Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +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/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h" |
| 28 | #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h" |
| 29 | #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h" |
| 30 | #include "arm_compute/core/GLES_COMPUTE/OpenGLES.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" |
| 36 | |
| 37 | using namespace arm_compute; |
Joel Liang | 1c5ffd6 | 2017-12-28 10:09:51 +0800 | [diff] [blame^] | 38 | using namespace arm_compute::gles_compute; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 39 | |
| 40 | GCGEMMInterleave4x4Kernel::GCGEMMInterleave4x4Kernel() |
| 41 | : _input(nullptr), _output(nullptr) |
| 42 | { |
| 43 | } |
| 44 | |
| 45 | void GCGEMMInterleave4x4Kernel::configure(const IGCTensor *input, IGCTensor *output) |
| 46 | { |
| 47 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 48 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
| 49 | |
| 50 | TensorShape output_shape = input->info()->tensor_shape(); |
| 51 | output_shape.set(0, input->info()->dimension(0) * 4); |
| 52 | output_shape.set(1, std::ceil(input->info()->dimension(1) / 4.0f)); |
| 53 | |
| 54 | // Output auto inizialitation if not yet initialized |
| 55 | auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); |
| 56 | |
| 57 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); |
| 58 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 59 | |
| 60 | _input = input; |
| 61 | _output = output; |
| 62 | |
| 63 | std::set<std::string> build_opts; |
| 64 | std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16"; |
| 65 | build_opts.emplace(("#define " + dt_name)); |
| 66 | build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)); |
| 67 | build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); |
| 68 | build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); |
| 69 | |
| 70 | // Create kernel |
| 71 | build_opts.emplace("#define GEMM_INTERLEAVE4x4"); |
| 72 | _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("gemm_interleave4x4", build_opts)); |
| 73 | |
| 74 | // Configure kernel window |
| 75 | const unsigned int num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input->info()->data_type()); |
| 76 | constexpr unsigned int num_elems_processed_per_iteration_y = 4; |
| 77 | const unsigned int num_elems_written_per_iteration = num_elems_processed_per_iteration_x * num_elems_processed_per_iteration_y; |
| 78 | |
| 79 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 80 | |
| 81 | AccessWindowRectangle input_access(input->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 82 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration, 1, 4.f, 0.25f); |
| 83 | |
| 84 | update_window_and_padding(win, input_access, output_access); |
| 85 | |
| 86 | output_access.set_valid_region(win, input->info()->valid_region()); |
| 87 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 88 | IGCKernel::configure(win); |
| 89 | } |
| 90 | |
| 91 | void GCGEMMInterleave4x4Kernel::run(const Window &window) |
| 92 | { |
| 93 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 94 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IGCKernel::window(), window); |
| 95 | |
| 96 | _kernel.use(); |
| 97 | |
| 98 | /* |
| 99 | * This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values) |
| 100 | * |a00 a01 a02 a03| |
| 101 | * |a10 a11 a12 a13| |
| 102 | * |a20 a21 a22 a23| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 | |
| 103 | * |a30 a31 a32 a33| |
| 104 | * |
| 105 | * After this operation, the output matrix will have the following shape: [ height * 4, width / 4 ] |
| 106 | */ |
| 107 | Window in_slice = window.first_slice_window_2D(); |
| 108 | Window out_slice = window.first_slice_window_2D(); |
| 109 | |
| 110 | // Change x and y steps for the slide of output tensor |
| 111 | out_slice.scale(Window::DimX, 4.f); |
| 112 | out_slice.scale(Window::DimY, 0.25f); |
| 113 | |
| 114 | do |
| 115 | { |
| 116 | unsigned int idx = 0; |
| 117 | add_2D_tensor_argument(idx, _input, 1, in_slice); |
| 118 | add_2D_tensor_argument(idx, _output, 2, out_slice); |
| 119 | |
| 120 | _kernel.update_shader_params(); |
| 121 | |
| 122 | enqueue(*this, in_slice); |
| 123 | } |
| 124 | while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice)); |
| 125 | } |