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 <math.h> |
| 10 | #include <stddef.h> |
| 11 | #include <stdint.h> |
| 12 | #include <stdlib.h> |
| 13 | |
| 14 | #include <xnnpack.h> |
| 15 | #include <xnnpack/allocator.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 16 | #include <xnnpack/log.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 17 | #include <xnnpack/operator.h> |
| 18 | #include <xnnpack/params-init.h> |
| 19 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | |
| 21 | |
| 22 | enum xnn_status xnn_create_clamp_nc_u8( |
| 23 | size_t channels, |
| 24 | size_t input_stride, |
| 25 | size_t output_stride, |
| 26 | uint8_t output_min, |
| 27 | uint8_t output_max, |
| 28 | uint32_t flags, |
| 29 | xnn_operator_t* clamp_op_out) |
| 30 | { |
| 31 | xnn_operator_t clamp_op = NULL; |
| 32 | enum xnn_status status = xnn_status_uninitialized; |
| 33 | |
| 34 | if (!xnn_params.initialized) { |
| 35 | xnn_log_error("failed to create Clamp operator: XNNPACK is not initialized"); |
| 36 | goto error; |
| 37 | } |
| 38 | |
| 39 | status = xnn_status_invalid_parameter; |
| 40 | |
| 41 | if (channels == 0) { |
| 42 | xnn_log_error( |
| 43 | "failed to create Clamp operator with %zu channels: number of channels must be non-zero", channels); |
| 44 | goto error; |
| 45 | } |
| 46 | |
| 47 | if (input_stride < channels) { |
| 48 | xnn_log_error( |
| 49 | "failed to create Clamp operator with input element stride of %zu: " |
| 50 | "stride must be at least as large as the number of channels (%zu)", |
| 51 | input_stride, channels); |
| 52 | goto error; |
| 53 | } |
| 54 | |
| 55 | if (output_stride < channels) { |
| 56 | xnn_log_error( |
| 57 | "failed to create Clamp operator with output element stride of %zu: " |
| 58 | "stride must be at least as large as the number of channels (%zu)", |
| 59 | output_stride, channels); |
| 60 | goto error; |
| 61 | } |
| 62 | |
| 63 | if (output_min >= output_max) { |
| 64 | xnn_log_error( |
| 65 | "failed to create Clamp operator with [%" PRIu8 ", %" PRIu8 "] output range: range min must be below range max", |
| 66 | output_min, output_max); |
| 67 | goto error; |
| 68 | } |
| 69 | |
| 70 | status = xnn_status_out_of_memory; |
| 71 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 72 | clamp_op = xnn_allocate_zero_simd_memory(sizeof(struct xnn_operator)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 73 | if (clamp_op == NULL) { |
| 74 | xnn_log_error("failed to allocate %zu bytes for Clamp operator descriptor", sizeof(struct xnn_operator)); |
| 75 | goto error; |
| 76 | } |
| 77 | |
| 78 | clamp_op->channels = channels; |
| 79 | clamp_op->input_pixel_stride = input_stride; |
| 80 | clamp_op->output_pixel_stride = output_stride; |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 81 | clamp_op->u8_output_params = xnn_init_u8_output_params(output_min, output_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 82 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 83 | clamp_op->type = xnn_operator_type_clamp_nc_u8; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | clamp_op->ukernel.type = xnn_ukernel_type_clamp; |
| 85 | |
| 86 | clamp_op->state = xnn_run_state_invalid; |
| 87 | |
| 88 | *clamp_op_out = clamp_op; |
| 89 | return xnn_status_success; |
| 90 | |
| 91 | error: |
| 92 | xnn_delete_operator(clamp_op); |
| 93 | return status; |
| 94 | } |
| 95 | |
| 96 | enum xnn_status xnn_create_clamp_nc_f32( |
| 97 | size_t channels, |
| 98 | size_t input_stride, |
| 99 | size_t output_stride, |
| 100 | float output_min, |
| 101 | float output_max, |
| 102 | uint32_t flags, |
| 103 | xnn_operator_t* clamp_op_out) |
| 104 | { |
| 105 | xnn_operator_t clamp_op = NULL; |
| 106 | enum xnn_status status = xnn_status_uninitialized; |
| 107 | |
| 108 | if (!xnn_params.initialized) { |
| 109 | xnn_log_error("failed to create Clamp operator: XNNPACK is not initialized"); |
| 110 | goto error; |
| 111 | } |
| 112 | |
| 113 | status = xnn_status_invalid_parameter; |
| 114 | |
| 115 | if (channels == 0) { |
