Lior Dekel | 489e40b | 2021-08-02 12:03:55 +0300 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019-2021 Arm Limited. All rights reserved. |
| 3 | * |
| 4 | * SPDX-License-Identifier: Apache-2.0 |
| 5 | * |
| 6 | * Licensed under the Apache License, Version 2.0 (the License); you may |
| 7 | * not use this file except in compliance with the License. |
| 8 | * You may obtain a copy of the License at |
| 9 | * |
| 10 | * www.apache.org/licenses/LICENSE-2.0 |
| 11 | * |
| 12 | * Unless required by applicable law or agreed to in writing, software |
| 13 | * distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 14 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 15 | * See the License for the specific language governing permissions and |
| 16 | * limitations under the License. |
| 17 | */ |
| 18 | |
| 19 | /**************************************************************************** |
| 20 | * Includes |
| 21 | ****************************************************************************/ |
| 22 | #include "tx_api.h" |
| 23 | |
| 24 | #include <inttypes.h> |
| 25 | #include <stdio.h> |
| 26 | #include <vector> |
| 27 | |
| 28 | #include "inference_process.hpp" |
| 29 | |
| 30 | // Model data (Defined & changable by modifiying compile definition in CMakeLists.txt) |
| 31 | #include "input.h" |
| 32 | #include "model.h" |
| 33 | #include "output.h" |
| 34 | |
| 35 | using namespace std; |
| 36 | using namespace InferenceProcess; |
| 37 | |
| 38 | /**************************************************************************** |
| 39 | * Defines |
| 40 | ****************************************************************************/ |
| 41 | // Nr. of threads to process inferences with. Thread reserves driver & runs inference (Normally 1 per NPU, but not a |
| 42 | // must) |
| 43 | #define NUM_INFERENCE_THREADS 1 |
| 44 | // Nr. of threads to create jobs and recieve responses |
| 45 | #define NUM_JOB_THREADS 2 |
| 46 | // Nr. of jobs to create per job thread |
| 47 | #define NUM_JOBS_PER_THREAD 1 |
| 48 | |
| 49 | #define PROCESS_THREAD_STACK_SIZE (16 * 1024) |
| 50 | #define SENDER_THREAD_STACK_SIZE (2 * 1024) |
| 51 | #define PROCESS_THREAD_CONTEXT_SIZE (sizeof(TX_THREAD)) |
| 52 | #define SENDER_THREAD_CONTEXT_SIZE (sizeof(TX_THREAD)) |
| 53 | |
| 54 | // Tensor arena size |
| 55 | #ifdef TENSOR_ARENA_SIZE // If defined in model.h |
| 56 | #define TENSOR_ARENA_SIZE_PER_INFERENCE TENSOR_ARENA_SIZE |
| 57 | #else // If not defined, use maximum available - 2M |
| 58 | #define TENSOR_ARENA_SIZE 2000000 |
| 59 | #define TENSOR_ARENA_SIZE_PER_INFERENCE (TENSOR_ARENA_SIZE / NUM_INFERENCE_THREADS) |
| 60 | #endif |
| 61 | |
| 62 | #define PROCESS_QUEUE_SIZE (NUM_JOBS_PER_THREAD * NUM_JOB_THREADS * sizeof(xInferenceJob *)) |
| 63 | #define SENDER_QUEUE_SIZE (NUM_JOBS_PER_THREAD * sizeof(xInferenceJob *)) |
| 64 | |
| 65 | /* BYTE_POOL_SIZE_OVERHEAD is used to increase the memory byte pool size, as the number of |
| 66 | allocatable bytes in a memory byte pool is slightly less than what was specified during creation */ |
| 67 | #define BYTE_POOL_SIZE_OVERHEAD (512) |
| 68 | #define BYTE_POOL_SIZE \ |
| 69 | (((PROCESS_THREAD_CONTEXT_SIZE + PROCESS_THREAD_STACK_SIZE) * NUM_INFERENCE_THREADS) + \ |
| 70 | (SENDER_THREAD_CONTEXT_SIZE + SENDER_THREAD_STACK_SIZE + SENDER_QUEUE_SIZE) * NUM_JOB_THREADS + \ |
| 71 | PROCESS_QUEUE_SIZE + BYTE_POOL_SIZE_OVERHEAD) |
| 72 | |
| 73 | /**************************************************************************** |
| 74 | * Structures |
| 75 | ****************************************************************************/ |
