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/CLGEMMLowpMatrixMultiplyKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 28 | #include "arm_compute/core/CL/ICLTensor.h" |
| 29 | #include "arm_compute/core/CL/OpenCL.h" |
| 30 | #include "arm_compute/core/Error.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/Types.h" |
| 33 | #include "arm_compute/core/Utils.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
Moritz Pflanzer | 05da6dd | 2017-07-04 12:08:41 +0100 | [diff] [blame] | 36 | #include "support/ToolchainSupport.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 37 | |
| 38 | #include <cstddef> |
| 39 | #include <cstdint> |
| 40 | #include <tuple> |
| 41 | |
| 42 | using namespace arm_compute; |
| 43 | |
| 44 | namespace arm_compute |
| 45 | { |
| 46 | class Coordinates; |
| 47 | } // namespace arm_compute |
| 48 | |
| 49 | CLGEMMLowpMatrixMultiplyKernel::CLGEMMLowpMatrixMultiplyKernel() |
| 50 | : _input0(nullptr), _input1(nullptr), _output(nullptr) |
| 51 | { |
| 52 | } |
| 53 | |
| 54 | void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, |
| 55 | int32_t a_offset, int32_t b_offset, int32_t output_offset, int32_t output_mult_int, int32_t shift) |
| 56 | { |
| 57 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::U8); |
| 58 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8); |
| 59 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); |
| 60 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); |
| 61 | |
| 62 | _input0 = input0; |
| 63 | _input1 = input1; |
| 64 | _output = output; |
| 65 | |
| 66 | // Create kernel and set static arguments |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 67 | std::set<std::string> build_opts = { ("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0))) }; |
| 68 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mm_interleaved_transposed_u8", build_opts)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 69 | unsigned int idx = 3 * num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 70 | _kernel.setArg<int32_t>(idx++, a_offset); |
| 71 | _kernel.setArg<int32_t>(idx++, b_offset); |
| 72 | _kernel.setArg<int32_t>(idx++, output_offset); |
| 73 | _kernel.setArg<int32_t>(idx++, output_mult_int); |
| 74 | _kernel.setArg<int32_t>(idx++, shift); |
| 75 | |
| 76 | // Configure window |
| 77 | constexpr unsigned int num_elems_processed_per_iteration_x = 16; |
| 78 | constexpr unsigned int num_elems_processed_per_iteration_y = 4; |
| 79 | constexpr unsigned int num_elems_read_per_iteration_input0 = 4; |
| 80 | constexpr unsigned int num_elems_read_per_iteration_input1 = 16; |
| 81 | |
| 82 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 83 | |
| 84 | AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_read_per_iteration_input0, 1); |
| 85 | AccessWindowRectangle input1_access(input1->info(), 0, 0, num_elems_read_per_iteration_input1, 1); |
| 86 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 87 | |
| 88 | update_window_and_padding(win, input0_access, input1_access, output_access); |
| 89 | |
| 90 | output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); |
| 91 | |
| 92 | ICLKernel::configure(win); |
| 93 | } |
| 94 | |
| 95 | void CLGEMMLowpMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue) |
| 96 | { |
| 97 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 98 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 99 | |
| 100 | Window slice = window.first_slice_window_2D(); |
| 101 | Window slice_matrix_b = slice; |
| 102 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, _input1->info()->dimension(0), 1)); |
| 103 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, _input1->info()->dimension(1), 1)); |
| 104 | slice_matrix_b.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 105 | |
| 106 | do |
| 107 | { |
| 108 | Window slice_b = slice; |
| 109 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 110 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 111 | if(_input1->info()->num_dimensions() < 3) |
| 112 | { |
| 113 | slice_b = slice_matrix_b; |
| 114 | } |
| 115 | |
| 116 | unsigned int idx = 0; |
| 117 | add_2D_tensor_argument(idx, _input0, slice); |
| 118 | add_2D_tensor_argument(idx, _input1, slice_b); |
| 119 | add_2D_tensor_argument(idx, _output, slice); |
| 120 | enqueue(queue, *this, slice); |
| 121 | } |
| 122 | while(window.slide_window_slice_2D(slice)); |
| 123 | } |