rsCpuScript.cpp revision 5e48002cc4a11e9ce63852a77488cfb9f6765c2b
1/* 2 * Copyright (C) 2011-2012 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 "rsCpuExecutable.h" 20 21#ifdef RS_COMPATIBILITY_LIB 22 #include <stdio.h> 23 #include <sys/stat.h> 24 #include <unistd.h> 25#else 26 #include "rsCppUtils.h" 27 28 #include <bcc/Config/Config.h> 29 #include <bcinfo/MetadataExtractor.h> 30 #include <cutils/properties.h> 31 32 #include <zlib.h> 33 #include <sys/file.h> 34 #include <sys/types.h> 35 #include <unistd.h> 36 37 #include <string> 38 #include <vector> 39#endif 40 41#include <set> 42#include <string> 43#include <dlfcn.h> 44#include <stdlib.h> 45#include <string.h> 46#include <iostream> 47#include <sstream> 48 49namespace { 50 51static const bool kDebugGlobalVariables = false; 52 53static bool allocationLODIsNull(const android::renderscript::Allocation *alloc) { 54 // Even if alloc != nullptr, mallocPtr could be null if 55 // IO_OUTPUT/IO_INPUT with no bound surface. 56 return alloc && alloc->mHal.drvState.lod[0].mallocPtr == nullptr; 57} 58 59#ifndef RS_COMPATIBILITY_LIB 60 61static bool is_force_recompile() { 62#ifdef RS_SERVER 63 return false; 64#else 65 char buf[PROPERTY_VALUE_MAX]; 66 67 // Re-compile if floating point precision has been overridden. 68 property_get("debug.rs.precision", buf, ""); 69 if (buf[0] != '\0') { 70 return true; 71 } 72 73 // Re-compile if debug.rs.forcerecompile is set. 74 property_get("debug.rs.forcerecompile", buf, "0"); 75 if ((::strcmp(buf, "1") == 0) || (::strcmp(buf, "true") == 0)) { 76 return true; 77 } else { 78 return false; 79 } 80#endif // RS_SERVER 81} 82 83static void setCompileArguments(std::vector<const char*>* args, 84 const std::string& bcFileName, 85 const char* cacheDir, const char* resName, 86 const char* core_lib, bool useRSDebugContext, 87 const char* bccPluginName, bool emitGlobalInfo, 88 int optLevel, bool emitGlobalInfoSkipConstant) { 89 rsAssert(cacheDir && resName && core_lib); 90 args->push_back(android::renderscript::RsdCpuScriptImpl::BCC_EXE_PATH); 91 args->push_back("-unroll-runtime"); 92 args->push_back("-scalarize-load-store"); 93 if (emitGlobalInfo) { 94 args->push_back("-rs-global-info"); 95 if (emitGlobalInfoSkipConstant) { 96 args->push_back("-rs-global-info-skip-constant"); 97 } 98 } 99 args->push_back("-o"); 100 args->push_back(resName); 101 args->push_back("-output_path"); 102 args->push_back(cacheDir); 103 args->push_back("-bclib"); 104 args->push_back(core_lib); 105 args->push_back("-mtriple"); 106 args->push_back(DEFAULT_TARGET_TRIPLE_STRING); 107 args->push_back("-O"); 108 109 switch (optLevel) { 110 case 0: 111 args->push_back("0"); 112 break; 113 case 3: 114 args->push_back("3"); 115 break; 116 default: 117 ALOGW("Expected optimization level of 0 or 3. Received %d", optLevel); 118 args->push_back("3"); 119 break; 120 } 121 122 // Enable workaround for A53 codegen by default. 123#if defined(__aarch64__) && !defined(DISABLE_A53_WORKAROUND) 124 args->push_back("-aarch64-fix-cortex-a53-835769"); 125#endif 126 127 // Execute the bcc compiler. 128 if (useRSDebugContext) { 129 args->push_back("-rs-debug-ctx"); 130 } else { 131 // Only load additional libraries for compiles that don't use 132 // the debug context. 133 if (bccPluginName && strlen(bccPluginName) > 0) { 134 args->push_back("-load"); 135 args->push_back(bccPluginName); 136 } 137 } 138 139 args->push_back("-fPIC"); 140 args->push_back("-embedRSInfo"); 141 142 args->push_back(bcFileName.c_str()); 143 args->push_back(nullptr); 144} 145 146static bool compileBitcode(const std::string &bcFileName, 147 const char *bitcode, 148 size_t bitcodeSize, 149 std::vector<const char *> &compileArguments) { 150 rsAssert(bitcode && bitcodeSize); 151 152 FILE *bcfile = fopen(bcFileName.c_str(), "w"); 153 if (!bcfile) { 154 ALOGE("Could not write to %s", bcFileName.c_str()); 155 return false; 156 } 157 size_t nwritten = fwrite(bitcode, 1, bitcodeSize, bcfile); 158 fclose(bcfile); 159 if (nwritten != bitcodeSize) { 160 ALOGE("Could