rsCpuScriptGroup2.cpp revision 1c20667f7a174a7c0a1599d34a40c524fe24c615
1#include "rsCpuScriptGroup2.h" 2 3#include <dlfcn.h> 4#include <stdio.h> 5#include <stdlib.h> 6#include <unistd.h> 7 8#include <set> 9#include <sstream> 10#include <string> 11#include <vector> 12 13#ifndef RS_COMPATIBILITY_LIB 14#include "bcc/Config/Config.h" 15#endif 16 17#include "cpu_ref/rsCpuCore.h" 18#include "rsClosure.h" 19#include "rsContext.h" 20#include "rsCpuCore.h" 21#include "rsCpuExecutable.h" 22#include "rsCpuScript.h" 23#include "rsScript.h" 24#include "rsScriptGroup2.h" 25#include "rsScriptIntrinsic.h" 26 27using std::string; 28using std::vector; 29 30namespace android { 31namespace renderscript { 32 33namespace { 34 35const size_t DefaultKernelArgCount = 2; 36 37void groupRoot(const RsExpandKernelDriverInfo *kinfo, uint32_t xstart, 38 uint32_t xend, uint32_t outstep) { 39 const List<CPUClosure*>& closures = *(List<CPUClosure*>*)kinfo->usr; 40 RsExpandKernelDriverInfo *mutable_kinfo = const_cast<RsExpandKernelDriverInfo *>(kinfo); 41 42 const size_t oldInLen = mutable_kinfo->inLen; 43 44 decltype(mutable_kinfo->inStride) oldInStride; 45 memcpy(&oldInStride, &mutable_kinfo->inStride, sizeof(oldInStride)); 46 47 for (CPUClosure* cpuClosure : closures) { 48 const Closure* closure = cpuClosure->mClosure; 49 50 // There had better be enough space in mutable_kinfo 51 rsAssert(closure->mNumArg <= RS_KERNEL_INPUT_LIMIT); 52 53 for (size_t i = 0; i < closure->mNumArg; i++) { 54 const void* arg = closure->mArgs[i]; 55 const Allocation* a = (const Allocation*)arg; 56 const uint32_t eStride = a->mHal.state.elementSizeBytes; 57 const uint8_t* ptr = (uint8_t*)(a->mHal.drvState.lod[0].mallocPtr) + 58 eStride * xstart; 59 if (kinfo->dim.y > 1) { 60 ptr += a->mHal.drvState.lod[0].stride * kinfo->current.y; 61 } 62 mutable_kinfo->inPtr[i] = ptr; 63 mutable_kinfo->inStride[i] = eStride; 64 } 65 mutable_kinfo->inLen = closure->mNumArg; 66 67 const Allocation* out = closure->mReturnValue; 68 const uint32_t ostep = out->mHal.state.elementSizeBytes; 69 const uint8_t* ptr = (uint8_t *)(out->mHal.drvState.lod[0].mallocPtr) + 70 ostep * xstart; 71 if (kinfo->dim.y > 1) { 72 ptr += out->mHal.drvState.lod[0].stride * kinfo->current.y; 73 } 74 75 rsAssert(kinfo->outLen <= 1); 76 mutable_kinfo->outPtr[0] = const_cast<uint8_t*>(ptr); 77 78 cpuClosure->mFunc(kinfo, xstart, xend, ostep); 79 } 80 81 mutable_kinfo->inLen = oldInLen; 82 memcpy(&mutable_kinfo->inStride, &oldInStride, sizeof(oldInStride)); 83} 84 85} // namespace 86 87Batch::Batch(CpuScriptGroup2Impl* group, const char* name) : 88 mGroup(group), mFunc(nullptr) { 89 mName = strndup(name, strlen(name)); 90} 91 92Batch::~Batch() { 93 for (CPUClosure* c : mClosures) { 94 delete c; 95 } 96 free(mName); 97} 98 99bool Batch::conflict(CPUClosure* cpuClosure) const { 100 if (mClosures.empty()) { 101 return false; 102 } 103 104 const Closure* closure = cpuClosure->mClosure; 105 106 if (!closure->mIsKernel || !mClosures.front()->mClosure->mIsKernel) { 107 // An invoke should be in a batch by itself, so it conflicts with any other 108 // closure. 