Sheri Zhang | 6d9c982 | 2021-09-24 16:02:57 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 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 "src/cpu/kernels/CpuDirectConv3dKernel.h" |
| 25 | |
| 26 | #include "src/core/NEON/kernels/detail/NEDirectConvolutionDetail.h" |
| 27 | #include "src/core/NEON/wrapper/wrapper.h" |
| 28 | |
| 29 | #include "arm_compute/core/Error.h" |
| 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/IAccessWindow.h" |
| 32 | #include "arm_compute/core/ITensor.h" |
| 33 | #include "arm_compute/core/Types.h" |
| 34 | #include "arm_compute/core/Utils.h" |
| 35 | #include "arm_compute/core/Validate.h" |
| 36 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 37 | #include "src/core/CPP/Validate.h" |
| 38 | #include "src/core/helpers/AutoConfiguration.h" |
| 39 | #include "src/core/helpers/WindowHelpers.h" |
| 40 | |
| 41 | #include <algorithm> |
| 42 | |
| 43 | using namespace arm_compute::detail; |
| 44 | |
| 45 | namespace arm_compute |
| 46 | { |
| 47 | namespace cpu |
| 48 | { |
| 49 | namespace kernels |
| 50 | { |
| 51 | namespace |
| 52 | { |
| 53 | Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv3dInfo &conv_info) |
| 54 | { |
| 55 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); |
| 56 | ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() != DataLayout::NDHWC); |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); |
| 58 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); |
| 59 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); |
| 60 | |
| 61 | const DataLayout data_layout = src->data_layout(); |
| 62 | const int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); |
| 63 | |
| 64 | // Weight layout is D, H, W, Cin, Cout |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 5); |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != src->dimension(channel_idx)); |
| 67 | |
| 68 | if(biases != nullptr) |
| 69 | { |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 71 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(0) != weights->dimension(0), |
| 72 | "biases size and number of output feature maps should match"); |
| 73 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1, "biases should be one dimensional"); |
| 74 | } |
| 75 | |
| 76 | // Checks performed when output is configured |
| 77 | if(dst->total_size() != 0) |
| 78 | { |
| 79 | TensorShape output_shape = misc::shape_calculator::compute_conv3d_shape(src->tensor_shape(), weights->tensor_shape(), conv_info); |
| 80 | |
| 81 | DataType data_type = src->data_type(); |
| 82 | |
| 83 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), output_shape); |
| 84 | ARM_COMPUTE_RETURN_ERROR_ON(dst->data_type() != data_type); |
| 85 | } |
| 86 | |
| 87 | return Status{}; |
| 88 | } |
| 89 | |
| 90 | /** Reduce a vector to be a scalar by accumulating all lanes in the vector |
| 91 | * |
| 92 | * @param[in] v Vector to be reduced. |
| 93 | * |
| 94 | * @return the wrapped-around number. |
| 95 | */ |
| 96 | auto vreduce(const float32x4_t &v) |
| 97 | { |
| 98 | auto v0 = wrapper::vgethigh(v); |
| 99 | auto v1 = wrapper::vgetlow(v); |
| 100 | auto v_out = wrapper::vadd(v0, v1); |
| 101 | |
| 102 | float a = wrapper::vgetlane(v_out, 0); |
| 103 | float b = wrapper::vgetlane(v_out, 1); |
| 104 | return a + b; |
| 105 | } |
| 106 | |
| 107 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 108 | auto vreduce(const float16x8_t &v) |
| 109 | { |
