rsCpuCore.cpp revision 247f8ee57196d6cf3264e6f7505f53e8f8a7860d
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 <string.h> 30#include <unistd.h> 31 32#include <stdio.h> 33#include <stdlib.h> 34#include <fcntl.h> 35 36#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB) 37#include <cutils/properties.h> 38#include "utils/StopWatch.h" 39#endif 40 41#ifdef RS_SERVER 42// Android exposes gettid(), standard Linux does not 43static pid_t gettid() { 44 return syscall(SYS_gettid); 45} 46#endif 47 48using namespace android; 49using namespace android::renderscript; 50 51typedef void (*outer_foreach_t)( 52 const RsExpandKernelDriverInfo *, 53 uint32_t x1, uint32_t x2, uint32_t outstep); 54 55 56static pthread_key_t gThreadTLSKey = 0; 57static uint32_t gThreadTLSKeyCount = 0; 58static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER; 59 60bool android::renderscript::gArchUseSIMD = false; 61 62RsdCpuReference::~RsdCpuReference() { 63} 64 65RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major, 66 uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn 67 , bcc::RSLinkRuntimeCallback pLinkRuntimeCallback, 68 RSSelectRTCallback pSelectRTCallback, 69 const char *pBccPluginName 70 ) { 71 72 RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc); 73 if (!cpu) { 74 return nullptr; 75 } 76 if (!cpu->init(version_major, version_minor, lfn, slfn)) { 77 delete cpu; 78 return nullptr; 79 } 80 81 cpu->setLinkRuntimeCallback(pLinkRuntimeCallback); 82 cpu->setSelectRTCallback(pSelectRTCallback); 83 if (pBccPluginName) { 84 cpu->setBccPluginName(pBccPluginName); 85 } 86 87 return cpu; 88} 89 90 91Context * RsdCpuReference::getTlsContext() { 92 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 93 return tls->mContext; 94} 95 96const Script * RsdCpuReference::getTlsScript() { 97 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 98 return tls->mScript; 99} 100 101pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 102 103//////////////////////////////////////////////////////////// 104/// 105 106RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) { 107 mRSC = rsc; 108 109 version_major = 0; 110 version_minor = 0; 111 mInForEach = false; 112 memset(&mWorkers, 0, sizeof(mWorkers)); 113 memset(&mTlsStruct, 0, sizeof(mTlsStruct)); 114 mExit = false; 115 mLinkRuntimeCallback = nullptr; 116 mSelectRTCallback = nullptr; 117 mSetupCompilerCallback = nullptr; 118} 119 120 121void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 122 RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 123 124 uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 125 126 //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 127 128 dc->mWorkers.mLaunchSignals[idx].init(); 129 dc->mWorkers.mNativeThreadId[idx] = gettid(); 130 131 memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 132 int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 133 if (status) { 134 ALOGE("pthread_setspecific %i", status); 135 } 136 137#if 0 138 typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 139 cpu_set_t cpuset; 140 memset(&cpuset, 0, sizeof(cpuset)); 141 cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 142 int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 143 sizeof(cpuset), &cpuset); 144 ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 145#endif 146 147 while (!dc->mExit) { 148 dc->mWorkers.mLaunchSignals[idx].wait(); 149 if (dc->mWorkers.mLaunchCallback) { 150 // idx +1 is used because the calling thread is always worker 0. 151 dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 152 } 153 __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 154 dc->mWorkers.mCompleteSignal.set(); 155 } 156 157 //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 158 return nullptr; 159} 160 161void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 162 mWorkers.mLaunchData = data; 163 mWorkers.mLaunchCallback = cbk; 164 165 // fast path for very small launches 166 MTLaunchStruct *mtls = (MTLaunchStruct *)data; 167 if (mtls && mtls->fep.dim.y <= 1 && mtls->end.x <= mtls->start.x + mtls->mSliceSize) { 168 if (mWorkers.mLaunchCallback) { 169 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 170 } 171 return; 172 } 173 174 mWorkers.mRunningCount = mWorkers.mCount; 175 __sync_synchronize(); 176 177 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 178 mWorkers.mLaunchSignals[ct].set(); 179 } 180 181 // We use the calling thread as one of the workers so we can start without 182 // the delay of the thread wakeup. 