| 76 | struct ProcessThreadParams { |
| 77 | ProcessThreadParams() : queueHandle(nullptr), tensorArena(nullptr), arenaSize(0) {} |
| 78 | ProcessThreadParams(TX_QUEUE *_queue, uint8_t *_tensorArena, size_t _arenaSize) : |
| 79 | queueHandle(_queue), tensorArena(_tensorArena), arenaSize(_arenaSize) {} |
| 80 | |
| 81 | TX_QUEUE *queueHandle; |
| 82 | uint8_t *tensorArena; |
| 83 | size_t arenaSize; |
| 84 | }; |
| 85 | |
| 86 | // Wrapper around InferenceProcess::InferenceJob. Adds responseQueue and status for ThreadX multi-threaded purposes. |
| 87 | struct xInferenceJob : public InferenceJob { |
| 88 | TX_QUEUE *responseQueue; |
| 89 | bool status; |
| 90 | |
| 91 | xInferenceJob() : InferenceJob(), responseQueue(nullptr), status(false) {} |
| 92 | xInferenceJob(const string &_name, |
| 93 | const DataPtr &_networkModel, |
| 94 | const vector<DataPtr> &_input, |
| 95 | const vector<DataPtr> &_output, |
| 96 | const vector<DataPtr> &_expectedOutput, |
| 97 | const size_t _numBytesToPrint, |
| 98 | const vector<uint8_t> &_pmuEventConfig, |
| 99 | const uint32_t _pmuCycleCounterEnable, |
| 100 | TX_QUEUE *_queue) : |
| 101 | InferenceJob(_name, |
| 102 | _networkModel, |
| 103 | _input, |
| 104 | _output, |
| 105 | _expectedOutput, |
| 106 | _numBytesToPrint, |
| 107 | _pmuEventConfig, |
| 108 | _pmuCycleCounterEnable), |
| 109 | responseQueue(_queue), status(false) {} |
| 110 | }; |
| 111 | |
| 112 | /**************************************************************************** |
| 113 | * Global and static variables |
| 114 | ****************************************************************************/ |
| 115 | namespace { |
| 116 | // Number of total completed jobs, needed to exit application correctly if NUM_JOB_THREADS > 1 |
| 117 | int totalCompletedJobs = 0; |
| 118 | |
| 119 | // TensorArena static initialisation |
| 120 | const size_t arenaSize = TENSOR_ARENA_SIZE_PER_INFERENCE; |
| 121 | |
| 122 | TX_QUEUE inferenceProcessQueue; |
| 123 | |
| 124 | ProcessThreadParams threadParams[NUM_INFERENCE_THREADS]; |
| 125 | |
| 126 | TX_BYTE_POOL bytePool; |
| 127 | ULONG memoryArea[BYTE_POOL_SIZE / sizeof(ULONG)]; |
| 128 | } // namespace |
| 129 | |
| 130 | __attribute__((section(".bss.tensor_arena"), aligned(16))) |
| 131 | uint8_t inferenceProcessTensorArena[NUM_INFERENCE_THREADS][arenaSize]; |
| 132 | |
| 133 | /**************************************************************************** |
| 134 | * Mutex & Semaphore |
| 135 | * Overrides weak-linked symbols in ethosu_driver.c to implement thread handling |
| 136 | ****************************************************************************/ |
| 137 | extern "C" { |
| 138 | void *ethosu_mutex_create(void) { |
| 139 | UINT status; |
| 140 | TX_MUTEX *mutex; |
| 141 | |
| 142 | mutex = new TX_MUTEX; |
| 143 | status = tx_mutex_create(mutex, "mutex 0", TX_NO_INHERIT); |
| 144 | if (status != TX_SUCCESS) { |
| 145 | printf("mutex create failed, error - %d\n", status); |
| 146 | } |
| 147 | return (void *)mutex; |
| 148 | } |
| 149 | |
| 150 | void ethosu_mutex_lock(void *mutex) { |
| 151 | UINT status; |
| 152 | status = tx_mutex_get(reinterpret_cast<TX_MUTEX *>(mutex), TX_WAIT_FOREVER); |
| 153 | if (status != TX_SUCCESS) { |
| 154 | printf("mutex get failed, error - %d\n", status); |
| 155 | } |
| 156 | return; |
| 157 | } |
| 158 | |
| 159 | void ethosu_mutex_unlock(void *mutex) { |
| 160 | UINT status; |
| 161 | status = tx_mutex_put(reinterpret_cast<TX_MUTEX *>(mutex)); |
| 162 | if (status != TX_SUCCESS) { |
| 163 | printf("mutex put failed, error - %d\n", status); |
| 164 | } |
| 165 | return; |