not write %zu bytes to %s", bitcodeSize, 161 bcFileName.c_str()); 162 return false; 163 } 164 165 return android::renderscript::rsuExecuteCommand( 166 android::renderscript::RsdCpuScriptImpl::BCC_EXE_PATH, 167 compileArguments.size()-1, compileArguments.data()); 168} 169 170// The checksum is unnecessary under a few conditions, since the primary 171// use-case for it is debugging. If we are loading something from the 172// system partition (read-only), we know that it was precompiled as part of 173// application ahead of time (and thus the checksum is completely 174// unnecessary). The checksum is also unnecessary on release (non-debug) 175// builds, as the only way to get a shared object is to have compiled the 176// script once already. On a release build, there is no way to adjust the 177// other libraries/dependencies, and so the only reason to recompile would 178// be for a source APK change or an OTA. In either case, the APK would be 179// reinstalled, which would already clear the code_cache/ directory. 180bool isChecksumNeeded(const char *cacheDir) { 181 if ((::strcmp(SYSLIBPATH, cacheDir) == 0) || 182 (::strcmp(SYSLIBPATH_VENDOR, cacheDir) == 0)) 183 return false; 184 char buf[PROPERTY_VALUE_MAX]; 185 property_get("ro.debuggable", buf, ""); 186 return (buf[0] == '1'); 187} 188 189bool addFileToChecksum(const char *fileName, uint32_t &checksum) { 190 int FD = open(fileName, O_RDONLY); 191 if (FD == -1) { 192 ALOGE("Cannot open file \'%s\' to compute checksum", fileName); 193 return false; 194 } 195 196 char buf[256]; 197 while (true) { 198 ssize_t nread = read(FD, buf, sizeof(buf)); 199 if (nread < 0) { // bail out on failed read 200 ALOGE("Error while computing checksum for file \'%s\'", fileName); 201 return false; 202 } 203 204 checksum = adler32(checksum, (const unsigned char *) buf, nread); 205 if (static_cast<size_t>(nread) < sizeof(buf)) // EOF 206 break; 207 } 208 209 if (close(FD) != 0) { 210 ALOGE("Cannot close file \'%s\' after computing checksum", fileName); 211 return false; 212 } 213 return true; 214} 215 216#endif // !defined(RS_COMPATIBILITY_LIB) 217} // namespace 218 219namespace android { 220namespace renderscript { 221 222#ifndef RS_COMPATIBILITY_LIB 223 224uint32_t constructBuildChecksum(uint8_t const *bitcode, size_t bitcodeSize, 225 const char *commandLine, 226 const char** bccFiles, size_t numFiles) { 227 uint32_t checksum = adler32(0L, Z_NULL, 0); 228 229 // include checksum of bitcode 230 if (bitcode != nullptr && bitcodeSize > 0) { 231 checksum = adler32(checksum, bitcode, bitcodeSize); 232 } 233 234 // include checksum of command line arguments 235 checksum = adler32(checksum, (const unsigned char *) commandLine, 236 strlen(commandLine)); 237 238 // include checksum of bccFiles 239 for (size_t i = 0; i < numFiles; i++) { 240 const char* bccFile = bccFiles[i]; 241 if (bccFile[0] != 0 && !addFileToChecksum(bccFile, checksum)) { 242 // return empty checksum instead of something partial/corrupt 243 return 0; 244 } 245 } 246 247 return checksum; 248} 249 250#endif // !RS_COMPATIBILITY_LIB 251 252RsdCpuScriptImpl::RsdCpuScriptImpl(RsdCpuReferenceImpl *ctx, const Script *s) { 253 mCtx = ctx; 254 mScript = s; 255 256 mScriptSO = nullptr; 257 258 mRoot = nullptr; 259 mRootExpand = nullptr; 260 mInit = nullptr; 261 mFreeChildren = nullptr; 262 mScriptExec = nullptr; 263 264 mBoundAllocs = nullptr; 265 mIntrinsicData = nullptr; 266 mIsThreadable = true; 267 268 mBuildChecksum = 0; 269 mChecksumNeeded = false; 270} 271 272bool RsdCpuScriptImpl::storeRSInfoFromSO() { 273 // The shared object may have an invalid build checksum. 274 // Validate and fail early. 