1 /* 2 * Copyright (C) 2012 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 #include "rsCpuCore.h" 18 #include "rsCpuScript.h" 19 #include "rsCpuScriptGroup.h" 20 #include "rsCpuScriptGroup2.h" 21 22 #include <malloc.h> 23 #include "rsContext.h" 24 25 #include <sys/types.h> 26 #include <sys/resource.h> 27 #include <sched.h> 28 #include <sys/syscall.h> 29 #include <stdio.h> 30 #include <string.h> 31 #include <unistd.h> 32 33 #if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB) 34 #include <cutils/properties.h> 35 #include "utils/StopWatch.h" 36 #endif 37 38 #ifdef RS_SERVER 39 // Android exposes gettid(), standard Linux does not 40 static pid_t gettid() { 41 return syscall(SYS_gettid); 42 } 43 #endif 44 45 using namespace android; 46 using namespace android::renderscript; 47 48 #define REDUCE_ALOGV(mtls, level, ...) do { if ((mtls)->logReduce >= (level)) ALOGV(__VA_ARGS__); } while(0) 49 50 static pthread_key_t gThreadTLSKey = 0; 51 static uint32_t gThreadTLSKeyCount = 0; 52 static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER; 53 54 bool android::renderscript::gArchUseSIMD = false; 55 56 RsdCpuReference::~RsdCpuReference() { 57 } 58 59 RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major, 60 uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn 61 , RSSelectRTCallback pSelectRTCallback, 62 const char *pBccPluginName 63 ) { 64 65 RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc); 66 if (!cpu) { 67 return nullptr; 68 } 69 if (!cpu->init(version_major, version_minor, lfn, slfn)) { 70 delete cpu; 71 return nullptr; 72 } 73 74 cpu->setSelectRTCallback(pSelectRTCallback); 75 if (pBccPluginName) { 76 cpu->setBccPluginName(pBccPluginName); 77 } 78 79 return cpu; 80 } 81 82 83 Context * RsdCpuReference::getTlsContext() { 84 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 85 return tls->mContext; 86 } 87 88 const Script * RsdCpuReference::getTlsScript() { 89 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 90 return tls->mScript; 91 } 92 93 pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 94 95 //////////////////////////////////////////////////////////// 96 /// 97 98 RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) { 99 mRSC = rsc; 100 101 version_major = 0; 102 version_minor = 0; 103 mInKernel = false; 104 memset(&mWorkers, 0, sizeof(mWorkers)); 105 memset(&mTlsStruct, 0, sizeof(mTlsStruct)); 106 mExit = false; 107 mSelectRTCallback = nullptr; 108 mEmbedGlobalInfo = true; 109 mEmbedGlobalInfoSkipConstant = true; 110 } 111 112 113 void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 114 RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 115 116 uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 117 118 //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 119 120 dc->mWorkers.mLaunchSignals[idx].init(); 121 dc->mWorkers.mNativeThreadId[idx] = gettid(); 122 123 memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 124 int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 125 if (status) { 126 ALOGE("pthread_setspecific %i", status); 127 } 128 129 #if 0 130 typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 131 cpu_set_t cpuset; 132 memset(&cpuset, 0, sizeof(cpuset)); 133 cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 134 int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 135 sizeof(cpuset), &cpuset); 136 ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 137 #endif 138 139 while (!dc->mExit) { 140 dc->mWorkers.mLaunchSignals[idx].wait(); 141 if (dc->mWorkers.mLaunchCallback) { 142 // idx +1 is used because the calling thread is always worker 0. 143 dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 144 } 145 __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 146 dc->mWorkers.mCompleteSignal.set(); 147 } 148 149 //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 150 return nullptr; 151 } 152 153 // Launch a kernel. 154 // The callback function is called to execute the kernel. 