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#define REDUCE_ALOGV(mtls, level, ...) do { if ((mtls)->logReduce >= (level)) ALOGV(__VA_ARGS__); } while(0) 34 35static pthread_key_t gThreadTLSKey = 0; 36static uint32_t gThreadTLSKeyCount = 0; 37static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER; 38 39namespace android { 40namespace renderscript { 41 42bool gArchUseSIMD = false; 43 44RsdCpuReference::~RsdCpuReference() { 45} 46 47RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major, 48 uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn 49 , RSSelectRTCallback pSelectRTCallback, 50 const char *pBccPluginName 51 ) { 52 53 RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc); 54 if (!cpu) { 55 return nullptr; 56 } 57 if (!cpu->init(version_major, version_minor, lfn, slfn)) { 58 delete cpu; 59 return nullptr; 60 } 61 62 cpu->setSelectRTCallback(pSelectRTCallback); 63 if (pBccPluginName) { 64 cpu->setBccPluginName(pBccPluginName); 65 } 66 67 return cpu; 68} 69 70 71Context * RsdCpuReference::getTlsContext() { 72 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 73 return tls->mContext; 74} 75 76const Script * RsdCpuReference::getTlsScript() { 77 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 78 return tls->mScript; 79} 80 81pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 82 83//////////////////////////////////////////////////////////// 84/// 85 86RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) { 87 mRSC = rsc; 88 89 version_major = 0; 90 version_minor = 0; 91 mInKernel = false; 92 memset(&mWorkers, 0, sizeof(mWorkers)); 93 memset(&mTlsStruct, 0, sizeof(mTlsStruct)); 94 mExit = false; 95 mSelectRTCallback = nullptr; 96 mEmbedGlobalInfo = true; 97 mEmbedGlobalInfoSkipConstant = true; 98} 99 100 101void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 102 RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 103 104 uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 105 106 //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 107 108 dc->mWorkers.mLaunchSignals[idx].init(); 109 dc->mWorkers.mNativeThreadId[idx] = gettid(); 110 111 memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 112 int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 113 if (status) { 114 ALOGE("pthread_setspecific %i", status); 115 } 116 117#if 0 118 typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 119 cpu_set_t cpuset; 120 memset(&cpuset, 0, sizeof(cpuset)); 121 cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 122 int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 123 sizeof(cpuset), &cpuset); 124 ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 125#endif 126 127 while (!dc->mExit) { 128 dc->mWorkers.mLaunchSignals[idx].wait(); 129 if (dc->mWorkers.mLaunchCallback) { 130 // idx +1 is used because the calling thread is always worker 0. 131 dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 132 } 133 __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 134 dc->mWorkers.mCompleteSignal.set(); 135 } 136 137 //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 138 return nullptr; 139} 140 141// Launch a kernel. 142// The callback function is called to execute the kernel. 143void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 144 mWorkers.mLaunchData = data; 145 mWorkers.mLaunchCallback = cbk; 146 147 // fast path for very small launches 148 MTLaunchStructCommon *mtls = (MTLaunchStructCommon *)data; 149 if (mtls && mtls->dimPtr->y <= 1 && mtls->end.x <= mtls->start.x + mtls->mSliceSize) { 150 if (mWorkers.mLaunchCallback) { 151 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 152 } 153 return; 154 } 155 156 mWorkers.mRunningCount = mWorkers.mCount; 157 __sync_synchronize(); 158 159 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 160 mWorkers.mLaunchSignals[ct].set(); 161 } 162 163 // We use the calling thread as one of the workers so we can start without 164 // the delay of the thread wakeup. 