109 return true; 110 } 111 112 const auto& globalDeps = closure->mGlobalDeps; 113 const auto& argDeps = closure->mArgDeps; 114 115 for (CPUClosure* c : mClosures) { 116 const Closure* batched = c->mClosure; 117 if (globalDeps.find(batched) != globalDeps.end()) { 118 return true; 119 } 120 const auto& it = argDeps.find(batched); 121 if (it != argDeps.end()) { 122 const auto& args = (*it).second; 123 for (const auto &p1 : *args) { 124 if (p1.second.get() != nullptr) { 125 return true; 126 } 127 } 128 } 129 } 130 131 // The compiler fusion pass in bcc expects that kernels chained up through 132 // (1st) input and output. 133 134 const Closure* lastBatched = mClosures.back()->mClosure; 135 const auto& it = argDeps.find(lastBatched); 136 137 if (it == argDeps.end()) { 138 return true; 139 } 140 141 const auto& args = (*it).second; 142 for (const auto &p1 : *args) { 143 if (p1.first == 0 && p1.second.get() == nullptr) { 144 // The new closure depends on the last batched closure's return 145 // value (fieldId being nullptr) for its first argument (argument 0) 146 return false; 147 } 148 } 149 150 return true; 151} 152 153CpuScriptGroup2Impl::CpuScriptGroup2Impl(RsdCpuReferenceImpl *cpuRefImpl, 154 const ScriptGroupBase *sg) : 155 mCpuRefImpl(cpuRefImpl), mGroup((const ScriptGroup2*)(sg)), 156 mExecutable(nullptr), mScriptObj(nullptr) { 157 rsAssert(!mGroup->mClosures.empty()); 158 159 Batch* batch = new Batch(this, "Batch0"); 160 int i = 0; 161 for (Closure* closure: mGroup->mClosures) { 162 CPUClosure* cc; 163 const IDBase* funcID = closure->mFunctionID.get(); 164 RsdCpuScriptImpl* si = 165 (RsdCpuScriptImpl *)mCpuRefImpl->lookupScript(funcID->mScript); 166 if (closure->mIsKernel) { 167 MTLaunchStruct mtls; 168 si->forEachKernelSetup(funcID->mSlot, &mtls); 169 cc = new CPUClosure(closure, si, (ExpandFuncTy)mtls.kernel); 170 } else { 171 cc = new CPUClosure(closure, si); 172 } 173 174 if (batch->conflict(cc)) { 175 mBatches.push_back(batch); 176 std::stringstream ss; 177 ss << "Batch" << ++i; 178 batch = new Batch(this, ss.str().c_str()); 179 } 180 181 batch->mClosures.push_back(cc); 182 } 183 184 rsAssert(!batch->mClosures.empty()); 185 mBatches.push_back(batch); 186 187#ifndef RS_COMPATIBILITY_LIB 188 compile(mGroup->mCacheDir); 189 if (mScriptObj != nullptr && mExecutable != nullptr) { 190 for (Batch* batch : mBatches) { 191 batch->resolveFuncPtr(mScriptObj); 192 } 193 } 194#endif // RS_COMPATIBILITY_LIB 195} 196 197void Batch::resolveFuncPtr(void* sharedObj) { 198 std::string funcName(mName); 199 if (mClosures.front()->mClosure->mIsKernel) { 200 funcName.append(".expand"); 201 } 202 mFunc = dlsym(sharedObj, funcName.c_str()); 203 rsAssert (mFunc != nullptr); 204} 205 206CpuScriptGroup2Impl::~CpuScriptGroup2Impl() { 207 for (Batch* batch : mBatches) { 208 delete batch; 209 } 210 delete mExecutable; 211 // TODO: move this dlclose into ~ScriptExecutable(). 212 if (mScriptObj != nullptr) { 213 dlclose(mScriptObj); 214 } 215} 216 217namespace { 218 219#ifndef RS_COMPATIBILITY_LIB 220 221string getCoreLibPath(Context* context, string* coreLibRelaxedPath) { 222 *coreLibRelaxedPath = ""; 223 224 // If we're debugging, use the debug library. 225 if (context->getContextType() == RS_CONTEXT_TYPE_DEBUG) { 226 return SYSLIBPATH"/libclcore_debug.bc"; 227 } 228 229 // Check for a platform specific library 230 231#if defined(ARCH_ARM_HAVE_NEON) && !defined(DISABLE_CLCORE_NEON) 232 // NEON-capable ARMv7a devices can use an accelerated math library 233 // for all reduced precision scripts. 234 // ARMv8 does not use NEON, as ASIMD can be used with all precision 235 // levels. 