| 116 | xnn_log_error( |
| 117 | "failed to create Clamp operator with %zu channels: number of channels must be non-zero", channels); |
| 118 | goto error; |
| 119 | } |
| 120 | |
| 121 | if (input_stride < channels) { |
| 122 | xnn_log_error( |
| 123 | "failed to create Clamp operator with input element stride of %zu: " |
| 124 | "stride must be at least as large as the number of channels (%zu)", |
| 125 | input_stride, channels); |
| 126 | goto error; |
| 127 | } |
| 128 | |
| 129 | if (output_stride < channels) { |
| 130 | xnn_log_error( |
| 131 | "failed to create Clamp operator with output element stride of %zu: " |
| 132 | "stride must be at least as large as the number of channels (%zu)", |
| 133 | output_stride, channels); |
| 134 | goto error; |
| 135 | } |
| 136 | |
| 137 | if (isnan(output_min)) { |
| 138 | xnn_log_error( |
| 139 | "failed to create Clamp operator with NaN output lower bound: lower bound must be non-NaN"); |
| 140 | goto error; |
| 141 | } |
| 142 | |
| 143 | if (isnan(output_max)) { |
| 144 | xnn_log_error( |
| 145 | "failed to create Clamp operator with NaN output upper bound: upper bound must be non-NaN"); |
| 146 | goto error; |
| 147 | } |
| 148 | |
| 149 | if (output_min >= output_max) { |
| 150 | xnn_log_error( |
| 151 | "failed to create Clamp operator with [%.7g, %.7g] output range: lower bound must be below upper bound", |
| 152 | output_min, output_max); |
| 153 | goto error; |
| 154 | } |
| 155 | |
| 156 | status = xnn_status_out_of_memory; |
| 157 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 158 | clamp_op = xnn_allocate_zero_simd_memory(sizeof(struct xnn_operator)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 159 | if (clamp_op == NULL) { |
| 160 | xnn_log_error("failed to allocate %zu bytes for Clamp operator descriptor", sizeof(struct xnn_operator)); |
| 161 | goto error; |
| 162 | } |
| 163 | |
| 164 | clamp_op->channels = channels; |
| 165 | clamp_op->input_pixel_stride = input_stride; |
| 166 | clamp_op->output_pixel_stride = output_stride; |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 167 | clamp_op->f32_output_params = xnn_init_f32_output_params(output_min, output_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 168 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 169 | clamp_op->type = xnn_operator_type_clamp_nc_f32; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 170 | clamp_op->ukernel.type = xnn_ukernel_type_clamp; |
| 171 | |
| 172 | clamp_op->state = xnn_run_state_invalid; |
| 173 | |
| 174 | *clamp_op_out = clamp_op; |
| 175 | return xnn_status_success; |
| 176 | |
| 177 | error: |
| 178 | xnn_delete_operator(clamp_op); |
| 179 | return status; |
| 180 | } |
| 181 | |
| 182 | enum xnn_status xnn_setup_clamp_nc_u8( |
| 183 | xnn_operator_t clamp_op, |
| 184 | size_t batch_size, |
| 185 | const uint8_t* input, |
| 186 | uint8_t* output, |
| 187 | pthreadpool_t threadpool) |
| 188 | { |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 189 | if (clamp_op->type != xnn_operator_type_clamp_nc_u8) { |
| 190 | xnn_log_error("failed to setup Clamp (NC, U8) operator: operator type mismatch"); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 191 | return xnn_status_invalid_parameter; |
| 192 | } |
| 193 | clamp_op->state = xnn_run_state_invalid; |
| 194 | |
| 195 | if (!xnn_params.initialized) { |
| 196 | xnn_log_error("failed to setup Clamp operator: XNNPACK is not initialized"); |
| 197 | return xnn_status_uninitialized; |
| 198 | } |
| 199 | |
| 200 | if (batch_size == 0) { |
| 201 | clamp_op->state = xnn_run_state_skip; |
| 202 | return xnn_status_success; |
| 203 | } |
| 204 | |
| 205 | const size_t channels = clamp_op->channels; |
| 206 | const size_t input_stride = clamp_op->input_pixel_stride; |
| 207 | const size_t output_stride = clamp_op->output_pixel_stride; |
| 208 | if ((((input_stride ^ channels) | (output_stride ^ channels)) == 0) || batch_size == 1) { |
| 209 | const size_t block_size = 4096; |
| 210 | clamp_op->context.univector_contiguous = (struct univector_contiguous_context) { |
