rsCpuCore.cpp revision 554054c2317cc38df1e57186caf171f1b07ec6e3
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 40static pid_t gettid() { 41 return syscall(SYS_gettid); 42} 43#endif 44 45using namespace android; 46using namespace android::renderscript; 47 48#define REDUCE_NEW_ALOGV(mtls, level, ...) do { if ((mtls)->logReduce >= (level)) ALOGV(__VA_ARGS__); } while(0) 49 50static pthread_key_t gThreadTLSKey = 0; 51static uint32_t gThreadTLSKeyCount = 0; 52static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER; 53 54bool android::renderscript::gArchUseSIMD = false; 55 56RsdCpuReference::~RsdCpuReference() { 57} 58 59RsdCpuReference * 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 83Context * RsdCpuReference::getTlsContext() { 84 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 85 return tls->mContext; 86} 87 88const Script * RsdCpuReference::getTlsScript() { 89 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 90 return tls->mScript; 91} 92 93pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 94 95//////////////////////////////////////////////////////////// 96/// 97 98RsdCpuReferenceImpl::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 113void * 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. 155void 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 187void RsdCpuReferenceImpl::lockMutex() { 188 pthread_mutex_lock(&gInitMutex); 189} 190 191void RsdCpuReferenceImpl::unlockMutex() { 192 pthread_mutex_unlock(&gInitMutex); 193} 194 195// Determine if the CPU we're running on supports SIMD instructions. 196static 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 220bool 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 300void 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 306RsdCpuReferenceImpl::~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 339static 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 357static inline void RedpPtrSetup(const MTLaunchStructReduceNew *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 364static 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 377static 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 389static 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 393static 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 420static 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 445static 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// 491static const int kFormatInBytesMax = 16; 492// ": " + 2 digits per byte + 1 separator between bytes + "..." + null 493typedef char FormatBuf[2 + kFormatInBytesMax*2 + (kFormatInBytesMax - 1) + 3 + 1]; 494static 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 511static void reduce_new_get_accumulator(uint8_t *&accumPtr, const MTLaunchStructReduceNew *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_NEW_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 535static void walk_1d_reduce_new(void *usr, uint32_t idx) { 536 const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr; 537 RsExpandKernelDriverInfo redp = mtls->redp; 538 539 // find accumulator 540 uint8_t *&accumPtr = mtls->accumPtr[idx]; 541 if (!accumPtr) { 542 reduce_new_get_accumulator(accumPtr, mtls, __func__, idx); 543 } 544 545 // accumulate 546 const ReduceNewAccumulatorFunc_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_NEW_ALOGV(mtls, 2, "walk_1d_reduce_new(%p): idx = %u, x in [%u, %u)%s", 570 mtls->accumFunc, idx, xStart, xEnd, fmt); 571 } 572} 573 574static void walk_2d_reduce_new(void *usr, uint32_t idx) { 575 const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr; 576 RsExpandKernelDriverInfo redp = mtls->redp; 577 578 // find accumulator 579 uint8_t *&accumPtr = mtls->accumPtr[idx]; 580 if (!accumPtr) { 581 