Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" |
| 25 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/AccessWindowStatic.h" |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/CPP/Validate.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/TensorInfo.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/Window.h" |
Gian Marco Iodice | 82d9dd1 | 2019-06-10 16:45:40 +0100 | [diff] [blame] | 37 | #include "arm_compute/core/utils/helpers/float_ops.h" |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 38 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 39 | #include "src/core/NEON/NEFixedPoint.h" |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 40 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | #include <arm_neon.h> |
| 42 | #include <cstddef> |
| 43 | #include <cstdint> |
| 44 | #include <tuple> |
| 45 | |
| 46 | using namespace arm_compute; |
| 47 | |
| 48 | namespace arm_compute |
| 49 | { |
| 50 | class Coordinates; |
| 51 | } // namespace arm_compute |
| 52 | |
| 53 | namespace |
| 54 | { |
| 55 | template <bool multiply_alpha> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 56 | void vector_matrix_multiply_f16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha) |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 57 | { |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 58 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 59 | const auto width_matrix_b = static_cast<int>(output->info()->dimension(0)); |
| 60 | const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type())); |
| 61 | const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0)); |
| 62 | |
| 63 | // The implementation computes 32 elements per iteration |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 64 | const int window_start_x = 32 * info.thread_id; |
| 65 | const int window_step_x = 32 * info.num_threads; |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 66 | const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; |
| 67 | ARM_COMPUTE_ERROR_ON_MSG((window_end_x - window_start_x) % window_step_x, " (window_end_x - window_start_x) must be multiple of window_step_x"); |
| 68 | |
| 69 | Window win_out(window); |
| 70 | win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); |
| 71 | win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 72 | |
| 73 | Window win_a(window); |
| 74 | win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 75 | win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 76 | |
| 77 | Window win_b; |
| 78 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 79 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 80 | if(input1->info()->num_dimensions() >= 3) |
| 81 | { |
| 82 | win_b = window; |
| 83 | } |
| 84 | win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); |
| 85 | win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 86 | |
| 87 | Iterator ina(input0, win_a); |
| 88 | Iterator inb(input1, win_b); |
| 89 | Iterator out(output, win_out); |
| 90 | |
| 91 | const float16x8_t alpha_f16 = vdupq_n_f16(alpha); |
| 92 | ARM_COMPUTE_UNUSED(alpha_f16); |
| 93 | |
| 94 | execute_window_loop(win_out, [&](const Coordinates & id) |
| 95 | { |
| 96 | if(id.x() > width_matrix_b) |
| 97 | { |
| 98 | return; |
| 99 | } |
| 100 | |
| 101 | float16x8_t acc0 = vdupq_n_f16(0.f); |
| 102 | float16x8_t acc1 = vdupq_n_f16(0.f); |
| 103 | float16x8_t acc2 = vdupq_n_f16(0.f); |
| 104 | float16x8_t acc3 = vdupq_n_f16(0.f); |
| 105 | |
| 106 | auto vec_a = reinterpret_cast<const float16_t *>(ina.ptr()); |
| 107 | auto matrix_b = reinterpret_cast<const float16_t *>(inb.ptr()); |
| 108 | |
| 109 | const float16_t *vec_a_end_addr = vec_a + num_elems_vec_a; |
| 110 | for(; vec_a <= (vec_a_end_addr - 4);) |
| 111 | { |
| 112 | const float16x4_t a0l = vld1_f16(vec_a); |
| 113 | |
| 114 | float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); |
| 115 | float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); |
| 116 | float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); |
| 117 | float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); |
| 118 | float16x8_t b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride); |
| 119 | float16x8_t b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride); |
| 120 | float16x8_t b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride); |
| 121 | float16x8_t b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride); |
| 122 | |
| 123 | acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 0)); |
| 124 | acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 0)); |
| 125 | acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 0)); |
| 126 | acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 0)); |
| 127 | acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 1)); |
| 128 | acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 1)); |
| 129 | acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 1)); |
| 130 | acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 1)); |
| 131 | |
| 132 | matrix_b += 2 * in_b_stride; |
| 133 | |
| 134 | b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); |
| 135 | b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); |
| 136 | b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); |
| 137 | b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); |
| 138 | b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride); |
| 139 | b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride); |
| 140 | b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride); |
| 141 | b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride); |
| 142 | |
| 143 | acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 2)); |
| 144 | acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 2)); |
| 145 | acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 2)); |
| 146 | acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 2)); |