155 void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 156 mWorkers.mLaunchData = data; 157 mWorkers.mLaunchCallback = cbk; 158 159 // fast path for very small launches 160 MTLaunchStructCommon *mtls = (MTLaunchStructCommon *)data; 161 if (mtls && mtls->dimPtr->y <= 1 && mtls->end.x <= mtls->start.x + mtls->mSliceSize) { 162 if (mWorkers.mLaunchCallback) { 163 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 164 } 165 return; 166 } 167 168 mWorkers.mRunningCount = mWorkers.mCount; 169 __sync_synchronize(); 170 171 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 172 mWorkers.mLaunchSignals[ct].set(); 173 } 174 175 // We use the calling thread as one of the workers so we can start without 176 // the delay of the thread wakeup. 177 if (mWorkers.mLaunchCallback) { 178 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 179 } 180 181 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 182 mWorkers.mCompleteSignal.wait(); 183 } 184 } 185 186 187 void RsdCpuReferenceImpl::lockMutex() { 188 pthread_mutex_lock(&gInitMutex); 189 } 190 191 void RsdCpuReferenceImpl::unlockMutex() { 192 pthread_mutex_unlock(&gInitMutex); 193 } 194 195 // Determine if the CPU we're running on supports SIMD instructions. 196 static void GetCpuInfo() { 197 // Read the CPU flags from /proc/cpuinfo. 198 FILE *cpuinfo = fopen("/proc/cpuinfo", "r"); 199 200 if (!cpuinfo) { 201 return; 202 } 203 204 char cpuinfostr[4096]; 205 // fgets() ends with newline or EOF, need to check the whole 206 // "cpuinfo" file to make sure we can use SIMD or not. 207 while (fgets(cpuinfostr, sizeof(cpuinfostr), cpuinfo)) { 208 #if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS) 209 gArchUseSIMD = strstr(cpuinfostr, " neon") || strstr(cpuinfostr, " asimd"); 210 #elif defined(ARCH_X86_HAVE_SSSE3) 211 gArchUseSIMD = strstr(cpuinfostr, " ssse3"); 212 #endif 213 if (gArchUseSIMD) { 214 break; 215 } 216 } 217 fclose(cpuinfo); 218 } 219 220 bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 221 sym_lookup_t lfn, script_lookup_t slfn) { 222 mSymLookupFn = lfn; 223 mScriptLookupFn = slfn; 224 225 lockMutex(); 226 if (!gThreadTLSKeyCount) { 227 int status = pthread_key_create(&gThreadTLSKey, nullptr); 228 if (status) { 229 ALOGE("Failed to init thread tls key."); 230 unlockMutex(); 231 return false; 232 } 233 } 234 gThreadTLSKeyCount++; 235 unlockMutex(); 236 237 mTlsStruct.mContext = mRSC; 238 mTlsStruct.mScript = nullptr; 239 int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 240 if (status) { 241 ALOGE("pthread_setspecific %i", status); 242 } 243 244 mPageSize = sysconf(_SC_PAGE_SIZE); 245 // ALOGV("page size = %ld", mPageSize); 246 247 GetCpuInfo(); 248 249 int cpu = sysconf(_SC_NPROCESSORS_CONF); 250 if(mRSC->props.mDebugMaxThreads) { 251 cpu = mRSC->props.mDebugMaxThreads; 252 } 253 if (cpu < 2) { 254 mWorkers.mCount = 0; 255 return true; 256 } 257 258 // Subtract one from the cpu count because we also use the command thread as a worker. 259 mWorkers.mCount = (uint32_t)(cpu - 1); 260 261 if (mRSC->props.mLogScripts) { 262 ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 263 } 264 265 mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 266 mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 267 mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 268 mWorkers.mLaunchCallback = nullptr; 269 270 mWorkers.mCompleteSignal.init(); 271 272 mWorkers.mRunningCount = mWorkers.mCount; 273 mWorkers.mLaunchCount = 0; 274 __sync_synchronize(); 275 276 pthread_attr_t threadAttr; 277 status = pthread_attr_init(&threadAttr); 278 if (status) { 279 ALOGE("Failed to init thread attribute."); 280 return false; 281 } 282 283 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 284 status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 285 if (status) { 286 mWorkers.mCount = ct; 287 ALOGE("Created fewer than expected number of RS threads."); 288 break; 289 } 290 } 291 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 292 usleep(100); 293 } 294 295 pthread_attr_destroy(&threadAttr); 296 return true; 297 } 298 299 300 void RsdCpuReferenceImpl::setPriority(int32_t priority) { 301 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 302 setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 303 } 304 } 305 306 RsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 307 mExit = true; 308 mWorkers.mLaunchData = nullptr; 309 mWorkers.mLaunchCallback = nullptr; 310 mWorkers.mRunningCount = mWorkers.mCount; 311 __sync_synchronize(); 312 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 313 mWorkers.mLaunchSignals[ct].set(); 314 } 315 void *res; 316 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 317 pthread_join(mWorkers.mThreadId[ct], &res); 318 } 319 rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 320 free(mWorkers.mThreadId); 321 free(mWorkers.mNativeThreadId); 322 delete[] mWorkers.mLaunchSignals; 323 324 // Global structure cleanup. 