reduce_new_get_accumulator(accumPtr, mtls, __func__, idx); 582 } 583 584 // accumulate 585 const ReduceNewAccumulatorFunc_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_NEW_ALOGV(mtls, 2, "walk_2d_reduce_new(%p): idx = %u, y in [%u, %u)%s", 609 mtls->accumFunc, idx, yStart, yEnd, fmt); 610 } 611} 612 613static void walk_3d_reduce_new(void *usr, uint32_t idx) { 614 const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr; 615 RsExpandKernelDriverInfo redp = mtls->redp; 616 617 // find accumulator 618 uint8_t *&accumPtr = mtls->accumPtr[idx]; 619 if (!accumPtr) { 620 reduce_new_get_accumulator(accumPtr, mtls, __func__, idx); 621 } 622 623 // accumulate 624 const ReduceNewAccumulatorFunc_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_NEW_ALOGV(mtls, 2, "walk_3d_reduce_new(%p): idx = %u, z = %u%s", 644 mtls->accumFunc, idx, redp.current.z, fmt); 645 } 646} 647 648// Launch a simple reduce-style kernel. 649// Inputs: 650// ain: The allocation that contains the input 651// aout: The allocation that will hold the output 652// mtls: Holds launch parameters 653void RsdCpuReferenceImpl::launchReduce(const Allocation *ain, 654 Allocation *aout, 655 MTLaunchStructReduce *mtls) { 656 const uint32_t xStart = mtls->start.x; 657 const uint32_t xEnd = mtls->end.x; 658 659 if (xStart >= xEnd) { 660 return; 661 } 662 663 const uint32_t startOffset = ain->getType()->getElementSizeBytes() * xStart; 664 mtls->kernel(&mtls->inBuf[startOffset], mtls->outBuf, xEnd - xStart); 665} 666 667// Launch a general reduce-style kernel. 668// Inputs: 669// ains[0..inLen-1]: Array of allocations that contain the inputs 670// aout: The allocation that will hold the output 671// mtls: Holds launch parameters 672void RsdCpuReferenceImpl::launchReduceNew(const Allocation ** ains, 673 uint32_t inLen, 674 Allocation * aout, 675 MTLaunchStructReduceNew *mtls) { 676 mtls->logReduce = mRSC->props.mLogReduce; 677 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 678 launchReduceNewParallel(ains, inLen, aout, mtls); 679 } else { 680 launchReduceNewSerial(ains, inLen, aout, mtls); 681 } 682} 683 684// Launch a general reduce-style kernel, single-threaded. 685// Inputs: 686// ains[0..inLen-1]: Array of allocations that contain the inputs 687// aout: The allocation that will hold the output 688// mtls: Holds launch parameters 689void RsdCpuReferenceImpl::launchReduceNewSerial(const Allocation ** ains, 690 uint32_t inLen, 691 Allocation * aout, 692 MTLaunchStructReduceNew *mtls) { 693 REDUCE_NEW_ALOGV(mtls, 1, "launchReduceNewSerial(%p): %u x %u x %u", mtls->accumFunc, 694 mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z); 695 696 // In the presence of outconverter, we allocate temporary memory for 697 // the accumulator. 698 // 699 // In the absence of outconverter, we use the output allocation as the 700 // accumulator. 701 uint8_t *const accumPtr = (mtls->outFunc 702 ? static_cast<uint8_t *>(malloc(mtls->accumSize)) 703 : mtls->redp.outPtr[0]); 704 705 // initialize 706 if (mtls->initFunc) { 707 mtls->initFunc(accumPtr); 708 } else { 709 memset(accumPtr, 0, mtls->accumSize); 710 } 711 712 // accumulate 713 const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc; 714 uint32_t slice = 0; 715 while (SelectOuterSlice(mtls, &mtls->redp, slice++)) { 716 for (mtls->redp.current.y = mtls->start.y; 717 mtls->redp.current.y < mtls->end.y; 718 mtls->redp.current.y++) { 719 RedpPtrSetup(mtls, &mtls->redp, mtls->start.x, mtls->redp.current.y, mtls->redp.current.z); 720 fn(&mtls->redp, mtls->start.x, mtls->end.x, accumPtr); 721 } 722 } 723 724 // outconvert 725 if (mtls->outFunc) { 726 mtls->outFunc(mtls->redp.outPtr[0], accumPtr); 727 free(accumPtr); 728 } 729} 730 731// Launch a general reduce-style kernel, multi-threaded. 