183 if (mWorkers.mLaunchCallback) { 184 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 185 } 186 187 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 188 mWorkers.mCompleteSignal.wait(); 189 } 190} 191 192 193void RsdCpuReferenceImpl::lockMutex() { 194 pthread_mutex_lock(&gInitMutex); 195} 196 197void RsdCpuReferenceImpl::unlockMutex() { 198 pthread_mutex_unlock(&gInitMutex); 199} 200 201static int 202read_file(const char* pathname, char* buffer, size_t buffsize) 203{ 204 int fd, len; 205 206 fd = open(pathname, O_RDONLY); 207 if (fd < 0) 208 return -1; 209 210 do { 211 len = read(fd, buffer, buffsize); 212 } while (len < 0 && errno == EINTR); 213 214 close(fd); 215 216 return len; 217} 218 219static void GetCpuInfo() { 220 char cpuinfo[4096]; 221 int cpuinfo_len; 222 223 cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo); 224 if (cpuinfo_len < 0) /* should not happen */ { 225 return; 226 } 227 228#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS) 229 gArchUseSIMD = (!!strstr(cpuinfo, " neon")) || 230 (!!strstr(cpuinfo, " asimd")); 231#elif defined(ARCH_X86_HAVE_SSSE3) 232 gArchUseSIMD = !!strstr(cpuinfo, " ssse3"); 233#endif 234} 235 236bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 237 sym_lookup_t lfn, script_lookup_t slfn) { 238 239 mSymLookupFn = lfn; 240 mScriptLookupFn = slfn; 241 242 lockMutex(); 243 if (!gThreadTLSKeyCount) { 244 int status = pthread_key_create(&gThreadTLSKey, nullptr); 245 if (status) { 246 ALOGE("Failed to init thread tls key."); 247 unlockMutex(); 248 return false; 249 } 250 } 251 gThreadTLSKeyCount++; 252 unlockMutex(); 253 254 mTlsStruct.mContext = mRSC; 255 mTlsStruct.mScript = nullptr; 256 int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 257 if (status) { 258 ALOGE("pthread_setspecific %i", status); 259 } 260 261 GetCpuInfo(); 262 263 int cpu = sysconf(_SC_NPROCESSORS_CONF); 264 if(mRSC->props.mDebugMaxThreads) { 265 cpu = mRSC->props.mDebugMaxThreads; 266 } 267 if (cpu < 2) { 268 mWorkers.mCount = 0; 269 return true; 270 } 271 272 // Subtract one from the cpu count because we also use the command thread as a worker. 273 mWorkers.mCount = (uint32_t)(cpu - 1); 274 275 ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 276 277 mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 278 mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 279 mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 280 mWorkers.mLaunchCallback = nullptr; 281 282 mWorkers.mCompleteSignal.init(); 283 284 mWorkers.mRunningCount = mWorkers.mCount; 285 mWorkers.mLaunchCount = 0; 286 __sync_synchronize(); 287 288 pthread_attr_t threadAttr; 289 status = pthread_attr_init(&threadAttr); 290 if (status) { 291 ALOGE("Failed to init thread attribute."); 292 return false; 293 } 294 295 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 296 status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 297 if (status) { 298 mWorkers.mCount = ct; 299 ALOGE("Created fewer than expected number of RS threads."); 300 break; 301 } 302 } 303 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 304 usleep(100); 305 } 306 307 pthread_attr_destroy(&threadAttr); 308 return true; 309} 310 311 312void RsdCpuReferenceImpl::setPriority(int32_t priority) { 313 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 314 setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 315 } 316} 317 318RsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 319 mExit = true; 320 mWorkers.mLaunchData = nullptr; 321 mWorkers.mLaunchCallback = nullptr; 322 mWorkers.mRunningCount = mWorkers.mCount; 323 __sync_synchronize(); 324 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 325 mWorkers.mLaunchSignals[ct].set(); 326 } 327 void *res; 328 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 329 pthread_join(mWorkers.mThreadId[ct], &res); 330 } 331 rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 332 free(mWorkers.mThreadId); 333 free(mWorkers.mNativeThreadId); 334 delete[] mWorkers.mLaunchSignals; 335 336 // Global structure cleanup. 337 lockMutex(); 338 --gThreadTLSKeyCount; 339 if (!gThreadTLSKeyCount) { 340 pthread_key_delete(gThreadTLSKey); 341 } 342 unlockMutex(); 343 344} 345 346static inline void FepPtrSetup(const MTLaunchStruct *mtls, RsExpandKernelDriverInfo *fep, 347 uint32_t x, uint32_t y, 348 uint32_t z = 0, uint32_t lod = 0, 349 RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X, 350 uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) { 351 352 for (uint32_t i = 0; i < fep->inLen; i++) { 353 fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 354 } 355 356 if (mtls->aout[0] != nullptr) { 357 fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 358 } 359} 360 361static uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) { 362 if (start >= end) { 363 *p = start; 364 return val; 365 } 366 367 uint32_t div = end - start; 368 369 uint32_t n = val / div; 370 *p = (val - (n * div)) + start; 371 return n; 372} 373 374static bool SelectOuterSlice(const MTLaunchStruct *mtls, RsExpandKernelDriverInfo* fep, uint32_t sliceNum) { 375 376 uint32_t r = sliceNum; 377 r = sliceInt(&fep->current.z, r, mtls->start.z, mtls->end.z); 378 r = sliceInt(&fep->current.lod, r, mtls->start.lod, mtls->end.lod); 379 r = sliceInt(&fep->current.face, r, mtls->start.face, mtls->end.face); 380 r = sliceInt(&fep->current.array[0], r, mtls->start.array[0], mtls->end.array[0]); 381 r = sliceInt(&fep->current.array[1], r, mtls->start.array[1], mtls->end.array[1]); 382 r = sliceInt(&fep->current.array[2], r, mtls->start.array[2], mtls->end.array[2]); 383 r = sliceInt(&fep->current.array[3], r, mtls->start.array[3], mtls->end.array[3]); 384 return r == 0; 385} 386 387 388static void walk_general(void *usr, uint32_t idx) { 389 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 390 RsExpandKernelDriverInfo fep = mtls->fep; 391 fep.lid = idx; 392 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 393 394 395 while(1) { 396 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 397 398 if (!SelectOuterSlice(mtls, &fep, slice)) { 399 return; 400 } 401 402 for (fep.current.y = mtls->start.y; fep.current.y < mtls->end.y; 403 fep.current.y++) { 404 405 FepPtrSetup(mtls, &fep, mtls->start.x, 406 fep.current.y, fep.current.z, fep.current.lod, 407 (RsAllocationCubemapFace)fep.current.face, 408 fep.current.array[0], fep.current.array[1], 409 fep.current.array[2], fep.current.array[3]); 410 411 fn(&fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 412 } 413 } 414 415} 416 417static void walk_2d(void *usr, uint32_t idx) { 418 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 419 RsExpandKernelDriverInfo fep = mtls->fep; 420 fep.lid = idx; 421 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 422 423 while (1) { 424 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 425 uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 426 uint32_t yEnd = yStart + mtls->mSliceSize; 427 428 yEnd = rsMin(yEnd, mtls->end.y); 429 430 if (yEnd <= yStart) { 431 return; 432 } 433 434 for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) { 435 FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y); 436 437 fn(&fep, mtls->start.x, mtls->end.x, fep.outStride[0]); 438 } 439 } 440} 441 442static void walk_1d(void *usr, uint32_t idx) { 443 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 444 RsExpandKernelDriverInfo fep = mtls->fep; 445 fep.lid = idx; 446 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 447 448 while (1) { 449 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 450 uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 451 uint32_t xEnd = xStart + mtls->mSliceSize; 452 453 xEnd = rsMin(xEnd, mtls->end.x); 454 455 if (xEnd <= xStart) { 456 return; 457 } 458 459 FepPtrSetup(mtls, &fep, xStart, 0); 460 461 fn(&fep, xStart, xEnd, fep.outStride[0]); 462 } 463} 464 465void RsdCpuReferenceImpl::launchThreads(const Allocation ** ains, 466 uint32_t inLen, 467 Allocation* aout, 468 const RsScriptCall* sc, 469 MTLaunchStruct* mtls) { 470 471 //android::StopWatch kernel_time("kernel time"); 472 473 bool outerDims = (mtls->start.z != mtls->end.z) || 474 (mtls->start.face != mtls->end.face) || 475 (mtls->start.lod != mtls->end.lod) || 476 (mtls->start.array[0] != mtls->end.array[0]) || 477 (mtls->start.array[1] != mtls->end.array[1]) || 478 (mtls->start.array[2] != mtls->end.array[2]) || 479 (mtls->start.array[3] != mtls->end.array[3]); 480 481 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) { 482 const size_t targetByteChunk = 16 * 1024; 483 mInForEach = true; 484 485 if (outerDims) { 486 // No fancy logic for chunk size 487 mtls->mSliceSize = 1; 488 launchThreads(walk_general, mtls); 489 } else if (mtls->fep.dim.y > 1) { 490 uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4); 491 uint32_t s2 = 0; 492 493 // This chooses our slice size to rate limit atomic ops to 494 // one per 16k bytes of reads/writes. 495 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) { 496 s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride; 497 } else if (mtls->ains[0]) { 498 s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride; 499 } else { 500 // Launch option only case 501 // Use s1 based only on the dimensions 502 s2 = s1; 503 } 504 mtls->mSliceSize = rsMin(s1, s2); 505 506 if(mtls->mSliceSize < 1) { 507 mtls->mSliceSize = 1; 508 } 509 510 launchThreads(walk_2d, mtls); 511 } else { 512 uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4); 513 uint32_t s2 = 0; 514 515 // This chooses our slice size to rate limit atomic ops to 516 // one per 16k bytes of reads/writes. 