| 147 | acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 3)); |
| 148 | acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 3)); |
| 149 | acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 3)); |
| 150 | acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 3)); |
| 151 | |
| 152 | vec_a += 4; |
| 153 | matrix_b += 2 * in_b_stride; |
| 154 | } |
| 155 | |
| 156 | for(; vec_a < vec_a_end_addr;) |
| 157 | { |
| 158 | const float16_t a0 = *vec_a; |
| 159 | const float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); |
| 160 | const float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); |
| 161 | const float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); |
| 162 | const float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); |
| 163 | |
| 164 | acc0 = vaddq_f16(acc0, vmulq_n_f16(b00, a0)); |
| 165 | acc1 = vaddq_f16(acc1, vmulq_n_f16(b01, a0)); |
| 166 | acc2 = vaddq_f16(acc2, vmulq_n_f16(b02, a0)); |
| 167 | acc3 = vaddq_f16(acc3, vmulq_n_f16(b03, a0)); |
| 168 | |
| 169 | vec_a += 1; |
| 170 | matrix_b += in_b_stride; |
| 171 | } |
| 172 | |
| 173 | // Multiply by the weight of matrix product (alpha) |
| 174 | if(multiply_alpha) |
| 175 | { |
| 176 | acc0 = vmulq_f16(acc0, alpha_f16); |
| 177 | acc1 = vmulq_f16(acc1, alpha_f16); |
| 178 | acc2 = vmulq_f16(acc2, alpha_f16); |
| 179 | acc3 = vmulq_f16(acc3, alpha_f16); |
| 180 | } |
| 181 | |
| 182 | const auto vec_out = reinterpret_cast<float16_t *>(out.ptr()); |
| 183 | |
| 184 | vst1q_f16(vec_out + 0, acc0); |
| 185 | vst1q_f16(vec_out + 8, acc1); |
| 186 | vst1q_f16(vec_out + 16, acc2); |
| 187 | vst1q_f16(vec_out + 24, acc3); |
| 188 | |
| 189 | }, |
| 190 | ina, inb, out); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 191 | #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Georgios Pinitas | 30f0215 | 2017-09-27 11:20:48 +0100 | [diff] [blame] | 192 | ARM_COMPUTE_UNUSED(input0); |
| 193 | ARM_COMPUTE_UNUSED(input1); |
| 194 | ARM_COMPUTE_UNUSED(output); |
| 195 | ARM_COMPUTE_UNUSED(window); |
| 196 | ARM_COMPUTE_UNUSED(info); |
| 197 | ARM_COMPUTE_UNUSED(alpha); |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 198 | ARM_COMPUTE_ERROR("Not implemented"); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 199 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 200 | } |
| 201 | |
| 202 | template <bool multiply_alpha> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 203 | void vector_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, const ThreadInfo &info, float alpha) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 204 | { |
| 205 | const auto width_matrix_b = static_cast<int>(output->info()->dimension(0)); |
| 206 | const auto in_b_stride = static_cast<int>(input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type())); |
| 207 | const auto num_elems_vec_a = static_cast<int>(input0->info()->dimension(0)); |
| 208 | |
| 209 | // The implementation computes 16 elements per iteration |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 210 | const int window_start_x = 16 * info.thread_id; |
| 211 | const int window_step_x = 16 * info.num_threads; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 212 | // Make sure (window_end_x - window_start_x) is a multiple of window_step_x |
| 213 | const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; |
| 214 | |
| 215 | Window win_out(window); |
| 216 | win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); |
| 217 | win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 218 | |
| 219 | Window win_a(window); |
| 220 | win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 221 | win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 222 | |
| 223 | Window win_b; |
| 224 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 225 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 226 | if(input1->info()->num_dimensions() >= 3) |
| 227 | { |
| 228 | win_b = window; |
| 229 | } |
| 230 | win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); |
| 231 | win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 232 | |
| 233 | Iterator ina(input0, win_a); |
| 234 | Iterator inb(input1, win_b); |
| 235 | Iterator out(output, win_out); |
| 236 | |
| 237 | execute_window_loop(win_out, [&](const Coordinates & id) |
| 238 | { |
| 239 | if(id.x() > width_matrix_b) |
| 240 | { |
| 241 | return; |
| 242 | } |
| 243 | |
| 244 | float32x4_t acc0 = vdupq_n_f32(0.f); |
| 245 | float32x4_t acc1 = vdupq_n_f32(0.f); |
| 246 | float32x4_t acc2 = vdupq_n_f32(0.f); |
| 247 | float32x4_t acc3 = vdupq_n_f32(0.f); |
| 248 | |
| 249 | auto vec_a = reinterpret_cast<const float *>(ina.ptr()); |
| 250 | auto matrix_b = reinterpret_cast<const float *>(inb.ptr()); |
| 251 | |
| 252 | #if __arm__ |
| 253 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(vec_a))); |
| 254 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b))); |
| 255 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + in_b_stride))); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 256 | #endif /* __arm__ */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 257 | |
| 258 | auto vec_a_end_addr = vec_a + num_elems_vec_a; |
| 259 | for(; vec_a <= (vec_a_end_addr - 4);) |
| 260 | { |
| 261 | float32x2_t a0l = vld1_f32(vec_a); |
| 262 | |
| 263 | float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); |
| 264 | float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); |
| 265 | float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); |
| 266 | float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); |
| 267 | |
| 268 | float32x4_t b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); |
| 269 | float32x4_t b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); |
| 270 | float32x4_t b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); |
| 271 | float32x4_t b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); |
| 272 | |
| 273 | #if __arm__ |