275 mScriptExec = ScriptExecutable::createFromSharedObject( 276 mScriptSO, mChecksumNeeded ? mBuildChecksum : 0); 277 278 if (mScriptExec == nullptr) { 279 return false; 280 } 281 282 mRoot = (RootFunc_t) dlsym(mScriptSO, "root"); 283 if (mRoot) { 284 //ALOGE("Found root(): %p", mRoot); 285 } 286 mRootExpand = (RootFunc_t) dlsym(mScriptSO, "root.expand"); 287 if (mRootExpand) { 288 //ALOGE("Found root.expand(): %p", mRootExpand); 289 } 290 mInit = (InitOrDtorFunc_t) dlsym(mScriptSO, "init"); 291 if (mInit) { 292 //ALOGE("Found init(): %p", mInit); 293 } 294 mFreeChildren = (InitOrDtorFunc_t) dlsym(mScriptSO, ".rs.dtor"); 295 if (mFreeChildren) { 296 //ALOGE("Found .rs.dtor(): %p", mFreeChildren); 297 } 298 299 size_t varCount = mScriptExec->getExportedVariableCount(); 300 if (varCount > 0) { 301 mBoundAllocs = new Allocation *[varCount]; 302 memset(mBoundAllocs, 0, varCount * sizeof(*mBoundAllocs)); 303 } 304 305 mIsThreadable = mScriptExec->getThreadable(); 306 //ALOGE("Script isThreadable? %d", mIsThreadable); 307 308 if (kDebugGlobalVariables) { 309 mScriptExec->dumpGlobalInfo(); 310 } 311 312 return true; 313} 314 315bool RsdCpuScriptImpl::init(char const *resName, char const *cacheDir, 316 uint8_t const *bitcode, size_t bitcodeSize, 317 uint32_t flags, char const *bccPluginName) { 318 //ALOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir, 319 // bitcode, bitcodeSize, flags, lookupFunc); 320 //ALOGE("rsdScriptInit %p %p", rsc, script); 321 322 mCtx->lockMutex(); 323#ifndef RS_COMPATIBILITY_LIB 324 bool useRSDebugContext = false; 325 326 bcinfo::MetadataExtractor bitcodeMetadata((const char *) bitcode, bitcodeSize); 327 if (!bitcodeMetadata.extract()) { 328 ALOGE("Could not extract metadata from bitcode"); 329 mCtx->unlockMutex(); 330 return false; 331 } 332 333 const char* core_lib = findCoreLib(bitcodeMetadata, (const char*)bitcode, bitcodeSize); 334 335 if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) { 336 useRSDebugContext = true; 337 } 338 339 int optLevel = mCtx->getContext()->getOptLevel(); 340 341 std::string bcFileName(cacheDir); 342 bcFileName.append("/"); 343 bcFileName.append(resName); 344 bcFileName.append(".bc"); 345 346 std::vector<const char*> compileArguments; 347 bool emitGlobalInfo = mCtx->getEmbedGlobalInfo(); 348 bool emitGlobalInfoSkipConstant = mCtx->getEmbedGlobalInfoSkipConstant(); 349 setCompileArguments(&compileArguments, bcFileName, cacheDir, resName, core_lib, 350 useRSDebugContext, bccPluginName, emitGlobalInfo, 351 optLevel, emitGlobalInfoSkipConstant); 352 353 mChecksumNeeded = isChecksumNeeded(cacheDir); 354 if (mChecksumNeeded) { 355 std::vector<const char *> bccFiles = { BCC_EXE_PATH, 356 core_lib, 357 }; 358 359 // The last argument of compileArguments is a nullptr, so remove 1 from 360 // the size. 361 std::unique_ptr<const char> compileCommandLine( 362 rsuJoinStrings(compileArguments.size()-1, compileArguments.data())); 363 364 mBuildChecksum = constructBuildChecksum(bitcode, bitcodeSize, 365 compileCommandLine.get(), 366 bccFiles.data(), bccFiles.size()); 367 368 if (mBuildChecksum == 0) { 369 // cannot compute checksum but verification is enabled 370 mCtx->unlockMutex(); 371 return false; 372 } 373 } 374 else { 375 // add a dummy/constant as a checksum if verification is disabled 376 mBuildChecksum = 0xabadcafe; 377 } 378 379 // Append build checksum to commandline 380 // Handle the terminal nullptr in compileArguments 381 compileArguments.pop_back(); 382 compileArguments.push_back("-build-checksum"); 383 std::stringstream ss; 384 ss << std::hex << mBuildChecksum; 385 compileArguments.push_back(ss.str().c_str()); 386 compileArguments.push_back(nullptr); 387 388 if (!is_force_recompile() && !useRSDebugContext) { 389 mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 390 391 // Read RS info from the shared object to detect checksum mismatch 392 if (mScriptSO != nullptr && !storeRSInfoFromSO()) { 393 dlclose(mScriptSO); 394 mScriptSO = nullptr; 395 } 396 } 397 398 // If we can't, it's either not there or out of date. We compile the bit code and try loading 399 // again. 