732// Inputs: 733// ains[0..inLen-1]: Array of allocations that contain the inputs 734// aout: The allocation that will hold the output 735// mtls: Holds launch parameters 736void RsdCpuReferenceImpl::launchReduceNewParallel(const Allocation ** ains, 737 uint32_t inLen, 738 Allocation * aout, 739 MTLaunchStructReduceNew *mtls) { 740 // For now, we don't know how to go parallel in the absence of a combiner. 741 if (!mtls->combFunc) { 742 launchReduceNewSerial(ains, inLen, aout, mtls); 743 return; 744 } 745 746 // Number of threads = "main thread" + number of other (worker) threads 747 const uint32_t numThreads = mWorkers.mCount + 1; 748 749 // In the absence of outconverter, we use the output allocation as 750 // an accumulator, and therefore need to allocate one fewer accumulator. 751 const uint32_t numAllocAccum = numThreads - (mtls->outFunc == nullptr); 752 753 // If mDebugReduceSplitAccum, then we want each accumulator to start 754 // on a page boundary. (TODO: Would some unit smaller than a page 755 // be sufficient to avoid false sharing?) 756 if (mRSC->props.mDebugReduceSplitAccum) { 757 // Round up accumulator size to an integral number of pages 758 mtls->accumStride = 759 (unsigned(mtls->accumSize) + unsigned(mPageSize)-1) & 760 ~(unsigned(mPageSize)-1); 761 // Each accumulator gets its own page. Alternatively, if we just 762 // wanted to make sure no two accumulators are on the same page, 763 // we could instead do 764 // allocSize = mtls->accumStride * (numAllocation - 1) + mtls->accumSize 765 const size_t allocSize = mtls->accumStride * numAllocAccum; 766 mtls->accumAlloc = static_cast<uint8_t *>(memalign(mPageSize, allocSize)); 767 } else { 768 mtls->accumStride = mtls->accumSize; 769 mtls->accumAlloc = static_cast<uint8_t *>(malloc(mtls->accumStride * numAllocAccum)); 770 } 771 772 const size_t accumPtrArrayBytes = sizeof(uint8_t *) * numThreads; 773 mtls->accumPtr = static_cast<uint8_t **>(malloc(accumPtrArrayBytes)); 774 memset(mtls->accumPtr, 0, accumPtrArrayBytes); 775 776 mtls->accumCount = 0; 777 778 rsAssert(!mInKernel); 779 mInKernel = true; 780 REDUCE_NEW_ALOGV(mtls, 1, "launchReduceNewParallel(%p): %u x %u x %u, %u threads, accumAlloc = %p", 781 mtls->accumFunc, 782 mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z, 783 numThreads, mtls->accumAlloc); 784 if (mtls->redp.dim.z > 1) { 785 mtls->mSliceSize = 1; 786 launchThreads(walk_3d_reduce_new, mtls); 787 } else if (mtls->redp.dim.y > 1) { 788 mtls->mSliceSize = rsMax(1U, mtls->redp.dim.y / (numThreads * 4)); 789 launchThreads(walk_2d_reduce_new, mtls); 790 } else { 791 mtls->mSliceSize = rsMax(1U, mtls->redp.dim.x / (numThreads * 4)); 792 launchThreads(walk_1d_reduce_new, mtls); 793 } 794 mInKernel = false; 795 796 // Combine accumulators and identify final accumulator 797 uint8_t *finalAccumPtr = (mtls->outFunc ? nullptr : mtls->redp.outPtr[0]); 798 // Loop over accumulators, combining into finalAccumPtr. If finalAccumPtr 799 // is null, then the first accumulator I find becomes finalAccumPtr. 800 for (unsigned idx = 0; idx < mtls->accumCount; ++idx) { 801 uint8_t *const thisAccumPtr = mtls->accumPtr[idx]; 802 if (finalAccumPtr) { 803 if (finalAccumPtr != thisAccumPtr) { 804 if (mtls->combFunc) { 805 if (mtls->logReduce >= 3) { 806 FormatBuf fmt; 807 REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): accumulating into%s", 808 mtls->accumFunc, 809 format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 810 REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): accumulator[%d]%s", 811 mtls->accumFunc, idx, 812 format_bytes(&fmt, thisAccumPtr, mtls->accumSize)); 813 } 814 mtls->combFunc(finalAccumPtr, thisAccumPtr); 815 } else { 816 rsAssert(!"expected combiner"); 817 } 818 } 819 } else { 820 