diff options
author | Luca Foschiani <luca.foschiani@arm.com> | 2020-02-17 17:02:49 +0000 |
---|---|---|
committer | Luca Foschiani <luca.foschiani@arm.com> | 2020-04-07 09:04:19 +0000 |
commit | fedefc3a8d76b9dea5945414324427ef5a01835d (patch) | |
tree | b2a2f6ab45d8a16ab26b5a99c832a18e207899aa /src | |
parent | 0d008f77b0085619c446d0ab5dc1228a80776706 (diff) | |
download | ComputeLibrary-fedefc3a8d76b9dea5945414324427ef5a01835d.tar.gz |
COMPMID-2765 Add support for QASYMM8_SIGNED in NEDeconvolutionLayer
Signed-off-by: Luca Foschiani <luca.foschiani@arm.com>
Change-Id: I8295fadee15311a9ab846aa24c031b82c0b799eb
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2952
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Diffstat (limited to 'src')
-rw-r--r-- | src/core/CPP/kernels/CPPFlipWeightsKernel.cpp | 113 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEReverseKernel.cpp | 69 | ||||
-rw-r--r-- | src/runtime/NEON/functions/NEDeconvolutionLayer.cpp | 24 |
3 files changed, 38 insertions, 168 deletions
diff --git a/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp b/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp deleted file mode 100644 index 2d4c0ce5c8..0000000000 --- a/src/core/CPP/kernels/CPPFlipWeightsKernel.cpp +++ /dev/null @@ -1,113 +0,0 @@ -/* - * Copyright (c) 2018 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/CPP/kernels/CPPFlipWeightsKernel.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" - -#include <cstddef> -#include <cstdint> - -using namespace arm_compute; - -CPPFlipWeightsKernel::CPPFlipWeightsKernel() - : _input(nullptr), _output(nullptr), _func(nullptr) -{ -} - -template <typename T> -void CPPFlipWeightsKernel::flip_weights(const Window &window_input) -{ - // Create iterators - Iterator in(_input, window_input); - - const DataLayout data_layout = _input->info()->data_layout(); - const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - const int kernel_width = _input->info()->dimension(idx_w); - const int kernel_height = _input->info()->dimension(idx_h); - - execute_window_loop(window_input, [&](const Coordinates & id) - { - const unsigned int x = kernel_width - id[idx_w] - 1; - const unsigned int y = kernel_height - id[idx_h] - 1; - Coordinates output_coord(id); - output_coord.set(idx_w, x); - output_coord.set(idx_h, y); - *(reinterpret_cast<T *>(_output->ptr_to_element(output_coord))) = *(reinterpret_cast<const T *>(in.ptr())); - }, - in); -} - -void CPPFlipWeightsKernel::configure(const ITensor *input, ITensor *output) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - - _input = input; - _output = output; - - // Configure kernel window - Window win = calculate_max_window(*input->info(), Steps()); - - // The CPPFlipWeightsKernel doesn't need padding so update_window_and_padding() can be skipped - Coordinates coord; - coord.set_num_dimensions(output->info()->num_dimensions()); - output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); - - ICPPKernel::configure(win); - - switch(input->info()->data_type()) - { - case DataType::F32: - _func = &CPPFlipWeightsKernel::flip_weights<float>; - break; - case DataType::F16: - _func = &CPPFlipWeightsKernel::flip_weights<half>; - break; - case DataType::QASYMM8: - _func = &CPPFlipWeightsKernel::flip_weights<uint8_t>; - break; - default: - ARM_COMPUTE_ERROR("Not supported"); - } -} - -void CPPFlipWeightsKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - - (this->*_func)(window); -} diff --git a/src/core/NEON/kernels/NEReverseKernel.cpp b/src/core/NEON/kernels/NEReverseKernel.cpp index 2f584164dc..5a8c446ddd 100644 --- a/src/core/NEON/kernels/NEReverseKernel.cpp +++ b/src/core/NEON/kernels/NEReverseKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 ARM Limited. + * Copyright (c) 2018-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -106,33 +106,20 @@ void run_reverse(const Window &window, const ITensor *input, const ITensor *axis } // Check if we need a left-over loop for the y dimension - const int window_step_x = 16 / input->info()->element_size(); - const int window_start_x = window.x().start(); - const int window_end_x = std::min(window.x().end(), static_cast<int>(input->info()->dimension(0))); - const int window_end_x_multiple_of = ((window_end_x - window_start_x) / window_step_x) * window_step_x; - bool left_over_loop_x = (((window_end_x - window_start_x) % window_step_x) != 0); + const int window_step_x = 16 / input->info()->element_size(); + const int window_start_x = window.x().start(); + const int window_end_x = window.x().end(); - Window slice = window.first_slice_window_4D(); + Window win(window); + win.set(Window::DimX, Window::Dimension(0, 1, 1)); - if(left_over_loop_x) + Iterator input_it(input, win); + execute_window_loop(win, [&](const Coordinates & id) { - // Check if window_end_y_multiple_of is greater than window_start_y - if(window_end_x_multiple_of > window_start_x) + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) { - slice.set(Window::DimX, Window::Dimension(window_start_x, window_end_x_multiple_of, window_step_x)); - } - else - { - slice.set(Window::DimX, Window::Dimension(0, 0, 1)); - } - } - - do - { - Iterator input_it(input, slice); - execute_window_loop(slice, [&](const