| 274 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(vec_a))); |
| 275 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 1 * in_b_stride))); |
| 276 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 2 * in_b_stride))); |
| 277 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 3 * in_b_stride))); |
| 278 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(matrix_b + 4 * in_b_stride))); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 279 | #endif /* __arm__ */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 280 | |
| 281 | acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); |
| 282 | acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); |
| 283 | acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); |
| 284 | acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); |
| 285 | |
| 286 | acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); |
| 287 | acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); |
| 288 | acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); |
| 289 | acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); |
| 290 | |
| 291 | vec_a += 2; |
| 292 | matrix_b += 2 * in_b_stride; |
| 293 | |
| 294 | a0l = vld1_f32(vec_a); |
| 295 | |
| 296 | b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); |
| 297 | b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); |
| 298 | b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); |
| 299 | b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); |
| 300 | |
| 301 | b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); |
| 302 | b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); |
| 303 | b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); |
| 304 | b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); |
| 305 | |
| 306 | acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); |
| 307 | acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); |
| 308 | acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); |
| 309 | acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); |
| 310 | |
| 311 | acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); |
| 312 | acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); |
| 313 | acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); |
| 314 | acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); |
| 315 | |
| 316 | vec_a += 2; |
| 317 | matrix_b += 2 * in_b_stride; |
| 318 | } |
| 319 | |
| 320 | for(; vec_a < vec_a_end_addr;) |
| 321 | { |
| 322 | const float a0 = *vec_a; |
| 323 | |
| 324 | const float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); |
| 325 | const float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); |
| 326 | const float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); |
| 327 | const float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); |
| 328 | |
| 329 | acc0 = vmlaq_n_f32(acc0, b00, a0); |
| 330 | acc1 = vmlaq_n_f32(acc1, b01, a0); |
| 331 | acc2 = vmlaq_n_f32(acc2, b02, a0); |
| 332 | acc3 = vmlaq_n_f32(acc3, b03, a0); |
| 333 | |
| 334 | vec_a += 1; |
| 335 | matrix_b += in_b_stride; |
| 336 | } |
| 337 | |
| 338 | // Multiply by the weight of matrix product (alpha) |
| 339 | if(multiply_alpha) |
| 340 | { |
| 341 | const float32x4_t alpha_f32 = vdupq_n_f32(alpha); |
| 342 | acc0 = vmulq_f32(acc0, alpha_f32); |
| 343 | acc1 = vmulq_f32(acc1, alpha_f32); |
| 344 | acc2 = vmulq_f32(acc2, alpha_f32); |
| 345 | acc3 = vmulq_f32(acc3, alpha_f32); |
| 346 | } |
| 347 | |
| 348 | const auto vec_out = reinterpret_cast<float *>(out.ptr()); |
| 349 | |
| 350 | vst1q_f32(vec_out + 0, acc0); |
| 351 | vst1q_f32(vec_out + 4, acc1); |
| 352 | vst1q_f32(vec_out + 8, acc2); |
| 353 | vst1q_f32(vec_out + 12, acc3); |
| 354 | }, |
| 355 | ina, inb, out); |
| 356 | } |
| 357 | |
| 358 | template <bool multiply_alpha> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 359 | void matrix_matrix_multiply_f32(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) |
| 360 | { |
| 361 | const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); |
| 362 | const size_t out_stride1 = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); |
| 363 | const size_t out_stride2 = out_stride1 * 2; |
| 364 | const size_t out_stride3 = out_stride1 * 3; |
| 365 | const int num_elems_matrix_b_x = input1->info()->dimension(0); |
| 366 | |
| 367 | // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix |
| 368 | Window win_a(window); |
| 369 | win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 370 | win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); |
| 371 | |
| 372 | Window win_b; |
| 373 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 374 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 375 | if(input1->info()->num_dimensions() >= 3) |
| 376 | { |
| 377 | win_b = window; |
| 378 | } |
| 379 | // Set step_x and step_y for matrix B. Scale by a factor of 4 the X range as the input transposed matrix A has 4 times less the cols of the output matrix |
| 380 | // The step along the x direction is 2 times the in_b_stride because for each iteration we compute 2 blocks of size 4x4 |
| 381 | win_b.set(Window::DimX, Window::Dimension(window.x().start() / 4, window.x().end() / 4, 2 * in_b_stride)); |
| 382 | win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 383 | |
| 384 | Iterator ina(input0, win_a); |
| 385 | Iterator inb(input1, win_b); |
| 386 | Iterator out(output, window); |
| 387 | |
| 388 | // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW |
| 389 | // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration |
| 390 | // All the values needed for computing a single 4x4 block will be read from consecutive memory positions |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 391 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 392 | { |
| 393 | auto mtx_a0 = reinterpret_cast<const float *>(ina.ptr()); |
| 394 | auto mtx_b0 = reinterpret_cast<const float *>(inb.ptr()); |