165 if (mWorkers.mLaunchCallback) { 166 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 167 } 168 169 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 170 mWorkers.mCompleteSignal.wait(); 171 } 172} 173 174 175void RsdCpuReferenceImpl::lockMutex() { 176 pthread_mutex_lock(&gInitMutex); 177} 178 179void RsdCpuReferenceImpl::unlockMutex() { 180 pthread_mutex_unlock(&gInitMutex); 181} 182 183// Determine if the CPU we're running on supports SIMD instructions. 184static void GetCpuInfo() { 185 // Read the CPU flags from /proc/cpuinfo. 186 FILE *cpuinfo = fopen("/proc/cpuinfo", "r"); 187 188 if (!cpuinfo) { 189 return; 190 } 191 192 char cpuinfostr[4096]; 193 // fgets() ends with newline or EOF, need to check the whole 194 // "cpuinfo" file to make sure we can use SIMD or not. 195 while (fgets(cpuinfostr, sizeof(cpuinfostr), cpuinfo)) { 196#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS) 197 gArchUseSIMD = strstr(cpuinfostr, " neon") || strstr(cpuinfostr, " asimd"); 198#elif defined(ARCH_X86_HAVE_SSSE3) 199 gArchUseSIMD = strstr(cpuinfostr, " ssse3"); 200#endif 201 if (gArchUseSIMD) { 202 break; 203 } 204 } 205 fclose(cpuinfo); 206} 207 208bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 209 sym_lookup_t lfn, script_lookup_t slfn) { 210 mSymLookupFn = lfn; 211 mScriptLookupFn = slfn; 212 213 lockMutex(); 214 if (!gThreadTLSKeyCount) { 215 int status = pthread_key_create(&gThreadTLSKey, nullptr); 216 if (status) { 217 ALOGE("Failed to init thread tls key."); 218 unlockMutex(); 219 return false; 220 } 221 } 222 gThreadTLSKeyCount++; 223 unlockMutex(); 224 225 mTlsStruct.mContext = mRSC; 226 mTlsStruct.mScript = nullptr; 227 int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 228 if (status) { 229 ALOGE("pthread_setspecific %i", status); 230 } 231 232 mPageSize = sysconf(_SC_PAGE_SIZE); 233 // ALOGV("page size = %ld", mPageSize); 234 235 GetCpuInfo(); 236 237 int cpu = sysconf(_SC_NPROCESSORS_CONF); 238 if(mRSC->props.mDebugMaxThreads) { 239 cpu = mRSC->props.mDebugMaxThreads; 240 } 241 if (cpu < 2) { 242 mWorkers.mCount = 0; 243 return true; 244 } 245 246 // Subtract one from the cpu count because we also use the command thread as a worker. 247 mWorkers.mCount = (uint32_t)(cpu - 1); 248 249 if (mRSC->props.mLogScripts) { 250 ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 251 } 252 253 mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 254 mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 255 mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 256 mWorkers.mLaunchCallback = nullptr; 257 258 mWorkers.mCompleteSignal.init(); 259 260 mWorkers.mRunningCount = mWorkers.mCount; 261 mWorkers.mLaunchCount = 0; 262 __sync_synchronize(); 263 264 pthread_attr_t threadAttr; 265 status = pthread_attr_init(&threadAttr); 266 if (status) { 267 ALOGE("Failed to init thread attribute."); 268 return false; 269 } 270 271 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 272 status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 273 if (status) { 274 mWorkers.mCount = ct; 275 ALOGE("Created fewer than expected number of RS threads."); 276 break; 277 } 278 } 279 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 280 usleep(100); 281 } 282 283 pthread_attr_destroy(&threadAttr); 284 return true; 285} 286 287 288void RsdCpuReferenceImpl::setPriority(int32_t priority) { 289 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 290 setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 291 } 292} 293 294RsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 295 mExit = true; 296 mWorkers.mLaunchData = nullptr; 297 mWorkers.mLaunchCallback = nullptr; 298 mWorkers.mRunningCount = mWorkers.mCount; 299 __sync_synchronize(); 300 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 301 mWorkers.mLaunchSignals[ct].set(); 302 } 303 void *res; 304 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 305 pthread_join(mWorkers.mThreadId[ct], &res); 306 } 307 // b/23109602 308 // TODO: Refactor the implementation with threadpool to 309 // fix the race condition in the destuctor. 310 // rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 311 free(mWorkers.mThreadId); 312 free(mWorkers.mNativeThreadId); 313 delete[] mWorkers.mLaunchSignals; 314 315 // Global structure cleanup. 