XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <assert.h> |
| 10 | #include <math.h> |
| 11 | #include <stddef.h> |
| 12 | #include <stdint.h> |
| 13 | #include <stdlib.h> |
| 14 | |
| 15 | #include <xnnpack.h> |
| 16 | #include <xnnpack/allocator.h> |
| 17 | #include <xnnpack/operator.h> |
| 18 | #include <xnnpack/requantization.h> |
| 19 | #include <xnnpack/log.h> |
| 20 | #include <xnnpack/params.h> |
| 21 | |
| 22 | |
| 23 | enum xnn_status xnn_create_add_nc_q8( |
| 24 | size_t channels, |
| 25 | size_t a_stride, |
| 26 | size_t b_stride, |
| 27 | size_t sum_stride, |
| 28 | uint8_t a_zero_point, |
| 29 | float a_scale, |
| 30 | uint8_t b_zero_point, |
| 31 | float b_scale, |
| 32 | uint8_t sum_zero_point, |
| 33 | float sum_scale, |
| 34 | uint8_t sum_min, |
| 35 | uint8_t sum_max, |
| 36 | uint32_t flags, |
| 37 | xnn_operator_t* add_op_out) |
| 38 | { |
| 39 | xnn_operator_t add_op = NULL; |
| 40 | enum xnn_status status = xnn_status_uninitialized; |
| 41 | |
| 42 | if (!xnn_params.initialized) { |
| 43 | xnn_log_error("failed to create Add operator: XNNPACK is not initialized"); |
| 44 | goto error; |
| 45 | } |
| 46 | |
| 47 | status = xnn_status_invalid_parameter; |
| 48 | |
| 49 | if (channels == 0) { |
| 50 | xnn_log_error( |
| 51 | "failed to create Add operator with %zu channels: number of channels must be non-zero", channels); |
| 52 | goto error; |
| 53 | } |
| 54 | |
| 55 | if (a_stride < channels) { |
| 56 | xnn_log_error( |
| 57 | "failed to create Add operator with A element stride of %zu: " |
| 58 | "stride must be at least as large as the number of channels (%zu)", |
| 59 | a_stride, channels); |
| 60 | goto error; |
| 61 | } |
| 62 | |
| 63 | if (b_stride < channels) { |
| 64 | xnn_log_error( |
| 65 | "failed to create Add operator with B element stride of %zu: " |
| 66 | "stride must be at least as large as the number of channels (%zu)", |
| 67 | b_stride, channels); |
| 68 | goto error; |
| 69 | } |
| 70 | |
| 71 | if (sum_stride < channels) { |
| 72 | xnn_log_error( |
| 73 | "failed to create Add operator with Sum element stride of %zu: " |
| 74 | "stride must be at least as large as the number of channels (%zu)", |
| 75 | sum_stride, channels); |
| 76 | goto error; |
| 77 | } |
| 78 | |
| 79 | if (a_scale <= 0.0f || !isnormal(a_scale)) { |
| 80 | xnn_log_error( |
| 81 | "failed to create Add operator with %.7g A scale: scale must be finite, normalized, and positive", a_scale); |
| 82 | goto error; |
| 83 | } |
| 84 | |
| 85 | if (b_scale <= 0.0f || !isnormal(b_scale)) { |
| 86 | xnn_log_error( |
| 87 | "failed to create Add operator with %.7g B scale: scale must be finite, normalized, and positive", b_scale); |
| 88 | goto error; |
| 89 | } |
| 90 | |
| 91 | if (sum_scale <= 0.0f || !isnormal(sum_scale)) { |
| 92 | xnn_log_error( |
| 93 | "failed to create Add operator with %.7g output scale: scale must be finite, normalized, and positive", |
| 94 | sum_scale); |
| 95 | goto error; |
| 96 | } |
| 97 | |
| 98 | if (sum_min >= sum_max) { |
| 99 | xnn_log_error( |
| 100 | "failed to create Add operator with [%" PRIu8 ", %" PRIu8 "] output range: range min must be below range max", |
| 101 | sum_min, sum_max); |
| 102 | goto error; |
| 103 | } |
| 104 | |
| 105 | status = xnn_status_unsupported_parameter; |
| 106 | |
| 107 | const float a_output_scale = a_scale / sum_scale; |
