rsCpuCore.cpp revision 8409d6414dd4a42aa59779fcfe9fce18648cb135
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 mEmbedGlobalInfo = true; 119 mEmbedGlobalInfoSkipConstant = true; 120} 121 122 123void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 124 RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 125 126 uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 127 128 //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 129 130 dc->mWorkers.mLaunchSignals[idx].init(); 131 dc->mWorkers.mNativeThreadId[idx] = gettid(); 132 133 memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 134 int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 135 if (status) { 136 ALOGE("pthread_setspecific %i", status); 137 } 138 139#if 0 140 typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 141 cpu_set_t cpuset; 142 memset(&cpuset, 0, sizeof(cpuset)); 143 cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 144 int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 145 sizeof(cpuset), &cpuset); 146 ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 147#endif 148 149 while (!dc->mExit) { 150 dc->mWorkers.mLaunchSignals[idx].wait(); 151 if (dc->mWorkers.mLaunchCallback) { 152 // idx +1 is used because the calling thread is always worker 0. 153 dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 154 } 155 __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 156 dc->mWorkers.mCompleteSignal.set(); 157 } 158 159 //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 160 return nullptr; 161} 162 163void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 164 mWorkers.mLaunchData = data; 165 mWorkers.mLaunchCallback = cbk; 166 167 // fast path for very small launches 168 MTLaunchStruct *mtls = (MTLaunchStruct *)data; 169 if (mtls && mtls->fep.dim.y <= 1 && mtls->end.x <= mtls->start.x + mtls->mSliceSize) { 170 if (mWorkers.mLaunchCallback) { 171 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 172 } 173 return; 174 } 175 176 mWorkers.mRunningCount = mWorkers.mCount; 177 __sync_synchronize(); 178 179 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 180 mWorkers.mLaunchSignals[ct].set(); 181 } 182 183 // We use the calling thread as one of the workers so we can start without 184 // the delay of the thread wakeup. 185 if (mWorkers.mLaunchCallback) { 186 mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 187 } 188 189 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 190 mWorkers.mCompleteSignal.wait(); 191 } 192} 193 194 195void RsdCpuReferenceImpl::lockMutex() { 196 pthread_mutex_lock(&gInitMutex); 197} 198 199void RsdCpuReferenceImpl::unlockMutex() { 200 pthread_mutex_unlock(&gInitMutex); 201} 202 203static int 204read_file(const char* pathname, char* buffer, size_t buffsize) 205{ 206 int fd, len; 207 208 fd = open(pathname, O_RDONLY); 209 if (fd < 0) 210 return -1; 211 212 do { 213 len = read(fd, buffer, buffsize); 214 } while (len < 0 && errno == EINTR); 215 216 close(fd); 217 218 return len; 219} 220 221static void GetCpuInfo() { 222 char cpuinfo[4096]; 223 int cpuinfo_len; 224 225 cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo); 226 if (cpuinfo_len < 0) /* should not happen */ { 227 return; 228 } 229 230#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS) 231 gArchUseSIMD = (!!strstr(cpuinfo, " neon")) || 232 (!!strstr(cpuinfo, " asimd")); 233#elif defined(ARCH_X86_HAVE_SSSE3) 234 gArchUseSIMD = !!strstr(cpuinfo, " ssse3"); 235#endif 236} 237 238bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 239 sym_lookup_t lfn, script_lookup_t slfn) { 240 241 mSymLookupFn = lfn; 242 mScriptLookupFn = slfn; 243 244 lockMutex(); 245 if (!gThreadTLSKeyCount) { 246 int status = pthread_key_create(&gThreadTLSKey, nullptr); 247 if (status) { 248 ALOGE("Failed to init thread tls key."); 249 unlockMutex(); 250 return false; 251 } 252 } 253 gThreadTLSKeyCount++; 254 unlockMutex(); 255 256 mTlsStruct.mContext = mRSC; 257 mTlsStruct.mScript = nullptr; 258 int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 259 if (status) { 260 ALOGE("pthread_setspecific %i", status); 261 } 262 263 GetCpuInfo(); 264 265 int cpu = sysconf(_SC_NPROCESSORS_CONF); 266 if(mRSC->props.mDebugMaxThreads) { 267 cpu = mRSC->props.mDebugMaxThreads; 268 } 269 if (cpu < 2) { 270 mWorkers.mCount = 0; 271 return true; 272 } 273 274 // Subtract one from the cpu count because we also use the command thread as a worker. 