| 395 | auto mtx_b1 = mtx_b0 + in_b_stride; |
| 396 | |
| 397 | float32x4_t acc00 = vdupq_n_f32(0.f); |
| 398 | float32x4_t acc10 = vdupq_n_f32(0.f); |
| 399 | float32x4_t acc20 = vdupq_n_f32(0.f); |
| 400 | float32x4_t acc30 = vdupq_n_f32(0.f); |
| 401 | |
| 402 | float32x4_t acc01 = vdupq_n_f32(0.f); |
| 403 | float32x4_t acc11 = vdupq_n_f32(0.f); |
| 404 | float32x4_t acc21 = vdupq_n_f32(0.f); |
| 405 | float32x4_t acc31 = vdupq_n_f32(0.f); |
| 406 | |
| 407 | #if __arm__ |
| 408 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_a0))); |
| 409 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0))); |
| 410 | asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1))); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 411 | #endif /* __arm__ */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 412 | |
| 413 | auto mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; |
| 414 | for(; mtx_b0 <= (mtx_b0_end_addr - 32);) |
| 415 | { |
| 416 | float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); |
| 417 | float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); |
| 418 | float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); |
| 419 | float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); |
| 420 | |
| 421 | float32x4_t b00 = vld1q_f32(mtx_b0); |
| 422 | float32x4_t b10 = vld1q_f32(mtx_b1); |
| 423 | float32x4_t b01 = vld1q_f32(mtx_b0 + 4); |
| 424 | float32x4_t b11 = vld1q_f32(mtx_b1 + 4); |
| 425 | |
| 426 | #if __arm__ |
| 427 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_a0))); |
| 428 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0))); |
| 429 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1))); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 430 | #endif /* __arm__ */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 431 | |
| 432 | // 4x4 block 0 |
| 433 | acc00 = vmlaq_f32(acc00, b00, a0); |
| 434 | acc10 = vmlaq_f32(acc10, b00, a1); |
| 435 | acc20 = vmlaq_f32(acc20, b00, a2); |
| 436 | acc30 = vmlaq_f32(acc30, b00, a3); |
| 437 | |
| 438 | float32x4_t a4 = vld1q_dup_f32(mtx_a0 + 4); |
| 439 | float32x4_t a5 = vld1q_dup_f32(mtx_a0 + 5); |
| 440 | float32x4_t a6 = vld1q_dup_f32(mtx_a0 + 6); |
| 441 | float32x4_t a7 = vld1q_dup_f32(mtx_a0 + 7); |
| 442 | |
| 443 | // 4x4 block 1 |
| 444 | acc01 = vmlaq_f32(acc01, b10, a0); |
| 445 | acc11 = vmlaq_f32(acc11, b10, a1); |
| 446 | acc21 = vmlaq_f32(acc21, b10, a2); |
| 447 | acc31 = vmlaq_f32(acc31, b10, a3); |
| 448 | |
| 449 | // 4x4 block 0 |
| 450 | acc00 = vmlaq_f32(acc00, b01, a4); |
| 451 | acc10 = vmlaq_f32(acc10, b01, a5); |
| 452 | acc20 = vmlaq_f32(acc20, b01, a6); |
| 453 | acc30 = vmlaq_f32(acc30, b01, a7); |
| 454 | |
| 455 | // 4x4 block 1 |
| 456 | acc01 = vmlaq_f32(acc01, b11, a4); |
| 457 | acc11 = vmlaq_f32(acc11, b11, a5); |
| 458 | acc21 = vmlaq_f32(acc21, b11, a6); |
| 459 | acc31 = vmlaq_f32(acc31, b11, a7); |
| 460 | |
| 461 | mtx_a0 += 8; |
| 462 | mtx_b0 += 8; |
| 463 | mtx_b1 += 8; |
| 464 | |
| 465 | a0 = vld1q_dup_f32(mtx_a0 + 0); |
| 466 | a1 = vld1q_dup_f32(mtx_a0 + 1); |
| 467 | a2 = vld1q_dup_f32(mtx_a0 + 2); |
| 468 | a3 = vld1q_dup_f32(mtx_a0 + 3); |
| 469 | |
| 470 | b00 = vld1q_f32(mtx_b0); |
| 471 | b10 = vld1q_f32(mtx_b1); |
| 472 | b01 = vld1q_f32(mtx_b0 + 4); |
| 473 | b11 = vld1q_f32(mtx_b1 + 4); |
| 474 | |
| 475 | // 4x4 block 0 |
| 476 | acc00 = vmlaq_f32(acc00, b00, a0); |
| 477 | acc10 = vmlaq_f32(acc10, b00, a1); |
| 478 | acc20 = vmlaq_f32(acc20, b00, a2); |
| 479 | acc30 = vmlaq_f32(acc30, b00, a3); |
| 480 | |
| 481 | a4 = vld1q_dup_f32(mtx_a0 + 4); |
| 482 | a5 = vld1q_dup_f32(mtx_a0 + 5); |
| 483 | a6 = vld1q_dup_f32(mtx_a0 + 6); |
| 484 | a7 = vld1q_dup_f32(mtx_a0 + 7); |
| 485 | |
| 486 | // 4x4 block 1 |
| 487 | acc01 = vmlaq_f32(acc01, b10, a0); |
| 488 | acc11 = vmlaq_f32(acc11, b10, a1); |
| 489 | acc21 = vmlaq_f32(acc21, b10, a2); |
| 490 | acc31 = vmlaq_f32(acc31, b10, a3); |
| 491 | |
| 492 | // 4x4 block 0 |
| 493 | acc00 = vmlaq_f32(acc00, b01, a4); |
| 494 | acc10 = vmlaq_f32(acc10, b01, a5); |
| 495 | acc20 = vmlaq_f32(acc20, b01, a6); |
| 496 | acc30 = vmlaq_f32(acc30, b01, a7); |
| 497 | |
| 498 | // 4x4 block 1 |
| 499 | acc01 = vmlaq_f32(acc01, b11, a4); |
| 500 | acc11 = vmlaq_f32(acc11, b11, a5); |
| 501 | acc21 = vmlaq_f32(acc21, b11, a6); |
| 502 | acc31 = vmlaq_f32(acc31, b11, a7); |
| 503 | |
| 504 | mtx_a0 += 8; |
| 505 | mtx_b0 += 8; |
| 506 | mtx_b1 += 8; |
| 507 | |
| 508 | a0 = vld1q_dup_f32(mtx_a0 + 0); |
| 509 | a1 = vld1q_dup_f32(mtx_a0 + 1); |
| 510 | a2 = vld1q_dup_f32(mtx_a0 + 2); |
| 511 | a3 = vld1q_dup_f32(mtx_a0 + 3); |
| 512 | b00 = vld1q_f32(mtx_b0); |
| 513 | b10 = vld1q_f32(mtx_b1); |
| 514 | b01 = vld1q_f32(mtx_b0 + 4); |
| 515 | b11 = vld1q_f32(mtx_b1 + 4); |
| 516 | |
| 517 | #if __arm__ |
| 518 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_a0))); |
| 519 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0))); |
| 520 | asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1))); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 521 | #endif /* __arm__ */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 522 | |
| 523 | // 4x4 block 0 |
| 524 | acc00 = vmlaq_f32(acc00, b00, a0); |
| 525 | acc10 = vmlaq_f32(acc10, b00, a1); |
| 526 | acc20 = vmlaq_f32(acc20, b00, a2); |
| 527 | acc30 = vmlaq_f32(acc30, b00, a3); |
| 528 | |
| 529 | a4 = vld1q_dup_f32(mtx_a0 + 4); |
| 530 | a5 = vld1q_dup_f32(mtx_a0 + 5); |
| 531 | a6 = vld1q_dup_f32(mtx_a0 + 6); |
| 532 | a7 = vld1q_dup_f32(mtx_a0 + 7); |
| 533 | |
| 534 | // 4x4 block 1 |
| 535 | acc01 = vmlaq_f32(acc01, b10, a0); |
| 536 | acc11 = vmlaq_f32(acc11, b10, a1); |
| 537 | acc21 = vmlaq_f32(acc21, b10, a2); |
| 538 | acc31 = vmlaq_f32(acc31, b10, a3); |
| 539 | |
| 540 | // 4x4 block 0 |
| 541 | acc00 = vmlaq_f32(acc00, b01, a4); |
| 542 | acc10 = vmlaq_f32(acc10, b01, a5); |
| 543 | acc20 = vmlaq_f32(acc20, b01, a6); |
| 544 | acc30 = vmlaq_f32(acc30, b01, a7); |