| 166 | } |
| 167 | |
| 168 | void *ethosu_semaphore_create(void) { |
| 169 | UINT status; |
| 170 | TX_SEMAPHORE *semaphore; |
| 171 | |
| 172 | semaphore = new TX_SEMAPHORE; |
| 173 | status = tx_semaphore_create(semaphore, "semaphore", 1); |
| 174 | |
| 175 | if (status != TX_SUCCESS) { |
| 176 | printf("Semaphore create failed, error - %d\n", status); |
| 177 | } |
| 178 | |
| 179 | return (void *)semaphore; |
| 180 | } |
| 181 | |
| 182 | void ethosu_semaphore_take(void *sem) { |
| 183 | UINT status; |
| 184 | |
| 185 | status = tx_semaphore_get(reinterpret_cast<TX_SEMAPHORE *>(sem), TX_WAIT_FOREVER); |
| 186 | |
| 187 | if (status != TX_SUCCESS) { |
| 188 | printf("Semaphore get/take, error - %d\n", status); |
| 189 | } |
| 190 | |
| 191 | return; |
| 192 | } |
| 193 | |
| 194 | void ethosu_semaphore_give(void *sem) { |
| 195 | UINT status; |
| 196 | |
| 197 | status = tx_semaphore_put(reinterpret_cast<TX_SEMAPHORE *>(sem)); |
| 198 | |
| 199 | if (status != TX_SUCCESS) { |
| 200 | printf("Semaphore put/give, error - %d\n", status); |
| 201 | } |
| 202 | |
| 203 | return; |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | /**************************************************************************** |
| 208 | * Functions |
| 209 | ****************************************************************************/ |
| 210 | // inferenceProcessThread - Run jobs from queue with available driver |
| 211 | void inferenceProcessThread(ULONG pvParameters) { |
| 212 | ProcessThreadParams params = *reinterpret_cast<ProcessThreadParams *>(pvParameters); |
| 213 | UINT tx_status = TX_QUEUE_ERROR; |
| 214 | |
| 215 | class InferenceProcess inferenceProcess(params.tensorArena, params.arenaSize); |
| 216 | |
| 217 | for (;;) { |
| 218 | xInferenceJob *xJob; |
| 219 | |
| 220 | // Get the job details from the process queue |
| 221 | tx_status = tx_queue_receive(params.queueHandle, &xJob, TX_WAIT_FOREVER); |
| 222 | if (tx_status != TX_SUCCESS) { |
| 223 | printf("process failed to receive from Queue, error - %d\n", tx_status); |
| 224 | exit(1); |
| 225 | } |
| 226 | |
| 227 | // run the job |
| 228 | bool status = inferenceProcess.runJob(*xJob); |
| 229 | xJob->status = status; |
| 230 | |
| 231 | // Send response for the job in the response queue |
| 232 | tx_status = tx_queue_send(xJob->responseQueue, &xJob, TX_WAIT_FOREVER); |
| 233 | if (tx_status != TX_SUCCESS) { |
| 234 | printf("process inferenceProcessThread failed to send to Queue, error - %d\n", tx_status); |
| 235 | exit(1); |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | tx_status = tx_thread_terminate(nullptr); |
| 240 | if (tx_status != TX_SUCCESS) { |
| 241 | printf("process inferenceProcessThread failed to terminate thread, error - %d\n", tx_status); |
| 242 | exit(1); |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | // inferenceSenderThread - Creates NUM_INFERNECE_JOBS jobs, queues them, and then listens for completion status |
| 247 | void inferenceSenderThread(ULONG pvParameters) { |
| 248 | int ret = 0; |
| 249 | TX_QUEUE senderQueue; |
| 250 | UINT status = TX_QUEUE_ERROR; |
| 251 | TX_QUEUE *inferenceProcessQueueLocal = reinterpret_cast<TX_QUEUE *>(pvParameters); |
| 252 | xInferenceJob jobs[NUM_JOBS_PER_THREAD]; |
| 253 | CHAR *senderQueuePtr = nullptr; |
| 254 | |
| 255 | /* Allocate memory for this inference sender thread responses queue */ |
| 256 | status = tx_byte_allocate(&bytePool, reinterpret_cast<VOID **>(&senderQueuePtr), SENDER_QUEUE_SIZE, TX_NO_WAIT); |
| 257 | if (status != TX_SUCCESS) { |
| 258 | printf("Sender thread failed to allocate bytes for Queue, error - %d\n", status); |