400 if (mScriptSO == nullptr) { 401 if (!compileBitcode(bcFileName, (const char*)bitcode, bitcodeSize, 402 compileArguments)) 403 { 404 ALOGE("bcc: FAILS to compile '%s'", resName); 405 mCtx->unlockMutex(); 406 return false; 407 } 408 409 if (!SharedLibraryUtils::createSharedLibrary(mCtx->getContext()->getDriverName(), 410 cacheDir, resName)) { 411 ALOGE("Linker: Failed to link object file '%s'", resName); 412 mCtx->unlockMutex(); 413 return false; 414 } 415 416 mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 417 if (mScriptSO == nullptr) { 418 ALOGE("Unable to load '%s'", resName); 419 mCtx->unlockMutex(); 420 return false; 421 } 422 423 // Read RS symbol information from the .so. 424 if (!storeRSInfoFromSO()) { 425 goto error; 426 } 427 } 428 429 mBitcodeFilePath.setTo(bcFileName.c_str()); 430 431#else // RS_COMPATIBILITY_LIB is defined 432 const char *nativeLibDir = mCtx->getContext()->getNativeLibDir(); 433 mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName, nativeLibDir); 434 435 if (!mScriptSO) { 436 goto error; 437 } 438 439 if (!storeRSInfoFromSO()) { 440 goto error; 441 } 442#endif 443 mCtx->unlockMutex(); 444 return true; 445 446error: 447 448 mCtx->unlockMutex(); 449 if (mScriptSO) { 450 dlclose(mScriptSO); 451 mScriptSO = nullptr; 452 } 453 return false; 454} 455 456#ifndef RS_COMPATIBILITY_LIB 457 458const char* RsdCpuScriptImpl::findCoreLib(const bcinfo::MetadataExtractor& ME, const char* bitcode, 459 size_t bitcodeSize) { 460 const char* defaultLib = SYSLIBPATH"/libclcore.bc"; 461 462 // If we're debugging, use the debug library. 463 if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) { 464 return SYSLIBPATH"/libclcore_debug.bc"; 465 } 466 467 if (ME.hasDebugInfo()) { 468 return SYSLIBPATH"/libclcore_g.bc"; 469 } 470 471 // If a callback has been registered to specify a library, use that. 472 RSSelectRTCallback selectRTCallback = mCtx->getSelectRTCallback(); 473 if (selectRTCallback != nullptr) { 474 return selectRTCallback((const char*)bitcode, bitcodeSize); 475 } 476 477 // Check for a platform specific library 478#if defined(ARCH_ARM_HAVE_NEON) && !defined(DISABLE_CLCORE_NEON) 479 enum bcinfo::RSFloatPrecision prec = ME.getRSFloatPrecision(); 480 if (prec == bcinfo::RS_FP_Relaxed) { 481 // NEON-capable ARMv7a devices can use an accelerated math library 482 // for all reduced precision scripts. 483 // ARMv8 does not use NEON, as ASIMD can be used with all precision 484 // levels. 485 return SYSLIBPATH"/libclcore_neon.bc"; 486 } else { 487 return defaultLib; 488 } 489#elif defined(__i386__) || defined(__x86_64__) 490 // x86 devices will use an optimized library. 491 return SYSLIBPATH"/libclcore_x86.bc"; 492#else 493 return defaultLib; 494#endif 495} 496 497#endif 498 499void RsdCpuScriptImpl::populateScript(Script *script) { 500 // Copy info over to runtime 501 script->mHal.info.exportedFunctionCount = mScriptExec->getExportedFunctionCount(); 502 script->mHal.info.exportedReduceCount = mScriptExec->getExportedReduceCount(); 503 script->mHal.info.exportedReduceNewCount = mScriptExec->getExportedReduceNewCount(); 504 script->mHal.info.exportedForEachCount = mScriptExec->getExportedForEachCount(); 505 script->mHal.info.exportedVariableCount = mScriptExec->getExportedVariableCount(); 506 script->mHal.info.exportedPragmaCount = mScriptExec->getPragmaCount();; 507 script->mHal.info.exportedPragmaKeyList = mScriptExec->getPragmaKeys(); 508 script->mHal.info.exportedPragmaValueList = mScriptExec->getPragmaValues(); 509 510 // Bug, need to stash in metadata 511 if (mRootExpand) { 512 script->mHal.info.root = mRootExpand; 513 } else { 514 script->mHal.info.root = mRoot; 515 } 516} 517 518// Set up the launch dimensions, and write the values of the launch 519// dimensions into the mtls start/end fields. 