275 mWorkers.mCount = (uint32_t)(cpu - 1); 276 277 ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 278 279 mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 280 mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 281 mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 282 mWorkers.mLaunchCallback = nullptr; 283 284 mWorkers.mCompleteSignal.init(); 285 286 mWorkers.mRunningCount = mWorkers.mCount; 287 mWorkers.mLaunchCount = 0; 288 __sync_synchronize(); 289 290 pthread_attr_t threadAttr; 291 status = pthread_attr_init(&threadAttr); 292 if (status) { 293 ALOGE("Failed to init thread attribute."); 294 return false; 295 } 296 297 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 298 status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 299 if (status) { 300 mWorkers.mCount = ct; 301 ALOGE("Created fewer than expected number of RS threads."); 302 break; 303 } 304 } 305 while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 306 usleep(100); 307 } 308 309 pthread_attr_destroy(&threadAttr); 310 return true; 311} 312 313 314void RsdCpuReferenceImpl::setPriority(int32_t priority) { 315 for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 316 setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 317 } 318} 319 320RsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 321 mExit = true; 322 mWorkers.mLaunchData = nullptr; 323 mWorkers.mLaunchCallback = nullptr; 324 mWorkers.mRunningCount = mWorkers.mCount; 325 __sync_synchronize(); 326 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 327 mWorkers.mLaunchSignals[ct].set(); 328 } 329 void *res; 330 for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 331 pthread_join(mWorkers.mThreadId[ct], &res); 332 } 333 rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 334 free(mWorkers.mThreadId); 335 free(mWorkers.mNativeThreadId); 336 delete[] mWorkers.mLaunchSignals; 337 338 // Global structure cleanup. 339 lockMutex(); 340 --gThreadTLSKeyCount; 341 if (!gThreadTLSKeyCount) { 342 pthread_key_delete(gThreadTLSKey); 343 } 344 unlockMutex(); 345 346} 347 348static inline void FepPtrSetup(const MTLaunchStruct *mtls, RsExpandKernelDriverInfo *fep, 349 uint32_t x, uint32_t y, 350 uint32_t z = 0, uint32_t lod = 0, 351 RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X, 352 uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) { 353 354 for (uint32_t i = 0; i < fep->inLen; i++) { 355 fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 356 } 357 358 if (mtls->aout[0] != nullptr) { 359 fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 360 } 361} 362 363static uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) { 364 if (start >= end) { 365 *p = start; 366 return val; 367 } 368 369 uint32_t div = end - start; 370 371 uint32_t n = val / div; 372 *p = (val - (n * div)) + start; 373 return n; 374} 375 376static bool SelectOuterSlice(const MTLaunchStruct *mtls, RsExpandKernelDriverInfo* fep, uint32_t sliceNum) { 377 378 uint32_t r = sliceNum; 379 r = sliceInt(&fep->current.z, r, mtls->start.z, mtls->end.z); 380 r = sliceInt(&fep->current.lod, r, mtls->start.lod, mtls->end.lod); 381 r = sliceInt(&fep->current.face, r, mtls->start.face, mtls->end.face); 382 r = sliceInt(&fep->current.array[0], r, mtls->start.array[0], mtls->end.array[0]); 383 r = sliceInt(&fep->current.array[1], r, mtls->start.array[1], mtls->end.array[1]); 384 r = sliceInt(&fep->current.array[2], r, mtls->start.array[2], mtls->end.array[2]); 385 r = sliceInt(&fep->current.array[3], r, mtls->start.array[3], mtls->end.array[3]); 386 return r == 0; 387} 388 389 390static void walk_general(void *usr, uint32_t idx) { 391 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 392 RsExpandKernelDriverInfo fep = mtls->fep; 393 fep.lid = idx; 394 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 395 396 397 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} 418 419static void walk_2d(void *usr, uint32_t idx) { 420 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 421 RsExpandKernelDriverInfo fep = mtls->fep; 422 fep.lid = idx; 423 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 424 425 while (1) { 426 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 427 uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 428 uint32_t yEnd = yStart + mtls->mSliceSize; 429 430 yEnd = rsMin(yEnd, mtls->end.y); 431 432 if (yEnd <= yStart) { 433 return; 434 } 435 436 for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) { 437 FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y); 438 439 fn(&fep, mtls->start.x, mtls->end.x, fep.outStride[0]); 440 } 441 } 442} 443 444static void walk_1d(void *usr, uint32_t idx) { 445 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 446 RsExpandKernelDriverInfo fep = mtls->fep; 447 fep.lid = idx; 448 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 449 450 while (1) { 451 uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 452 uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 453 uint32_t xEnd = xStart + mtls->mSliceSize; 454 455 xEnd = rsMin(xEnd, mtls->end.x); 456 457 if (xEnd <= xStart) { 458 return; 459 } 460 461 FepPtrSetup(mtls, &fep, xStart, 0); 462 463 fn(&fep, xStart, xEnd, fep.outStride[0]); 464 } 465} 466 467void RsdCpuReferenceImpl::launchThreads(const Allocation ** ains, 468 uint32_t inLen, 469 Allocation* aout, 470 const RsScriptCall* sc, 471 MTLaunchStruct* mtls) { 472 473 //android::StopWatch kernel_time("kernel time"); 474 475 bool outerDims = (mtls->start.z != mtls->end.z) || 476 (mtls->start.face != mtls->end.face) || 477 (mtls->start.lod != mtls->end.lod) || 478 (mtls->start.array[0] != mtls->end.array[0]) || 479 (mtls->start.array[1] != mtls->end.array[1]) || 480 (mtls->start.array[2] != mtls->end.array[2]) || 481 (mtls->start.array[3] != mtls->end.array[3]); 482 483 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) { 484 const size_t targetByteChunk = 16 * 1024; 485 mInForEach = true; 486 487 if (outerDims) { 488 // No fancy logic for chunk size 489 mtls->mSliceSize = 1; 490 launchThreads(walk_general, mtls); 491 } else if (mtls->fep.dim.y > 1) { 492 uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4); 493 uint32_t s2 = 0; 494 495 // This chooses our slice size to rate limit atomic ops to 496 // one per 16k bytes of reads/writes. 497 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) { 498 s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride; 499 } else if (mtls->ains[0]) { 500 s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride; 501 } else { 502 // Launch option only case 503 // Use s1 based only on the dimensions 504 s2 = s1; 505 } 506 mtls->mSliceSize = rsMin(s1, s2); 507 508 if(mtls->mSliceSize < 1) { 509 mtls->mSliceSize = 1; 510 } 511 512 launchThreads(walk_2d, mtls); 513 } else { 514 uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4); 515 uint32_t s2 = 0; 516 517 // This chooses our slice size to rate limit atomic ops to 518 // one per 16k bytes of reads/writes. 519 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) { 520 s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes(); 521 } else if (mtls->ains[0]) { 522 s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes(); 523 } else { 524 // Launch option only case 525 // Use s1 based only on the dimensions 526 s2 = s1; 527 } 528 mtls->mSliceSize = rsMin(s1, s2); 529 530 if (mtls->mSliceSize < 1) { 531 mtls->mSliceSize = 1; 532 } 533 534 launchThreads(walk_1d, mtls); 535 } 536 mInForEach = false; 537 538 } else { 539 outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 540 uint32_t slice = 0; 541 542 543 while(SelectOuterSlice(mtls, &mtls->fep, slice++)) { 544 for (mtls->fep.current.y = mtls->start.y; 545 mtls->fep.current.y < mtls->end.y; 546 mtls->fep.current.y++) { 547 548 FepPtrSetup(mtls, &mtls->fep, mtls->start.x, 549 mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod, 550 (RsAllocationCubemapFace) mtls->fep.current.face, 551 mtls->fep.current.array[0], mtls->fep.current.array[1], 552 mtls->fep.current.array[2], mtls->fep.current.array[3]); 553 554 fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 555 } 556 } 557 } 558} 559 560RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 561 //ALOGE("setTls %p", sc); 562 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 563 rsAssert(tls); 564 RsdCpuScriptImpl *old = tls->mImpl; 565 tls->mImpl = sc; 566 tls->mContext = mRSC; 567 if (sc) { 568 tls->mScript = sc->getScript(); 569 } else { 570 tls->mScript = nullptr; 571 } 572 return old; 573} 574 575const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 576 return mSymLookupFn(mRSC, name); 577} 578 579 580RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 581 char const *resName, char const *cacheDir, 582 uint8_t const *bitcode, size_t bitcodeSize, 583 uint32_t flags) { 584 585 RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 586 if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 587 , getBccPluginName() 588 )) { 589 delete i; 590 return nullptr; 591 } 592 return i; 593} 594 595extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 596 const Script *s, const Element *e); 597extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 598 const Script *s, const Element *e); 599extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 600 const Script *s, const Element *e); 601extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 602 const Script *s, const Element *e); 603extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 604 const Script *s, const Element *e); 605extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 606 const Script *s, const Element *e); 607extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 608 const Script *s, const Element *e); 609extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 610 const Script *s, const Element *e); 611extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 612 const Script *s, const Element *e); 613extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 614 const Script *s, const Element *e); 615extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx, 616 const Script *s, const Element *e); 617 618RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 619 RsScriptIntrinsicID iid, Element *e) { 620 621 RsdCpuScriptImpl *i = nullptr; 622 switch (iid) { 623 case RS_SCRIPT_INTRINSIC_ID_3DLUT: 624 i = rsdIntrinsic_3DLUT(this, s, e); 625 break; 626 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 627 i = rsdIntrinsic_Convolve3x3(this, s, e); 628 break; 629 case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 630 i = rsdIntrinsic_ColorMatrix(this, s, e); 631 break; 632 case RS_SCRIPT_INTRINSIC_ID_LUT: 633 i = rsdIntrinsic_LUT(this, s, e); 634 break; 635 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 636 i = rsdIntrinsic_Convolve5x5(this, s, e); 637 break; 638 case RS_SCRIPT_INTRINSIC_ID_BLUR: 639 i = rsdIntrinsic_Blur(this, s, e); 640 break; 641 case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 642 i = rsdIntrinsic_YuvToRGB(this, s, e); 643 break; 644 case RS_SCRIPT_INTRINSIC_ID_BLEND: 645 i = rsdIntrinsic_Blend(this, s, e); 646 break; 647 case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 648 i = rsdIntrinsic_Histogram(this, s, e); 649 break; 650 case RS_SCRIPT_INTRINSIC_ID_RESIZE: 651 i = rsdIntrinsic_Resize(this, s, e); 652 break; 653#if !defined(RS_COMPATIBILITY_LIB) 654 case RS_SCRIPT_INTRINSIC_ID_BLAS: 655 i = rsdIntrinsic_BLAS(this, s, e); 656 break; 657#endif 658 659 default: 660 rsAssert(0); 661 } 662 663 return i; 664} 665 666void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) { 667 switch (sg->getApiVersion()) { 668 case ScriptGroupBase::SG_V1: { 669 CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 670 if (!sgi->init()) { 671 delete sgi; 672 return nullptr; 673 } 674 return sgi; 675 } 676 case ScriptGroupBase::SG_V2: { 677 return new CpuScriptGroup2Impl(this, sg); 678 } 679 } 680 return nullptr; 681} 682