src/runtime/CL/functions/CLConvolutionLayer.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/CL/functions/CLConvolutionLayer.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/CL/CLScheduler.h"
  30
  31 #include <cmath>
  32 #include <tuple>
  33
  34 using namespace arm_compute;
  35
  36 CLConvolutionLayer::CLConvolutionLayer()
  37     : _input_im2col_kernel(), _weights_reshape_kernel(), _input_interleave_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), _is_fc(false)
  39 {
  40 }
  41
  42 void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
  43 {
  44     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
  45     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F16, DataType::F32);
  46     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, 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::F16, 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     bool is_same_dimension = true;
  68     // Make sure the input and weights have same low three dimensions
  69     for(int i = 0; i < 3; i++)
  70     {
  71         is_same_dimension = (is_same_dimension) && (input->info()->dimension(i) == weights->info()->dimension(i));
  72     }
  73
  74     // Run the fully connected path if is_same_dimension is true and conv_stride_x/conv_stride_y are 1, and conv_pad_x/conv_pad_y are 0 and skip col2im
  75     _is_fc = (is_same_dimension) && ((stride_x & stride_y) == 1) && ((pad_x | pad_y) == 0);
  76
  77     // Get convolved dimensions
  78     unsigned int conv_w = 0;
  79     unsigned int conv_h = 0;
  80     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0),
  81                                                  stride_x, stride_y, pad_x, pad_y, conv_info.round());
  82     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");
  83
  84     // Create tensor to store the reshaped weights
  85     const size_t      mat_weights_cols = weights->info()->dimension(3);
  86     const size_t      mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0);
  87     const TensorShape shape_wr(mat_weights_cols, mat_weights_rows);
  88     _weights_reshaped.allocator()->init(TensorInfo(shape_wr, 1, weights->info()->data_type()));
  89
  90     // Create tensor to store transposed weights
  91     TensorShape shape_wt(mat_weights_rows * 4, static_cast<size_t>(std::ceil(mat_weights_cols / 4.f)));
  92     TensorInfo  info_wt(shape_wt, 1, weights->info()->data_type());
  93     _weights_transposed.allocator()->init(info_wt);
  94
  95     // Create tensor to store im2col reshaped inputs
  96     const size_t mat_input_cols = mat_weights_rows;
  97     const size_t mat_input_rows = _is_fc ? (input->info()->dimension(3)) : (conv_w * conv_h);
  98     TensorShape  shape_im2col   = input->info()->tensor_shape();
  99     shape_im2col.set(0, mat_input_cols);
 100     shape_im2col.set(1, mat_input_rows);
 101     shape_im2col.set(2, 1);
 102     if(_is_fc)
 103     {
 104         shape_im2col.set(3, 1);
 105     }
 106     _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
 107
 108     // Create tensor to prepare input tensor for GEMM
 109     TensorShape shape_interleaved = shape_im2col;
 110     shape_interleaved.set(0, shape_interleaved.x() * 4);
 111     shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4));
 112     _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type()));
 113
 114     // Create GEMM output tensor
 115     TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
 116     shape_gemm.set(0, mat_weights_cols);
 117     shape_gemm.set(1, mat_input_rows);
 118     _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type()));
 119
 120     // Configure kernels
 121     _input_im2col_kernel.configure(input, &_input_im2col_reshaped, std::make_pair(conv_w, conv_h), conv_info, _has_bias);
 122     _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
 123     _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
 124     _weights_transposed_kernel.configure(&_weights_reshaped, &_weights_transposed);
 125     if(_is_fc)
 126     {
 127         _mm_kernel.configure(&_input_interleaved_reshaped, &_weights_transposed, output, 1.0f);
 128     }
 129     else
 130     {
 131         _mm_kernel.configure(&_input_interleaved_reshaped, &_weights_transposed, &_gemm_output, 1.0f);
 132         _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
 133     }
 134
 135     // Allocate intermediate tensors
 136     _weights_reshaped.allocator()->allocate();
 137     _weights_transposed.allocator()->allocate();
 138     _input_im2col_reshaped.allocator()->allocate();
 139     _input_interleaved_reshaped.allocator()->allocate();
 140     _gemm_output.allocator()->allocate();
 141 }
 142
 143 void CLConvolutionLayer::run()
 144 {
 145     // Run weights reshaping (Runs once for every configure)
 146     if(_is_first_run)
 147     {
 148         _is_first_run = false;
 149         CLScheduler::get().enqueue(_weights_reshape_kernel);
 150         CLScheduler::get().enqueue(_weights_transposed_kernel);
 151     }
 152
 153     // Run input reshaping
 154     CLScheduler::get().enqueue(_input_im2col_kernel);
 155     CLScheduler::get().enqueue(_input_interleave_kernel);
 156
 157     // Runs matrix multiply on reshaped matrices
 158     CLScheduler::get().enqueue(_mm_kernel);
 159
 160     // Reshape output matrix
 161
 162     if(!_is_fc)
 163     {
 164         CLScheduler::get().enqueue(_output_col2im_kernel, false);
 165     }
 166 }