520// 521// Inputs: 522// baseDim - base shape of the input 523// sc - used to constrain the launch dimensions 524// 525// Returns: 526// True on success, false on failure to set up 527bool RsdCpuScriptImpl::setUpMtlsDimensions(MTLaunchStructCommon *mtls, 528 const RsLaunchDimensions &baseDim, 529 const RsScriptCall *sc) { 530 rsAssert(mtls); 531 532#define SET_UP_DIMENSION(DIM_FIELD, SC_FIELD) do { \ 533 if (!sc || (sc->SC_FIELD##End == 0)) { \ 534 mtls->end.DIM_FIELD = baseDim.DIM_FIELD; \ 535 } else { \ 536 mtls->start.DIM_FIELD = \ 537 rsMin(baseDim.DIM_FIELD, sc->SC_FIELD##Start); \ 538 mtls->end.DIM_FIELD = \ 539 rsMin(baseDim.DIM_FIELD, sc->SC_FIELD##End); \ 540 if (mtls->start.DIM_FIELD >= mtls->end.DIM_FIELD) { \ 541 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, \ 542 "Failed to launch kernel; Invalid " \ 543 #SC_FIELD "Start or " #SC_FIELD "End."); \ 544 return false; \ 545 } \ 546 }} while(0) 547 548 SET_UP_DIMENSION(x, x); 549 SET_UP_DIMENSION(y, y); 550 SET_UP_DIMENSION(z, z); 551 SET_UP_DIMENSION(array[0], array); 552 SET_UP_DIMENSION(array[1], array2); 553 SET_UP_DIMENSION(array[2], array3); 554 SET_UP_DIMENSION(array[3], array4); 555#undef SET_UP_DIMENSION 556 557 return true; 558} 559 560// Preliminary work to prepare a simple reduce-style kernel for launch. 561bool RsdCpuScriptImpl::reduceMtlsSetup(const Allocation *ain, 562 const Allocation *aout, 563 const RsScriptCall *sc, 564 MTLaunchStructReduce *mtls) { 565 rsAssert(ain && aout); 566 memset(mtls, 0, sizeof(MTLaunchStructReduce)); 567 mtls->dimPtr = &mtls->inputDim; 568 569 if (allocationLODIsNull(ain) || allocationLODIsNull(aout)) { 570 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 571 "reduce called with a null allocation"); 572 return false; 573 } 574 575 // Set up the dimensions of the input. 576 const Type *inType = ain->getType(); 577 mtls->inputDim.x = inType->getDimX(); 578 rsAssert(inType->getDimY() == 0); 579 580 if (!setUpMtlsDimensions(mtls, mtls->inputDim, sc)) { 581 return false; 582 } 583 584 mtls->rs = mCtx; 585 // Currently not threaded. 586 mtls->isThreadable = false; 587 mtls->mSliceNum = -1; 588 589 // Set up input and output. 590 mtls->inBuf = static_cast<uint8_t *>(ain->getPointerUnchecked(0, 0)); 591 mtls->outBuf = static_cast<uint8_t *>(aout->getPointerUnchecked(0, 0)); 592 593 rsAssert(mtls->inBuf && mtls->outBuf); 594 595 return true; 596} 597 598// Preliminary work to prepare a general reduce-style kernel for launch. 599bool RsdCpuScriptImpl::reduceNewMtlsSetup(const Allocation ** ains, 600 uint32_t inLen, 601 const Allocation * aout, 602 const RsScriptCall *sc, 603 MTLaunchStructReduceNew *mtls) { 604 rsAssert(ains && (inLen >= 1) && aout); 605 memset(mtls, 0, sizeof(MTLaunchStructReduceNew)); 606 mtls->dimPtr = &mtls->redp.dim; 607 608 for (int index = inLen; --index >= 0;) { 609 if (allocationLODIsNull(ains[index])) { 610 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 611 "reduce called with null in allocations"); 612 return false; 613 } 614 } 615 616 if (allocationLODIsNull(aout)) { 617 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 618 "reduce called with null out allocation"); 619 return false; 620 } 621 622 const Allocation *ain0 = ains[0]; 623 const Type *inType = ain0->getType(); 624 625 mtls->redp.dim.x = inType->getDimX(); 626 mtls->redp.dim.y = inType->getDimY(); 627 mtls->redp.dim.z = inType->getDimZ(); 628 629 for (int Index = inLen; --Index >= 1;) { 630 if (!ain0->hasSameDims(ains[Index])) { 631 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 632 "Failed to launch reduction kernel;" 633 "dimensions of input allocations do not match."); 634 return false; 635 } 636 } 637 638 if (!setUpMtlsDimensions(mtls, mtls->redp.dim, sc)) { 639 return false; 640 } 641 642 // The X & Y walkers always want 0-1 min even if dim is not present 643 mtls->end.x = rsMax((uint32_t)1, mtls->end.x); 644 mtls->end.y = rsMax((uint32_t)1, mtls->end.y); 645 646 mtls->rs = mCtx; 647 648 mtls->mSliceNum = 0; 649 mtls->mSliceSize = 1; 650 mtls->isThreadable = mIsThreadable; 651 652 // Set up output, 653 mtls->redp.outLen = 1; 654 mtls->redp.outPtr[0] = (uint8_t *)aout->mHal.drvState.lod[0].mallocPtr; 655 mtls->redp.outStride[0] = aout->getType()->getElementSizeBytes(); 656 657 // Set up input. 