Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 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/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.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/GLES_COMPUTE/GCHelpers.h" |
| 30 | #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h" |
| 31 | #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h" |
| 32 | #include "arm_compute/core/GLES_COMPUTE/OpenGLES.h" |
| 33 | #include "arm_compute/core/Helpers.h" |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 34 | #include "arm_compute/core/TensorInfo.h" |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 35 | #include "arm_compute/core/Types.h" |
| 36 | #include "arm_compute/core/Utils.h" |
| 37 | #include "arm_compute/core/Validate.h" |
| 38 | #include "arm_compute/core/Window.h" |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 39 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 40 | |
| 41 | #include <set> |
| 42 | #include <string> |
| 43 | |
| 44 | using namespace arm_compute; |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 45 | using namespace arm_compute::misc::shape_calculator; |
| 46 | |
| 47 | namespace |
| 48 | { |
| 49 | using ElementsProcessed = Steps; |
| 50 | |
| 51 | inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info) |
| 52 | { |
| 53 | ARM_COMPUTE_UNUSED(reshape_info); |
| 54 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 55 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32); |
| 56 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3"); |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 58 | |
| 59 | if(!is_interleaved_transposed) |
| 60 | { |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 61 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1)); |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 62 | |
| 63 | if(output->total_size() != 0) |
| 64 | { |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != output->dimension(0)); |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != output->dimension(1)); |
| 67 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
| 68 | } |
| 69 | } |
| 70 | else |
| 71 | { |
| 72 | const int m = reshape_info.m(); |
| 73 | const int n = reshape_info.n(); |
| 74 | const int k = reshape_info.k(); |
| 75 | const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width(); |
| 76 | const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height(); |
| 77 | |
| 78 | TensorShape tensor_shape0{ input0->tensor_shape() }; |
| 79 | tensor_shape0.set(0, k); |
| 80 | tensor_shape0.set(1, m); |
| 81 | |
| 82 | TensorShape tensor_shape1{ input1->tensor_shape() }; |
| 83 | tensor_shape1.set(0, n); |
| 84 | tensor_shape1.set(1, k); |
| 85 | |
| 86 | const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0); |
| 87 | const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1); |
| 88 | |
| 89 | const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height)); |
| 90 | const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width)); |
| 91 | |
| 92 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0); |
| 93 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1); |
| 94 | |
| 95 | if(output->total_size() != 0) |
| 96 | { |
| 97 | ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != static_cast<size_t>(n)); |
| 98 | ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != static_cast<size_t>(m)); |
| 99 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
| 100 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input0, output); |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | return Status{}; |
| 105 | } |
| 106 | |
| 107 | inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, |
| 108 | bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, |
| 109 | GPUTarget gpu_target, ElementsProcessed &num_elements_processed) |
| 110 | { |
| 111 | ARM_COMPUTE_UNUSED(gpu_target); |
| 112 | |
| 113 | // Output tensor auto inizialitation if not yet initialized |
| 114 | TensorShape tensor_shape{ input0->tensor_shape() }; |
| 115 | tensor_shape.set(0, is_interleaved_transposed ? reshape_info.n() : input1->dimension(0)); |
| 116 | tensor_shape.set(1, is_interleaved_transposed ? reshape_info.m() : input0->dimension(1)); |
| 117 | |
| 118 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(tensor_shape)); |
| 119 | |
| 120 | bool window_changed = false; |
| 121 | Window win{}; |
| 122 | |
| 123 | const DataType data_type = input0->data_type(); |
| 124 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 125 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
| 126 | |
| 127 | if(is_interleaved_transposed) |
| 128 | { |
| 129 | // Configure window kernel |
| 130 | num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(data_type); |
| 131 | num_elems_processed_per_iteration_y = 4; |
| 132 | |
| 133 | win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 134 | |
| 135 | AccessWindowRectangle input0_access(input0, 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f); |