325 lockMutex(); 326 --gThreadTLSKeyCount; 327 if (!gThreadTLSKeyCount) { 328 pthread_key_delete(gThreadTLSKey); 329 } 330 unlockMutex(); 331 332 } 333 334 // Set up the appropriate input and output pointers to the kernel driver info structure. 335 // Inputs: 336 // mtls - The MTLaunchStruct holding information about the kernel launch 337 // fep - The forEach parameters (driver info structure) 338 // x, y, z, lod, face, a1, a2, a3, a4 - The start offsets into each dimension 339 static inline void FepPtrSetup(const MTLaunchStructForEach *mtls, RsExpandKernelDriverInfo *fep, 340 uint32_t x, uint32_t y, 341 uint32_t z = 0, uint32_t lod = 0, 342 RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X, 343 uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) { 344 for (uint32_t i = 0; i < fep->inLen; i++) { 345 fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 346 } 347 if (mtls->aout[0] != nullptr) { 348 fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 349 } 350 } 351 352 // Set up the appropriate input and output pointers to the kernel driver info structure. 353 // Inputs: 354 // mtls - The MTLaunchStruct holding information about the kernel launch 355 // redp - The reduce parameters (driver info structure) 356 // x, y, z - The start offsets into each dimension 357 static inline void RedpPtrSetup(const MTLaunchStructReduce *mtls, RsExpandKernelDriverInfo *redp, 358 uint32_t x, uint32_t y, uint32_t z) { 359 for (uint32_t i = 0; i < redp->inLen; i++) { 360 redp->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z); 361 } 362 } 363 364 static uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) { 365 if (start >= end) { 366 *p = start; 367 return val; 368 } 369 370 uint32_t div = end - start; 371 372 uint32_t n = val / div; 373 *p = (val - (n * div)) + start; 374 return n; 375 } 376 377 static bool SelectOuterSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) { 378 uint32_t r = sliceNum; 379 r = sliceInt(&info->current.z, r, mtls->start.z, mtls->end.z); 380 r = sliceInt(&info->current.lod, r, mtls->start.lod, mtls->end.lod); 381 r = sliceInt(&info->current.face, r, mtls->start.face, mtls->end.face); 382 r = sliceInt(&info->current.array[0], r, mtls->start.array[0], mtls->end.array[0]); 383 r = sliceInt(&info->current.array[1], r, mtls->start.array[1], mtls->end.array[1]); 384 r = sliceInt(&info->current.array[2], r, mtls->start.array[2], mtls->end.array[2]); 385 r = sliceInt(&info->current.array[3], r, mtls->start.array[3], mtls->end.array[3]); 386 return r == 0; 387 } 388 389 static bool SelectZSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) { 390 return sliceInt(&info->current.z, sliceNum, mtls->start.z, mtls->end.z) == 0; 391 } 392 393 static void walk_general_foreach(void *usr, uint32_t idx) { 394 MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 395 RsExpandKernelDriverInfo fep = mtls->fep; 396 fep.lid = idx; 397 ForEachFunc_t fn = mtls->kernel; 398 399 while(1) { 400 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 401 402 if (!SelectOuterSlice(mtls, &fep, slice)) { 403 return; 404 } 405 406 for (fep.current.y = mtls->start.y; fep.current.y < mtls->end.y; 407 fep.current.y++) { 408 409 FepPtrSetup(mtls, &fep, mtls->start.x, 410 fep.current.y, fep.current.z, fep.current.lod, 411 (RsAllocationCubemapFace)fep.current.face, 412 fep.current.array[0], fep.current.array[1], 413 fep.current.array[2], fep.current.array[3]); 414 415 fn(&fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 416 } 417 } 418 } 419 420 