658 memcpy(mtls->ains, ains, inLen * sizeof(ains[0])); 659 mtls->redp.inLen = inLen; 660 for (int index = inLen; --index >= 0;) { 661 mtls->redp.inPtr[index] = (const uint8_t*)ains[index]->mHal.drvState.lod[0].mallocPtr; 662 mtls->redp.inStride[index] = ains[index]->getType()->getElementSizeBytes(); 663 } 664 665 // All validation passed, ok to launch threads 666 return true; 667} 668 669// Preliminary work to prepare a forEach-style kernel for launch. 670bool RsdCpuScriptImpl::forEachMtlsSetup(const Allocation ** ains, 671 uint32_t inLen, 672 Allocation * aout, 673 const void * usr, uint32_t usrLen, 674 const RsScriptCall *sc, 675 MTLaunchStructForEach *mtls) { 676 if (ains == nullptr && inLen != 0) { 677 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 678 "rsForEach called with none-zero inLen with null in allocations"); 679 return false; 680 } 681 682 memset(mtls, 0, sizeof(MTLaunchStructForEach)); 683 mtls->dimPtr = &mtls->fep.dim; 684 685 for (int index = inLen; --index >= 0;) { 686 if (allocationLODIsNull(ains[index])) { 687 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 688 "rsForEach called with null in allocations"); 689 return false; 690 } 691 } 692 693 if (allocationLODIsNull(aout)) { 694 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 695 "rsForEach called with null out allocations"); 696 return false; 697 } 698 699 if (inLen > 0) { 700 const Allocation *ain0 = ains[0]; 701 const Type *inType = ain0->getType(); 702 703 mtls->fep.dim.x = inType->getDimX(); 704 mtls->fep.dim.y = inType->getDimY(); 705 mtls->fep.dim.z = inType->getDimZ(); 706 707 for (int Index = inLen; --Index >= 1;) { 708 if (!ain0->hasSameDims(ains[Index])) { 709 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 710 "Failed to launch kernel; dimensions of input" 711 "allocations do not match."); 712 return false; 713 } 714 } 715 } else if (aout != nullptr) { 716 const Type *outType = aout->getType(); 717 718 mtls->fep.dim.x = outType->getDimX(); 719 mtls->fep.dim.y = outType->getDimY(); 720 mtls->fep.dim.z = outType->getDimZ(); 721 722 } else if (sc != nullptr) { 723 mtls->fep.dim.x = sc->xEnd; 724 mtls->fep.dim.y = sc->yEnd; 725 mtls->fep.dim.z = 0; 726 } else { 727 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 728 "rsForEach called with null allocations"); 729 return false; 730 } 731 732 if (inLen > 0 && aout != nullptr) { 733 if (!ains[0]->hasSameDims(aout)) { 734 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 735 "Failed to launch kernel; dimensions of input and output allocations do not match."); 736 737 return false; 738 } 739 } 740 741 if (!setUpMtlsDimensions(mtls, mtls->fep.dim, sc)) { 742 return false; 743 } 744 745 // The X & Y walkers always want 0-1 min even if dim is not present 746 mtls->end.x = rsMax((uint32_t)1, mtls->end.x); 747 mtls->end.y = rsMax((uint32_t)1, mtls->end.y); 748 mtls->rs = mCtx; 749 if (ains) { 750 memcpy(mtls->ains, ains, inLen * sizeof(ains[0])); 751 } 752 mtls->aout[0] = aout; 753 mtls->fep.usr = usr; 754 mtls->fep.usrLen = usrLen; 755 mtls->mSliceSize = 1; 756 mtls->mSliceNum = 0; 757 758 mtls->isThreadable = mIsThreadable; 759 760 if (inLen > 0) { 761 mtls->fep.inLen = inLen; 762 for (int index = inLen; --index >= 0;) { 763 mtls->fep.inPtr[index] = (const uint8_t*)ains[index]->mHal.drvState.lod[0].mallocPtr; 764 mtls->fep.inStride[index] = ains[index]->getType()->getElementSizeBytes(); 765 } 766 } 767 768 if (aout != nullptr) { 769 mtls->fep.outPtr[0] = (uint8_t *)aout->mHal.drvState.lod[0].mallocPtr; 770 mtls->fep.outStride[0] = aout->getType()->getElementSizeBytes(); 771 } 772 773 // All validation passed, ok to launch threads 774 return true; 775} 776 777 778void RsdCpuScriptImpl::invokeForEach(uint32_t slot, 779 const Allocation ** ains, 780 uint32_t inLen, 781 Allocation * aout, 782 const void * usr, 783 uint32_t usrLen, 784 const RsScriptCall *sc) { 785 786 MTLaunchStructForEach mtls; 787 788 if (forEachMtlsSetup(ains, inLen, aout, usr, usrLen, sc, &mtls)) { 789 forEachKernelSetup(slot, &mtls); 790 791 RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this); 792 mCtx->launchForEach(ains, inLen, aout, sc, &mtls); 793 mCtx->setTLS(oldTLS); 794 } 795} 796 797void RsdCpuScriptImpl::invokeReduce(uint32_t