| 545 | |
| 546 | // 4x4 block 1 |
| 547 | acc01 = vmlaq_f32(acc01, b11, a4); |
| 548 | acc11 = vmlaq_f32(acc11, b11, a5); |
| 549 | acc21 = vmlaq_f32(acc21, b11, a6); |
| 550 | acc31 = vmlaq_f32(acc31, b11, a7); |
| 551 | |
| 552 | mtx_a0 += 8; |
| 553 | mtx_b0 += 8; |
| 554 | mtx_b1 += 8; |
| 555 | |
| 556 | a0 = vld1q_dup_f32(mtx_a0 + 0); |
| 557 | a1 = vld1q_dup_f32(mtx_a0 + 1); |
| 558 | a2 = vld1q_dup_f32(mtx_a0 + 2); |
| 559 | a3 = vld1q_dup_f32(mtx_a0 + 3); |
| 560 | b00 = vld1q_f32(mtx_b0); |
| 561 | b10 = vld1q_f32(mtx_b1); |
| 562 | b01 = vld1q_f32(mtx_b0 + 4); |
| 563 | b11 = vld1q_f32(mtx_b1 + 4); |
| 564 | |
| 565 | // 4x4 block 0 |
| 566 | acc00 = vmlaq_f32(acc00, b00, a0); |
| 567 | acc10 = vmlaq_f32(acc10, b00, a1); |
| 568 | acc20 = vmlaq_f32(acc20, b00, a2); |
| 569 | acc30 = vmlaq_f32(acc30, b00, a3); |
| 570 | |
| 571 | a4 = vld1q_dup_f32(mtx_a0 + 4); |
| 572 | a5 = vld1q_dup_f32(mtx_a0 + 5); |
| 573 | a6 = vld1q_dup_f32(mtx_a0 + 6); |
| 574 | a7 = vld1q_dup_f32(mtx_a0 + 7); |
| 575 | |
| 576 | // 4x4 block 1 |
| 577 | acc01 = vmlaq_f32(acc01, b10, a0); |
| 578 | acc11 = vmlaq_f32(acc11, b10, a1); |
| 579 | acc21 = vmlaq_f32(acc21, b10, a2); |
| 580 | acc31 = vmlaq_f32(acc31, b10, a3); |
| 581 | |
| 582 | // 4x4 block 0 |
| 583 | acc00 = vmlaq_f32(acc00, b01, a4); |
| 584 | acc10 = vmlaq_f32(acc10, b01, a5); |
| 585 | acc20 = vmlaq_f32(acc20, b01, a6); |
| 586 | acc30 = vmlaq_f32(acc30, b01, a7); |
| 587 | |
| 588 | // 4x4 block 1 |
| 589 | acc01 = vmlaq_f32(acc01, b11, a4); |
| 590 | acc11 = vmlaq_f32(acc11, b11, a5); |
| 591 | acc21 = vmlaq_f32(acc21, b11, a6); |
| 592 | acc31 = vmlaq_f32(acc31, b11, a7); |
| 593 | |
| 594 | mtx_a0 += 8; |
| 595 | mtx_b0 += 8; |
| 596 | mtx_b1 += 8; |
| 597 | } |
| 598 | |
| 599 | for(; mtx_b0 < mtx_b0_end_addr;) |
| 600 | { |
| 601 | float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); |
| 602 | float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); |
| 603 | float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); |
| 604 | float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); |
| 605 | float32x4_t b00 = vld1q_f32(mtx_b0); |
| 606 | float32x4_t b10 = vld1q_f32(mtx_b1); |
| 607 | |
| 608 | #if __arm__ |
| 609 | asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_a0))); |
| 610 | asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b0))); |
| 611 | asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast<const uint8_t *>(mtx_b1))); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 612 | #endif /* __arm__ */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 613 | // 4x4 block 0 |
| 614 | acc00 = vmlaq_f32(acc00, b00, a0); |
| 615 | acc10 = vmlaq_f32(acc10, b00, a1); |
| 616 | acc20 = vmlaq_f32(acc20, b00, a2); |
| 617 | acc30 = vmlaq_f32(acc30, b00, a3); |
| 618 | |
| 619 | // 4x4 block 1 |
| 620 | acc01 = vmlaq_f32(acc01, b10, a0); |
| 621 | acc11 = vmlaq_f32(acc11, b10, a1); |
| 622 | acc21 = vmlaq_f32(acc21, b10, a2); |
| 623 | acc31 = vmlaq_f32(acc31, b10, a3); |
| 624 | |
| 625 | mtx_a0 += 4; |
| 626 | mtx_b0 += 4; |
| 627 | mtx_b1 += 4; |
| 628 | } |
| 629 | |
| 630 | // Multiply by the weight of matrix product (alpha) |
| 631 | if(multiply_alpha) |
| 632 | { |
| 633 | const float32x4_t alpha_f32 = vdupq_n_f32(alpha); |
| 634 | acc00 = vmulq_f32(acc00, alpha_f32); |
| 635 | acc10 = vmulq_f32(acc10, alpha_f32); |
| 636 | acc20 = vmulq_f32(acc20, alpha_f32); |
| 637 | acc30 = vmulq_f32(acc30, alpha_f32); |
| 638 | acc01 = vmulq_f32(acc01, alpha_f32); |
| 639 | acc11 = vmulq_f32(acc11, alpha_f32); |
| 640 | acc21 = vmulq_f32(acc21, alpha_f32); |
| 641 | acc31 = vmulq_f32(acc31, alpha_f32); |
| 642 | } |
| 643 | |
| 644 | const auto mtx_out0 = reinterpret_cast<float *>(out.ptr()); |
| 645 | const auto mtx_out1 = mtx_out0 + 4; |
| 646 | |
| 647 | // Store the 4 blocks |
| 648 | vst1q_f32(mtx_out0, acc00); |
| 649 | vst1q_f32(mtx_out1, acc01); |
| 650 | vst1q_f32(mtx_out0 + out_stride1, acc10); |
| 651 | vst1q_f32(mtx_out1 + out_stride1, acc11); |
| 652 | vst1q_f32(mtx_out0 + out_stride2, acc20); |
| 653 | vst1q_f32(mtx_out1 + out_stride2, acc21); |
| 654 | vst1q_f32(mtx_out0 + out_stride3, acc30); |
| 655 | vst1q_f32(mtx_out1 + out_stride3, acc31); |
| 656 | }, |
| 657 | ina, inb, out); |
| 658 | } |
| 659 | |
| 660 | template <bool multiply_alpha> |
| 661 | void matrix_matrix_multiply_f16(const ITensor *input0, const ITensor *input1, ITensor *output, const Window &window, float alpha) |
| 662 | { |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 663 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 664 | const size_t in_b_stride = input1->info()->strides_in_bytes()[1] / data_size_from_type(input1->info()->data_type()); |
| 665 | const size_t out_stride = output->info()->strides_in_bytes()[1] / data_size_from_type(output->info()->data_type()); |
| 666 | const int num_elems_matrix_b_x = input1->info()->dimension(0); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 667 | |
| 668 | // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix |
| 669 | Window win_a(window); |
| 670 | win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 671 | win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); |
| 672 | |
| 673 | Window win_b; |
| 674 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 675 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 676 | if(input1->info()->num_dimensions() >= 3) |
| 677 | { |
| 678 | win_b = window; |
| 679 | } |
| 680 | // Set step_x and step_y for matrix B. Scale by a factor of 8 the X range as the input transposed matrix A has 8 times less the cols of the output matrix |
| 681 | win_b.set(Window::DimX, Window::Dimension(window.x().start() / 8, window.x().end() / 8, in_b_stride)); |
| 682 | win_b.set(Window::DimY, Window::Dimension(0, 1, 0)); |
| 683 | |
| 684 | Iterator ina(input0, win_a); |
| 685 | Iterator inb(input1, win_b); |
| 686 | Iterator out(output, window); |