finalAccumPtr = thisAccumPtr; 821 } 822 } 823 rsAssert(finalAccumPtr != nullptr); 824 if (mtls->logReduce >= 3) { 825 FormatBuf fmt; 826 REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): final accumulator%s", 827 mtls->accumFunc, format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 828 } 829 830 // Outconvert 831 if (mtls->outFunc) { 832 mtls->outFunc(mtls->redp.outPtr[0], finalAccumPtr); 833 if (mtls->logReduce >= 3) { 834 FormatBuf fmt; 835 REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): final outconverted result%s", 836 mtls->accumFunc, 837 format_bytes(&fmt, mtls->redp.outPtr[0], mtls->redp.outStride[0])); 838 } 839 } 840 841 // Clean up 842 free(mtls->accumPtr); 843 free(mtls->accumAlloc); 844} 845 846 847void RsdCpuReferenceImpl::launchForEach(const Allocation ** ains, 848 uint32_t inLen, 849 Allocation* aout, 850 const RsScriptCall* sc, 851 MTLaunchStructForEach* mtls) { 852 853 //android::StopWatch kernel_time("kernel time"); 854 855 bool outerDims = (mtls->start.z != mtls->end.z) || 856 (mtls->start.face != mtls->end.face) || 857 (mtls->start.lod != mtls->end.lod) || 858 (mtls->start.array[0] != mtls->end.array[0]) || 859 (mtls->start.array[1] != mtls->end.array[1]) || 860 (mtls->start.array[2] != mtls->end.array[2]) || 861 (mtls->start.array[3] != mtls->end.array[3]); 862 863 if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 864 const size_t targetByteChunk = 16 * 1024; 865 mInKernel = true; // NOTE: The guard immediately above ensures this was !mInKernel 866 867 if (outerDims) { 868 // No fancy logic for chunk size 869 mtls->mSliceSize = 1; 870 launchThreads(walk_general_foreach, mtls); 871 } else if (mtls->fep.dim.y > 1) { 872 uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4); 873 uint32_t s2 = 0; 874 875 // This chooses our slice size to rate limit atomic ops to 876 // one per 16k bytes of reads/writes. 877 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) { 878 s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride; 879 } else if (mtls->ains[0]) { 880 s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride; 881 } else { 882 // Launch option only case 883 // Use s1 based only on the dimensions 884 s2 = s1; 885 } 886 mtls->mSliceSize = rsMin(s1, s2); 887 888 if(mtls->mSliceSize < 1) { 889 mtls->mSliceSize = 1; 890 } 891 892 launchThreads(walk_2d_foreach, mtls); 893 } else { 894 uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4); 895 uint32_t s2 = 0; 896 897 // This chooses our slice size to rate limit atomic ops to 898 // one per 16k bytes of reads/writes. 899 if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) { 900 s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes(); 901 } else if (mtls->ains[0]) { 902 s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes(); 903 } else { 904 // Launch option only case 905 // Use s1 based only on the dimensions 906 s2 = s1; 907 } 908 mtls->mSliceSize = rsMin(s1, s2); 909 910 if (mtls->mSliceSize < 1) { 911 mtls->mSliceSize = 1; 912 } 913 914 launchThreads(walk_1d_foreach, mtls); 915 } 916 mInKernel = false; 917 918 } else { 919 ForEachFunc_t fn = mtls->kernel; 920 uint32_t slice = 0; 921 922 923 while(SelectOuterSlice(mtls, &mtls->fep, slice++)) { 924 for (mtls->fep.current.y = mtls->start.y; 925 mtls->fep.current.y < mtls->end.y; 926 mtls->fep.current.y++) { 927 928 FepPtrSetup(mtls, &mtls->fep, mtls->start.x, 929 mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod, 930 (RsAllocationCubemapFace) mtls->fep.current.face, 931 mtls->fep.current.array[0], mtls->fep.current.array[1], 932 mtls->fep.current.array[2], mtls->fep.current.array[3]); 933 934 fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 935 } 936 } 937 } 938} 939 940RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 