slot, 798 const Allocation *ain, 799 Allocation *aout, 800 const RsScriptCall *sc) { 801 MTLaunchStructReduce mtls; 802 803 if (reduceMtlsSetup(ain, aout, sc, &mtls)) { 804 reduceKernelSetup(slot, &mtls); 805 RsdCpuScriptImpl *oldTLS = mCtx->setTLS(this); 806 mCtx->launchReduce(ain, aout, &mtls); 807 mCtx->setTLS(oldTLS); 808 } 809} 810 811void RsdCpuScriptImpl::invokeReduceNew(uint32_t slot, 812 const Allocation ** ains, uint32_t inLen, 813 Allocation *aout, 814 const RsScriptCall *sc) { 815 MTLaunchStructReduceNew mtls; 816 817 if (reduceNewMtlsSetup(ains, inLen, aout, sc, &mtls)) { 818 reduceNewKernelSetup(slot, &mtls); 819 RsdCpuScriptImpl *oldTLS = mCtx->setTLS(this); 820 mCtx->launchReduceNew(ains, inLen, aout, &mtls); 821 mCtx->setTLS(oldTLS); 822 } 823} 824 825void RsdCpuScriptImpl::forEachKernelSetup(uint32_t slot, MTLaunchStructForEach *mtls) { 826 mtls->script = this; 827 mtls->fep.slot = slot; 828 mtls->kernel = mScriptExec->getForEachFunction(slot); 829 rsAssert(mtls->kernel != nullptr); 830} 831 832void RsdCpuScriptImpl::reduceKernelSetup(uint32_t slot, MTLaunchStructReduce *mtls) { 833 mtls->script = this; 834 mtls->kernel = mScriptExec->getReduceFunction(slot); 835 rsAssert(mtls->kernel != nullptr); 836} 837 838void RsdCpuScriptImpl::reduceNewKernelSetup(uint32_t slot, MTLaunchStructReduceNew *mtls) { 839 mtls->script = this; 840 mtls->redp.slot = slot; 841 842 const ReduceNewDescription *desc = mScriptExec->getReduceNewDescription(slot); 843 mtls->accumFunc = desc->accumFunc; 844 mtls->initFunc = desc->initFunc; // might legally be nullptr 845 mtls->combFunc = desc->combFunc; // might legally be nullptr 846 mtls->outFunc = desc->outFunc; // might legally be nullptr 847 mtls->accumSize = desc->accumSize; 848 849 rsAssert(mtls->accumFunc != nullptr); 850} 851 852int RsdCpuScriptImpl::invokeRoot() { 853 RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this); 854 int ret = mRoot(); 855 mCtx->setTLS(oldTLS); 856 return ret; 857} 858 859void RsdCpuScriptImpl::invokeInit() { 860 if (mInit) { 861 mInit(); 862 } 863} 864 865void RsdCpuScriptImpl::invokeFreeChildren() { 866 if (mFreeChildren) { 867 mFreeChildren(); 868 } 869} 870 871void RsdCpuScriptImpl::invokeFunction(uint32_t slot, const void *params, 872 size_t paramLength) { 873 //ALOGE("invoke %i %p %zu", slot, params, paramLength); 874 void * ap = nullptr; 875 876#if defined(__x86_64__) 877 // The invoked function could have input parameter of vector type for example float4 which 878 // requires void* params to be 16 bytes aligned when using SSE instructions for x86_64 platform. 879 // So try to align void* params before passing them into RS exported function. 880 881 if ((uint8_t)(uint64_t)params & 0x0F) { 882 if ((ap = (void*)memalign(16, paramLength)) != nullptr) { 883 memcpy(ap, params, paramLength); 884 } else { 885 ALOGE("x86_64: invokeFunction memalign error, still use params which" 886 " is not 16 bytes aligned."); 887 } 888 } 889#endif 890 891 RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this); 892 reinterpret_cast<void (*)(const void *, uint32_t)>( 893 mScriptExec->getInvokeFunction(slot))(ap? (const void *) ap: params, paramLength); 894 895#if defined(__x86_64__) 896 free(ap); 897#endif 898 899 mCtx->setTLS(oldTLS); 900} 901 902void RsdCpuScriptImpl::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) { 903 //rsAssert(!script->mFieldIsObject[slot]); 904 //ALOGE("setGlobalVar %i %p %zu", slot, data, dataLength); 905 906 //if (mIntrinsicID) { 907 //mIntrinsicFuncs.setVar(dc, script, drv->mIntrinsicData, slot, data, dataLength); 908 //return; 909 //} 910 911 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 912 if (!destPtr) { 913 //ALOGV("Calling setVar on slot = %i which is null", slot); 914 return; 915 } 916 917 memcpy(destPtr, data, dataLength); 918} 919 920void RsdCpuScriptImpl::getGlobalVar(uint32_t slot, void *data, size_t dataLength) { 921 //rsAssert(!script->mFieldIsObject[slot]); 922 //ALOGE("getGlobalVar %i %p %zu", slot, data, dataLength); 