static void walk_2d_foreach(void *usr, uint32_t idx) { 421 MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 422 RsExpandKernelDriverInfo fep = mtls->fep; 423 fep.lid = idx; 424 ForEachFunc_t fn = mtls->kernel; 425 426 while (1) { 427 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 428 uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 429 uint32_t yEnd = yStart + mtls->mSliceSize; 430 431 yEnd = rsMin(yEnd, mtls->end.y); 432 433 if (yEnd <= yStart) { 434 return; 435 } 436 437 for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) { 438 FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y); 439 440 fn(&fep, mtls->start.x, mtls->end.x, fep.outStride[0]); 441 } 442 } 443 } 444 445 static void walk_1d_foreach(void *usr, uint32_t idx) { 446 MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 447 RsExpandKernelDriverInfo fep = mtls->fep; 448 fep.lid = idx; 449 ForEachFunc_t fn = mtls->kernel; 450 451 while (1) { 452 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 453 uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 454 uint32_t xEnd = xStart + mtls->mSliceSize; 455 456 xEnd = rsMin(xEnd, mtls->end.x); 457 458 if (xEnd <= xStart) { 459 return; 460 } 461 462 FepPtrSetup(mtls, &fep, xStart, 0); 463 464 fn(&fep, xStart, xEnd, fep.outStride[0]); 465 } 466 } 467 468 // The function format_bytes() is an auxiliary function to assist in logging. 469 // 470 // Bytes are read from an input (inBuf) and written (as pairs of hex digits) 471 // to an output (outBuf). 472 // 473 // Output format: 474 // - starts with ": " 475 // - each input byte is translated to a pair of hex digits 476 // - bytes are separated by "." except that every fourth separator is "|" 477 // - if the input is sufficiently long, the output is truncated and terminated with "..." 478 // 479 // Arguments: 480 // - outBuf -- Pointer to buffer of type "FormatBuf" into which output is written 481 // - inBuf -- Pointer to bytes which are to be formatted into outBuf 482 // - inBytes -- Number of bytes in inBuf 483 // 484 // Constant: 485 // - kFormatInBytesMax -- Only min(kFormatInBytesMax, inBytes) bytes will be read 486 // from inBuf 487 // 488 // Return value: 489 // - pointer (const char *) to output (which is part of outBuf) 490 // 491 static const int kFormatInBytesMax = 16; 492 // ": " + 2 digits per byte + 1 separator between bytes + "..." + null 493 typedef char FormatBuf[2 + kFormatInBytesMax*2 + (kFormatInBytesMax - 1) + 3 + 1]; 494 static const char *format_bytes(FormatBuf *outBuf, const uint8_t *inBuf, const int inBytes) { 495 strcpy(*outBuf, ": "); 496 int pos = 2; 497 const int lim = std::min(kFormatInBytesMax, inBytes); 498 for (int i = 0; i < lim; ++i) { 499 if (i) { 500 sprintf(*outBuf + pos, (i % 4 ? "." : "|")); 501 ++pos; 502 } 503 sprintf(*outBuf + pos, "%02x", inBuf[i]); 504 pos += 2; 505 } 506 if (kFormatInBytesMax < inBytes) 507 strcpy(*outBuf + pos, "..."); 508 return *outBuf; 509 } 510 511 static void reduce_get_accumulator(uint8_t *&accumPtr, const MTLaunchStructReduce *mtls, 512 const char *walkerName, uint32_t threadIdx) { 513 rsAssert(!accumPtr); 514 515 uint32_t accumIdx = (uint32_t)__sync_fetch_and_add(&mtls->accumCount, 1); 516 if (mtls->outFunc) { 517 accumPtr = mtls->accumAlloc + mtls->accumStride * accumIdx; 518 } else { 519 if (accumIdx == 0) { 520 accumPtr = mtls->redp.outPtr[0]; 521 } else { 522 accumPtr = mtls->accumAlloc + mtls->accumStride * (accumIdx - 1); 523 } 524 } 525 REDUCE_ALOGV(mtls, 2, "%s(%p): idx = %u got accumCount %u and accumPtr %p", 526 walkerName, mtls->accumFunc, threadIdx, accumIdx, accumPtr); 527 // initialize accumulator 528 if (mtls->initFunc) { 529 mtls->initFunc(accumPtr); 530 } else { 531 memset(accumPtr, 0, mtls->accumSize); 532 } 533 } 534 535 static void walk_1d_reduce(void *usr, uint32_t idx) { 536 const MTLaunchStructReduce *mtls = (const MTLaunchStructReduce *)usr; 537 RsExpandKernelDriverInfo redp = mtls->redp; 538 539 // find accumulator 540 uint8_t *&accumPtr = mtls->accumPtr[idx]; 541 if (!accumPtr) { 542 reduce_get_accumulator(accumPtr, mtls, __func__, idx); 543 } 544 545 // accumulate 546 