| 136 | AccessWindowTranspose input1_access(input1, 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f); |
| 137 | AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 138 | |
| 139 | update_window_and_padding(win, input0_access, input1_access, output_access); |
| 140 | |
| 141 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); |
| 142 | } |
| 143 | else // The input tensors have not been reshaped |
| 144 | { |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 145 | // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. |
| 146 | num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4); |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 147 | |
| 148 | switch(data_type) |
| 149 | { |
| 150 | case DataType::F16: |
| 151 | num_elems_processed_per_iteration_x = 4; |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 152 | break; |
| 153 | |
| 154 | case DataType::F32: |
| 155 | num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(data_type); |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 156 | break; |
| 157 | |
| 158 | default: |
| 159 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 160 | break; |
| 161 | } |
| 162 | |
| 163 | win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 164 | |
| 165 | AccessWindowStatic input0_access(input0, 0, 0, ceil_to_multiple(input0->dimension(0), 8), ceil_to_multiple(input0->dimension(1), num_elems_processed_per_iteration_y)); |
| 166 | AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), input1->dimension(1)); |
| 167 | AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 168 | |
| 169 | update_window_and_padding(win, input0_access, input1_access, output_access); |
| 170 | |
| 171 | Coordinates coord; |
| 172 | coord.set_num_dimensions(output->num_dimensions()); |
| 173 | output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape())); |
| 174 | } |
| 175 | |
| 176 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 177 | return std::make_pair(err, win); |
| 178 | } |
| 179 | } // namespace |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 180 | |
| 181 | GCGEMMMatrixMultiplyKernel::GCGEMMMatrixMultiplyKernel() |
| 182 | : _input0(nullptr), _input1(nullptr), _output(nullptr) |
| 183 | { |
| 184 | } |
| 185 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 186 | void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 187 | { |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 188 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 189 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 190 | // Perform validate step |
| 191 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info)); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 192 | |
| 193 | _input0 = input0; |
| 194 | _input1 = input1; |
| 195 | _output = output; |
| 196 | |
Michele Di Giorgio | b8fc60f | 2018-04-25 11:58:07 +0100 | [diff] [blame] | 197 | // Get target architecture |
| 198 | GPUTarget gpu_target = get_target(); |
| 199 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 200 | ElementsProcessed num_elements_processed{}; |
| 201 | |
| 202 | // Configure kernel window |
Michele Di Giorgio | b8fc60f | 2018-04-25 11:58:07 +0100 | [diff] [blame] | 203 | auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info, gpu_target, num_elements_processed); |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 204 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 205 | IGCKernel::configure(win_config.second); |
| 206 | |
| 207 | // Create build options |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 208 | std::set<std::string> build_opts; |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 209 | std::string kernel_name; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 210 | |
| 211 | build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)); |
| 212 | build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); |
| 213 | build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); |
| 214 | build_opts.emplace("#define COLS_A " + support::cpp11::to_string(input0->info()->dimension(0))); |
| 215 | build_opts.emplace("#define COLS_B " + support::cpp11::to_string(input1->info()->dimension(0))); |
| 216 | build_opts.emplace("#define ALPHA " + float_to_string_with_full_precision(alpha)); |
| 217 | |
| 218 | // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication |
| 219 | if(is_interleaved_transposed) |
| 220 | { |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 221 | const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width(); |
| 222 | const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height(); |
| 223 | |
| 224 | build_opts.emplace("#define MULT_TRANSPOSE1XW_WIDTH " + support::cpp11::to_string(mult_transpose1xW_width)); |
| 225 | build_opts.emplace("#define MULT_INTERLEAVE4X4_HEIGHT " + support::cpp11::to_string(mult_interleave4x4_height)); |
| 226 | |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 227 | switch(input0->info()->data_type()) |