923 924 int32_t *srcPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 925 if (!srcPtr) { 926 //ALOGV("Calling setVar on slot = %i which is null", slot); 927 return; 928 } 929 memcpy(data, srcPtr, dataLength); 930} 931 932 933void RsdCpuScriptImpl::setGlobalVarWithElemDims(uint32_t slot, const void *data, size_t dataLength, 934 const Element *elem, 935 const uint32_t *dims, size_t dimLength) { 936 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 937 if (!destPtr) { 938 //ALOGV("Calling setVar on slot = %i which is null", slot); 939 return; 940 } 941 942 // We want to look at dimension in terms of integer components, 943 // but dimLength is given in terms of bytes. 944 dimLength /= sizeof(int); 945 946 // Only a single dimension is currently supported. 947 rsAssert(dimLength == 1); 948 if (dimLength == 1) { 949 // First do the increment loop. 950 size_t stride = elem->getSizeBytes(); 951 const char *cVal = reinterpret_cast<const char *>(data); 952 for (uint32_t i = 0; i < dims[0]; i++) { 953 elem->incRefs(cVal); 954 cVal += stride; 955 } 956 957 // Decrement loop comes after (to prevent race conditions). 958 char *oldVal = reinterpret_cast<char *>(destPtr); 959 for (uint32_t i = 0; i < dims[0]; i++) { 960 elem->decRefs(oldVal); 961 oldVal += stride; 962 } 963 } 964 965 memcpy(destPtr, data, dataLength); 966} 967 968void RsdCpuScriptImpl::setGlobalBind(uint32_t slot, Allocation *data) { 969 970 //rsAssert(!script->mFieldIsObject[slot]); 971 //ALOGE("setGlobalBind %i %p", slot, data); 972 973 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 974 if (!destPtr) { 975 //ALOGV("Calling setVar on slot = %i which is null", slot); 976 return; 977 } 978 979 void *ptr = nullptr; 980 mBoundAllocs[slot] = data; 981 if (data) { 982 ptr = data->mHal.drvState.lod[0].mallocPtr; 983 } 984 memcpy(destPtr, &ptr, sizeof(void *)); 985} 986 987void RsdCpuScriptImpl::setGlobalObj(uint32_t slot, ObjectBase *data) { 988 989 //rsAssert(script->mFieldIsObject[slot]); 990 //ALOGE("setGlobalObj %i %p", slot, data); 991 992 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 993 if (!destPtr) { 994 //ALOGV("Calling setVar on slot = %i which is null", slot); 995 return; 996 } 997 998 rsrSetObject(mCtx->getContext(), (rs_object_base *)destPtr, data); 999} 1000 1001const char* RsdCpuScriptImpl::getFieldName(uint32_t slot) const { 1002 return mScriptExec->getFieldName(slot); 1003} 1004 1005RsdCpuScriptImpl::~RsdCpuScriptImpl() { 1006 delete mScriptExec; 1007 delete[] mBoundAllocs; 1008 if (mScriptSO) { 1009 dlclose(mScriptSO); 1010 } 1011} 1012 1013Allocation * RsdCpuScriptImpl::getAllocationForPointer(const void *ptr) const { 1014 if (!ptr) { 1015 return nullptr; 1016 } 1017 1018 for (uint32_t ct=0; ct < mScript->mHal.info.exportedVariableCount; ct++) { 1019 Allocation *a = mBoundAllocs[ct]; 1020 if (!a) continue; 1021 if (a->mHal.drvState.lod[0].mallocPtr == ptr) { 1022 return a; 1023 } 1024 } 1025 ALOGE("rsGetAllocation, failed to find %p", ptr); 1026 return nullptr; 1027} 1028 1029int RsdCpuScriptImpl::getGlobalEntries() const { 1030 return mScriptExec->getGlobalEntries(); 1031} 1032 1033const char * RsdCpuScriptImpl::getGlobalName(int i) const { 1034 return mScriptExec->getGlobalName(i); 1035} 1036 1037const void * RsdCpuScriptImpl::getGlobalAddress(int i) const { 1038 return mScriptExec->getGlobalAddress(i); 1039} 1040 1041size_t RsdCpuScriptImpl::getGlobalSize(int i) const { 1042 return mScriptExec->getGlobalSize(i); 1043} 1044 1045uint32_t RsdCpuScriptImpl::getGlobalProperties(int i) const { 1046 return mScriptExec->getGlobalProperties(i); 1047} 1048 1049void RsdCpuScriptImpl::preLaunch(uint32_t slot, const Allocation ** ains, 1050 uint32_t inLen, Allocation * aout, 1051 const void * usr, uint32_t usrLen, 1052 const RsScriptCall *sc) {} 1053 1054void RsdCpuScriptImpl::postLaunch(uint32_t slot, const Allocation ** ains, 1055 uint32_t inLen, Allocation * aout, 1056 const void * usr, uint32_t usrLen, 1057 const RsScriptCall *sc) {} 1058 1059 1060} 1061} 1062