| 259 | exit(1); |
| 260 | } |
| 261 | |
| 262 | /* Create responses queue for this inference sender thread */ |
| 263 | status = tx_queue_create( |
| 264 | &senderQueue, "senderQueue", sizeof(xInferenceJob *) / sizeof(uint32_t), senderQueuePtr, SENDER_QUEUE_SIZE); |
| 265 | |
| 266 | if (status != TX_SUCCESS) { |
| 267 | printf("Sender thread failed to create Queue, error - %d\n", status); |
| 268 | exit(1); |
| 269 | } |
| 270 | |
| 271 | /* Create the jobs and queue them in the inference process queue */ |
| 272 | for (int n = 0; n < NUM_JOBS_PER_THREAD; n++) { |
| 273 | |
| 274 | // Create job |
| 275 | xInferenceJob *job = &jobs[n]; |
| 276 | job->name = string(modelName); |
| 277 | job->networkModel = DataPtr(networkModelData, sizeof(networkModelData)); |
| 278 | job->input.push_back(DataPtr(inputData, sizeof(inputData))); |
| 279 | job->expectedOutput.push_back(DataPtr(expectedOutputData, sizeof(expectedOutputData))); |
| 280 | job->responseQueue = &senderQueue; |
| 281 | |
| 282 | // queue job |
| 283 | status = tx_queue_send(inferenceProcessQueueLocal, &job, TX_WAIT_FOREVER); |
| 284 | if (status != TX_SUCCESS) { |
| 285 | printf("Sender thread failed to send to Queue, error - %d\n", status); |
| 286 | exit(1); |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | /* Listen for completion status on the response queue */ |
| 291 | do { |
| 292 | xInferenceJob *pSendJob; |
| 293 | |
| 294 | status = tx_queue_receive(&senderQueue, &pSendJob, TX_WAIT_FOREVER); |
| 295 | if (status != TX_SUCCESS) { |
| 296 | printf("Sender thread failed to receive from Queue, error - %d\n", status); |
| 297 | exit(1); |
| 298 | } |
| 299 | |
| 300 | totalCompletedJobs++; |
| 301 | ret = (pSendJob->status); |
| 302 | if (pSendJob->status != 0) { |
| 303 | break; |
| 304 | } |
| 305 | } while (totalCompletedJobs < NUM_JOBS_PER_THREAD * NUM_JOB_THREADS); |
| 306 | |
| 307 | /* delete the response queue */ |
| 308 | status = tx_queue_delete(&senderQueue); |
| 309 | if (status != TX_SUCCESS) { |
| 310 | printf("Sender thread failed to delete Queue, error - %d\n", status); |
| 311 | exit(1); |
| 312 | } |
| 313 | |
| 314 | exit(ret); |
| 315 | } |
| 316 | |
| 317 | /**************************************************************************** |
| 318 | * Application |
| 319 | ****************************************************************************/ |
| 320 | int main() { |
| 321 | /* Enter the ThreadX kernel. */ |
| 322 | tx_kernel_enter(); |
| 323 | return 0; |
| 324 | } |
| 325 | |
| 326 | void tx_application_define(void *first_unused_memory) { |
| 327 | UINT status; |
| 328 | CHAR *senderThreadStackPtr[NUM_JOB_THREADS] = {nullptr}; |
| 329 | CHAR *processThreadStackPtr[NUM_INFERENCE_THREADS] = {nullptr}; |
| 330 | CHAR *processQueuePtr = nullptr; |
| 331 | CHAR *senderThreadPtr[NUM_JOB_THREADS] = {nullptr}; |
| 332 | CHAR *processThreadPtr[NUM_INFERENCE_THREADS] = {nullptr}; |
| 333 | |
| 334 | /* Create a byte memory pool from which to allocate the threads stacks and queues. */ |
| 335 | status = tx_byte_pool_create(&bytePool, "byte pool", memoryArea, BYTE_POOL_SIZE); |
| 336 | if (status != TX_SUCCESS) { |
| 337 | printf("Main failed to allocate pool of bytes, error - %d\n", status); |
| 338 | exit(1); |
| 339 | } |
| 340 | |
| 341 | /* Allocate memory for the inference process queue */ |
| 342 | status = tx_byte_allocate(&bytePool, reinterpret_cast<VOID **>(&processQueuePtr), PROCESS_QUEUE_SIZE, TX_NO_WAIT); |
| 343 | if (status != TX_SUCCESS) { |
| 344 | printf("Main failed to allocate bytes for process queue, error - %d\n", status); |
