Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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/CLGEMMMatrixMultiplyKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/AccessWindowTranspose.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/CL/CLHelpers.h" |
| 29 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 30 | #include "arm_compute/core/CL/ICLTensor.h" |
| 31 | #include "arm_compute/core/CL/OpenCL.h" |
| 32 | #include "arm_compute/core/Error.h" |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 33 | #include "arm_compute/core/FixedPoint.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 34 | #include "arm_compute/core/Helpers.h" |
| 35 | #include "arm_compute/core/Types.h" |
| 36 | #include "arm_compute/core/Utils.h" |
| 37 | #include "arm_compute/core/Validate.h" |
| 38 | #include "arm_compute/core/Window.h" |
| 39 | |
| 40 | #include <set> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | #include <string> |
| 42 | |
| 43 | using namespace arm_compute; |
| 44 | |
| 45 | CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel() |
| 46 | : _input0(nullptr), _input1(nullptr), _output(nullptr) |
| 47 | { |
| 48 | } |
| 49 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 50 | void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, bool is_interleaved_transposed) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 51 | { |
Gian Marco Iodice | 8a38369 | 2017-07-03 17:41:47 +0100 | [diff] [blame] | 52 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 53 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); |
Gian Marco Iodice | 3a3066b | 2017-06-23 13:38:14 +0100 | [diff] [blame] | 54 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 55 | if(!is_interleaved_transposed) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 56 | { |
| 57 | ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); |
| 58 | } |
| 59 | |
| 60 | _input0 = input0; |
| 61 | _input1 = input1; |
| 62 | _output = output; |
| 63 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 64 | const DataType data_type = input0->info()->data_type(); |
| 65 | const int fp_pos = input0->info()->fixed_point_position(); |
| 66 | |
| 67 | // Get target architecture |
| 68 | GPUTarget arch_target = get_arch_from_target(get_target()); |
| 69 | |
| 70 | // Configure LWS hint |
| 71 | _lws_hint = (output->info()->dimension(1) == 196) ? cl::NDRange(1, 7) : cl::NDRange(8, 8); |
| 72 | |
| 73 | // Create build options |
| 74 | CLBuildOptions build_opts; |
| 75 | build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(fp_pos)); |
| 76 | |
| 77 | const bool multiply_alpha = std::abs(1.0f - alpha) > 0.00001f; |
| 78 | |
| 79 | // Only define ALPHA when alpha is not 1.0f. This avoids performing unnecessary multiplications. |
| 80 | if(multiply_alpha) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 81 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 82 | build_opts.add_option_if_else(is_data_type_fixed_point(data_type), |
| 83 | "-DALPHA=" + support::cpp11::to_string((data_type == DataType::QS8 ? sqcvt_qs8_f32(alpha, fp_pos) : sqcvt_qs16_f32(alpha, fp_pos))), |
| 84 | "-DALPHA=" + float_to_string_with_full_precision(alpha)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 85 | } |
| 86 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 87 | std::string kernel_name; |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 88 | if(is_interleaved_transposed) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 89 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 90 | build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0))); |
| 91 | if(data_type == DataType::F32) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 92 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 93 | kernel_name = "gemm_mm_interleaved_transposed_f32_" + string_from_target(arch_target); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 94 | } |
| 95 | else |
| 96 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 97 | kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 98 | } |
| 99 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 100 | // Configure kernel window |
| 101 | const unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 102 | constexpr unsigned int num_elems_processed_per_iteration_y = 4; |
| 103 | |
| 104 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 105 | |
| 106 | AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f); |
| 107 | AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f); |
| 108 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 109 | |
| 110 | update_window_and_padding(win, input0_access, input1_access, output_access); |
| 111 | |
| 112 | output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); |
| 113 | |
| 114 | ICLKernel::configure(win); |
| 115 | } |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 116 | else // The input tensors have not been reshaped |
| 117 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 118 | build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0))); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 119 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 120 | // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x is set up for the default case. |
