src/runtime/NEON/functions/NEConvolutionLayer.cpp

   1 /*
   2  * Copyright (c) 2017 ARM Limited.
   3  *
   4  * SPDX-License-Identifier: MIT
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to
   8  * deal in the Software without restriction, including without limitation the
   9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  10  * sell copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in all
  14  * copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  */
  24 #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h"
  25
  26 #include "arm_compute/core/PixelValue.h"
  27 #include "arm_compute/core/Utils.h"
  28 #include "arm_compute/core/Validate.h"
  29 #include "arm_compute/runtime/NEON/NEScheduler.h"
  30
  31 #include <cmath>
  32 #include <tuple>
  33
  34 using namespace arm_compute;
  35
  36 NEConvolutionLayer::NEConvolutionLayer()
  37     : _input_im2col_kernel(), _input_interleave_kernel(), _weights_reshape_kernel(), _weights_transposed_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(),
  38       _input_interleaved_reshaped(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _is_first_run(false), _has_bias(false)
  39 {
  40 }
  41
  42 void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
  43 {
  44     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
  45     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
  46     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
  47     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
  48     ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2));
  49     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
  50
  51     if(biases != nullptr)
  52     {
  53         ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32);
  54         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
  55         ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
  56         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
  57     }
  58
  59     _has_bias     = (biases != nullptr);
  60     _is_first_run = true;
  61
  62     // Get parameters for conv_info
  63     unsigned int stride_x, stride_y, pad_x, pad_y = 0;
  64     std::tie(stride_x, stride_y) = conv_info.stride();
  65     std::tie(pad_x, pad_y)       = conv_info.pad();
  66
  67     // Get convolved dimensions
  68     unsigned int conv_w = 0;
  69     unsigned int conv_h = 0;
  70     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0),
  71                                                  stride_x, stride_y, pad_x, pad_y, conv_info.round());
  72     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");
  73
  74     // Create tensor to store the reshaped weights
  75     const size_t      mat_weights_cols = weights->info()->dimension(3);
  76     const size_t      mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0);
  77     const TensorShape shape_wr(mat_weights_cols, mat_weights_rows);
  78     TensorInfo        info_wr(shape_wr, 1, weights->info()->data_type());
  79     _weights_reshaped.allocator()->init(info_wr);
  80
  81     // Create tensor to store transposed weights
  82     TensorShape shape_wt(mat_weights_rows * 4, static_cast<size_t>(std::ceil(mat_weights_cols / 4.f)));
  83     TensorInfo  info_wt(shape_wt, 1, weights->info()->data_type());
  84     _weights_transposed.allocator()->init(info_wt);
  85
  86     // Create tensor to store im2col reshaped inputs
  87     const size_t mat_input_cols = mat_weights_rows;
  88     const size_t mat_input_rows = conv_w * conv_h;
  89     TensorShape  shape_im2col   = input->info()->tensor_shape();
  90     shape_im2col.set(0, mat_input_cols);
  91     shape_im2col.set(1, mat_input_rows);
  92     shape_im2col.set(2, 1);
  93     TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type());
  94     _input_im2col_reshaped.allocator()->init(info_im2col);
  95
  96     // Create tensor to prepare input tensor for GEMM
  97     TensorShape shape_interleaved = shape_im2col;
  98     shape_interleaved.set(0, shape_interleaved.x() * 4);
  99     shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4));
 100     TensorInfo info_interleaved(shape_interleaved, 1, input->info()->data_type());
 101     _input_interleaved_reshaped.allocator()->init(info_interleaved);
 102
 103     // Create GEMM output tensor
 104     TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
 105     shape_gemm.set(0, mat_weights_cols);
 106     shape_gemm.set(1, mat_input_rows);
 107     TensorInfo info_gemm(shape_gemm, 1, input->info()->data_type());
 108     _gemm_output.allocator()->init(info_gemm);
 109
 110     // Configure kernels
 111     _input_im2col_kernel.configure(input, &_input_im2col_reshaped, std::make_pair(conv_w, conv_h), conv_info, _has_bias);
 112     _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
 113     _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
 114     _weights_transposed_kernel.configure(&_weights_reshaped, &_weights_transposed);
 115     _mm_kernel.configure(&_input_interleaved_reshaped, &_weights_transposed, &_gemm_output, 1.0f);
 116     _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
 117
 118     // Allocate the tensors once the all configure methods have been called
 119     _weights_reshaped.allocator()->allocate();
 120     _weights_transposed.allocator()->allocate();
 121     _input_im2col_reshaped.allocator()->allocate();
 122     _input_interleaved_reshaped.allocator()->allocate();
 123     _gemm_output.allocator()->allocate();
 124 }
 125
 126 void NEConvolutionLayer::run()
 127 {
 128     // Run weights reshaping (Runs once for every configure)
 129     if(_is_first_run)
 130     {
 131         _is_first_run = false;
 132         NEScheduler::get().multithread(&_weights_reshape_kernel, 3);
 133         NEScheduler::get().multithread(&_weights_transposed_kernel);
 134     }
 135
 136     // Run input reshaping
 137     NEScheduler::get().multithread(&_input_im2col_kernel);
 138
 139     // Run interleave
 140     NEScheduler::get().multithread(&_input_interleave_kernel);
 141
 142     // Runs GEMM on reshaped matrices
 143     NEScheduler::get().multithread(&_mm_kernel);
 144
 145     // Reshape output matrix
 146     NEScheduler::get().multithread(&_output_col2im_kernel);
 147 }