diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2017-11-23 12:10:21 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:41:04 +0000 |
commit | b7b31538eb9137e4d3e8de6d381dcbe9fc58df94 (patch) | |
tree | 7cca8c388cfb15867b9d92c6fd793ca1588b6526 /src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp | |
parent | 02bf80d4554cfc824a76008905921cb564bee999 (diff) | |
download | ComputeLibrary-b7b31538eb9137e4d3e8de6d381dcbe9fc58df94.tar.gz |
COMPMID-464 Implement Depthwise separable convolution on NEON
Change-Id: Iccd686be18381e96bcf09b14c7017c6dda0f38d8
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/109824
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp')
-rw-r--r-- | src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp | 95 |
1 files changed, 95 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp new file mode 100644 index 0000000000..6deda506ab --- /dev/null +++ b/src/core/NEON/kernels/NEDepthwiseVectorToTensorKernel.cpp @@ -0,0 +1,95 @@ +/* + * Copyright (c) 2017 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/NEDepthwiseVectorToTensorKernel.h" + +#include "arm_compute/core/AccessWindowTranspose.h" +#include "arm_compute/core/Coordinates.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/INEKernel.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" + +using namespace arm_compute; + +NEDepthwiseVectorToTensorKernel::NEDepthwiseVectorToTensorKernel() + : _input(nullptr), _output(nullptr), _conv_dims() +{ +} + +void NEDepthwiseVectorToTensorKernel::configure(const ITensor *input, ITensor *output, size_t conv_w, size_t conv_h) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); + + _input = input; + _output = output; + _conv_dims = std::pair<size_t, size_t>(conv_w, conv_h); + + // Configure kernel window + Window win = calculate_max_window(*input->info(), Steps()); + // The NEDepthwisevectorToTensorKernel doesn't need padding so update_window_and_padding() can be skipped + output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); + + INEKernel::configure(win); +} + +void NEDepthwiseVectorToTensorKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + + // const int input_w = _input->info()->dimension(0); + const int output_stride_x = _output->info()->strides_in_bytes().x(); + const int output_stride_y = _output->info()->strides_in_bytes().y(); + const int output_stride_z = _output->info()->strides_in_bytes().z(); + + // Setup output window + Window window_out(window); + window_out.set(Window::DimX, Window::Dimension(0, 0, 0)); + window_out.set(Window::DimY, Window::Dimension(0, 0, 0)); + window_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); + + Iterator in(_input, window); + Iterator out(_output, window_out); + + const int patch_size = _conv_dims.first * _conv_dims.second; + + execute_window_loop(window, [&](const Coordinates & id) + { + const int z = id.x() / patch_size; + const int index2D = id.x() - z * patch_size; + + auto input_ptr = reinterpret_cast<float *>(in.ptr()); + auto output_ptr = reinterpret_cast<float *>(out.ptr() + index2D % _conv_dims.first * output_stride_x + index2D / _conv_dims.first * output_stride_y + z * output_stride_z); + + *output_ptr = *input_ptr; + }, + in, out); +} |