| 108 | if (a_output_scale < 0x1.0p-14f || a_output_scale >= 0x1.0p+8f) { |
| 109 | xnn_log_error( |
| 110 | "failed to create Add operator with %.7g A-to-output scale ratio: scale ratio must be in [2**-14, 2**8) range", |
| 111 | a_output_scale); |
| 112 | goto error; |
| 113 | } |
| 114 | |
| 115 | const float b_output_scale = b_scale / sum_scale; |
| 116 | if (b_output_scale < 0x1.0p-14f || b_output_scale >= 0x1.0p+8f) { |
| 117 | xnn_log_error( |
| 118 | "failed to create Add operator with %.7g A-to-output scale ratio: scale ratio must be in [2**-14, 2**8) range", |
| 119 | b_output_scale); |
| 120 | goto error; |
| 121 | } |
| 122 | |
| 123 | status = xnn_status_out_of_memory; |
| 124 | |
| 125 | add_op = xnn_allocate_zero_memory(sizeof(struct xnn_operator)); |
| 126 | if (add_op == NULL) { |
| 127 | xnn_log_error("failed to allocate %zu bytes for Add operator descriptor", sizeof(struct xnn_operator)); |
| 128 | goto error; |
| 129 | } |
| 130 | |
| 131 | add_op->channels = channels; |
| 132 | add_op->input_pixel_stride = a_stride; |
| 133 | add_op->input2_pixel_stride = b_stride; |
| 134 | add_op->output_pixel_stride = sum_stride; |
| 135 | add_op->q8_add_params = |
| 136 | xnn_compute_q8_add_params( |
| 137 | a_zero_point, b_zero_point, sum_zero_point, |
| 138 | a_scale / sum_scale, b_scale / sum_scale, |
| 139 | sum_min, sum_max); |
| 140 | |
| 141 | add_op->type = xnn_operator_type_add_q8; |
| 142 | add_op->ukernel.type = xnn_ukernel_type_add; |
| 143 | |
| 144 | add_op->state = xnn_run_state_invalid; |
| 145 | |
| 146 | *add_op_out = add_op; |
| 147 | return xnn_status_success; |
| 148 | |
| 149 | error: |
| 150 | xnn_delete_operator(add_op); |
| 151 | return status; |
| 152 | } |
| 153 | |
| 154 | enum xnn_status xnn_create_add_nc_f32( |
| 155 | size_t channels, |
| 156 | size_t a_stride, |
| 157 | size_t b_stride, |
| 158 | size_t sum_stride, |
| 159 | float sum_min, |
| 160 | float sum_max, |
| 161 | uint32_t flags, |
| 162 | xnn_operator_t* add_op_out) |
| 163 | { |
| 164 | xnn_operator_t add_op = NULL; |
| 165 | enum xnn_status status = xnn_status_uninitialized; |
| 166 | |
| 167 | if (!xnn_params.initialized) { |
| 168 | xnn_log_error("failed to create Add operator: XNNPACK is not initialized"); |
| 169 | goto error; |
| 170 | } |
| 171 | |
| 172 | status = xnn_status_invalid_parameter; |
| 173 | |
| 174 | if (channels == 0) { |
| 175 | xnn_log_error( |
| 176 | "failed to create add operator with %zu channels: number of channels must be non-zero", channels); |
| 177 | goto error; |
| 178 | } |
| 179 | |
| 180 | if (a_stride < channels) { |
| 181 | xnn_log_error( |
| 182 | "failed to create Add operator with A element stride of %zu: " |
| 183 | "stride must be at least as large as the number of channels (%zu)", |
| 184 | a_stride, channels); |
| 185 | goto error; |
| 186 | } |
| 187 | |
| 188 | if (b_stride < channels) { |
| 189 | xnn_log_error( |
| 190 | "failed to create Add operator with B element stride of %zu: " |
| 191 | "stride must be at least as large as the number of channels (%zu)", |
| 192 | b_stride, channels); |
| 193 | goto error; |
| 194 | } |
| 195 | |
| 196 | if (sum_stride < channels) { |
| 197 | xnn_log_error( |
| 198 | "failed to create Add operator with Sum element stride of %zu: " |
| 199 | "stride must be at least as large as the number of channels (%zu)", |
