Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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/NEGEMMMatrixVectorMultiplyKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/NEON/INEKernel.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/Window.h" |
| 34 | |
| 35 | #include <arm_neon.h> |
| 36 | #include <cstddef> |
| 37 | #include <cstdint> |
| 38 | #include <tuple> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | NEGEMMMatrixVectorMultiplyKernel::NEGEMMMatrixVectorMultiplyKernel() |
| 43 | : _input0(nullptr), _input1(nullptr), _output(nullptr) |
| 44 | { |
| 45 | } |
| 46 | |
| 47 | void NEGEMMMatrixVectorMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output) |
| 48 | { |
| 49 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32); |
| 50 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); |
| 51 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output); |
| 52 | ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1)); |
| 53 | |
| 54 | _input0 = input0; |
| 55 | _input1 = input1; |
| 56 | _output = output; |
| 57 | |
| 58 | // Configure kernel window |
| 59 | const unsigned int num_elems_read_per_iteration = 4; |
| 60 | |
| 61 | Window win = calculate_max_window(*input0->info(), Steps(num_elems_read_per_iteration)); |
| 62 | |
| 63 | AccessWindowHorizontal input0_access(input0->info(), 0, num_elems_read_per_iteration); |
| 64 | AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_read_per_iteration); |
Michalis Spyrou | 4db041e | 2017-11-29 11:56:49 +0000 | [diff] [blame] | 65 | AccessWindowStatic output_access(output->info(), 0, 0, output->info()->dimension(0), output->info()->dimension(1)); |
Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 66 | |
| 67 | update_window_and_padding(win, input0_access, input1_access, output_access); |
| 68 | |
| 69 | _output->info()->set_valid_region(ValidRegion(Coordinates(), _output->info()->tensor_shape())); |
| 70 | |
| 71 | INEKernel::configure(win); |
| 72 | } |
| 73 | |
| 74 | void NEGEMMMatrixVectorMultiplyKernel::run(const Window &window, const ThreadInfo &info) |
| 75 | { |
| 76 | ARM_COMPUTE_UNUSED(info); |
| 77 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 78 | |
| 79 | Window window_slice = window.first_slice_window_3D(); |
| 80 | |
| 81 | Window window_in(window); |
| 82 | Window window_weights(window_slice); |
| 83 | Window window_out(window); |
| 84 | |
| 85 | // Setup input0 slice |
| 86 | window_in.set(Window::DimX, Window::Dimension(0, _input0->info()->dimension(0), _input0->info()->dimension(0))); |
| 87 | window_in.set(Window::DimY, Window::Dimension(0, _input0->info()->dimension(1), 1)); |
| 88 | window_in.set(Window::DimZ, Window::Dimension(0, _input0->info()->dimension(2), 1)); |
| 89 | |
| 90 | // Setup input1 and output slice. Their dimensions are increased in the kernel. |
| 91 | window_weights.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 92 | window_weights.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 93 | window_weights.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 94 | |
| 95 | window_out.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 96 | window_out.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 97 | window_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 98 | |
| 99 | Iterator in(_input0, window_in); |
| 100 | Iterator in2(_input1, window_weights); |
| 101 | Iterator out(_output, window_out); |
| 102 | |
| 103 | const int input_w = _input0->info()->dimension(0); |
| 104 | const int input_h = _input0->info()->dimension(1); |
| 105 | const int input_stride_x = _input0->info()->strides_in_bytes().x(); |
| 106 | const int weights_stride_x = _input1->info()->strides_in_bytes().x(); |
| 107 | const int weights_stride_y = _input1->info()->strides_in_bytes().y(); |
| 108 | const int output_stride_x = _output->info()->strides_in_bytes().x(); |
| 109 | |
| 110 | execute_window_loop(window_in, [&](const Coordinates & id) |
| 111 | { |
| 112 | // Get pointers |
| 113 | const uint8_t *const input_ptr = in.ptr(); |
| 114 | const uint8_t *const weights_ptr = in2.ptr() + id.z() * weights_stride_y; |
| 115 | auto output_ptr = reinterpret_cast<float *>(out.ptr() + (id.y() + id.z() * input_h) * output_stride_x); |
| 116 | |
| 117 | float32x4_t row_dot = vdupq_n_f32(0.f); |
| 118 | for(int i = 0; i < input_w; i += 4) |
| 119 | { |
| 120 | const auto input = vld1q_f32(reinterpret_cast<const float *>(input_ptr + i * input_stride_x)); |
| 121 | const auto weights = vld1q_f32(reinterpret_cast<const float *>(weights_ptr + i * weights_stride_x)); |
| 122 | row_dot = vaddq_f32(row_dot, vmulq_f32(input, weights)); |
| 123 | } |
| 124 | |
| 125 | auto temp = vadd_f32(vget_high_f32(row_dot), vget_low_f32(row_dot)); |
| 126 | temp = vpadd_f32(temp, temp); |
| 127 | |
| 128 | *output_ptr = vget_lane_f32(temp, 0); |
| 129 | }, |
| 130 | in, in2, out); |
| 131 | } |