Coordinates & id) - { - auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr())); + auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr()) + x); // Reverse 0 axis if(axis_bit & 0x1) @@ -141,39 +128,29 @@ void run_reverse(const Window &window, const ITensor *input, const ITensor *axis in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in)); } - const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - id.x() - window_step_x : id.x(); + const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - window_step_x : x; const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y(); const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z(); const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3]; auto out_ptr = reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))); wrapper::vstore(out_ptr, in); - }, - input_it); + } - if(left_over_loop_x) + // Compute left-over elements + for(; x < window_end_x; ++x) { - slice.set(Window::DimX, Window::Dimension(window_end_x_multiple_of, window_end_x, 1)); + const auto in = *(reinterpret_cast<T *>(input_it.ptr()) + x); - Iterator input_it(input, slice); - - // Compute left-over elements along the y dimension (1x1) - execute_window_loop(slice, [&](const Coordinates & id) - { - const auto in = *reinterpret_cast<T *>(input_it.ptr()); - - const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - id.x() - 1 : id.x(); - const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y(); - const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z(); - const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3]; + const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - 1 : x; + const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y(); + const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z(); + const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3]; - *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) = in; - }, - input_it); + *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) = in; } - - } - while(window.slide_window_slice_4D(slice)); + }, + input_it); } void NEReverseKernel::run(const Window &window, const ThreadInfo &info) diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp index 0411b41220..06885d59e5 100644 --- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,7 @@ NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memor _permuted_input(), _permuted_weights(), _permuted_output(), + _flip_axis(), _is_nchw(false), _original_weights(nullptr), _input(nullptr), @@ -57,7 +58,7 @@ NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memor Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8, DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, input); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(weights, input); const unsigned int width_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH); @@ -122,6 +123,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con _info = info; _is_prepared = false; _is_nchw = data_layout == DataLayout::NCHW; + _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32)); const unsigned int pad_left = info.pad_left(); const unsigned int pad_right = info.pad_right(); @@ -139,6 +141,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con // Output auto initialization if not yet initialized auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info()); + _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32)); _memory_group.manage(&_scaled_output); if(!_is_nchw) @@ -185,7 +188,7 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con _weights_flipped.allocator()->init(*_permuted_weights.info()->clone()); _weights_flipped.info()->set_quantization_info(weights->info()->quantization_info()); - _flip_weights.configure(&_permuted_weights, &_weights_flipped); + _flip_weights.configure(&_permuted_weights, &_weights_flipped, &_flip_axis); // setup the function to convolve the upscaled output const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL); @@ -230,13 +233,19 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con _upsample_f.configure(input, &_scaled_output, upsample_info); _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout)); - _flip_weights.configure(weights, &_weights_flipped); + _flip_weights.configure(weights, &_weights_flipped, &_flip_axis); // setup the function to convolve the upscaled output const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL); _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info); } _scaled_output.allocator()->allocate(); + + // Setup flip axis data + _flip_axis.allocator()->allocate(); + auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer()); + axis_data[0] = 0; + axis_data[1] = 1; } void NEDeconvolutionLayer::run() @@ -276,16 +285,13 @@ void NEDeconvolutionLayer::prepare() // Run weights flipping and mark original weights tensor as unused _weights_flipped.allocator()->allocate(); - NEScheduler::get().schedule(&_flip_weights, Window::DimZ); + _flip_weights.run(); _original_weights->mark_as_unused(); // Prepare convolution _conv_f.prepare(); - if(!_weights_flipped.is_used()) - { - _weights_flipped.allocator()->free(); - } + // Unused weights are already released in _conv_f if(!_is_nchw) { |