| 110 | auto v0 = wrapper::vgethigh(v); |
| 111 | auto v1 = wrapper::vgetlow(v); |
| 112 | auto v_out = wrapper::vadd(v0, v1); |
| 113 | |
| 114 | float16_t a = wrapper::vgetlane(v_out, 0); |
| 115 | float16_t b = wrapper::vgetlane(v_out, 1); |
| 116 | float16_t c = wrapper::vgetlane(v_out, 2); |
| 117 | float16_t d = wrapper::vgetlane(v_out, 3); |
| 118 | return a + b + c + d; |
| 119 | } |
| 120 | #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 121 | } |
| 122 | |
| 123 | template <typename T> |
| 124 | void CpuDirectConv3dKernel::convolve_ndhwc(const Window &window, const ITensor *src, const ITensor *weights, const ITensor *biases, ITensor *dst) |
| 125 | { |
| 126 | using vtype = wrapper::traits::neon_bitvector<T, wrapper::traits::BitWidth::W128>; |
| 127 | using vector_type = typename vtype::type; |
| 128 | using tag_type = typename vtype::tag_type; |
| 129 | constexpr int num_elems_read_per_iteration = 16 / sizeof(T); |
| 130 | |
| 131 | // Scalar quantities (N D H W Cin) |
| 132 | const int element_size = src->info()->element_size(); |
| 133 | const int input_stride_w = src->info()->strides_in_bytes().y() / element_size; |
| 134 | const int input_stride_h = src->info()->strides_in_bytes().z() / element_size; |
| 135 | const int input_stride_d = src->info()->strides_in_bytes()[3] / element_size; |
| 136 | const int input_stride_n = src->info()->strides_in_bytes()[4] / element_size; |
| 137 | const int input_dim_w = src->info()->dimension(1); |
| 138 | const int input_dim_h = src->info()->dimension(2); |
| 139 | const int input_dim_d = src->info()->dimension(3); |
| 140 | |
| 141 | // Kernel info (D H W Cin Cout) |
| 142 | const unsigned int kernel_stride_w = weights->info()->strides_in_bytes()[2] / element_size; |
| 143 | const unsigned int kernel_stride_h = weights->info()->strides_in_bytes()[3] / element_size; |
| 144 | const unsigned int kernel_stride_d = weights->info()->strides_in_bytes()[4] / element_size; |
| 145 | const int kernel_dim_w = weights->info()->dimension(2); |
| 146 | const int kernel_dim_h = weights->info()->dimension(3); |
| 147 | const int kernel_dim_d = weights->info()->dimension(4); |
| 148 | |
| 149 | // Convolution padding and stride |
| 150 | const int conv_pad_top = _conv_info.padding.top; |
| 151 | const int conv_pad_left = _conv_info.padding.left; |
| 152 | const int conv_pad_front = _conv_info.padding.front; |
| 153 | const int conv_stride_w = _conv_info.stride.width; |
| 154 | const int conv_stride_h = _conv_info.stride.height; |
| 155 | const int conv_stride_d = _conv_info.stride.depth; |
| 156 | |
| 157 | // Setup input window for the output iterator |
| 158 | Window window_out = window; |
| 159 | window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 160 | |
| 161 | // Setup input window for the weights iterator |
| 162 | Window window_w = calculate_max_window(*weights->info(), Steps()); |
| 163 | window_w.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 164 | window_w.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 165 | window_w.set(Window::DimW, Window::Dimension(0, 1, 1)); |
| 166 | window_w.set(4, Window::Dimension(0, 1, 1)); |
| 167 | |
| 168 | Iterator out(dst, window_out); |
| 169 | Iterator wei(weights, window_w); |
| 170 | |
| 171 | const T *biases_ptr = nullptr; |
| 172 | if(biases) |
| 173 | { |
| 174 | biases_ptr = reinterpret_cast<T *>(biases->buffer() + biases->info()->offset_first_element_in_bytes()); |
| 175 | } |
