler; /**< Kernel to handle borders */ Tensor _input_squared; /**< The intermediate buffer which stores results of squaring input */ }; } diff --git a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp index e5f6e4f41a..dd98d74260 100644 --- a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,8 +35,8 @@ #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" -using namespace arm_compute; - +namespace arm_compute +{ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *input_squared, const ITensorInfo *output, const NormalizationLayerInfo &norm_info) @@ -60,58 +60,13 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *input_squ return Status{}; } -std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *input_squared, ITensorInfo *output, const NormalizationLayerInfo &norm_info) -{ - // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, *input->clone()); - - const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); - - const unsigned int norm_idx = get_normalization_dimension_index(input->data_layout(), norm_info); - const bool is_norm_accross_width = norm_idx == 0; - - const unsigned int border_width = is_norm_accross_width ? num_elems_processed_per_iteration - 1 : 0; - const BorderSize border_size = BorderSize(0, border_width); - - // Configure window - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); - bool window_changed = false; - - if(is_norm_accross_width) - { - AccessWindowStatic input_access(input, -border_size.left, 0, input->dimension(0) + border_size.right, 0); - AccessWindowStatic input_squared_access(input_squared, -border_size.left, 0, input->dimension(0) + border_size.right, 0); - window_changed = window_changed || update_window_and_padding(win, input_access, input_squared_access); - } - else - { - AccessWindowHorizontal input_access(input, -border_size.left, num_elems_processed_per_iteration); - AccessWindowHorizontal input_squared_access(input_squared, -border_size.left, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win, input_access, input_squared_access); - } - - if(output->total_size() != 0) - { - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win, output_access); - output_access.set_valid_region(win, input->valid_region()); - } - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} } // namespace NENormalizationLayerKernel::NENormalizationLayerKernel() - : _func(nullptr), _input(nullptr), _input_squared(nullptr), _output(nullptr), _norm_info(NormType::IN_MAP_1D), _border_size() + : _func(nullptr), _input(nullptr), _input_squared(nullptr), _output(nullptr), _norm_info(NormType::IN_MAP_1D) { } -BorderSize NENormalizationLayerKernel::border_size() const -{ - return _border_size; -} - void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor *input_squared, ITensor *output, NormalizationLayerInfo norm_info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_squared, output); @@ -121,17 +76,12 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * // Perform validation step ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), input_squared->info(), output->info(), norm_info)); - const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); - - const unsigned int norm_idx = get_normalization_dimension_index(input->info()->data_layout(), norm_info); - const bool is_norm_accross_width = norm_idx == 0; - const unsigned int border_width = is_norm_accross_width ? num_elems_processed_per_iteration - 1 : 0; + const unsigned int norm_idx = get_normalization_dimension_index(input->info()->data_layout(), norm_info); _input = input; _input_squared = input_squared; _output = output; _norm_info = norm_info; - _border_size = BorderSize(0, border_width); switch(_input->info()->data_type()) { @@ -210,9 +160,11 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor * } // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), input_squared->info(), output->info(), norm_info); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - INEKernel::configure(win_config.second); + Window win = calculate_max_window(*input->info(), Steps()); + Coordinates coord; + coord.set_num_dimensions(output->info()->num_dimensions()); + output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); + INEKernel::configure(win); } template @@ -221,15 +173,23 @@ void NENormalizationLayerKernel::normalize_float(const Window &window) /** NEON vector tag type. */ using ExactTagType = typename wrapper::traits::neon_vector::tag_type; - Iterator input(_input, window); - Iterator input_squared(_input_squared, window); - Iterator output(_output, window); + Window win(window); + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const int window_step_x = S; + + Iterator input(_input, win); + Iterator input_squared(_input_squared, win); + Iterator output(_output, win); + + const int dim_y = _input->info()->data_layout() == DataLayout::NCHW ? 1 : 2; + const int radius = _norm_info.norm_size() / 2; + const int input_squared_stride_x = _input_squared->info()->strides_in_bytes()[0]; + const int input_squared_stride_slice = _input_squared->info()->strides_in_bytes()[dim]; + const int input_squared_stride_row = _input_squared->info()->strides_in_bytes()[dim_y]; - const int dim_y = _input->info()->data_layout() == DataLayout::NCHW ? 