| 200 | sum_stride, channels); |
| 201 | goto error; |
| 202 | } |
| 203 | |
| 204 | if (isnan(sum_min)) { |
| 205 | xnn_log_error( |
| 206 | "failed to create Add operator with NaN output lower bound: lower bound must be non-NaN"); |
| 207 | goto error; |
| 208 | } |
| 209 | |
| 210 | if (isnan(sum_max)) { |
| 211 | xnn_log_error( |
| 212 | "failed to create Add operator with NaN output upper bound: upper bound must be non-NaN"); |
| 213 | goto error; |
| 214 | } |
| 215 | |
| 216 | if (sum_min >= sum_max) { |
| 217 | xnn_log_error( |
| 218 | "failed to create Add operator with [%.7g, %.7g] output range: lower bound must be below upper bound", |
| 219 | sum_min, sum_max); |
| 220 | goto error; |
| 221 | } |
| 222 | |
| 223 | status = xnn_status_out_of_memory; |
| 224 | |
| 225 | add_op = xnn_allocate_zero_memory(sizeof(struct xnn_operator)); |
| 226 | if (add_op == NULL) { |
| 227 | xnn_log_error("failed to allocate %zu bytes for Add operator descriptor", sizeof(struct xnn_operator)); |
| 228 | goto error; |
| 229 | } |
| 230 | |
| 231 | add_op->channels = channels; |
| 232 | add_op->input_pixel_stride = a_stride; |
| 233 | add_op->input2_pixel_stride = b_stride; |
| 234 | add_op->output_pixel_stride = sum_stride; |
| 235 | add_op->f32_output_params = xnn_compute_f32_output_params(sum_min, sum_max); |
| 236 | |
| 237 | add_op->type = xnn_operator_type_add_f32; |
| 238 | add_op->ukernel.type = xnn_ukernel_type_add; |
| 239 | |
| 240 | add_op->state = xnn_run_state_invalid; |
| 241 | |
| 242 | *add_op_out = add_op; |
| 243 | return xnn_status_success; |
| 244 | |
| 245 | error: |
| 246 | xnn_delete_operator(add_op); |
| 247 | return status; |
| 248 | } |
| 249 | |
| 250 | enum xnn_status xnn_setup_add_nc_q8( |
| 251 | xnn_operator_t add_op, |
| 252 | size_t batch_size, |
| 253 | const uint8_t* a, |
| 254 | const uint8_t* b, |
| 255 | uint8_t* sum, |
| 256 | pthreadpool_t threadpool) |
| 257 | { |
| 258 | if (add_op->type != xnn_operator_type_add_q8) { |
| 259 | xnn_log_error("failed to setup Add (Q8) operator: operator type mismatch"); |
| 260 | return xnn_status_invalid_parameter; |
| 261 | } |
| 262 | add_op->state = xnn_run_state_invalid; |
| 263 | |
| 264 | if (!xnn_params.initialized) { |
| 265 | xnn_log_error("failed to setup Add operator: XNNPACK is not initialized"); |
| 266 | return xnn_status_uninitialized; |
| 267 | } |
| 268 | |
| 269 | if (batch_size == 0) { |
| 270 | add_op->state = xnn_run_state_skip; |
| 271 | return xnn_status_success; |
| 272 | } |
| 273 | |
| 274 | const size_t channels = add_op->channels; |
| 275 | const size_t a_stride = add_op->input_pixel_stride; |
| 276 | const size_t b_stride = add_op->input2_pixel_stride; |
| 277 | const size_t sum_stride = add_op->output_pixel_stride; |
| 278 | if ((((a_stride ^ channels) | (b_stride ^ channels) | (sum_stride ^ channels)) == 0) || batch_size == 1) { |
| 279 | const size_t block_size = 4096; |
| 280 | add_op->context.add_contiguous = (struct add_contiguous_context) { |
| 281 | .a = a, |
| 282 | .b = b, |
| 283 | .y = sum, |
| 284 | .params.q8 = add_op->q8_add_params, |
| 285 | .ukernel = xnn_params.q8.vadd, |
| 286 | }; |
| 287 | add_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 288 | add_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_add_contiguous; |
| 289 | add_op->compute.range[0] = batch_size * channels * sizeof(uint8_t); |