| 176 | execute_window_loop(window_out, [&](const Coordinates & id) |
| 177 | { |
| 178 | // We are computing the theoretical input starting points |
| 179 | const int in_w_start_t = static_cast<int>(id.y()) * conv_stride_w - conv_pad_left; |
| 180 | const int in_h_start_t = static_cast<int>(id.z()) * conv_stride_h - conv_pad_top; |
| 181 | const int in_d_start_t = static_cast<int>(id[3]) * conv_stride_d - conv_pad_front; |
| 182 | const int in_w_end_t = in_w_start_t + kernel_dim_w; |
| 183 | const int in_h_end_t = in_h_start_t + kernel_dim_h; |
| 184 | const int in_d_end_t = in_d_start_t + kernel_dim_d; |
| 185 | |
| 186 | // We are computing the valid initial and ending input points by checking the borders |
| 187 | const int in_w_start = std::max(in_w_start_t, 0); |
| 188 | const int in_h_start = std::max(in_h_start_t, 0); |
| 189 | const int in_d_start = std::max(in_d_start_t, 0); |
| 190 | const int in_w_end = std::min(in_w_end_t, input_dim_w); |
| 191 | const int in_h_end = std::min(in_h_end_t, input_dim_h); |
| 192 | const int in_d_end = std::min(in_d_end_t, input_dim_d); |
| 193 | |
| 194 | // We use the input points to select the valid weight points to use |
| 195 | const int wei_w_start = in_w_start - in_w_start_t; |
| 196 | const int wei_h_start = in_h_start - in_h_start_t; |
| 197 | const int wei_d_start = in_d_start - in_d_start_t; |
| 198 | const int wei_w_end = kernel_dim_w - (in_w_end_t - in_w_end); |
| 199 | const int wei_h_end = kernel_dim_h - (in_h_end_t - in_h_end); |
| 200 | const int wei_d_end = kernel_dim_d - (in_d_end_t - in_d_end); |
| 201 | |
| 202 | const int index_c_out_end = weights->info()->dimension(0); |
| 203 | const int index_c_in_end = weights->info()->dimension(1); |
| 204 | const T *const in_ptr_start = reinterpret_cast<const T *>(src->buffer() + src->info()->offset_first_element_in_bytes()) + id[4] * input_stride_n; |
| 205 | |
| 206 | execute_window_loop(window_w, [&](const Coordinates & id_w) |
| 207 | { |
| 208 | /* |
| 209 | * This is the loop in the weights, and it goes along OFM (output feature map) |
| 210 | */ |
| 211 | const auto weights_ptr_start = reinterpret_cast<const T *>(wei.ptr()); |
| 212 | T out_temp = static_cast<T>(0); |
| 213 | T *out_ptr = reinterpret_cast<T *>(out.ptr()); |
| 214 | for(int index_wei_d = wei_d_start, index_in_d = in_d_start; index_wei_d < wei_d_end; ++index_wei_d, ++index_in_d) |
| 215 | { |
| 216 | const auto in_ptr_d = in_ptr_start + index_in_d * input_stride_d; |
| 217 | const auto weights_ptr_d = weights_ptr_start + index_wei_d * kernel_stride_d; |
| 218 | 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) |
| 219 | { |
| 220 | const T *const in_ptr_row = in_ptr_d + index_in_h * input_stride_h; |
| 221 | const T *const weights_ptr_row = weights_ptr_d + index_wei_h * kernel_stride_h; |
| 222 | 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) |
| 223 | { |
| 224 | const T *in_ptr_mover = in_ptr_row + index_in_w * input_stride_w; |
| 225 | const T *weights_ptr_mover = weights_ptr_row + index_wei_w * kernel_stride_w; |
| 226 | int index_c_in = 0; |
| 227 | vector_type out_temp_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); |
| 228 | vector_type w_vec = wrapper::vdup_n(static_cast<T>(0), tag_type()); |
| 229 | for(; index_c_in <= index_c_in_end - num_elems_read_per_iteration; |
| 230 | index_c_in += num_elems_read_per_iteration, in_ptr_mover += num_elems_read_per_iteration) |
| 231 | { |