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 547 while (1) { 548 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 549 uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 550 uint32_t xEnd = xStart + mtls->mSliceSize; 551 552 xEnd = rsMin(xEnd, mtls->end.x); 553 554 if (xEnd <= xStart) { 555 return; 556 } 557 558 RedpPtrSetup(mtls, &redp, xStart, 0, 0); 559 fn(&redp, xStart, xEnd, accumPtr); 560 561 // Emit log line after slice has been run, so that we can include 562 // the results of the run on that line. 563 FormatBuf fmt; 564 if (mtls->logReduce >= 3) { 565 format_bytes(&fmt, accumPtr, mtls->accumSize); 566 } else { 567 fmt[0] = 0; 568 } 569 REDUCE_ALOGV(mtls, 2, "walk_1d_reduce(%p): idx = %u, x in [%u, %u)%s", 570 mtls->accumFunc, idx, xStart, xEnd, fmt); 571 } 572 } 573 574 static void walk_2d_reduce(void *usr, uint32_t idx) { 575 const MTLaunchStructReduce *mtls = (const MTLaunchStructReduce *)usr; 576 RsExpandKernelDriverInfo redp = mtls->redp; 577 578 // find accumulator 579 uint8_t *&accumPtr = mtls->accumPtr[idx]; 580 if (!accumPtr) { 581 reduce_get_accumulator(accumPtr, mtls, __func__, idx); 582 } 583 584 // accumulate 585 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 586 while (1) { 587 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 588 uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 589 uint32_t yEnd = yStart + mtls->mSliceSize; 590 591 yEnd = rsMin(yEnd, mtls->end.y); 592 593 if (yEnd <= yStart) { 594 return; 595 } 596 597 for (redp.current.y = yStart; redp.current.y < yEnd; redp.current.y++) { 598 RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, 0); 599 fn(&redp, mtls->start.x, mtls->end.x, accumPtr); 600 } 601 602 FormatBuf fmt; 603 if (mtls->logReduce >= 3) { 604 format_bytes(&fmt, accumPtr, mtls->accumSize); 605 } else { 606 fmt[0] = 0; 607 } 608 REDUCE_ALOGV(mtls, 2, "walk_2d_reduce(%p): idx = %u, y in [%u, %u)%s", 609 mtls->accumFunc, idx, yStart, yEnd, fmt); 610 } 611 } 612 613 static void walk_3d_reduce(void *usr, uint32_t idx) { 614 const MTLaunchStructReduce *mtls = (const MTLaunchStructReduce *)usr; 615 RsExpandKernelDriverInfo redp = mtls->redp; 616 617 // find accumulator 618 uint8_t *&accumPtr = mtls->accumPtr[idx]; 619 if (!accumPtr) { 620 reduce_get_accumulator(accumPtr, mtls, __func__, idx); 621 } 622 623 // accumulate 624 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 625 while (1) { 626 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 627 628 if (!SelectZSlice(mtls, &redp, slice)) { 629 return; 630 } 631 632 for (redp.current.y = mtls->start.y; redp.current.y < mtls->end.y; redp.current.y++) { 633 RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, redp.current.z); 634 fn(&redp, mtls->start.x, mtls->end.x, accumPtr); 635 } 636 637 FormatBuf fmt; 638 if (mtls->logReduce >= 3) { 639 format_bytes(&fmt, accumPtr, mtls->accumSize); 640 } else { 641 fmt[0] = 0; 642 } 643 REDUCE_ALOGV(mtls, 2, "walk_3d_reduce(%p): idx = %u, z = %u%s", 644 mtls->accumFunc, idx, redp.current.z, fmt); 645 } 646 } 647 648 // Launch a general reduce-style kernel. 649 // Inputs: 650 // ains[0..inLen-1]: Array of allocations that contain the inputs 651 // aout: The allocation that will hold the output 652 // mtls: Holds launch parameters 653 void RsdCpuReferenceImpl::launchReduce(const Allocation ** ains, 654 uint32_t inLen, 655 Allocation * aout, 656 MTLaunchStructReduce *mtls) { 657 mtls->logReduce = mRSC->props.mLogReduce; 658 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 659 launchReduceParallel(ains, inLen, aout, mtls); 660 } else { 661 launchReduceSerial(ains, inLen, aout, mtls); 662 } 663 } 664 665 // Launch a general reduce-style kernel, single-threaded. 666 // Inputs: 667 // ains[0..inLen-1]: Array of allocations that contain the inputs 668 // aout: The allocation that will hold the output 669 // mtls: Holds launch parameters 670 void RsdCpuReferenceImpl::launchReduceSerial(const Allocation ** ains, 671 uint32_t inLen, 672 Allocation * aout, 673 MTLaunchStructReduce *mtls) { 674 REDUCE_ALOGV(mtls, 1, "launchReduceSerial(%p): %u x %u x %u", mtls->accumFunc, 675 mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z); 676 677 // In the presence of outconverter, we allocate temporary memory for 678 // the accumulator. 