| 345 | exit(1); |
| 346 | } |
| 347 | |
| 348 | status = tx_queue_create(&inferenceProcessQueue, |
| 349 | "inferenceProcessQueue", |
| 350 | sizeof(xInferenceJob *) / sizeof(uint32_t), |
| 351 | processQueuePtr, |
| 352 | PROCESS_QUEUE_SIZE); |
| 353 | if (status != TX_SUCCESS) { |
| 354 | printf("Main failed to create Queue, error - %d\n", status); |
| 355 | exit(1); |
| 356 | } |
| 357 | |
| 358 | /* inferenceSender threads to create and queue the jobs */ |
| 359 | for (int n = 0; n < NUM_JOB_THREADS; n++) { |
| 360 | |
| 361 | /* Allocate the thread context for the inference sender thread. */ |
| 362 | status = |
| 363 | tx_byte_allocate(&bytePool, reinterpret_cast<VOID **>(&senderThreadPtr[n]), sizeof(TX_THREAD), TX_NO_WAIT); |
| 364 | if (status != TX_SUCCESS) { |
| 365 | printf("Main failed to allocate bytes for sender tread, error - %d\n", status); |
| 366 | exit(1); |
| 367 | } |
| 368 | |
| 369 | /* Allocate the stack for the inference sender thread. */ |
| 370 | status = tx_byte_allocate( |
| 371 | &bytePool, reinterpret_cast<VOID **>(&senderThreadStackPtr[n]), SENDER_THREAD_STACK_SIZE, TX_NO_WAIT); |
| 372 | if (status != TX_SUCCESS) { |
| 373 | printf("Main failed to allocate bytes for sender tread stack, error - %d\n", status); |
| 374 | exit(1); |
| 375 | } |
| 376 | |
| 377 | /* Create the inference sender thread. */ |
| 378 | status = tx_thread_create(reinterpret_cast<TX_THREAD *>(senderThreadPtr[n]), |
| 379 | "senderThread", |
| 380 | inferenceSenderThread, |
| 381 | reinterpret_cast<ULONG>(&inferenceProcessQueue), |
| 382 | senderThreadStackPtr[n], |
| 383 | SENDER_THREAD_STACK_SIZE, |
| 384 | 1, |
| 385 | 1, |
| 386 | TX_NO_TIME_SLICE, |
| 387 | TX_AUTO_START); |
| 388 | if (status != TX_SUCCESS) { |
| 389 | printf("Main failed to create Thread, error - %d\n", status); |
| 390 | exit(1); |
| 391 | } |
| 392 | } |
| 393 | |
| 394 | /* Create inferenceProcess threads to process the queued jobs */ |
| 395 | for (int n = 0; n < NUM_INFERENCE_THREADS; n++) { |
| 396 | |
| 397 | /* Allocate the thread context for the inference process thread. */ |
| 398 | status = |
| 399 | tx_byte_allocate(&bytePool, reinterpret_cast<VOID **>(&processThreadPtr[n]), sizeof(TX_THREAD), TX_NO_WAIT); |
| 400 | if (status != TX_SUCCESS) { |
| 401 | printf("Main failed to allocate bytes for process tread, error - %d\n", status); |
| 402 | exit(1); |
| 403 | } |
| 404 | |
| 405 | /* Allocate the stack for the inference process thread. */ |
| 406 | status = tx_byte_allocate( |
| 407 | &bytePool, reinterpret_cast<VOID **>(&processThreadStackPtr[n]), PROCESS_THREAD_STACK_SIZE, TX_NO_WAIT); |
| 408 | if (status != TX_SUCCESS) { |
| 409 | printf("Main failed to allocate bytes for process stack, error - %d\n", status); |
| 410 | exit(1); |
| 411 | } |
| 412 | |
| 413 | threadParams[n] = ProcessThreadParams( |
| 414 | &inferenceProcessQueue, inferenceProcessTensorArena[n], reinterpret_cast<size_t>(arenaSize)); |
| 415 | |
| 416 | /* Create the inference process thread. */ |
| 417 | status = tx_thread_create(reinterpret_cast<TX_THREAD *>(processThreadPtr[n]), |
| 418 | "processThread", |
| 419 | inferenceProcessThread, |
| 420 | reinterpret_cast<ULONG>(&threadParams[n]), |
| 421 | processThreadStackPtr[n], |
| 422 | PROCESS_THREAD_STACK_SIZE, |
| 423 | 1, |
| 424 | 1, |
| 425 | TX_NO_TIME_SLICE, |
| 426 | TX_AUTO_START); |
| 427 | if (status != TX_SUCCESS) { |
| 428 | printf("Main failed to create thread, error - %d\n", status); |
| 429 | exit(1); |
| 430 | } |
| 431 | } |
| 432 | |
| 433 | printf("ThreadX application initialisation - Done \n"); |
| 434 | return; |
| 435 | } |