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/CLGEMMTranspose1xWKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowTranspose.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/CL/OpenCL.h" |
| 31 | #include "arm_compute/core/Error.h" |
| 32 | #include "arm_compute/core/Helpers.h" |
| 33 | #include "arm_compute/core/Types.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
| 36 | |
| 37 | #include <cmath> |
| 38 | |
| 39 | using namespace arm_compute; |
| 40 | |
| 41 | void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *output) |
| 42 | { |
| 43 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::F16, DataType::F32); |
| 44 | ARM_COMPUTE_ERROR_ON(output == nullptr); |
| 45 | |
| 46 | TensorShape output_shape{ input->info()->tensor_shape() }; |
| 47 | const size_t transpose_w = 16 / input->info()->element_size(); |
| 48 | output_shape.set(0, input->info()->dimension(1) * transpose_w); |
| 49 | output_shape.set(1, static_cast<size_t>(std::ceil((input->info()->dimension(0) / static_cast<float>(transpose_w))))); |
| 50 | |
| 51 | // Output tensor auto inizialitation if not yet initialized |
| 52 | auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); |
| 53 | |
| 54 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 55 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); |
| 56 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 57 | const unsigned int num_elems_processed_per_iteration = max_cl_vector_width / data_size_from_type(input->info()->data_type()); |
Georgios Pinitas | 4cbee6e | 2017-06-19 13:02:56 +0100 | [diff] [blame^] | 58 | const float scale_x = num_elems_processed_per_iteration; |
| 59 | ARM_COMPUTE_ERROR_ON((0 == static_cast<int>(input->info()->dimension(0) * (1.f / scale_x)))); |
| 60 | |
| 61 | _input = input; |
| 62 | _output = output; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 63 | |
| 64 | /* |
| 65 | * Following an example of how the transposition1xW works when the input data type is F32 |
| 66 | * |
| 67 | * |a00 a01 a02 a03| |
| 68 | * |a10 a11 a12 a13| |
| 69 | * |a20 a21 a22 a23| = | a00 a01 a02 a03 || a10 a11 a12 a13 || a20 a21 a22 a23 || a30 a31 a32 a33 | |
| 70 | * |a30 a31 a32 a33| |
| 71 | * |
| 72 | * If the input data type is F32, the output matrix will have the following shape: [ height * 4, width / 4 ] |
| 73 | * If the input data type is F16, the output matrix will have the following shape: [ height * 8, width / 8 ] |
| 74 | */ |
| 75 | // Create kernel |
| 76 | std::string data_type_name = lower_string(string_from_data_type(input->info()->data_type())); |
| 77 | std::string kernel_name = "gemm_transpose1x" + val_to_string(num_elems_processed_per_iteration) + "_" + data_type_name; |
| 78 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name)); |
| 79 | |
| 80 | // Configure window |
| 81 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 82 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 83 | AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); |
| 84 | AccessWindowTranspose output_access(output->info(), 0, 0, num_elems_processed_per_iteration, 1, scale_x, 1.f / scale_x); |
| 85 | |
| 86 | update_window_and_padding(win, input_access, output_access); |
| 87 | |
Georgios Pinitas | 4cbee6e | 2017-06-19 13:02:56 +0100 | [diff] [blame^] | 88 | output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->info()->tensor_shape())); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 89 | |
| 90 | ICLKernel::configure(win); |
| 91 | } |
| 92 | |
| 93 | void CLGEMMTranspose1xWKernel::run(const Window &window, cl::CommandQueue &queue) |
| 94 | { |
| 95 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 96 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 97 | |
| 98 | // Output is transposed |
| 99 | Window out_window(window); |
| 100 | out_window.set(Window::DimX, window.y()); |
| 101 | out_window.set(Window::DimY, window.x()); |
| 102 | |
| 103 | Window in_slice = window.first_slice_window_2D(); |
| 104 | Window out_slice = out_window.first_slice_window_2D(); |
| 105 | |
| 106 | do |
| 107 | { |
| 108 | unsigned int idx = 0; |
| 109 | add_2D_tensor_argument(idx, _input, in_slice); |
| 110 | add_2D_tensor_argument(idx, _output, out_slice); |
| 111 | enqueue(queue, *this, in_slice, _lws_hint); |
| 112 | } |
| 113 | while(window.slide_window_slice_2D(in_slice) && out_window.slide_window_slice_2D(out_slice)); |
| 114 | } |