| 228 | { |
| 229 | case DataType::F16: |
| 230 | build_opts.emplace("#define DATA_TYPE_FP16"); |
| 231 | break; |
| 232 | |
| 233 | case DataType::F32: |
| 234 | build_opts.emplace("#define DATA_TYPE_FP32"); |
| 235 | break; |
| 236 | |
| 237 | default: |
| 238 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 239 | break; |
| 240 | } |
| 241 | |
| 242 | build_opts.emplace("#define GEMM_MM_INTERLEAVED_TRANSPOSED"); |
| 243 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 244 | kernel_name = "gemm_mm_interleaved_transposed"; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 245 | } |
| 246 | else |
| 247 | { |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 248 | // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 249 | |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 250 | GPUTarget arch_target = get_arch_from_target(gpu_target); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 251 | switch(input0->info()->data_type()) |
| 252 | { |
| 253 | case DataType::F16: |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 254 | build_opts.emplace("#define DATA_TYPE_FP16"); |
Frank Lei | b9d38ee | 2017-12-05 10:43:33 +0800 | [diff] [blame] | 255 | build_opts.emplace("#define MM_PROCESS_4X_OPTIMIZED"); |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 256 | build_opts.emplace("#define GEMM_MM_FLOATING_POINT"); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 257 | break; |
| 258 | |
| 259 | case DataType::F32: |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 260 | build_opts.emplace("#define DATA_TYPE_FP32"); |
Michele Di Giorgio | f6f08da | 2018-04-26 10:24:30 +0100 | [diff] [blame] | 261 | |
| 262 | if(arch_target == GPUTarget::BIFROST && input0->info()->num_dimensions() != 1) |
| 263 | { |
| 264 | build_opts.emplace("#define GEMM_MM_FLOATING_POINT_BIFROST"); |
| 265 | } |
| 266 | else |
| 267 | { |
| 268 | build_opts.emplace("#define GEMM_MM_FLOATING_POINT"); |
| 269 | } |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 270 | break; |
| 271 | |
| 272 | default: |
| 273 | ARM_COMPUTE_ERROR("Current data type is not supported"); |
| 274 | break; |
| 275 | } |
| 276 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 277 | build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_X " + support::cpp11::to_string(num_elements_processed.x())); |
| 278 | build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_Y " + support::cpp11::to_string(num_elements_processed.y())); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 279 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 280 | kernel_name = "gemm_mm_floating_point"; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 281 | } |
| 282 | |
Michele Di Giorgio | 164b65d | 2018-04-13 14:28:08 +0100 | [diff] [blame] | 283 | // Create kernel |
| 284 | _kernel = GCKernelLibrary::get().create_kernel(kernel_name, build_opts); |
| 285 | } |
| 286 | |
| 287 | Status GCGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved_transposed, |
| 288 | const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target) |
| 289 | { |
| 290 | ARM_COMPUTE_UNUSED(alpha); |
| 291 | ElementsProcessed num_elements_processed{}; |
| 292 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, is_interleaved_transposed, reshape_info)); |
| 293 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), |
| 294 | input1->clone().get(), |
| 295 | output->clone().get(), |
| 296 | is_interleaved_transposed, |
| 297 | reshape_info, |
| 298 | gpu_target, |
| 299 | num_elements_processed) |
| 300 | .first); |
| 301 | return Status{}; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 302 | } |
| 303 | |
| 304 | void GCGEMMMatrixMultiplyKernel::run(const Window &window) |
| 305 | { |
| 306 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 307 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IGCKernel::window(), window); |
| 308 | |
| 309 | _kernel.use(); |
| 310 | |
| 311 | Window slice = window.first_slice_window_2D(); |
| 312 | Window slice_matrix_b = slice; |
| 313 | |
| 314 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 315 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 316 | |
| 317 | do |
| 318 | { |
| 319 | Window slice_b = slice; |
| 320 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 321 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 322 | if(_input1->info()->num_dimensions() < 3) |
| 323 | { |
| 324 | slice_b = slice_matrix_b; |
| 325 | } |
| 326 | |
| 327 | unsigned int idx = 0; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 328 | |
Joel Liang | abd03cf | 2018-01-08 15:20:48 +0800 | [diff] [blame] | 329 | add_2D_tensor_argument(idx, _input0, 1, slice); |
| 330 | add_2D_tensor_argument(idx, _input1, 2, slice_b); |
| 331 | add_2D_tensor_argument(idx, _output, 3, slice); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 332 | _kernel.update_shader_params(); |
| 333 | enqueue(*this, slice); |
| 334 | } |
| 335 | while(window.slide_window_slice_2D(slice)); |
| 336 | } |