316 lockMutex(); 317 --gThreadTLSKeyCount; 318 if (!gThreadTLSKeyCount) { 319 pthread_key_delete(gThreadTLSKey); 320 } 321 unlockMutex(); 322 323} 324 325// Set up the appropriate input and output pointers to the kernel driver info structure. 326// Inputs: 327// mtls - The MTLaunchStruct holding information about the kernel launch 328// fep - The forEach parameters (driver info structure) 329// x, y, z, lod, face, a1, a2, a3, a4 - The start offsets into each dimension 330static inline void FepPtrSetup(const MTLaunchStructForEach *mtls, RsExpandKernelDriverInfo *fep, 331 uint32_t x, uint32_t y, 332 uint32_t z = 0, uint32_t lod = 0, 333 RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X, 334 uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) { 335 // When rsForEach passes a null input allocation (as opposed to no input), 336 // fep->inLen can be 1 with mtls->ains[0] being null. 337 // This should only happen on old style kernels. 338 for (uint32_t i = 0; i < fep->inLen; i++) { 339 if (mtls->ains[i] == nullptr) { 340 rsAssert(fep->inLen == 1); 341 continue; 342 } 343 fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 344 } 345 if (mtls->aout[0] != nullptr) { 346 fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 347 } 348} 349 350// Set up the appropriate input and output pointers to the kernel driver info structure. 351// Inputs: 352// mtls - The MTLaunchStruct holding information about the kernel launch 353// redp - The reduce parameters (driver info structure) 354// x, y, z - The start offsets into each dimension 355static inline void RedpPtrSetup(const MTLaunchStructReduce *mtls, RsExpandKernelDriverInfo *redp, 356 uint32_t x, uint32_t y, uint32_t z) { 357 for (uint32_t i = 0; i < redp->inLen; i++) { 358 redp->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z); 359 } 360} 361 362static uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) { 363 if (start >= end) { 364 *p = start; 365 return val; 366 } 367 368 uint32_t div = end - start; 369 370 uint32_t n = val / div; 371 *p = (val - (n * div)) + start; 372 return n; 373} 374 375static bool SelectOuterSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) { 376 uint32_t r = sliceNum; 377 r = sliceInt(&info->current.z, r, mtls->start.z, mtls->end.z); 378 r = sliceInt(&info->current.lod, r, mtls->start.lod, mtls->end.lod); 379 r = sliceInt(&info->current.face, r, mtls->start.face, mtls->end.face); 380 r = sliceInt(&info->current.array[0], r, mtls->start.array[0], mtls->end.array[0]); 381 r = sliceInt(&info->current.array[1], r, mtls->start.array[1], mtls->end.array[1]); 382 r = sliceInt(&info->current.array[2], r, mtls->start.array[2], mtls->end.array[2]); 383 r = sliceInt(&info->current.array[3], r, mtls->start.array[3], mtls->end.array[3]); 384 return r == 0; 385} 386 387static bool SelectZSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) { 388 return sliceInt(&info->current.z, sliceNum, mtls->start.z, mtls->end.z) == 0; 389} 390 391static void walk_general_foreach(void *usr, uint32_t idx) { 392 MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 393 RsExpandKernelDriverInfo fep = mtls->fep; 394 fep.lid = idx; 395 ForEachFunc_t fn = mtls->kernel; 396 397 while(1) { 398 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 399 400 if (!SelectOuterSlice(mtls, &fep, slice)) { 401 return; 402 } 403 404 for (fep.current.y = mtls->start.y; fep.current.y < mtls->end.y; 405 fep.current.y++) { 406 407 FepPtrSetup(mtls, &fep, mtls->start.x, 408 fep.current.y, fep.current.z, fep.current.lod, 409 (RsAllocationCubemapFace)fep.current.face, 410 fep.current.array[0], fep.current.array[1], 411 fep.current.array[2], fep.current.array[3]); 412 413 fn(&fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 414 } 415 } 416} 417 418static void walk_2d_foreach(void *usr, uint32_t idx) { 419 MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 420 RsExpandKernelDriverInfo fep = mtls->fep; 421 fep.lid = idx; 422 ForEachFunc_t fn = mtls->kernel; 423 424 while (1) { 425 