679 // 680 // In the absence of outconverter, we use the output allocation as the 681 // accumulator. 682 uint8_t *const accumPtr = (mtls->outFunc 683 ? static_cast<uint8_t *>(malloc(mtls->accumSize)) 684 : mtls->redp.outPtr[0]); 685 686 // initialize 687 if (mtls->initFunc) { 688 mtls->initFunc(accumPtr); 689 } else { 690 memset(accumPtr, 0, mtls->accumSize); 691 } 692 693 // accumulate 694 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 695 uint32_t slice = 0; 696 while (SelectOuterSlice(mtls, &mtls->redp, slice++)) { 697 for (mtls->redp.current.y = mtls->start.y; 698 mtls->redp.current.y < mtls->end.y; 699 mtls->redp.current.y++) { 700 RedpPtrSetup(mtls, &mtls->redp, mtls->start.x, mtls->redp.current.y, mtls->redp.current.z); 701 fn(&mtls->redp, mtls->start.x, mtls->end.x, accumPtr); 702 } 703 } 704 705 // outconvert 706 if (mtls->outFunc) { 707 mtls->outFunc(mtls->redp.outPtr[0], accumPtr); 708 free(accumPtr); 709 } 710 } 711 712 // Launch a general reduce-style kernel, multi-threaded. 713 // Inputs: 714 // ains[0..inLen-1]: Array of allocations that contain the inputs 715 // aout: The allocation that will hold the output 716 // mtls: Holds launch parameters 717 void RsdCpuReferenceImpl::launchReduceParallel(const Allocation ** ains, 718 uint32_t inLen, 719 Allocation * aout, 720 MTLaunchStructReduce *mtls) { 721 // For now, we don't know how to go parallel in the absence of a combiner. 722 if (!mtls->combFunc) { 723 launchReduceSerial(ains, inLen, aout, mtls); 724 return; 725 } 726 727 // Number of threads = "main thread" + number of other (worker) threads 728 const uint32_t numThreads = mWorkers.mCount + 1; 729 730 // In the absence of outconverter, we use the output allocation as 731 // an accumulator, and therefore need to allocate one fewer accumulator. 732 const uint32_t numAllocAccum = numThreads - (mtls->outFunc == nullptr); 733 734 // If mDebugReduceSplitAccum, then we want each accumulator to start 735 // on a page boundary. (TODO: Would some unit smaller than a page 736 // be sufficient to avoid false sharing?) 737 if (mRSC->props.mDebugReduceSplitAccum) { 738 // Round up accumulator size to an integral number of pages 739 mtls->accumStride = 740 (unsigned(mtls->accumSize) + unsigned(mPageSize)-1) & 741 ~(unsigned(mPageSize)-1); 742 // Each accumulator gets its own page. Alternatively, if we just 743 // wanted to make sure no two accumulators are on the same page, 744 // we could instead do 745 // allocSize = mtls->accumStride * (numAllocation - 1) + mtls->accumSize 746 const size_t allocSize = mtls->accumStride * numAllocAccum; 747 mtls->accumAlloc = static_cast<uint8_t *>(memalign(mPageSize, allocSize)); 748 } else { 749 mtls->accumStride = mtls->accumSize; 750 mtls->accumAlloc = static_cast<uint8_t *>(malloc(mtls->accumStride * numAllocAccum)); 751 } 752 753 const size_t accumPtrArrayBytes = sizeof(uint8_t *) * numThreads; 754 mtls->accumPtr = static_cast<uint8_t **>(malloc(accumPtrArrayBytes)); 755 memset(mtls->accumPtr, 0, accumPtrArrayBytes); 756 757 mtls->accumCount = 0; 758 759 rsAssert(!mInKernel); 760 mInKernel = true; 761 REDUCE_ALOGV(mtls, 1, "launchReduceParallel(%p): %u x %u x %u, %u threads, accumAlloc = %p", 762 mtls->accumFunc, 763 mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z, 764 numThreads, mtls->accumAlloc); 765 if (mtls->redp.dim.z > 1) { 766 mtls->mSliceSize = 1; 767 launchThreads(walk_3d_reduce, mtls); 768 } else if (mtls->redp.dim.y > 1) { 769 mtls->mSliceSize = rsMax(1U, mtls->redp.dim.y / (numThreads * 4)); 770 launchThreads(walk_2d_reduce, mtls); 771 } else { 772 mtls->mSliceSize = rsMax(1U, mtls->redp.dim.x / (numThreads * 4)); 773 launchThreads(walk_1d_reduce, mtls); 774 } 775 mInKernel = false; 776 777 // Combine accumulators and identify final accumulator 778 uint8_t *finalAccumPtr = (mtls->outFunc ? nullptr : mtls->redp.outPtr[0]); 779 // Loop over accumulators, combining into finalAccumPtr. If finalAccumPtr 780 // is null, then the first accumulator I find becomes finalAccumPtr. 