Moritz Pflanzer | 80373f6 | 2017-09-15 10:42:58 +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/NEON/kernels/arm32/NEGEMMAArch32Kernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/AccessWindowTranspose.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/IAccessWindow.h" |
| 31 | #include "arm_compute/core/ITensor.h" |
| 32 | #include "arm_compute/core/NEON/NEFixedPoint.h" |
| 33 | #include "arm_compute/core/TensorInfo.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/core/Utils.h" |
| 36 | #include "arm_compute/core/Validate.h" |
| 37 | #include "arm_compute/core/Window.h" |
| 38 | #include "support/ToolchainSupport.h" |
| 39 | |
| 40 | namespace arm_compute |
| 41 | { |
| 42 | #include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" |
| 43 | #include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp" |
| 44 | } // namespace arm_compute |
| 45 | |
| 46 | #include <arm_neon.h> |
| 47 | #include <cstddef> |
| 48 | #include <cstdint> |
| 49 | #include <tuple> |
| 50 | |
| 51 | namespace arm_compute |
| 52 | { |
| 53 | void NEGEMMAArch32Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1) |
| 54 | { |
| 55 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32); |
| 56 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); |
| 57 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); |
| 58 | |
| 59 | _input0 = input0; |
| 60 | _input1 = input1; |
| 61 | _output = output; |
| 62 | _workspace = workspace; |
| 63 | _alpha = alpha; |
| 64 | _beta = beta; |
| 65 | _transform_0 = transform_0; |
| 66 | _transform_1 = transform_1; |
| 67 | |
| 68 | // Configure kernel window |
| 69 | Window win = calculate_max_window(*output->info()); |
| 70 | |
| 71 | AccessWindowRectangle output_access(output->info(), 0, 0, 8, 6); |
| 72 | |
| 73 | const int input0_access_end = ceil_to_multiple(input0->info()->tensor_shape().x(), 6); |
| 74 | const int input1_access_end = ceil_to_multiple(input1->info()->tensor_shape().x(), 8); |
| 75 | |
| 76 | update_window_and_padding(win, |
| 77 | AccessWindowStatic(input0->info(), 0, 0, input0_access_end, input0->info()->tensor_shape().y()), |
| 78 | AccessWindowStatic(input1->info(), 0, 0, input1_access_end, input1->info()->tensor_shape().y()), |
| 79 | output_access); |
| 80 | |
| 81 | INEKernel::configure(win); |
| 82 | } |
| 83 | |
| 84 | void NEGEMMAArch32Kernel::run(const Window &window, const ThreadInfo &info) |
| 85 | { |
| 86 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 87 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 88 | |
| 89 | const int lda = _input0->info()->strides_in_bytes().y() / sizeof(float); |
| 90 | const int ldb = _input1->info()->strides_in_bytes().y() / sizeof(float); |
| 91 | const int ldc = _output->info()->strides_in_bytes().y() / sizeof(float); |
| 92 | |
| 93 | const auto in1_ptr = reinterpret_cast<const float *>(_input1->buffer()); |
| 94 | |
| 95 | const int M = std::min(_output->info()->tensor_shape().y(), static_cast<size_t>(window.y().end())) - window.y().start(); |
| 96 | const int N = _output->info()->tensor_shape().x(); |
| 97 | const int K = _input0->info()->tensor_shape().x(); |
| 98 | |
| 99 | // Only iterate over batches |
| 100 | Window win(window); |
| 101 | win.set(0, Window::Dimension(0, 1, 1)); |
| 102 | win.set(1, Window::Dimension(0, 1, 1)); |
| 103 | |
| 104 | Iterator in0(_input0, window); |
| 105 | Iterator out(_output, window); |
| 106 | |
| 107 | GemmInterleaved<sgemm_8x6, float, float> gemm(&info.cpu_info, M, N, K, !_transform_0, !_transform_1); |
| 108 | constexpr size_t alignment = 4096; |
| 109 | const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id; |
| 110 | void *workspace = _workspace->buffer() + offset; |
| 111 | size_t workspace_size = _workspace->info()->total_size(); |
| 112 | |
| 113 | if(support::cpp11::align(alignment, gemm.get_working_size(), workspace, workspace_size) == nullptr) |
| 114 | { |
| 115 | ARM_COMPUTE_ERROR("Not enough space to align buffer!"); |
| 116 | } |
| 117 | |
| 118 | execute_window_loop(win, [&](const Coordinates & id) |
| 119 | { |
| 120 | gemm.execute(reinterpret_cast<const float *>(in0.ptr()), lda, |
| 121 | reinterpret_cast<const float *>(in1_ptr), ldb, |
| 122 | reinterpret_cast<float *>(out.ptr()), ldc, |
| 123 | _alpha, _beta, workspace); |
| 124 | }, |
| 125 | in0, out); |
| 126 | } |
| 127 | } // namespace arm_compute |