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 426 uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 427 uint32_t yEnd = yStart + mtls->mSliceSize; 428 429 yEnd = rsMin(yEnd, mtls->end.y); 430 431 if (yEnd <= yStart) { 432 return; 433 } 434 435 for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) { 436 FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y); 437 438 fn(&fep, mtls->start.x, mtls->end.x, fep.outStride[0]); 439 } 440 } 441} 442 443static void walk_1d_foreach(void *usr, uint32_t idx) { 444 MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 445 RsExpandKernelDriverInfo fep = mtls->fep; 446 fep.lid = idx; 447 ForEachFunc_t fn = mtls->kernel; 448 449 while (1) { 450 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 451 uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 452 uint32_t xEnd = xStart + mtls->mSliceSize; 453 454 xEnd = rsMin(xEnd, mtls->end.x); 455 456 if (xEnd <= xStart) { 457 return; 458 } 459 460 FepPtrSetup(mtls, &fep, xStart, 0); 461 462 fn(&fep, xStart, xEnd, fep.outStride[0]); 463 } 464} 465 466// The function format_bytes() is an auxiliary function to assist in logging. 467// 468// Bytes are read from an input (inBuf) and written (as pairs of hex digits) 469// to an output (outBuf). 470// 471// Output format: 472// - starts with ": " 473// - each input byte is translated to a pair of hex digits 474// - bytes are separated by "." except that every fourth separator is "|" 475// - if the input is sufficiently long, the output is truncated and terminated with "..." 476// 477// Arguments: 478// - outBuf -- Pointer to buffer of type "FormatBuf" into which output is written 479// - inBuf -- Pointer to bytes which are to be formatted into outBuf 480// - inBytes -- Number of bytes in inBuf 481// 482// Constant: 483// - kFormatInBytesMax -- Only min(kFormatInBytesMax, inBytes) bytes will be read 484// from inBuf 485// 486// Return value: 487// - pointer (const char *) to output (which is part of outBuf) 488// 489static const int kFormatInBytesMax = 16; 490// ": " + 2 digits per byte + 1 separator between bytes + "..." + null 491typedef char FormatBuf[2 + kFormatInBytesMax*2 + (kFormatInBytesMax - 1) + 3 + 1]; 492static const char *format_bytes(FormatBuf *outBuf, const uint8_t *inBuf, const int inBytes) { 493 strlcpy(*outBuf, ": ", sizeof(*outBuf)); 494 int pos = 2; 495 const int lim = std::min(kFormatInBytesMax, inBytes); 496 for (int i = 0; i < lim; ++i) { 497 if (i) { 498 sprintf(*outBuf + pos, (i % 4 ? "." : "|")); 499 ++pos; 500 } 501 sprintf(*outBuf + pos, "%02x", inBuf[i]); 502 pos += 2; 503 } 504 if (kFormatInBytesMax < inBytes) 505 strlcpy(*outBuf + pos, "...", sizeof(FormatBuf) - pos); 506 return *outBuf; 507} 508 509static void reduce_get_accumulator(uint8_t *&accumPtr, const MTLaunchStructReduce *mtls, 510 const char *walkerName, uint32_t threadIdx) { 511 rsAssert(!accumPtr); 512 513 uint32_t accumIdx = (uint32_t)__sync_fetch_and_add(&mtls->accumCount, 1); 514 if (mtls->outFunc) { 515 accumPtr = mtls->accumAlloc + mtls->accumStride * accumIdx; 516 } else { 517 if (accumIdx == 0) { 518 accumPtr = mtls->redp.outPtr[0]; 519 } else { 520 accumPtr = mtls->accumAlloc + mtls->accumStride * (accumIdx - 1); 521 } 522 } 523 REDUCE_ALOGV(mtls, 2, "%s(%p): idx = %u got accumCount %u and accumPtr %p", 524 walkerName, mtls->accumFunc, threadIdx, accumIdx, accumPtr); 525 // initialize accumulator 526 if (mtls->initFunc) { 527 mtls->initFunc(accumPtr); 528 } else { 529 memset(accumPtr, 0, mtls->accumSize); 530 } 531} 532 533static void walk_1d_reduce(void *usr, uint32_t idx) { 534 const MTLaunchStructReduce *mtls = (const MTLaunchStructReduce *)usr; 535 RsExpandKernelDriverInfo redp = mtls->redp; 536 537 // find accumulator 538 uint8_t *&accumPtr = mtls->accumPtr[idx]; 539 if (!accumPtr) { 540 reduce_get_accumulator(accumPtr, mtls, __func__, idx); 541 } 542 543 // accumulate 544 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 545 while (1) { 546 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 547 uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 548 uint32_t