From 6ff3b19ee6120edf015fad8caab2991faa3070af Mon Sep 17 00:00:00 2001 From: Anthony Barbier Date: Mon, 4 Sep 2017 18:44:23 +0100 Subject: COMPMID-344 Updated doxygen Change-Id: I32f7b84daa560e460b77216add529c8fa8b327ae --- .../NEON/functions/NELocallyConnectedLayer.cpp | 131 +++++++++++++++++++++ 1 file changed, 131 insertions(+) create mode 100644 src/runtime/NEON/functions/NELocallyConnectedLayer.cpp (limited to 'src/runtime/NEON/functions/NELocallyConnectedLayer.cpp') diff --git a/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp b/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp new file mode 100644 index 0000000000..85d7ba3650 --- /dev/null +++ b/src/runtime/NEON/functions/NELocallyConnectedLayer.cpp @@ -0,0 +1,131 @@ +/* + * 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/runtime/NEON/functions/NELocallyConnectedLayer.h" + +#include "arm_compute/core/PixelValue.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" + +#include +#include + +using namespace arm_compute; + +NELocallyConnectedLayer::NELocallyConnectedLayer() + : _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(), _is_first_run(false) +{ +} + +void NELocallyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); + ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2)); + + if(biases != nullptr) + { + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); + ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 2); + } + + bool _has_bias = (biases != nullptr); + _is_first_run = true; + + // Get parameters for conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + unsigned int pad_x = 0; + unsigned int pad_y = 0; + std::tie(stride_x, stride_y) = conv_info.stride(); + std::tie(pad_x, pad_y) = conv_info.pad(); + + // Get convolved dimensions + unsigned int conv_w = 0; + unsigned int conv_h = 0; + std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), + stride_x, stride_y, pad_x, pad_y, conv_info.round()); + + ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); + ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(4) != (conv_w * conv_h), "Weights shape does not match the expected one"); + + // Create tensor to store the reshaped weights + const size_t mat_weights_cols = weights->info()->dimension(3); + const size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0); + const size_t mat_weights_num = weights->info()->dimension(4); + + const TensorShape shape_wr(mat_weights_cols, mat_weights_rows, mat_weights_num); + + _weights_reshaped.allocator()->init(TensorInfo(shape_wr, 1, weights->info()->data_type())); + + // Create tensor to store im2col reshaped inputs + const size_t mat_input_cols = mat_weights_rows; + const size_t mat_input_rows = conv_w * conv_h; + TensorShape shape_im2col = input->info()->tensor_shape(); + shape_im2col.set(0, mat_input_cols); + shape_im2col.set(1, mat_input_rows); + shape_im2col.set(2, 1); + + _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type())); + + // Create locally connected layer output tensor + TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, mat_input_rows); + _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type())); + + // Configure kernels + _input_im2col_kernel.configure(input, &_input_im2col_reshaped, std::make_pair(conv_w, conv_h), conv_info, _has_bias); + _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped); + _mm_kernel.configure(&_input_im2col_reshaped, &_weights_reshaped, &_gemm_output); + _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h)); + + // Allocate intermediate tensors + _weights_reshaped.allocator()->allocate(); + _input_im2col_reshaped.allocator()->allocate(); + _gemm_output.allocator()->allocate(); +} + +void NELocallyConnectedLayer::run() +{ + // Run weights reshaping (Runs once for every configure) + if(_is_first_run) + { + _is_first_run = false; + NEScheduler::get().schedule(&_weights_reshape_kernel, 3); + } + + // Run input reshaping + NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); + + // Runs GEMM on reshaped matrices + NEScheduler::get().schedule(&_mm_kernel, Window::DimX); + + // Reshape output matrix + NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); +} -- cgit v1.2.1