| 290 | add_op->compute.tile[0] = block_size; |
| 291 | } else { |
| 292 | add_op->context.add_strided = (struct add_strided_context) { |
| 293 | .a = a, |
| 294 | .a_stride = a_stride * sizeof(uint8_t), |
| 295 | .b = b, |
| 296 | .b_stride = b_stride * sizeof(uint8_t), |
| 297 | .y = sum, |
| 298 | .y_stride = sum_stride * sizeof(uint8_t), |
| 299 | .n = channels, |
| 300 | .params.q8 = add_op->q8_add_params, |
| 301 | .ukernel = xnn_params.q8.vadd, |
| 302 | }; |
| 303 | add_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 304 | add_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_add_strided; |
| 305 | add_op->compute.range[0] = batch_size; |
| 306 | add_op->compute.tile[0] = 1; |
| 307 | } |
| 308 | add_op->state = xnn_run_state_ready; |
| 309 | |
| 310 | return xnn_status_success; |
| 311 | } |
| 312 | |
| 313 | enum xnn_status xnn_setup_add_nc_f32( |
| 314 | xnn_operator_t add_op, |
| 315 | size_t batch_size, |
| 316 | const float* a, |
| 317 | const float* b, |
| 318 | float* sum, |
| 319 | pthreadpool_t threadpool) |
| 320 | { |
| 321 | if (add_op->type != xnn_operator_type_add_f32) { |
| 322 | xnn_log_error("failed to setup Add (F32) operator: operator type mismatch"); |
| 323 | return xnn_status_invalid_parameter; |
| 324 | } |
| 325 | add_op->state = xnn_run_state_invalid; |
| 326 | |
| 327 | if (!xnn_params.initialized) { |
| 328 | xnn_log_error("failed to setup Add operator: XNNPACK is not initialized"); |
| 329 | return xnn_status_uninitialized; |
| 330 | } |
| 331 | |
| 332 | if (batch_size == 0) { |
| 333 | add_op->state = xnn_run_state_skip; |
| 334 | return xnn_status_success; |
| 335 | } |
| 336 | |
| 337 | const size_t channels = add_op->channels; |
| 338 | const size_t a_stride = add_op->input_pixel_stride; |
| 339 | const size_t b_stride = add_op->input2_pixel_stride; |
| 340 | const size_t sum_stride = add_op->output_pixel_stride; |
| 341 | if ((((a_stride ^ channels) | (b_stride ^ channels) | (sum_stride ^ channels)) == 0) || batch_size == 1) { |
| 342 | const size_t block_size = 4096; |
| 343 | add_op->context.add_contiguous = (struct add_contiguous_context) { |
| 344 | .a = a, |
| 345 | .b = b, |
| 346 | .y = sum, |
| 347 | .params.f32 = add_op->f32_output_params, |
| 348 | .ukernel = xnn_params.f32.vadd, |
| 349 | }; |
| 350 | add_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 351 | add_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_add_contiguous; |
| 352 | add_op->compute.range[0] = batch_size * channels * sizeof(float); |
| 353 | add_op->compute.tile[0] = block_size; |
| 354 | } else { |
| 355 | add_op->context.add_strided = (struct add_strided_context) { |
| 356 | .a = a, |
| 357 | .a_stride = a_stride * sizeof(float), |
| 358 | .b = b, |
| 359 | .b_stride = b_stride * sizeof(float), |
| 360 | .y = sum, |
| 361 | .y_stride = sum_stride * sizeof(float), |
| 362 | .n = channels * sizeof(float), |
| 363 | .params.f32 = add_op->f32_output_params, |
| 364 | .ukernel = xnn_params.f32.vadd, |
| 365 | }; |
| 366 | add_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 367 | add_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_add_strided; |
| 368 | add_op->compute.range[0] = batch_size; |
| 369 | add_op->compute.tile[0] = 1; |
| 370 | } |
| 371 | add_op->state = xnn_run_state_ready; |
| 372 | |
| 373 | return xnn_status_success; |
| 374 | } |