| 687 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 688 | const float16x8_t alpha_f16 = vdupq_n_f16(alpha); |
| 689 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 690 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 691 | { |
| 692 | const auto *mtx_a0 = reinterpret_cast<const float16_t *>(ina.ptr()); |
| 693 | const auto *mtx_b0 = reinterpret_cast<const float16_t *>(inb.ptr()); |
| 694 | auto *mtx_out = reinterpret_cast<float16_t *>(out.ptr()); |
| 695 | float16x8x4_t c = |
| 696 | { |
| 697 | { |
| 698 | vdupq_n_f16(0.f), |
| 699 | vdupq_n_f16(0.f), |
| 700 | vdupq_n_f16(0.f), |
| 701 | vdupq_n_f16(0.f) |
| 702 | } |
| 703 | }; |
| 704 | |
| 705 | /* |
| 706 | This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values) |
| 707 | |a00 a01 a02 a03 | a04 a05 a06 a07| |
| 708 | |a10 a11 a12 a13 | a14 a15 a16 a17| |
| 709 | |a20 a21 a22 a23 | a24 a25 a26 a27| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 | a40 a50 a60 a70 | ... |
| 710 | |a30 a31 a32 a33 | a34 a35 a36 a37| | a04 a14 a24 a34 || a05 a15 a25 a35 || a06 a15 a26 a36 || a07 a17 a27 a37 | a44 a54 a64 a74 | ... |
| 711 | |a40 a41 a42 a43 | a44 a45 a46 a47| |
| 712 | |a50 a51 a52 a53 | a54 a55 a56 a57| |
| 713 | |a60 a61 a62 a63 | a64 a65 a66 a67| |
| 714 | |a70 a71 a72 a73 | a74 a75 a76 a77| |
| 715 | |
| 716 | After this operation, the output matrix will have the following shape: [ height * 4, width / 4 ] |
| 717 | |
| 718 | B Matrix has been transposed as shown below |
| 719 | |
| 720 | |b00 b01 b02 b03 b04 b05 b06 b07| |
| 721 | |b10 b11 b12 b13 b14 b15 b16 b17| |
| 722 | |b20 b21 b22 b23 b24 b25 b26 b27| |
| 723 | |b30 b31 b32 b33 b34 b35 b36 b37| |
| 724 | -------------------> |
| 725 | |
| 726 | |b00 b01 b02 b03 b04 b05 b06 b07||b10 b11 b12 b13 b14 b15 b16 b17||b20 b21 b22 b23 b24 b25 b26 b27||b30 b31 b32 b33 b34 b35 b36 b37| |
| 727 | |
| 728 | c.val[0][0] = a00*b00 + a01*b10 + a02*b20 + a03*b30 |
| 729 | c.val[0][1] = a00*b01 + a01*b11 + a02*b21 + a03*b31 |
| 730 | |
| 731 | The size of the output tensor's XY-plane must be the following shape [ width * 8, height / 8 ]. All other dimensions must have the same size. |
| 732 | */ |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 733 | const float16_t *mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; |
| 734 | |
| 735 | for(; mtx_b0 <= (mtx_b0_end_addr - 32);) |
| 736 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 737 | { |
| 738 | const float16x8_t p00 = vld1q_f16(mtx_a0); |
| 739 | const float16x8_t p02 = vld1q_f16(mtx_a0 + 8); |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 740 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 741 | const float16x8_t q00 = vld1q_f16(mtx_b0); |
| 742 | const float16x8_t q02 = vld1q_f16(mtx_b0 + 8); |
| 743 | const float16x8_t q04 = vld1q_f16(mtx_b0 + 16); |
| 744 | const float16x8_t q06 = vld1q_f16(mtx_b0 + 24); |
| 745 | |
| 746 | c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vgetq_lane_f16(p00, 0))); |
| 747 | c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vgetq_lane_f16(p00, 1))); |
| 748 | c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vgetq_lane_f16(p00, 2))); |
| 749 | c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vgetq_lane_f16(p00, 3))); |
| 750 | |
| 751 | c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q02, vgetq_lane_f16(p00, 4))); |
| 752 | c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q02, vgetq_lane_f16(p00, 5))); |
| 753 | c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q02, vgetq_lane_f16(p00, 6))); |
| 754 | c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q02, vgetq_lane_f16(p00, 7))); |
| 755 | |
| 756 | c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q04, vgetq_lane_f16(p02, 0))); |
| 757 | c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q04, vgetq_lane_f16(p02, 1))); |
| 758 | c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q04, vgetq_lane_f16(p02, 2))); |
| 759 | c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q04, vgetq_lane_f16(p02, 3))); |
| 760 | |
| 761 | c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q06, vgetq_lane_f16(p02, 4))); |
| 762 | c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q06, vgetq_lane_f16(p02, 5))); |
| 763 | c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q06, vgetq_lane_f16(p02, 6))); |
| 764 | c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q06, vgetq_lane_f16(p02, 7))); |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 765 | |
| 766 | mtx_a0 += 16; |
| 767 | mtx_b0 += 32; |
| 768 | } |
| 769 | |
| 770 | for(; mtx_b0 < mtx_b0_end_addr;) |
| 771 | |
| 772 | { |
| 773 | const float16x4_t p00 = vld1_f16(mtx_a0); |
| 774 | const float16x8_t q00 = vld1q_f16(mtx_b0); |
| 775 | |
| 776 | c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vget_lane_f16(p00, 0))); |
| 777 | c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vget_lane_f16(p00, 1))); |
| 778 | c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vget_lane_f16(p00, 2))); |
| 779 | c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vget_lane_f16(p00, 3))); |
| 780 | |
| 781 | mtx_a0 += 4; |
| 782 | mtx_b0 += 8; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 783 | } |
| 784 | |
| 785 | if(multiply_alpha) |
| 786 | { |
| 787 | c.val[0] = vmulq_f16(c.val[0], alpha_f16); |
| 788 | c.val[1] = vmulq_f16(c.val[1], alpha_f16); |
| 789 | c.val[2] = vmulq_f16(c.val[2], alpha_f16); |
| 790 | c.val[3] = vmulq_f16(c.val[3], alpha_f16); |
| 791 | } |
| 792 | |
| 793 | vst1q_f16(mtx_out + 0 * out_stride, c.val[0]); |
| 794 | vst1q_f16(mtx_out + 1 * out_stride, c.val[1]); |
| 795 | vst1q_f16(mtx_out + 2 * out_stride, c.val[2]); |
| 796 | vst1q_f16(mtx_out + 3 * out_stride, c.val[3]); |
| 797 | }, |
| 798 | ina, inb, out); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 799 | #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Georgios Pinitas | 30f0215 | 2017-09-27 11:20:48 +0100 | [diff] [blame] | 800 | ARM_COMPUTE_UNUSED(input0); |