xEnd = xStart + mtls->mSliceSize; 549 550 xEnd = rsMin(xEnd, mtls->end.x); 551 552 if (xEnd <= xStart) { 553 return; 554 } 555 556 RedpPtrSetup(mtls, &redp, xStart, 0, 0); 557 fn(&redp, xStart, xEnd, accumPtr); 558 559 // Emit log line after slice has been run, so that we can include 560 // the results of the run on that line. 561 FormatBuf fmt; 562 if (mtls->logReduce >= 3) { 563 format_bytes(&fmt, accumPtr, mtls->accumSize); 564 } else { 565 fmt[0] = 0; 566 } 567 REDUCE_ALOGV(mtls, 2, "walk_1d_reduce(%p): idx = %u, x in [%u, %u)%s", 568 mtls->accumFunc, idx, xStart, xEnd, fmt); 569 } 570} 571 572static void walk_2d_reduce(void *usr, uint32_t idx) { 573 const MTLaunchStructReduce *mtls = (const MTLaunchStructReduce *)usr; 574 RsExpandKernelDriverInfo redp = mtls->redp; 575 576 // find accumulator 577 uint8_t *&accumPtr = mtls->accumPtr[idx]; 578 if (!accumPtr) { 579 reduce_get_accumulator(accumPtr, mtls, __func__, idx); 580 } 581 582 // accumulate 583 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 584 while (1) { 585 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 586 uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 587 uint32_t yEnd = yStart + mtls->mSliceSize; 588 589 yEnd = rsMin(yEnd, mtls->end.y); 590 591 if (yEnd <= yStart) { 592 return; 593 } 594 595 for (redp.current.y = yStart; redp.current.y < yEnd; redp.current.y++) { 596 RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, 0); 597 fn(&redp, mtls->start.x, mtls->end.x, accumPtr); 598 } 599 600 FormatBuf fmt; 601 if (mtls->logReduce >= 3) { 602 format_bytes(&fmt, accumPtr, mtls->accumSize); 603 } else { 604 fmt[0] = 0; 605 } 606 REDUCE_ALOGV(mtls, 2, "walk_2d_reduce(%p): idx = %u, y in [%u, %u)%s", 607 mtls->accumFunc, idx, yStart, yEnd, fmt); 608 } 609} 610 611static void walk_3d_reduce(void *usr, uint32_t idx) { 612 const MTLaunchStructReduce *mtls = (const MTLaunchStructReduce *)usr; 613 RsExpandKernelDriverInfo redp = mtls->redp; 614 615 // find accumulator 616 uint8_t *&accumPtr = mtls->accumPtr[idx]; 617 if (!accumPtr) { 618 reduce_get_accumulator(accumPtr, mtls, __func__, idx); 619 } 620 621 // accumulate 622 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 623 while (1) { 624 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 625 626 if (!SelectZSlice(mtls, &redp, slice)) { 627 return; 628 } 629 630 for (redp.current.y = mtls->start.y; redp.current.y < mtls->end.y; redp.current.y++) { 631 RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, redp.current.z); 632 fn(&redp, mtls->start.x, mtls->end.x, accumPtr); 633 } 634 635 FormatBuf fmt; 636 if (mtls->logReduce >= 3) { 637 format_bytes(&fmt, accumPtr, mtls->accumSize); 638 } else { 639 fmt[0] = 0; 640 } 641 REDUCE_ALOGV(mtls, 2, "walk_3d_reduce(%p): idx = %u, z = %u%s", 642 mtls->accumFunc, idx, redp.current.z, fmt); 643 } 644} 645 646// Launch a general reduce-style kernel. 647// Inputs: 648// ains[0..inLen-1]: Array of allocations that contain the inputs 649// aout: The allocation that will hold the output 650// mtls: Holds launch parameters 651void RsdCpuReferenceImpl::launchReduce(const Allocation ** ains, 652 uint32_t inLen, 653 Allocation * aout, 654 MTLaunchStructReduce *mtls) { 655 mtls->logReduce = mRSC->props.mLogReduce; 656 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 657 launchReduceParallel(ains, inLen, aout, mtls); 658 } else { 659 launchReduceSerial(ains, inLen, aout, mtls); 660 } 661} 662 663// Launch a general reduce-style kernel, single-threaded. 664// Inputs: 665// ains[0..inLen-1]: Array of allocations that contain the inputs 666// aout: The allocation that will hold the output 667// mtls: Holds launch parameters 668void RsdCpuReferenceImpl::launchReduceSerial(const Allocation ** ains, 669 uint32_t inLen, 670 Allocation * aout, 671 MTLaunchStructReduce *mtls) { 672 REDUCE_ALOGV(mtls, 1, "launchReduceSerial(%p): %u x %u x %u", mtls->accumFunc, 673 mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z); 674 675 // In the presence of outconverter, we allocate temporary memory for 676 // the accumulator. 