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/CLGEMMInterleave4x4Kernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.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" |
| 36 | |
| 37 | using namespace arm_compute; |
| 38 | |
| 39 | CLGEMMInterleave4x4Kernel::CLGEMMInterleave4x4Kernel() |
| 40 | : _input(nullptr), _output(nullptr) |
| 41 | { |
| 42 | } |
| 43 | |
| 44 | void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *output) |
| 45 | { |
Gian Marco Iodice | 8a38369 | 2017-07-03 17:41:47 +0100 | [diff] [blame^] | 46 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16, |
| 47 | DataType::F32); |
Gian Marco Iodice | 9f89bae | 2017-06-22 12:09:49 +0100 | [diff] [blame] | 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); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 58 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
Gian Marco Iodice | 9f89bae | 2017-06-22 12:09:49 +0100 | [diff] [blame] | 59 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 60 | |
| 61 | _input = input; |
| 62 | _output = output; |
| 63 | |
| 64 | // Create kernel |
| 65 | std::string data_type_name; |
Moritz Pflanzer | 05da6dd | 2017-07-04 12:08:41 +0100 | [diff] [blame] | 66 | data_type_name = support::cpp11::to_string(input->info()->element_size() * 8) + "bit"; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 67 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_interleave4x4_" + data_type_name)); |
| 68 | |
| 69 | // Configure kernel window |
| 70 | const unsigned int num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(input->info()->data_type()); |
| 71 | constexpr unsigned int num_elems_processed_per_iteration_y = 4; |
| 72 | const unsigned int num_elems_written_per_iteration = num_elems_processed_per_iteration_x * num_elems_processed_per_iteration_y; |
| 73 | |
| 74 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 75 | |
| 76 | AccessWindowRectangle input_access(input->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 77 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration, 1, 4.f, 0.25f); |
| 78 | |
| 79 | update_window_and_padding(win, input_access, output_access); |
| 80 | |
| 81 | output_access.set_valid_region(win, input->info()->valid_region()); |
| 82 | |
| 83 | ICLKernel::configure(win); |
| 84 | } |
| 85 | |
| 86 | void CLGEMMInterleave4x4Kernel::run(const Window &window, cl::CommandQueue &queue) |
| 87 | { |
| 88 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 89 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 90 | |
| 91 | /* |
| 92 | * This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values) |
| 93 | * |a00 a01 a02 a03| |
| 94 | * |a10 a11 a12 a13| |
| 95 | * |a20 a21 a22 a23| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 | |
| 96 | * |a30 a31 a32 a33| |
| 97 | * |
| 98 | * After this operation, the output matrix will have the following shape: [ height * 4, width / 4 ] |
| 99 | */ |
| 100 | Window in_slice = window.first_slice_window_2D(); |
| 101 | Window out_slice = window.first_slice_window_2D(); |
| 102 | |
| 103 | // Change x and y steps for the slide of output tensor |
| 104 | out_slice.scale(Window::DimX, 4.f); |
| 105 | out_slice.scale(Window::DimY, 0.25f); |
| 106 | |
| 107 | do |
| 108 | { |
| 109 | unsigned int idx = 0; |
| 110 | add_2D_tensor_argument(idx, _input, in_slice); |
| 111 | add_2D_tensor_argument(idx, _output, out_slice); |
| 112 | enqueue(queue, *this, in_slice); |
| 113 | } |
| 114 | while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice)); |
| 115 | } |