1 : 2; - const int radius = _norm_info.norm_size() / 2; - const int input_squared_stride = _input_squared->info()->strides_in_bytes()[dim]; - // We account padding across X only and we iterate over rows - const int min_left = (dim == 2) ? 0 : -static_cast(border_size().left); const int max_right = _input->info()->dimension(dim) - 1; const int max_bottom = _input->info()->dimension(dim_y) - 1; @@ -237,33 +197,80 @@ 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{}); - execute_window_loop(window, [&](const Coordinates & id) + 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) { - // Get range to normalize - const int current_row = do_2D_norm ? id[dim_y] : 0; - const int current_slice = id[dim]; - 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; - const int first_slice = std::max(current_slice - radius, min_left); + 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_start_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++) + auto accu = static_cast(0.f); + for(int j = first_row; j <= last_row; ++j) { // Compute row displacement - const int row = (j - current_row) * _input_squared->info()->strides_in_bytes()[dim_y]; - const uint8_t *const input_squared_ptr = input_squared.ptr() + row - (current_slice * input_squared_stride); + 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 * input_squared_stride))); + accu += *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(reinterpret_cast(input.ptr())), wrapper::vinv(normalized)); - wrapper::vstore(reinterpret_cast(output.ptr()), normalized_pixel); + 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()); + + // 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); + } + + // 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 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); + } + + // 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); } @@ -271,7 +278,6 @@ void NENormalizationLayerKernel::normalize_float(const Window &window) 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)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), input_squared->clone().get(), output->clone().get(), norm_info).first); return Status{}; } @@ -286,3 +292,4 @@ void NENormalizationLayerKernel::run(const Window &window, const ThreadInfo &inf // Run function (this->*_func)(window); } +} // namespace arm_compute \ No newline at end of file diff --git a/src/runtime/NEON/functions/NENormalizationLayer.cpp b/src/runtime/NEON/functions/NENormalizationLayer.cpp index d52e92828e..f3a3ac6322 100644 --- a/src/runtime/NEON/functions/NENormalizationLayer.cpp +++ b/src/runtime/NEON/functions/NENormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -30,10 +30,10 @@ #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" -using namespace arm_compute; - +namespace arm_compute +{ NENormalizationLayer::NENormalizationLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _norm_kernel(), _multiply_kernel(), _border_handler(), _input_squared() + : _memory_group(std::move(memory_manager)), _norm_kernel(), _multiply_kernel(), _input_squared() { } @@ -50,7 +50,6 @@ void NENormalizationLayer::configure(const ITensor *input, ITensor *output, cons // Configure kernels _norm_kernel.configure(input, &_input_squared, output, norm_info); _multiply_kernel.configure(input, input, &_input_squared, 1.0f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); - _border_handler.configure(&_input_squared, _norm_kernel.border_size(), BorderMode::CONSTANT, PixelValue(0.0f)); // Allocate the tensor once the configure methods have been called _input_squared.allocator()->allocate(); @@ -72,6 +71,6 @@ void NENormalizationLayer::run() MemoryGroupResourceScope scope_mg(_memory_group); NEScheduler::get().schedule(&_multiply_kernel, Window::DimY); - NEScheduler::get().schedule(&_border_handler, Window::DimY); NEScheduler::get().schedule(&_norm_kernel, Window::DimY); } +} \ No newline at end of file diff --git a/tests/validation/NEON/NormalizationLayer.cpp b/tests/validation/NEON/NormalizationLayer.cpp index 20dcafb719..f72d156e9f 100644 --- a/tests/validation/NEON/NormalizationLayer.cpp +++ b/tests/validation/NEON/NormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -61,8 +61,6 @@ const auto NormalizationDatasetFP32 = combine(combine(combine(datasets::Normaliz TEST_SUITE(NEON) TEST_SUITE(NormalizationLayer) -//TODO(COMPMID-415): Missing configuration? - // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( @@ -96,6 +94,27 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( // clang-format on // *INDENT-ON* +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)), + shape, data_type) +{ + NormalizationLayerInfo info(NormType::IN_MAP_1D, 3U, 5.0f, 2.0f, 1.f, false); + + // Create tensors + Tensor src = create_tensor(shape, data_type); + Tensor dst = create_tensor(shape, data_type); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Create and configure function + NENormalizationLayer norm; + norm.configure(&src, &dst, info); + + // To enable check on src as soon as NEPixelWiseMultiplicationKernel stops using padding anymore: COMPMID-3477 + //validate(src.info()->padding(), PaddingSize(0,0,0,0)); + validate(dst.info()->padding(), PaddingSize()); +} + template using NENormalizationLayerFixture = NormalizationValidationFixture; -- cgit v1.2.1