steniu01 | 27b386c | 2017-07-18 17:37:43 +0100 | [diff] [blame^] | 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/core/CL/kernels/CLDirectConvolutionLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/CL/CLHelpers.h" |
| 28 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 29 | #include "arm_compute/core/CL/ICLTensor.h" |
| 30 | #include "arm_compute/core/Error.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/IAccessWindow.h" |
| 33 | #include "arm_compute/core/ITensor.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/core/Validate.h" |
| 36 | #include "support/ToolchainSupport.h" |
| 37 | |
| 38 | using namespace arm_compute; |
| 39 | |
| 40 | template <unsigned int kernel_size> |
| 41 | CLDirectConvolutionLayerKernel<kernel_size>::CLDirectConvolutionLayerKernel() |
| 42 | : _input(nullptr), _biases(nullptr), _weights(nullptr), _output(nullptr), _border_size(0), _conv_pad_x(0), _conv_pad_y(0), _conv_stride_x(0), _conv_stride_y(0) |
| 43 | { |
| 44 | } |
| 45 | |
| 46 | template <unsigned int kernel_size> |
| 47 | BorderSize CLDirectConvolutionLayerKernel<kernel_size>::border_size() const |
| 48 | { |
| 49 | return _border_size; |
| 50 | } |
| 51 | |
| 52 | template <unsigned int kernel_size> |
| 53 | void CLDirectConvolutionLayerKernel<kernel_size>::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info) |
| 54 | { |
| 55 | static_assert(kernel_size == 3, "Currently only 3x3 direct convolution is supported!"); |
| 56 | |
| 57 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 58 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 59 | ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2)); |
| 60 | ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != weights->info()->dimension(1)); |
| 61 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 62 | ARM_COMPUTE_ERROR_ON_MSG((kernel_size == 3 && std::get<0>(conv_info.stride()) > 2), "Strides larger than 2 not supported in 3x3 direct convolution!"); |
| 63 | |
| 64 | ARM_COMPUTE_ERROR_ON(kernel_size != weights->info()->dimension(0)); |
| 65 | |
| 66 | if(biases != nullptr) |
| 67 | { |
| 68 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 69 | ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); |
| 70 | ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); |
| 71 | } |
| 72 | |
| 73 | _conv_stride_x = std::get<0>(conv_info.stride()); |
| 74 | _conv_stride_y = std::get<1>(conv_info.stride()); |
| 75 | _conv_pad_x = std::get<0>(conv_info.pad()); |
| 76 | _conv_pad_y = std::get<1>(conv_info.pad()); |
| 77 | |
| 78 | _input = input; |
| 79 | _weights = weights; |
| 80 | _output = output; |
| 81 | _biases = biases; |
| 82 | _border_size = BorderSize(_conv_pad_y, _conv_pad_x); |
| 83 | |
| 84 | std::stringstream kernel_name; |
| 85 | std::set<std::string> options; |
| 86 | kernel_name << "direct_convolution" << kernel_size << "x" << kernel_size; |
| 87 | |
| 88 | options.insert("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); |
| 89 | |
| 90 | options.emplace("-DSTRIDE_X=" + support::cpp11::to_string(_conv_stride_x)); |
| 91 | |
| 92 | if(_biases != nullptr) |
| 93 | { |
| 94 | options.emplace("-DHAS_BIAS"); |
| 95 | } |
| 96 | |
| 97 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name.str(), options)); |
| 98 | |
| 99 | unsigned int idx = (_biases == nullptr) ? 3 * num_arguments_per_3D_tensor() : (num_arguments_per_1D_tensor() + 3 * num_arguments_per_3D_tensor()); |
| 100 | _kernel.setArg<cl_uint>(idx++, _weights->info()->strides_in_bytes()[3]); // weights_stride_w |
| 101 | _kernel.setArg<cl_uint>(idx++, _weights->info()->dimension(2)); // filter depth |
| 102 | |
| 103 | // Using this local workgroup size gives better performance over others that have been tried. |
| 104 | _lws_hint = cl::NDRange(4, 1, 8); |
| 105 | |
| 106 | // Configure kernel window |
| 107 | Window win = calculate_max_window(*output->info()); |
| 108 | |
| 109 | unsigned int num_elems_read_per_iteration = 16 * _conv_stride_x; |
| 110 | unsigned int num_elems_written_per_iteration = 8; |
| 111 | |
| 112 | // Calculate right and bottom border |
| 113 | const int input_width = input->info()->dimension(0); |
| 114 | const int input_height = input->info()->dimension(1); |
| 115 | const int upper_bound_w = ceil_to_multiple(((output->info()->dimension(0) - 1) * _conv_stride_x + kernel_size), num_elems_read_per_iteration) - _conv_pad_x - input_width; |
| 116 | const int upper_bound_h = ((output->info()->dimension(1) - 1) * _conv_stride_y - _conv_pad_y + kernel_size) - input_height; |
| 117 | const int padding_right = std::max(upper_bound_w, static_cast<int>(kernel_size)); |
| 118 | const int padding_bottom = std::max(upper_bound_h, static_cast<int>(kernel_size)); |
| 119 | |
| 120 | // Create window and update padding |
| 121 | win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration)); |
| 122 | AccessWindowStatic input_access(input->info(), -_conv_pad_x, -_conv_pad_y, input_width + padding_right, input_height + padding_bottom); |
| 123 | |
| 124 | AccessWindowStatic weights_access(weights->info(), 0, 0, kernel_size, kernel_size); |
| 125 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 126 | update_window_and_padding(win, input_access, weights_access, output_access); |
| 127 | |
| 128 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); |
| 129 | |
| 130 | ICLKernel::configure(win); |
| 131 | } |
| 132 | |
| 133 | template <unsigned int kernel_size> |
| 134 | void CLDirectConvolutionLayerKernel<kernel_size>::run(const Window &window, cl::CommandQueue &queue) |
| 135 | { |
| 136 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 137 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 138 | |
| 139 | // Get initial windows |
| 140 | Window slice = window.first_slice_window_3D(); |
| 141 | Window win_in = window; |
| 142 | |
| 143 | win_in.adjust(Window::DimX, -_conv_pad_x, true); |
| 144 | win_in.adjust(Window::DimY, -_conv_pad_y, true); |
| 145 | win_in.set_dimension_step(Window::DimX, window.x().step() * _conv_stride_x); |
| 146 | win_in.set_dimension_step(Window::DimY, window.y().step() * _conv_stride_y); |
| 147 | |
| 148 | Window slice_in = win_in.first_slice_window_3D(); |
| 149 | |
| 150 | unsigned int idx1 = 2 * num_arguments_per_3D_tensor(); |
| 151 | add_3D_tensor_argument(idx1, _weights, slice); |
| 152 | |
| 153 | if(_biases != nullptr) |
| 154 | { |
| 155 | Window slice_biases; |
| 156 | slice_biases.use_tensor_dimensions(_biases->info()); |
| 157 | add_1D_tensor_argument(idx1, _biases, slice_biases); |
| 158 | } |
| 159 | |
| 160 | do |
| 161 | { |
| 162 | unsigned int idx = 0; |
| 163 | add_3D_tensor_argument(idx, _input, slice_in); |
| 164 | add_3D_tensor_argument(idx, _output, slice); |
| 165 | |
| 166 | enqueue(queue, *this, slice, _lws_hint); |
| 167 | } |
| 168 | while(window.slide_window_slice_3D(slice) && win_in.slide_window_slice_3D(slice_in)); |
| 169 | } |
| 170 | |
| 171 | template class arm_compute::CLDirectConvolutionLayerKernel<3>; |