| 211 | .x = input, |
| 212 | .x_stride = input_stride * sizeof(uint8_t), |
| 213 | .y = output, |
| 214 | .y_stride = output_stride * sizeof(uint8_t), |
| 215 | .ukernel = xnn_params.u8.clamp, |
| 216 | .params.u8_output = clamp_op->u8_output_params, |
| 217 | }; |
| 218 | clamp_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 219 | clamp_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_univector_contiguous; |
| 220 | clamp_op->compute.range[0] = batch_size * channels * sizeof(uint8_t); |
| 221 | clamp_op->compute.tile[0] = block_size; |
| 222 | } else { |
| 223 | clamp_op->context.univector_strided = (struct univector_strided_context) { |
| 224 | .n = channels * sizeof(uint8_t), |
| 225 | .x = input, |
| 226 | .x_stride = input_stride * sizeof(uint8_t), |
| 227 | .y = output, |
| 228 | .y_stride = output_stride * sizeof(uint8_t), |
| 229 | .ukernel = xnn_params.u8.clamp, |
| 230 | .params.u8_output = clamp_op->u8_output_params, |
| 231 | }; |
| 232 | clamp_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 233 | clamp_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_univector_strided; |
| 234 | clamp_op->compute.range[0] = batch_size; |
| 235 | clamp_op->compute.tile[0] = 1; |
| 236 | } |
| 237 | clamp_op->state = xnn_run_state_ready; |
| 238 | |
| 239 | return xnn_status_success; |
| 240 | } |
| 241 | |
| 242 | enum xnn_status xnn_setup_clamp_nc_f32( |
| 243 | xnn_operator_t clamp_op, |
| 244 | size_t batch_size, |
| 245 | const float* input, |
| 246 | float* output, |
| 247 | pthreadpool_t threadpool) |
| 248 | { |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 249 | if (clamp_op->type != xnn_operator_type_clamp_nc_f32) { |
| 250 | xnn_log_error("failed to setup Clamp (NC, F32) operator: operator type mismatch"); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 251 | return xnn_status_invalid_parameter; |
| 252 | } |
| 253 | clamp_op->state = xnn_run_state_invalid; |
| 254 | |
| 255 | if (!xnn_params.initialized) { |
| 256 | xnn_log_error("failed to setup Clamp operator: XNNPACK is not initialized"); |
| 257 | return xnn_status_uninitialized; |
| 258 | } |
| 259 | |
| 260 | if (batch_size == 0) { |
| 261 | clamp_op->state = xnn_run_state_skip; |
| 262 | return xnn_status_success; |
| 263 | } |
| 264 | |
| 265 | const size_t channels = clamp_op->channels; |
| 266 | const size_t input_stride = clamp_op->input_pixel_stride; |
| 267 | const size_t output_stride = clamp_op->output_pixel_stride; |
| 268 | if ((((input_stride ^ channels) | (output_stride ^ channels)) == 0) || batch_size == 1) { |
| 269 | const size_t block_size = 4096; |
| 270 | clamp_op->context.univector_contiguous = (struct univector_contiguous_context) { |
| 271 | .x = input, |
| 272 | .x_stride = input_stride * sizeof(float), |
| 273 | .y = output, |
| 274 | .y_stride = output_stride * sizeof(float), |
| 275 | .ukernel = xnn_params.f32.clamp, |
| 276 | .params.f32_output = clamp_op->f32_output_params, |
| 277 | }; |
| 278 | clamp_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 279 | clamp_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_univector_contiguous; |
| 280 | clamp_op->compute.range[0] = batch_size * channels * sizeof(float); |
| 281 | clamp_op->compute.tile[0] = block_size; |
| 282 | } else { |
| 283 | clamp_op->context.univector_strided = (struct univector_strided_context) { |
| 284 | .n = channels * sizeof(float), |
| 285 | .x = input, |
| 286 | .x_stride = input_stride * sizeof(float), |
| 287 | .y = output, |
| 288 | .y_stride = output_stride * sizeof(float), |
| 289 | .ukernel = xnn_params.f32.clamp, |
| 290 | .params.f32_output = clamp_op->f32_output_params, |
| 291 | }; |
| 292 | clamp_op->compute.type = xnn_parallelization_type_1d_tile_1d; |
| 293 | clamp_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_univector_strided; |
| 294 | clamp_op->compute.range[0] = batch_size; |
| 295 | clamp_op->compute.tile[0] = 1; |
| 296 | } |
| 297 | clamp_op->state = xnn_run_state_ready; |
| 298 | |
| 299 | return xnn_status_success; |
| 300 | } |