517 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) { 518 s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes(); 519 } else if (mtls->ains[0]) { 520 s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes(); 521 } else { 522 // Launch option only case 523 // Use s1 based only on the dimensions 524 s2 = s1; 525 } 526 mtls->mSliceSize = rsMin(s1, s2); 527 528 if (mtls->mSliceSize < 1) { 529 mtls->mSliceSize = 1; 530 } 531 532 launchThreads(walk_1d, mtls); 533 } 534 mInForEach = false; 535 536 } else { 537 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 538 uint32_t slice = 0; 539 540 541 while(SelectOuterSlice(mtls, &mtls->fep, slice++)) { 542 for (mtls->fep.current.y = mtls->start.y; 543 mtls->fep.current.y < mtls->end.y; 544 mtls->fep.current.y++) { 545 546 FepPtrSetup(mtls, &mtls->fep, mtls->start.x, 547 mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod, 548 (RsAllocationCubemapFace) mtls->fep.current.face, 549 mtls->fep.current.array[0], mtls->fep.current.array[1], 550 mtls->fep.current.array[2], mtls->fep.current.array[3]); 551 552 fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 553 } 554 } 555 } 556} 557 558RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 559 //ALOGE("setTls %p", sc); 560 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 561 rsAssert(tls); 562 RsdCpuScriptImpl *old = tls->mImpl; 563 tls->mImpl = sc; 564 tls->mContext = mRSC; 565 if (sc) { 566 tls->mScript = sc->getScript(); 567 } else { 568 tls->mScript = nullptr; 569 } 570 return old; 571} 572 573const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 574 return mSymLookupFn(mRSC, name); 575} 576 577 578RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 579 char const *resName, char const *cacheDir, 580 uint8_t const *bitcode, size_t bitcodeSize, 581 uint32_t flags) { 582 583 RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 584 if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 585 , getBccPluginName() 586 )) { 587 delete i; 588 return nullptr; 589 } 590 return i; 591} 592 593extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 594 const Script *s, const Element *e); 595extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 596 const Script *s, const Element *e); 597extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 598 const Script *s, const Element *e); 599extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 600 const Script *s, const Element *e); 601extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 602 const Script *s, const Element *e); 603extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 604 const Script *s, const Element *e); 605extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 606 const Script *s, const Element *e); 607extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 608 const Script *s, const Element *e); 609extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 610 const Script *s, const Element *e); 611extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 612 const Script *s, const Element *e); 613extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx, 614 const Script *s, const Element *e); 615 616RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 617 RsScriptIntrinsicID iid, Element *e) { 618 619 RsdCpuScriptImpl *i = nullptr; 620 switch (iid) { 621 case RS_SCRIPT_INTRINSIC_ID_3DLUT: 622 i = rsdIntrinsic_3DLUT(this, s, e); 623 break; 624 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 625 i = rsdIntrinsic_Convolve3x3(this, s, e); 626 break; 627 case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 628 i = rsdIntrinsic_ColorMatrix(this, s, e); 629 break; 630 case RS_SCRIPT_INTRINSIC_ID_LUT: 631 i = rsdIntrinsic_LUT(this, s, e); 632 break; 633 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 634 i = rsdIntrinsic_Convolve5x5(this, s, e); 635 break; 636 case RS_SCRIPT_INTRINSIC_ID_BLUR: 637 i = rsdIntrinsic_Blur(this, s, e); 638 break; 639 case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 640 i = rsdIntrinsic_YuvToRGB(this, s, e); 641 break; 642 case RS_SCRIPT_INTRINSIC_ID_BLEND: 643 i = rsdIntrinsic_Blend(this, s, e); 644 break; 645 case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 646 i = rsdIntrinsic_Histogram(this, s, e); 647 break; 648 case RS_SCRIPT_INTRINSIC_ID_RESIZE: 649 i = rsdIntrinsic_Resize(this, s, e); 650 break; 651#if !defined(RS_COMPATIBILITY_LIB) 652 case RS_SCRIPT_INTRINSIC_ID_BLAS: 653 i = rsdIntrinsic_BLAS(this, s, e); 654 break; 655#endif 656 657 default: 658 rsAssert(0); 659 } 660 661 return i; 662} 663 664void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) { 665 switch (sg->getApiVersion()) { 666 case ScriptGroupBase::SG_V1: { 667 CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 668 if (!sgi->init()) { 669 delete sgi; 670 return nullptr; 671 } 672 return sgi; 673 } 674 case ScriptGroupBase::SG_V2: { 675 return new CpuScriptGroup2Impl(this, sg); 676 } 677 } 678 return nullptr; 679} 680