| 801 | ARM_COMPUTE_UNUSED(input1); |
| 802 | ARM_COMPUTE_UNUSED(output); |
| 803 | ARM_COMPUTE_UNUSED(window); |
| 804 | ARM_COMPUTE_UNUSED(alpha); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 805 | ARM_COMPUTE_ERROR("Not implemented"); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 806 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 807 | } |
| 808 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 809 | inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info) |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 810 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 811 | ARM_COMPUTE_UNUSED(alpha); |
| 812 | |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 813 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input0); |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 814 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32); |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 815 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 816 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 817 | if(!is_interleaved) |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 818 | { |
| 819 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 820 | |
| 821 | if(output->total_size() != 0) |
| 822 | { |
| 823 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != output->dimension(0)); |
| 824 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != output->dimension(1)); |
| 825 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 826 | } |
| 827 | } |
| 828 | else |
| 829 | { |
| 830 | const int m = reshape_info.m(); |
| 831 | const int n = reshape_info.n(); |
| 832 | const int k = reshape_info.k(); |
| 833 | const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width(); |
| 834 | const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height(); |
| 835 | |
| 836 | /* Interleave */ |
| 837 | TensorShape tensor_shape0{ input0->tensor_shape() }; |
| 838 | tensor_shape0.set(0, k); |
| 839 | tensor_shape0.set(1, m); |
| 840 | |
| 841 | const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0); |
| 842 | const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(misc::shape_calculator::compute_interleaved_shape(tensor_info0, mult_interleave4x4_height)); |
| 843 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0); |
| 844 | |
| 845 | if(n != 0) /* Transpose */ |
| 846 | { |
| 847 | TensorShape tensor_shape1{ input1->tensor_shape() }; |
| 848 | tensor_shape1.set(0, n); |
| 849 | tensor_shape1.set(1, k); |
| 850 | |
| 851 | const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1); |
| 852 | const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(misc::shape_calculator::compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width)); |
| 853 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1); |
| 854 | } |
| 855 | |
| 856 | if(output->total_size() != 0) |
| 857 | { |
| 858 | if(n != 0) |
| 859 | { |
| 860 | ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != static_cast<size_t>(n)); |
| 861 | } |
| 862 | ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != static_cast<size_t>(m)); |
| 863 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 864 | } |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 865 | } |
| 866 | |
| 867 | return Status{}; |
| 868 | } |
| 869 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 870 | inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output) |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 871 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 872 | bool window_changed{}; |
| 873 | Window win{}; |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 874 | |
| 875 | unsigned int num_elems_processed_per_iteration_x = 0; |
| 876 | const unsigned int num_elems_processed_per_iteration_y = 4; |
| 877 | |
| 878 | // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication |
| 879 | if((output->dimension(1) == 1)) |
| 880 | { |
| 881 | switch(input0->data_type()) |
| 882 | { |
| 883 | case DataType::F32: |
| 884 | { |
| 885 | num_elems_processed_per_iteration_x = 16; |
| 886 | break; |
| 887 | } |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 888 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 889 | case DataType::F16: |
| 890 | { |
| 891 | num_elems_processed_per_iteration_x = 32; |
| 892 | break; |
| 893 | } |
| 894 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 895 | default: |
| 896 | { |
| 897 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 898 | break; |
| 899 | } |
| 900 | } |
| 901 | |
| 902 | // Configure kernel window |
| 903 | win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x)); |
| 904 | |
| 905 | AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x); |
| 906 | |
| 907 | window_changed = update_window_and_padding(win, |
| 908 | AccessWindowStatic(input0, 0, 0, input0->tensor_shape().x(), 1), |
| 909 | AccessWindowHorizontal(input1, 0, num_elems_processed_per_iteration_x), |
| 910 | output_access); |
| 911 | |
| 912 | Coordinates coord; |
| 913 | coord.set_num_dimensions(output->num_dimensions()); |
| 914 | output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape())); |
| 915 | } |
| 916 | else |
| 917 | { |
| 918 | switch(input0->data_type()) |
| 919 | { |
| 920 | case DataType::F32: |
| 921 | { |
| 922 | num_elems_processed_per_iteration_x = 8; |
| 923 | break; |
| 924 | } |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 925 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 926 | case DataType::F16: |
| 927 | { |
| 928 | num_elems_processed_per_iteration_x = 8; |
| 929 | break; |
| 930 | } |
| 931 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 932 | default: |
| 933 | { |
| 934 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 935 | break; |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | // Configure kernel window |
| 940 | win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 941 | |