677 // 678 // In the absence of outconverter, we use the output allocation as the 679 // accumulator. 680 uint8_t *const accumPtr = (mtls->outFunc 681 ? static_cast<uint8_t *>(malloc(mtls->accumSize)) 682 : mtls->redp.outPtr[0]); 683 684 // initialize 685 if (mtls->initFunc) { 686 mtls->initFunc(accumPtr); 687 } else { 688 memset(accumPtr, 0, mtls->accumSize); 689 } 690 691 // accumulate 692 const ReduceAccumulatorFunc_t fn = mtls->accumFunc; 693 uint32_t slice = 0; 694 while (SelectOuterSlice(mtls, &mtls->redp, slice++)) { 695 for (mtls->redp.current.y = mtls->start.y; 696 mtls->redp.current.y < mtls->end.y; 697 mtls->redp.current.y++) { 698 RedpPtrSetup(mtls, &mtls->redp, mtls->start.x, mtls->redp.current.y, mtls->redp.current.z); 699 fn(&mtls->redp, mtls->start.x, mtls->end.x, accumPtr); 700 } 701 } 702 703 // outconvert 704 if (mtls->outFunc) { 705 mtls->outFunc(mtls->redp.outPtr[0], accumPtr); 706 free(accumPtr); 707 } 708} 709 710// Launch a general reduce-style kernel, multi-threaded. 711// Inputs: 712// ains[0..inLen-1]: Array of allocations that contain the inputs 713// aout: The allocation that will hold the output 714// mtls: Holds launch parameters 715void RsdCpuReferenceImpl::launchReduceParallel(const Allocation ** ains, 716 uint32_t inLen, 717 Allocation * aout, 718 MTLaunchStructReduce *mtls) { 719 // For now, we don't know how to go parallel in the absence of a combiner. 720 if (!mtls->combFunc) { 721 launchReduceSerial(ains, inLen, aout, mtls); 722 return; 723 } 724 725 // Number of threads = "main thread" + number of other (worker) threads 726 const uint32_t numThreads = mWorkers.mCount + 1; 727 728 // In the absence of outconverter, we use the output allocation as 729 // an accumulator, and therefore need to allocate one fewer accumulator. 730 const uint32_t numAllocAccum = numThreads - (mtls->outFunc == nullptr); 731 732 // If mDebugReduceSplitAccum, then we want each accumulator to start 733 // on a page boundary. (TODO: Would some unit smaller than a page 734 // be sufficient to avoid false sharing?) 735 if (mRSC->props.mDebugReduceSplitAccum) { 736 // Round up accumulator size to an integral number of pages 737 mtls->accumStride = 738 (unsigned(mtls->accumSize) + unsigned(mPageSize)-1) & 739 ~(unsigned(mPageSize)-1); 740 // Each accumulator gets its own page. Alternatively, if we just 741 // wanted to make sure no two accumulators are on the same page, 742 // we could instead do 743 // allocSize = mtls->accumStride * (numAllocation - 1) + mtls->accumSize 744 const size_t allocSize = mtls->accumStride * numAllocAccum; 745 mtls->accumAlloc = static_cast<uint8_t *>(memalign(mPageSize, allocSize)); 746 } else { 747 mtls->accumStride = mtls->accumSize; 748 mtls->accumAlloc = static_cast<uint8_t *>(malloc(mtls->accumStride * numAllocAccum)); 749 } 750 751 const size_t accumPtrArrayBytes = sizeof(uint8_t *) * numThreads; 752 mtls->accumPtr = static_cast<uint8_t **>(malloc(accumPtrArrayBytes)); 753 memset(mtls->accumPtr, 0, accumPtrArrayBytes); 754 755 mtls->accumCount = 0; 756 757 rsAssert(!mInKernel); 758 mInKernel = true; 759 REDUCE_ALOGV(mtls, 1, "launchReduceParallel(%p): %u x %u x %u, %u threads, accumAlloc = %p", 760 mtls->accumFunc, 761 mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z, 762 numThreads, mtls->accumAlloc); 763 if (mtls->redp.dim.z > 1) { 764 mtls->mSliceSize = 1; 765 launchThreads(walk_3d_reduce, mtls); 766 } else if (mtls->redp.dim.y > 1) { 767 mtls->mSliceSize = rsMax(1U, mtls->redp.dim.y / (numThreads * 4)); 768 launchThreads(walk_2d_reduce, mtls); 769 } else { 770 mtls->mSliceSize = rsMax(1U, mtls->redp.dim.x / (numThreads * 4)); 771 launchThreads(walk_1d_reduce, mtls); 772 } 773 mInKernel = false; 774 775 // Combine accumulators and identify final accumulator 776 uint8_t *finalAccumPtr = (mtls->outFunc ? nullptr : mtls->redp.outPtr[0]); 777 // Loop over accumulators, combining into finalAccumPtr. If finalAccumPtr 778 // is null, then the first accumulator I find becomes finalAccumPtr. 