781 for (unsigned idx = 0; idx < mtls->accumCount; ++idx) { 782 uint8_t *const thisAccumPtr = mtls->accumPtr[idx]; 783 if (finalAccumPtr) { 784 if (finalAccumPtr != thisAccumPtr) { 785 if (mtls->combFunc) { 786 if (mtls->logReduce >= 3) { 787 FormatBuf fmt; 788 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): accumulating into%s", 789 mtls->accumFunc, 790 format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 791 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): accumulator[%d]%s", 792 mtls->accumFunc, idx, 793 format_bytes(&fmt, thisAccumPtr, mtls->accumSize)); 794 } 795 mtls->combFunc(finalAccumPtr, thisAccumPtr); 796 } else { 797 rsAssert(!"expected combiner"); 798 } 799 } 800 } else { 801 finalAccumPtr = thisAccumPtr; 802 } 803 } 804 rsAssert(finalAccumPtr != nullptr); 805 if (mtls->logReduce >= 3) { 806 FormatBuf fmt; 807 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): final accumulator%s", 808 mtls->accumFunc, format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 809 } 810 811 // Outconvert 812 if (mtls->outFunc) { 813 mtls->outFunc(mtls->redp.outPtr[0], finalAccumPtr); 814 if (mtls->logReduce >= 3) { 815 FormatBuf fmt; 816 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): final outconverted result%s", 817 mtls->accumFunc, 818 format_bytes(&fmt, mtls->redp.outPtr[0], mtls->redp.outStride[0])); 819 } 820 } 821 822 // Clean up 823 free(mtls->accumPtr); 824 free(mtls->accumAlloc); 825 } 826 827 828 void RsdCpuReferenceImpl::launchForEach(const Allocation ** ains, 829 uint32_t inLen, 830 Allocation* aout, 831 const RsScriptCall* sc, 832 MTLaunchStructForEach* mtls) { 833 834 //android::StopWatch kernel_time("kernel time"); 835 836 bool outerDims = (mtls->start.z != mtls->end.z) || 837 (mtls->start.face != mtls->end.face) || 838 (mtls->start.lod != mtls->end.lod) || 839 (mtls->start.array[0] != mtls->end.array[0]) || 840 (mtls->start.array[1] != mtls->end.array[1]) || 841 (mtls->start.array[2] != mtls->end.array[2]) || 842 (mtls->start.array[3] != mtls->end.array[3]); 843 844 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 845 const size_t targetByteChunk = 16 * 1024; 846 mInKernel = true; // NOTE: The guard immediately above ensures this was !mInKernel 847 848 if (outerDims) { 849 // No fancy logic for chunk size 850 mtls->mSliceSize = 1; 851 launchThreads(walk_general_foreach, mtls); 852 } else if (mtls->fep.dim.y > 1) { 853 uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4); 854 uint32_t s2 = 0; 855 856 // This chooses our slice size to rate limit atomic ops to 857 // one per 16k bytes of reads/writes. 858 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) { 859 s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride; 860 } else if (mtls->ains[0]) { 861 s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride; 862 } else { 863 // Launch option only case 864 // Use s1 based only on the dimensions 865 s2 = s1; 866 } 867 mtls->mSliceSize = rsMin(s1, s2); 868 869 if(mtls->mSliceSize < 1) { 870 mtls->mSliceSize = 1; 871 } 872 873 launchThreads(walk_2d_foreach, mtls); 874 } else { 875 uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4); 876 uint32_t s2 = 0; 877 878 // This chooses our slice size to rate limit atomic ops to 879 // one per 16k bytes of reads/writes. 880 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) { 881 s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes(); 882 } else if (mtls->ains[0]) { 883 s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes(); 884 } else { 885 // Launch option only case 886 // Use s1 based only on the dimensions 887 s2 = s1; 888 } 889 mtls->mSliceSize = rsMin(s1, s2); 890 891 if (mtls->mSliceSize < 1) { 892 mtls->mSliceSize = 1; 893 } 894 895 launchThreads(walk_1d_foreach, mtls); 896 } 897 mInKernel = false; 898 899 } else { 900 ForEachFunc_t fn = mtls->kernel; 901 uint32_t slice = 0; 902 903 904 while(SelectOuterSlice(mtls, &mtls->fep, slice++)) { 905 for (mtls->fep.current.y = mtls->start.y; 906 mtls->fep.current.y < mtls->end.y; 907 mtls->fep.current.y++) { 908 909 FepPtrSetup(mtls, &mtls->fep, mtls->start.x, 