alerah01 | c9e519d | 2022-01-31 19:04:10 +0200 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018-2022 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 | |
| 25 | #include "src/cpu/kernels/directconv2d/nhwc/neon/impl.h" |
| 26 | |
| 27 | #include "src/core/NEON/kernels/detail/NEDirectConvolutionDetail.h" |
| 28 | #include "src/core/NEON/wrapper/wrapper.h" |
| 29 | |
| 30 | #include "arm_compute/core/Error.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/IAccessWindow.h" |
| 33 | #include "arm_compute/core/ITensor.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/core/Utils.h" |
| 36 | #include "src/core/helpers/WindowHelpers.h" |
| 37 | |
| 38 | #include <algorithm> |
| 39 | |
| 40 | using namespace arm_compute::detail; |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | namespace cpu |
| 45 | { |
| 46 | namespace kernels |
| 47 | { |
| 48 | namespace |
| 49 | { |
| 50 | bool have_zero_x_internal_padding(ITensorInfo *src, const ITensorInfo *weights) |
| 51 | { |
| 52 | return (src->padding().left == 0 && weights->padding().left == 0 && src->padding().right == 0 && weights->padding().right == 0); |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | template <typename T> |
| 57 | void convolve_nhwc(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info) |
| 58 | { |
| 59 | // Declare useful types |
| 60 | using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>; |
| 61 | using vector_type = typename vtype::type; |
| 62 | using tag_type = typename vtype::tag_type; |
| 63 | |
| 64 | // Scalar quantities |
| 65 | const int element_size = src->info()->element_size(); |
| 66 | const int input_stride_w = src->info()->strides_in_bytes().y() / element_size; |
| 67 | const int input_stride_h = src->info()->strides_in_bytes().z() / element_size; |
| 68 | const int input_stride_n = src->info()->strides_in_bytes()[3] / element_size; |
| 69 | const int input_dim_w = src->info()->dimension(1); |
| 70 | const int input_dim_h = src->info()->dimension(2); |
| 71 | |
| 72 | const int output_stride_c = dst->info()->strides_in_bytes().x(); |
| 73 | |
| 74 | const unsigned int kernel_stride_w = weights->info()->strides_in_bytes().y() / element_size; |
| 75 | const unsigned int kernel_stride_h = weights->info()->strides_in_bytes().z() / element_size; |
| 76 | const int kernel_dim_w = weights->info()->dimension(1); |
| 77 | const int kernel_dim_h = weights->info()->dimension(2); |
| 78 | |
| 79 | const int conv_pad_top = conv_info.pad_top(); |
| 80 | const int conv_pad_left = conv_info.pad_left(); |
| 81 | const int conv_stride_w = std::get<0>(conv_info.stride()); |
| 82 | const int conv_stride_h = std::get<1>(conv_info.stride()); |
| 83 | |
| 84 | // Setup input window for the output iterator |
| 85 | Window window_out = window; |
| 86 | window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 87 | |
| 88 | // Setup input window for the weights iterator |
| 89 | Window window_w = calculate_max_window(*weights->info(), Steps()); |
| 90 | window_w.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 91 | window_w.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 92 | window_w.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 93 | |
| 94 | Iterator out(dst, window_out); |
| 95 | Iterator wei(weights, window_w); |
| 96 | |
| 97 | constexpr int num_elems_read_per_iteration = 16 / sizeof(T); |
| 98 | |
| 99 | // nhwc optimized |
| 100 | if(have_zero_x_internal_padding(src->info(), weights->info())) |
| 101 | { |
| 102 | // This function assumes that input and weights have not padding in channel |
| 103 | |
| 104 | /* |
| 105 | * This implementation parallelize the full WC plane of input and weights by |
| 106 | * treating them as series of elements. So for example, a 3x3 weights and |
