From c9564cb3850b6675cef663d7cc0722567b55cc25 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Fri, 13 Sep 2019 10:20:25 +0100 Subject: COMPMID-2257: Implement NEGenerateProposals. Change-Id: I8d751f8b09f842a214c305a4530a71d82f16db8f Signed-off-by: Pablo Tello Reviewed-on: https://review.mlplatform.org/c/1943 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Michele Di Giorgio --- src/core/NEON/kernels/NECopyKernel.cpp | 116 ++++++++++++++---- .../kernels/NEGenerateProposalsLayerKernel.cpp | 131 +++++++++++++++++++++ 2 files changed, 227 insertions(+), 20 deletions(-) create mode 100644 src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp (limited to 'src/core/NEON') diff --git a/src/core/NEON/kernels/NECopyKernel.cpp b/src/core/NEON/kernels/NECopyKernel.cpp index 4722c05507..83f3dded4f 100644 --- a/src/core/NEON/kernels/NECopyKernel.cpp +++ b/src/core/NEON/kernels/NECopyKernel.cpp @@ -29,28 +29,88 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" -using namespace arm_compute; +namespace arm_compute +{ +namespace +{ +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList()) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4); + + // Validate output if initialized + if(output->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape()); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + } + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) +{ + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output, *input); + return std::make_pair(Status{}, calculate_max_window(*output)); +} + +std::pair validate_and_configure_window_with_padding(ITensorInfo *input, ITensorInfo *output, const PaddingList &padding) +{ + const TensorShape input_shape = input->tensor_shape(); + const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input_shape, padding); + auto_init_if_empty(*output, input->clone()->set_tensor_shape(padded_shape)); + // Configure window + const Window win = calculate_max_window(*output, output->dimension(0)); + return std::make_pair(Status{}, win); +} + +} // namespace NECopyKernel::NECopyKernel() - : _input(nullptr), _output(nullptr) + : _input(nullptr), _output(nullptr), _padding() { } -void NECopyKernel::configure(const ITensor *input, ITensor *output) +void NECopyKernel::configure(const ITensor *input, ITensor *output, const PaddingList &padding) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding)); + + _input = input; + _output = output; + _padding = padding; - _input = input; - _output = output; + std::pair win_config; - INEKernel::configure(calculate_max_window(*output->info())); + if(padding.empty()) + { + win_config = validate_and_configure_window(input->info(), output->info()); + } + else + { + win_config = validate_and_configure_window_with_padding(input->info(), output->info(), padding); + } + + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); } -Status NECopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output) +Status NECopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding)); + + if(padding.empty()) + { + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_with_padding(input->clone().get(), output->clone().get(), padding).first); + } + return Status{}; } @@ -60,22 +120,38 @@ void NECopyKernel::run(const Window &window, const ThreadInfo &info) ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - Window output_window{ window }; - output_window.set(Window::DimX, Window::Dimension(output_window.x().start(), output_window.x().end(), _input->info()->dimension(0))); - - Window out_slice = output_window.first_slice_window_1D(); + if(_padding.empty()) + { + Window output_window{ window }; + output_window.set(Window::DimX, Window::Dimension(output_window.x().start(), output_window.x().end(), _input->info()->dimension(0))); + Window out_slice = output_window.first_slice_window_1D(); + do + { + Iterator input_it(_input, out_slice); + Iterator output_it(_output, out_slice); - do + execute_window_loop(out_slice, [&](const Coordinates &) + { + memcpy(output_it.ptr(), input_it.ptr(), _output->info()->dimension(0) * _output->info()->element_size()); + }, + input_it, output_it); + } + while(output_window.slide_window_slice_1D(out_slice)); + } + else { - Iterator input_it(_input, out_slice); - Iterator output_it(_output, out_slice); + Window input_window{ window }; + input_window.set(Window::DimX, Window::Dimension(0, window.x().end() - _padding[0].first, _input->info()->dimension(0))); - execute_window_loop(out_slice, [&](const Coordinates &) + Iterator input_it(_input, input_window); + Iterator output_it(_output, window); + const size_t row_size_in_bytes = _input->info()->dimension(0) * _input->info()->element_size(); + execute_window_loop(window, [&](const Coordinates &) { - memcpy(output_it.ptr(), input_it.ptr(), _output->info()->dimension(0) * _output->info()->element_size()); + auto dst_ptr = output_it.ptr() + _padding[0].first * _output->info()->element_size(); + std::memcpy(dst_ptr, input_it.ptr(), row_size_in_bytes); }, input_it, output_it); - } - while(output_window.slide_window_slice_1D(out_slice)); } +} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp b/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp new file mode 100644 index 0000000000..4a585b70fd --- /dev/null +++ b/src/core/NEON/kernels/NEGenerateProposalsLayerKernel.cpp @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2019 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/NEON/kernels/NEGenerateProposalsLayerKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/CPP/Validate.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Window.h" + +namespace arm_compute +{ +namespace +{ +Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(anchors); + ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi()); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2); + if(all_anchors->total_size() > 0) + { + const size_t feature_height = info.feat_height(); + const size_t feature_width = info.feat_width(); + const size_t num_anchors = anchors->dimension(1); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(all_anchors, anchors); + ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2); + ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi()); + ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors); + } + return Status{}; +} + +} // namespace + +NEComputeAllAnchorsKernel::NEComputeAllAnchorsKernel() + : _anchors(nullptr), _all_anchors(nullptr), _anchors_info(0.f, 0.f, 0.f) +{ +} + +void NEComputeAllAnchorsKernel::configure(const ITensor *anchors, ITensor *all_anchors, const ComputeAnchorsInfo &info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(anchors, all_anchors); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(anchors->info(), all_anchors->info(), info)); + + // Metadata + const size_t num_anchors = anchors->info()->dimension(1); + const DataType data_type = anchors->info()->data_type(); + const float width = info.feat_width(); + const float height = info.feat_height(); + + // Initialize the output if empty + const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors); + auto_init_if_empty(*all_anchors->info(), output_shape, 1, data_type); + + // Set instance variables + _anchors = anchors; + _all_anchors = all_anchors; + _anchors_info = info; + + Window win = calculate_max_window(*all_anchors->info(), Steps(info.values_per_roi())); + + INEKernel::configure(win); +} + +Status NEComputeAllAnchorsKernel::validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(anchors, all_anchors, info)); + return Status{}; +} + +void NEComputeAllAnchorsKernel::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); + + Iterator all_anchors_it(_all_anchors, window); + Iterator anchors_it(_all_anchors, window); + + const size_t num_anchors = _anchors->info()->dimension(1); + const float stride = 1.f / _anchors_info.spatial_scale(); + const size_t feat_width = _anchors_info.feat_width(); + + execute_window_loop(window, [&](const Coordinates & id) + { + const size_t anchor_offset = id.y() % num_anchors; + + const auto out_anchor_ptr = reinterpret_cast(all_anchors_it.ptr()); + const auto anchor_ptr = reinterpret_cast(_anchors->ptr_to_element(Coordinates(0, anchor_offset))); + + *out_anchor_ptr = *anchor_ptr; + *(1 + out_anchor_ptr) = *(1 + anchor_ptr); + *(2 + out_anchor_ptr) = *(2 + anchor_ptr); + *(3 + out_anchor_ptr) = *(3 + anchor_ptr); + + const size_t shift_idy = id.y() / num_anchors; + const float shiftx = (shift_idy % feat_width) * stride; + const float shifty = (shift_idy / feat_width) * stride; + + *out_anchor_ptr += shiftx; + *(out_anchor_ptr + 1) += shifty; + *(out_anchor_ptr + 2) += shiftx; + *(out_anchor_ptr + 3) += shifty; + }, + all_anchors_it); +} +} // namespace arm_compute -- cgit v1.2.1