941 //ALOGE("setTls %p", sc); 942 ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 943 rsAssert(tls); 944 RsdCpuScriptImpl *old = tls->mImpl; 945 tls->mImpl = sc; 946 tls->mContext = mRSC; 947 if (sc) { 948 tls->mScript = sc->getScript(); 949 } else { 950 tls->mScript = nullptr; 951 } 952 return old; 953} 954 955const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 956 return mSymLookupFn(mRSC, name); 957} 958 959 960RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 961 char const *resName, char const *cacheDir, 962 uint8_t const *bitcode, size_t bitcodeSize, 963 uint32_t flags) { 964 965 RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 966 if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 967 , getBccPluginName() 968 )) { 969 delete i; 970 return nullptr; 971 } 972 return i; 973} 974 975extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 976 const Script *s, const Element *e); 977extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 978 const Script *s, const Element *e); 979extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 980 const Script *s, const Element *e); 981extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 982 const Script *s, const Element *e); 983extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 984 const Script *s, const Element *e); 985extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 986 const Script *s, const Element *e); 987extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 988 const Script *s, const Element *e); 989extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 990 const Script *s, const Element *e); 991extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 992 const Script *s, const Element *e); 993extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 994 const Script *s, const Element *e); 995extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx, 996 const Script *s, const Element *e); 997 998RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 999 RsScriptIntrinsicID iid, Element *e) { 1000 1001 RsdCpuScriptImpl *i = nullptr; 1002 switch (iid) { 1003 case RS_SCRIPT_INTRINSIC_ID_3DLUT: 1004 i = rsdIntrinsic_3DLUT(this, s, e); 1005 break; 1006 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 1007 i = rsdIntrinsic_Convolve3x3(this, s, e); 1008 break; 1009 case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 1010 i = rsdIntrinsic_ColorMatrix(this, s, e); 1011 break; 1012 case RS_SCRIPT_INTRINSIC_ID_LUT: 1013 i = rsdIntrinsic_LUT(this, s, e); 1014 break; 1015 case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 1016 i = rsdIntrinsic_Convolve5x5(this, s, e); 1017 break; 1018 case RS_SCRIPT_INTRINSIC_ID_BLUR: 1019 i = rsdIntrinsic_Blur(this, s, e); 1020 break; 1021 case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 1022 i = rsdIntrinsic_YuvToRGB(this, s, e); 1023 break; 1024 case RS_SCRIPT_INTRINSIC_ID_BLEND: 1025 i = rsdIntrinsic_Blend(this, s, e); 1026 break; 1027 case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 1028 i = rsdIntrinsic_Histogram(this, s, e); 1029 break; 1030 case RS_SCRIPT_INTRINSIC_ID_RESIZE: 1031 i = rsdIntrinsic_Resize(this, s, e); 1032 break; 1033 case RS_SCRIPT_INTRINSIC_ID_BLAS: 1034 i = rsdIntrinsic_BLAS(this, s, e); 1035 break; 1036 1037 default: 1038 rsAssert(0); 1039 } 1040 1041 return i; 1042} 1043 1044void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) { 1045 switch (sg->getApiVersion()) { 1046 case ScriptGroupBase::SG_V1: { 1047 CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 1048 if (!sgi->init()) { 1049 delete sgi; 1050 return nullptr; 1051 } 1052 return sgi; 1053 } 1054 case ScriptGroupBase::SG_V2: { 1055 return new CpuScriptGroup2Impl(this, sg); 1056 } 1057 } 1058 return nullptr; 1059} 1060