| 107 | * floating point vector operations of 4 elements per time, the first 3 |
| 108 | * channel elements of the first row would be taken and additionally the first |
| 109 | * element of the second row. The 9 elements in each single WC weight plane |
| 110 | * would require 2 4-element vector operations and a last single element operation. |
| 111 | * |
| 112 | * This works since when we create the input vector to multiply with the weights, |
| 113 | * the exact required elements are loaded in the same order. Therefore the |
| 114 | * multiplication works on the correct input/weight elements. |
| 115 | */ |
| 116 | execute_window_loop( |
| 117 | window_out, [&](const Coordinates & id) |
| 118 | { |
| 119 | /* |
| 120 | * In here we create theoretical indexes which then we validate for both |
| 121 | * inputs and weights. |
| 122 | * As a reminder, this loop take each output point in NHW, C is treated |
| 123 | * in the weights loop. |
| 124 | */ |
| 125 | // We are computing the theoretical starting input starting points |
| 126 | const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; |
| 127 | const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; |
| 128 | const int in_w_end_t = in_w_start_t + kernel_dim_w; |
| 129 | const int in_h_end_t = in_h_start_t + kernel_dim_h; |
| 130 | |
| 131 | // We are computing the valid initial and ending input points by checking the borders |
| 132 | const int in_w_start = std::max(in_w_start_t, 0); |
| 133 | const int in_h_start = std::max(in_h_start_t, 0); |
| 134 | const int in_w_end = std::min(in_w_end_t, input_dim_w); |
| 135 | const int in_h_end = std::min(in_h_end_t, input_dim_h); |
| 136 | |
| 137 | // We use the input points to select the valid weight points to use |
| 138 | const int index_wc_start = (in_w_start - in_w_start_t) * kernel_stride_w; |
| 139 | const int index_h_start = in_h_start - in_h_start_t; |
| 140 | const int index_wc_end = (kernel_dim_w - (in_w_end_t - in_w_end)) * kernel_stride_w; |
| 141 | const int index_h_end = kernel_dim_h - (in_h_end_t - in_h_end); |
| 142 | |
| 143 | execute_window_loop( |
| 144 | window_w, [&](const Coordinates & id_w) |
| 145 | { |
| 146 | /* |
| 147 | * This is the loop in the weights, and it goes along N (the batches) |
| 148 | * As a reminder, the batches of the weights are translated into the |
| 149 | * channels of the output |
| 150 | */ |
| 151 | const T *in_ptr_row = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) |
| 152 | + id[3] * input_stride_n + in_w_start * input_stride_w + in_h_start * input_stride_h; |
| 153 | const T *weights_ptr_row = reinterpret_cast<const T *>(wei.ptr()) + index_h_start * kernel_stride_h; |
| 154 | uint8_t *out_ptr = out.ptr() + id_w[3] * output_stride_c; |
| 155 | |
| 156 | T out_temp = static_cast<T>(0); |
| 157 | for(int index_h = index_h_start; index_h < index_h_end; ++index_h, in_ptr_row += input_stride_h, weights_ptr_row += kernel_stride_h) |
| 158 | { |
| 159 | const T *in_ptr_mover = in_ptr_row; |
| 160 | int index_wc = index_wc_start; |
| 161 | vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); |
| 162 | for(; index_wc <= index_wc_end - num_elems_read_per_iteration; index_wc += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration) |
| 163 | { |
| 164 | const auto src_vec = wrapper::vloadq(in_ptr_mover); |
| 165 | const auto w_vec = wrapper::vloadq(weights_ptr_row + index_wc); |
| 166 | out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec); |
| 167 | } |
| 168 | out_temp += vreduce(out_temp_vec); |
| 169 | for(; index_wc < index_wc_end; ++index_wc, ++in_ptr_mover) |
| 170 | { |
| 171 | const auto src_val = *(in_ptr_mover); |
| 172 | const auto w_val = *(weights_ptr_row + index_wc); |