| 232 | const auto src_vec = wrapper::vloadq(in_ptr_mover); |
| 233 | //Load Cin weights |
| 234 | for(unsigned int k = 0; k < num_elems_read_per_iteration; ++k, weights_ptr_mover += index_c_out_end) |
| 235 | { |
| 236 | w_vec = wrapper::vsetlane(*weights_ptr_mover, w_vec, k); |
| 237 | } |
| 238 | out_temp_vec = wrapper::vmla(out_temp_vec, w_vec, src_vec); |
| 239 | } |
| 240 | out_temp += vreduce(out_temp_vec); |
| 241 | for(; index_c_in < index_c_in_end; ++index_c_in, ++in_ptr_mover, weights_ptr_mover += index_c_out_end) |
| 242 | { |
| 243 | const auto src_val = *(in_ptr_mover); |
| 244 | const auto w_val = *(weights_ptr_mover); |
| 245 | out_temp += src_val * w_val; |
| 246 | } |
| 247 | } |
| 248 | } |
| 249 | } |
| 250 | *(reinterpret_cast<T *>(out_ptr + id_w[0])) = (biases) ? out_temp + biases_ptr[id_w[0]] : out_temp; |
| 251 | }, |
| 252 | wei); |
| 253 | }, |
| 254 | out); |
| 255 | } |
| 256 | |
| 257 | void CpuDirectConv3dKernel::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const Conv3dInfo &conv_info) |
| 258 | { |
| 259 | ARM_COMPUTE_UNUSED(biases); |
| 260 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); |
| 261 | |
| 262 | _conv_info = conv_info; |
| 263 | |
| 264 | // Get convolved dimensions |
| 265 | TensorShape output_shape = misc::shape_calculator::compute_conv3d_shape(src->tensor_shape(), weights->tensor_shape(), conv_info); |
| 266 | |
| 267 | DataType data_type = src->data_type(); |
| 268 | |
| 269 | // Output auto inizialitation if not yet initialized |
| 270 | auto_init_if_empty(*dst, output_shape, 1, data_type); |
| 271 | |
| 272 | // Perform validation step |
| 273 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, dst, conv_info)); |
| 274 | |
| 275 | // Configure kernel window |
| 276 | Window win = calculate_max_window(*dst, Steps()); |
| 277 | ICpuKernel::configure(win); |
| 278 | } |
| 279 | |
| 280 | Status CpuDirectConv3dKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv3dInfo &conv_info) |
| 281 | { |
| 282 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info)); |
| 283 | |
| 284 | return Status{}; |
| 285 | } |
| 286 | |
| 287 | void CpuDirectConv3dKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) |
| 288 | { |
| 289 | ARM_COMPUTE_UNUSED(info); |
| 290 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 291 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); |
| 292 | |
| 293 | auto src = tensors.get_const_tensor(TensorType::ACL_SRC_0); |
| 294 | auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1); |
| 295 | auto biases = tensors.get_const_tensor(TensorType::ACL_SRC_2); |
| 296 | auto dst = tensors.get_tensor(TensorType::ACL_DST); |
| 297 | |
| 298 | switch(src->info()->data_type()) |
| 299 | { |
| 300 | case DataType::F32: |
| 301 | { |
| 302 | convolve_ndhwc<float>(window, src, weights, biases, dst); |
| 303 | break; |
| 304 | } |
| 305 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 306 | case DataType::F16: |
| 307 | { |
| 308 | convolve_ndhwc<float16_t>(window, src, weights, biases, dst); |
| 309 | break; |
| 310 | } |
| 311 | #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 312 | default: |
| 313 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 314 | break; |
| 315 | } |
| 316 | } |
| 317 | |
| 318 | const char *CpuDirectConv3dKernel::name() const |
| 319 | { |
| 320 | return "CpuDirectConv3dKernel"; |
| 321 | } |
| 322 | } // namespace kernels |
| 323 | } // namespace cpu |
| 324 | } // namespace arm_compute |