910 mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod, 911 (RsAllocationCubemapFace) mtls->fep.current.face, 912 mtls->fep.current.array[0], mtls->fep.current.array[1], 913 mtls->fep.current.array[2], mtls->fep.current.array[3]); 914 915 fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 916 } 917 } 918 } 919 } 920 921 RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 922 //ALOGE("setTls %p", sc); 923 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 924 rsAssert(tls); 925 RsdCpuScriptImpl *old = tls->mImpl; 926 tls->mImpl = sc; 927 tls->mContext = mRSC; 928 if (sc) { 929 tls->mScript = sc->getScript(); 930 } else { 931 tls->mScript = nullptr; 932 } 933 return old; 934 } 935 936 const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 937 return mSymLookupFn(mRSC, name); 938 } 939 940 941 RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 942 char const *resName, char const *cacheDir, 943 uint8_t const *bitcode, size_t bitcodeSize, 944 uint32_t flags) { 945 946 RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 947 if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 948 , getBccPluginName() 949 )) { 950 delete i; 951 return nullptr; 952 } 953 return i; 954 } 955 956 extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 957 const Script *s, const Element *e); 958 extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 959 const Script *s, const Element *e); 960 extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 961 const Script *s, const Element *e); 962 extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 963 const Script *s, const Element *e); 964 extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 965 const Script *s, const Element *e); 966 extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 967 const Script *s, const Element *e); 968 extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 969 const Script *s, const Element *e); 970 extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 971 const Script *s, const Element *e); 972 extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 973 const Script *s, const Element *e); 974 extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 975 const Script *s, const Element *e); 976 extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx, 977 const Script *s, const Element *e); 978 979 RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 980 RsScriptIntrinsicID iid, Element *e) { 981 982 RsdCpuScriptImpl *i = nullptr; 983 switch (iid) { 984 case RS_SCRIPT_INTRINSIC_ID_3DLUT: 985 i = rsdIntrinsic_3DLUT(this, s, e); 986 break; 987 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 988 i = rsdIntrinsic_Convolve3x3(this, s, e); 989 break; 990 case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 991 i = rsdIntrinsic_ColorMatrix(this, s, e); 992 break; 993 case RS_SCRIPT_INTRINSIC_ID_LUT: 994 i = rsdIntrinsic_LUT(this, s, e); 995 break; 996 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 997 i = rsdIntrinsic_Convolve5x5(this, s, e); 998 break; 999 case RS_SCRIPT_INTRINSIC_ID_BLUR: 1000 i = rsdIntrinsic_Blur(this, s, e); 1001 break; 1002 case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 1003 i = rsdIntrinsic_YuvToRGB(this, s, e); 1004 break; 1005 case RS_SCRIPT_INTRINSIC_ID_BLEND: 1006 i = rsdIntrinsic_Blend(this, s, e); 1007 break; 1008 case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 1009 i = rsdIntrinsic_Histogram(this, s, e); 1010 break; 1011 case RS_SCRIPT_INTRINSIC_ID_RESIZE: 1012 i = rsdIntrinsic_Resize(this, s, e); 1013 break; 1014 case RS_SCRIPT_INTRINSIC_ID_BLAS: 1015 i = rsdIntrinsic_BLAS(this, s, e); 1016 break; 1017 1018 default: 1019 rsAssert(0); 1020 } 1021 1022 return i; 1023 } 1024 1025 void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) { 1026 switch (sg->getApiVersion()) { 1027 case ScriptGroupBase::SG_V1: { 1028 CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 1029 if (!sgi->init()) { 1030 delete sgi; 1031 return nullptr; 1032 } 1033 return sgi; 1034 } 1035 case ScriptGroupBase::SG_V2: { 1036 return new CpuScriptGroup2Impl(this, sg); 1037 } 1038 } 1039 return nullptr; 1040 } 1041