| 173 | out_temp += src_val * w_val; |
| 174 | } |
| 175 | } |
| 176 | *(reinterpret_cast<T *>(out_ptr)) = out_temp; |
| 177 | }, |
| 178 | wei); |
| 179 | }, |
| 180 | out); |
| 181 | } |
| 182 | else // nhwc non optimized |
| 183 | { |
| 184 | execute_window_loop( |
| 185 | window_out, [&](const Coordinates & id) |
| 186 | { |
| 187 | // We are computing the theoretical starting input starting points |
| 188 | const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; |
| 189 | const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; |
| 190 | const int in_w_end_t = in_w_start_t + kernel_dim_w; |
| 191 | const int in_h_end_t = in_h_start_t + kernel_dim_h; |
| 192 | |
| 193 | // We are computing the valid initial and ending input points by checking the borders |
| 194 | const int in_w_start = std::max(in_w_start_t, 0); |
| 195 | const int in_h_start = std::max(in_h_start_t, 0); |
| 196 | const int in_w_end = std::min(in_w_end_t, input_dim_w); |
| 197 | const int in_h_end = std::min(in_h_end_t, input_dim_h); |
| 198 | |
| 199 | // We use the input points to select the valid weight points to use |
| 200 | const int wei_w_start = in_w_start - in_w_start_t; |
| 201 | const int wei_h_start = in_h_start - in_h_start_t; |
| 202 | const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end); |
| 203 | const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end); |
| 204 | |
| 205 | const int index_c_end = weights->info()->dimension(0); |
| 206 | const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[3] * input_stride_n; |
| 207 | |
| 208 | execute_window_loop( |
| 209 | window_w, [&](const Coordinates & id_w) |
| 210 | { |
| 211 | const T *const weights_ptr_start = reinterpret_cast<const T *>(wei.ptr()); |
| 212 | uint8_t *out_ptr = out.ptr() + id_w[3] * output_stride_c; |
| 213 | |
| 214 | T out_temp = static_cast<T>(0); |
| 215 | for(int index_wei_h = wei_h_start, index_in_h = in_h_start; index_wei_h < wei_h_end; ++index_wei_h, ++index_in_h) |
| 216 | { |
| 217 | const T *const in_ptr_row = in_ptr_start + index_in_h * input_stride_h; |
| 218 | const T *const weights_ptr_row = weights_ptr_start + index_wei_h * kernel_stride_h; |
| 219 | for(int index_wei_w = wei_w_start, index_in_w = in_w_start; index_wei_w < wei_w_end; ++index_wei_w, ++index_in_w) |
| 220 | { |
| 221 | const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w; |
| 222 | const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w; |
| 223 | int index_c = 0; |
| 224 | vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); |
| 225 | for(; index_c <= index_c_end - num_elems_read_per_iteration; index_c += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration, weights_ptr_mover += num_elems_read_per_iteration) |
| 226 | { |
| 227 | const auto src_vec = wrapper::vloadq(in_ptr_mover); |
| 228 | const auto w_vec = wrapper::vloadq(weights_ptr_mover); |
| 229 | out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec); |
| 230 | } |
| 231 | out_temp += vreduce(out_temp_vec); |
| 232 | for(; index_c < index_c_end; ++index_c, ++in_ptr_mover, ++weights_ptr_mover) |
| 233 | { |
| 234 | const auto src_val = *(in_ptr_mover); |
| 235 | const auto w_val = *(weights_ptr_mover); |
| 236 | out_temp += src_val * w_val; |
| 237 | } |
| 238 | } |
| 239 | } |
| 240 | *(reinterpret_cast<T *>(out_ptr)) = out_temp; |
| 241 | }, |
| 242 | wei); |
| 243 | }, |
| 244 | out); |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | template void convolve_nhwc<float>(const Window &window, const ITensor *src, const ITensor *weights, ITensor *dst, const PadStrideInfo &conv_info); |
| 249 | |
| 250 | } // namespace kernels |
| 251 | } // namespace cpu |
| 252 | } // namespace arm_compute |