| 942 | AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 943 | |
| 944 | window_changed = update_window_and_padding(win, |
| 945 | AccessWindowRectangle(input0, 0, 0, 4, 1, 1.f, 0.25f), |
| 946 | AccessWindowStatic(input1, 0, 0, input1->tensor_shape().x(), ceil_to_multiple(input1->tensor_shape().y(), 4)), |
| 947 | output_access); |
| 948 | |
| 949 | output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape())); |
| 950 | } |
| 951 | |
| 952 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 953 | return std::make_pair(err, win); |
| 954 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 955 | } // namespace |
| 956 | |
| 957 | NEGEMMMatrixMultiplyKernel::NEGEMMMatrixMultiplyKernel() |
| 958 | : _input0(nullptr), _input1(nullptr), _output(nullptr), _alpha(1.0f) |
| 959 | { |
| 960 | } |
| 961 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 962 | void NEGEMMMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 963 | { |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 964 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 965 | |
| 966 | // Output tensor auto inizialitation if not yet initialized |
| 967 | TensorShape tensor_shape{ input0->info()->tensor_shape() }; |
| 968 | tensor_shape.set(0, is_interleaved ? reshape_info.n() : input1->info()->dimension(0)); |
| 969 | tensor_shape.set(1, is_interleaved ? reshape_info.m() : input0->info()->dimension(1)); |
| 970 | |
| 971 | auto_init_if_empty(*output->info(), input0->info()->clone()->set_tensor_shape(tensor_shape)); |
| 972 | |
| 973 | // Perform validate step |
| 974 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, is_interleaved, reshape_info)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 975 | |
| 976 | _input0 = input0; |
| 977 | _input1 = input1; |
| 978 | _output = output; |
| 979 | _alpha = alpha; |
| 980 | |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 981 | // Configure kernel window |
| 982 | auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info()); |
| 983 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 984 | INEKernel::configure(win_config.second); |
| 985 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 986 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 987 | Status NEGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved, |
| 988 | const GEMMReshapeInfo &reshape_info) |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 989 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame] | 990 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, is_interleaved, reshape_info)); |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 991 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), input1->clone().get(), output->clone().get()).first); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 992 | |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 993 | return Status{}; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 994 | } |
| 995 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 996 | void NEGEMMMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 997 | { |
| 998 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 999 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1000 | |
Gian Marco Iodice | 82d9dd1 | 2019-06-10 16:45:40 +0100 | [diff] [blame] | 1001 | const bool multiply_alpha = !(helpers::float_ops::is_one(_alpha)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1002 | |
Gian Marco Iodice | bdb6b0b | 2017-06-30 12:21:00 +0100 | [diff] [blame] | 1003 | // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1004 | if((_output->info()->dimension(1) == 1)) |
| 1005 | { |
| 1006 | switch(_input0->info()->data_type()) |
| 1007 | { |
| 1008 | case DataType::F32: |
| 1009 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1010 | multiply_alpha ? vector_matrix_multiply_f32<true>(_input0, _input1, _output, window, info, _alpha) : |
| 1011 | vector_matrix_multiply_f32<false>(_input0, _input1, _output, window, info, _alpha); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1012 | break; |
| 1013 | } |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 1014 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 1015 | case DataType::F16: |
| 1016 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1017 | multiply_alpha ? vector_matrix_multiply_f16<true>(_input0, _input1, _output, window, info, _alpha) : |
| 1018 | vector_matrix_multiply_f16<false>(_input0, _input1, _output, window, info, _alpha); |
Pablo Tello | 221f381 | 2017-06-28 17:27:56 +0100 | [diff] [blame] | 1019 | break; |
| 1020 | } |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 1021 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1022 | default: |
| 1023 | { |
| 1024 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 1025 | break; |
| 1026 | } |
| 1027 | } |
| 1028 | } |
| 1029 | else |
| 1030 | { |
| 1031 | switch(_input0->info()->data_type()) |
| 1032 | { |
| 1033 | case DataType::F32: |
| 1034 | { |
| 1035 | multiply_alpha ? matrix_matrix_multiply_f32<true>(_input0, _input1, _output, window, _alpha) : |
| 1036 | matrix_matrix_multiply_f32<false>(_input0, _input1, _output, window, _alpha); |
| 1037 | break; |
| 1038 | } |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 1039 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1040 | case DataType::F16: |
| 1041 | { |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1042 | multiply_alpha ? matrix_matrix_multiply_f16<true>(_input0, _input1, _output, window, _alpha) : |
| 1043 | matrix_matrix_multiply_f16<false>(_input0, _input1, _output, window, _alpha); |
| 1044 | break; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1045 | } |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 1046 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1047 | default: |
| 1048 | { |
| 1049 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 1050 | break; |
| 1051 | } |
| 1052 | } |
| 1053 | } |
| 1054 | } |