779 for (unsigned idx = 0; idx < mtls->accumCount; ++idx) { 780 uint8_t *const thisAccumPtr = mtls->accumPtr[idx]; 781 if (finalAccumPtr) { 782 if (finalAccumPtr != thisAccumPtr) { 783 if (mtls->combFunc) { 784 if (mtls->logReduce >= 3) { 785 FormatBuf fmt; 786 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): accumulating into%s", 787 mtls->accumFunc, 788 format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 789 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): accumulator[%d]%s", 790 mtls->accumFunc, idx, 791 format_bytes(&fmt, thisAccumPtr, mtls->accumSize)); 792 } 793 mtls->combFunc(finalAccumPtr, thisAccumPtr); 794 } else { 795 rsAssert(!"expected combiner"); 796 } 797 } 798 } else { 799 finalAccumPtr = thisAccumPtr; 800 } 801 } 802 rsAssert(finalAccumPtr != nullptr); 803 if (mtls->logReduce >= 3) { 804 FormatBuf fmt; 805 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): final accumulator%s", 806 mtls->accumFunc, format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 807 } 808 809 // Outconvert 810 if (mtls->outFunc) { 811 mtls->outFunc(mtls->redp.outPtr[0], finalAccumPtr); 812 if (mtls->logReduce >= 3) { 813 FormatBuf fmt; 814 REDUCE_ALOGV(mtls, 3, "launchReduceParallel(%p): final outconverted result%s", 815 mtls->accumFunc, 816 format_bytes(&fmt, mtls->redp.outPtr[0], mtls->redp.outStride[0])); 817 } 818 } 819 820 // Clean up 821 free(mtls->accumPtr); 822 free(mtls->accumAlloc); 823} 824 825 826void RsdCpuReferenceImpl::launchForEach(const Allocation ** ains, 827 uint32_t inLen, 828 Allocation* aout, 829 const RsScriptCall* sc, 830 MTLaunchStructForEach* mtls) { 831 832 //android::StopWatch kernel_time("kernel time"); 833 834 bool outerDims = (mtls->start.z != mtls->end.z) || 835 (mtls->start.face != mtls->end.face) || 836 (mtls->start.lod != mtls->end.lod) || 837 (mtls->start.array[0] != mtls->end.array[0]) || 838 (mtls->start.array[1] != mtls->end.array[1]) || 839 (mtls->start.array[2] != mtls->end.array[2]) || 840 (mtls->start.array[3] != mtls->end.array[3]); 841 842 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 843 const size_t targetByteChunk = 16 * 1024; 844 mInKernel = true; // NOTE: The guard immediately above ensures this was !mInKernel 845 846 if (outerDims) { 847 // No fancy logic for chunk size 848 mtls->mSliceSize = 1; 849 launchThreads(walk_general_foreach, mtls); 850 } else if (mtls->fep.dim.y > 1) { 851 uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4); 852 uint32_t s2 = 0; 853 854 // This chooses our slice size to rate limit atomic ops to 855 // one per 16k bytes of reads/writes. 856 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) { 857 s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride; 858 } else if (mtls->ains[0]) { 859 s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride; 860 } else { 861 // Launch option only case 862 // Use s1 based only on the dimensions 863 s2 = s1; 864 } 865 mtls->mSliceSize = rsMin(s1, s2); 866 867 if(mtls->mSliceSize < 1) { 868 mtls->mSliceSize = 1; 869 } 870 871 launchThreads(walk_2d_foreach, mtls); 872 } else { 873 uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4); 874 uint32_t s2 = 0; 875 876 // This chooses our slice size to rate limit atomic ops to 877 // one per 16k bytes of reads/writes. 878 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) { 879 s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes(); 880 } else if (mtls->ains[0]) { 881 s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes(); 882 } else { 883 // Launch option only case 884 // Use s1 based only on the dimensions 885 s2 = s1; 886 } 887 mtls->mSliceSize = rsMin(s1, s2); 888 889 if (mtls->mSliceSize < 1) { 890 mtls->mSliceSize = 1; 891 } 892 893 launchThreads(walk_1d_foreach, mtls); 894 } 895 mInKernel = false; 896 897 } else { 898 ForEachFunc_t fn = mtls->kernel; 899 uint32_t slice = 0; 900 901 902 while(SelectOuterSlice(mtls, &mtls->fep, slice++)) { 903 for (mtls->fep.current.y = mtls->start.y; 904 mtls->fep.current.y < mtls->end.y; 905 mtls->fep.current.y++) { 906 907 FepPtrSetup(mtls, &mtls->fep, mtls->start.x, 908 mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod, 909 (RsAllocationCubemapFace) mtls->fep.current.face, 910 mtls->fep.current.array[0], mtls->fep.current.array[1], 911 mtls->fep.current.array[2], mtls->fep.current.array[3]); 912 913 fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 914 } 915 } 916 } 917} 918 919RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 920 //ALOGE("setTls %p", sc); 921 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 922 rsAssert(tls); 923 RsdCpuScriptImpl *old = tls->mImpl; 924 tls->mImpl = sc; 925 tls->mContext = mRSC; 926 if (sc) { 927 tls->mScript = sc->getScript(); 928 } else { 929 tls->mScript = nullptr; 930 } 931 return old; 932} 933 934const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 935 return mSymLookupFn(mRSC, name); 936} 937 938 939RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 940 char const *resName, char const *cacheDir, 941 uint8_t const *bitcode, size_t bitcodeSize, 942 uint32_t flags) { 943 944 RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 945 if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 946 , getBccPluginName() 947 )) { 948 delete i; 949 return nullptr; 950 } 951 return i; 952} 953 954extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 955 const Script *s, const Element *e); 956extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 957 const Script *s, const Element *e); 958extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 959 const Script *s, const Element *e); 960extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 961 const Script *s, const Element *e); 962extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 963 const Script *s, const Element *e); 964extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 965 const Script *s, const Element *e); 966extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 967 const Script *s, const Element *e); 968extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 969 const Script *s, const Element *e); 970extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 971 const Script *s, const Element *e); 972extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 973 const Script *s, const Element *e); 974extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx, 975 const Script *s, const Element *e); 976 977RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 978 RsScriptIntrinsicID iid, Element *e) { 979 980 RsdCpuScriptImpl *i = nullptr; 981 switch (iid) { 982 case RS_SCRIPT_INTRINSIC_ID_3DLUT: 983 i = rsdIntrinsic_3DLUT(this, s, e); 984 break; 985 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 986 i = rsdIntrinsic_Convolve3x3(this, s, e); 987 break; 988 case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 989 i = rsdIntrinsic_ColorMatrix(this, s, e); 990 break; 991 case RS_SCRIPT_INTRINSIC_ID_LUT: 992 i = rsdIntrinsic_LUT(this, s, e); 993 break; 994 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 995 i = rsdIntrinsic_Convolve5x5(this, s, e); 996 break; 997 case RS_SCRIPT_INTRINSIC_ID_BLUR: 998 i = rsdIntrinsic_Blur(this, s, e); 999 break; 1000 case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 1001 i = rsdIntrinsic_YuvToRGB(this, s, e); 1002 break; 1003 case RS_SCRIPT_INTRINSIC_ID_BLEND: 1004 i = rsdIntrinsic_Blend(this, s, e); 1005 break; 1006 case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 1007 i = rsdIntrinsic_Histogram(this, s, e); 1008 break; 1009 case RS_SCRIPT_INTRINSIC_ID_RESIZE: 1010 i = rsdIntrinsic_Resize(this, s, e); 1011 break; 1012 case RS_SCRIPT_INTRINSIC_ID_BLAS: 1013 i = rsdIntrinsic_BLAS(this, s, e); 1014 break; 1015 1016 default: 1017 rsAssert(0); 1018 } 1019 1020 return i; 1021} 1022 1023void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) { 1024 switch (sg->getApiVersion()) { 1025 case ScriptGroupBase::SG_V1: { 1026 CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 1027 if (!sgi->init()) { 1028 delete sgi; 1029 return nullptr; 1030 } 1031 return sgi; 1032 } 1033 case ScriptGroupBase::SG_V2: { 1034 return new CpuScriptGroup2Impl(this, sg); 1035 } 1036 } 1037 return nullptr; 1038} 1039 1040} // namespace renderscript 1041} // namespace android 1042