| 121 | unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 122 | const unsigned int num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4); |
| 123 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 124 | // Create kernels according to the architecture, data type and input size. |
| 125 | if(arch_target == GPUTarget::BIFROST && data_type == DataType::F32) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 126 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 127 | // The first kernel is optimized for the case of 1000 or less output elements (e.g. FC8 of AlexNet and VGG-16, and |
| 128 | // FC1 of Inception v3). The second kernel is optimized for the case of greater than 1000 output elements (e.g. |
| 129 | // FC6 and FC7 of AlexNet and VGG-16). |
| 130 | if(input1->info()->dimension(0) <= 1000) |
| 131 | { |
| 132 | // Each work-item processes 2 elements in the X dimension. |
| 133 | num_elems_processed_per_iteration_x = 2; |
| 134 | kernel_name = "gemm_mm_floating_point_f32_bifrost_1000"; |
| 135 | } |
| 136 | else |
| 137 | { |
| 138 | // Each work-item processes 4 elements in the X dimension (as in the default case). |
| 139 | num_elems_processed_per_iteration_x = 4; |
| 140 | kernel_name = "gemm_mm_floating_point_f32_bifrost"; |
| 141 | } |
| 142 | // The work-group size equal to the Bifrost quad size has been proved to be optimal for these kernels |
| 143 | // via exhaustive autotuning over a range of representative layer configurations. |
| 144 | _lws_hint = cl::NDRange(4); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 145 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 146 | else if(is_data_type_fixed_point(data_type)) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 147 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 148 | kernel_name = "gemm_mm_" + lower_string(string_from_data_type(data_type)); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 149 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 150 | else // (MIDGARD and F32) or (F16) |
| 151 | { |
| 152 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); |
| 153 | kernel_name = "gemm_mm_floating_point"; |
| 154 | } |
| 155 | build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elems_processed_per_iteration_y)); |
| 156 | build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elems_processed_per_iteration_x)); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 157 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 158 | // Configure window |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 159 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 160 | |
| 161 | AccessWindowStatic input0_access(input0->info(), 0, 0, input0->info()->dimension(0), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y)); |
| 162 | AccessWindowStatic input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1)); |
| 163 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 164 | |
| 165 | update_window_and_padding(win, input0_access, input1_access, output_access); |
| 166 | |
| 167 | Coordinates coord; |
| 168 | coord.set_num_dimensions(output->info()->num_dimensions()); |
| 169 | output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape())); |
| 170 | |
| 171 | ICLKernel::configure(win); |
| 172 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame^] | 173 | |
| 174 | // Create kernel |
| 175 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |
| 176 | |
| 177 | // Set config_id for enabling LWS tuning |
| 178 | _config_id = "gemm_"; |
| 179 | _config_id += (is_interleaved_transposed ? "reshaped_" : ""); |
| 180 | _config_id += lower_string(string_from_data_type(input0->info()->data_type())); |
| 181 | _config_id += "_"; |
| 182 | _config_id += support::cpp11::to_string(output->info()->dimension(1)); |
| 183 | _config_id += "_"; |
| 184 | _config_id += support::cpp11::to_string(output->info()->dimension(0)); |
| 185 | _config_id += "_"; |
| 186 | _config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1))); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 187 | } |
| 188 | |
| 189 | void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue) |
| 190 | { |
| 191 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 192 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 193 | |
| 194 | Window slice = window.first_slice_window_2D(); |
| 195 | Window slice_matrix_b = slice; |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 196 | |
| 197 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 198 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 199 | |
| 200 | do |
| 201 | { |
| 202 | Window slice_b = slice; |
| 203 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 204 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 205 | if(_input1->info()->num_dimensions() < 3) |
| 206 | { |
| 207 | slice_b = slice_matrix_b; |
| 208 | } |
| 209 | |
| 210 | unsigned int idx = 0; |
| 211 | add_2D_tensor_argument(idx, _input0, slice); |
| 212 | add_2D_tensor_argument(idx, _input1, slice_b); |
| 213 | add_2D_tensor_argument(idx, _output, slice); |
| 214 | enqueue(queue, *this, slice, _lws_hint); |
| 215 | } |
| 216 | while(window.slide_window_slice_2D(slice)); |
| 217 | } |