236 *coreLibRelaxedPath = SYSLIBPATH"/libclcore_neon.bc"; 237#endif 238 239#if defined(__i386__) || defined(__x86_64__) 240 // x86 devices will use an optimized library. 241 return SYSLIBPATH"/libclcore_x86.bc"; 242#else 243 return SYSLIBPATH"/libclcore.bc"; 244#endif 245} 246 247void setupCompileArguments( 248 const vector<const char*>& inputs, const vector<string>& kernelBatches, 249 const vector<string>& invokeBatches, 250 const char* outputDir, const char* outputFileName, 251 const char* coreLibPath, const char* coreLibRelaxedPath, 252 const bool emitGlobalInfo, const bool emitGlobalInfoSkipConstant, 253 vector<const char*>* args) { 254 args->push_back(RsdCpuScriptImpl::BCC_EXE_PATH); 255 args->push_back("-fPIC"); 256 args->push_back("-embedRSInfo"); 257 if (emitGlobalInfo) { 258 args->push_back("-rs-global-info"); 259 if (emitGlobalInfoSkipConstant) { 260 args->push_back("-rs-global-info-skip-constant"); 261 } 262 } 263 args->push_back("-mtriple"); 264 args->push_back(DEFAULT_TARGET_TRIPLE_STRING); 265 args->push_back("-bclib"); 266 args->push_back(coreLibPath); 267 args->push_back("-bclib_relaxed"); 268 args->push_back(coreLibRelaxedPath); 269 for (const char* input : inputs) { 270 args->push_back(input); 271 } 272 for (const string& batch : kernelBatches) { 273 args->push_back("-merge"); 274 args->push_back(batch.c_str()); 275 } 276 for (const string& batch : invokeBatches) { 277 args->push_back("-invoke"); 278 args->push_back(batch.c_str()); 279 } 280 args->push_back("-output_path"); 281 args->push_back(outputDir); 282 args->push_back("-o"); 283 args->push_back(outputFileName); 284} 285 286void generateSourceSlot(RsdCpuReferenceImpl* ctxt, 287 const Closure& closure, 288 const std::vector<const char*>& inputs, 289 std::stringstream& ss) { 290 const IDBase* funcID = (const IDBase*)closure.mFunctionID.get(); 291 const Script* script = funcID->mScript; 292 293 rsAssert (!script->isIntrinsic()); 294 295 const RsdCpuScriptImpl *cpuScript = 296 (const RsdCpuScriptImpl *)ctxt->lookupScript(script); 297 const string& bitcodeFilename = cpuScript->getBitcodeFilePath(); 298 299 const int index = find(inputs.begin(), inputs.end(), bitcodeFilename) - 300 inputs.begin(); 301 302 ss << index << "," << funcID->mSlot << "."; 303} 304 305#endif // RS_COMPATIBILTY_LIB 306 307} // anonymous namespace 308 309void CpuScriptGroup2Impl::compile(const char* cacheDir) { 310#ifndef RS_COMPATIBILITY_LIB 311 if (mGroup->mClosures.size() < 2) { 312 return; 313 } 314 315 auto comparator = [](const char* str1, const char* str2) -> bool { 316 return strcmp(str1, str2) < 0; 317 }; 318 std::set<const char*, decltype(comparator)> inputSet(comparator); 319 320 for (Closure* closure : mGroup->mClosures) { 321 const Script* script = closure->mFunctionID.get()->mScript; 322 323 // If any script is an intrinsic, give up trying fusing the kernels. 324 if (script->isIntrinsic()) { 325 return; 326 } 327 328 const RsdCpuScriptImpl *cpuScript = 329 (const RsdCpuScriptImpl *)mCpuRefImpl->lookupScript(script); 330 331 const char* bitcodeFilename = cpuScript->getBitcodeFilePath(); 332 inputSet.insert(bitcodeFilename); 333 } 334 335 std::vector<const char*> inputs(inputSet.begin(), inputSet.end()); 336 337 std::vector<string> kernelBatches; 338 std::vector<string> invokeBatches; 339 340 int i = 0; 341 for (const auto& batch : mBatches) { 342 rsAssert(batch->size() > 0); 343 344 std::stringstream ss; 345 ss << batch->mName << ":"; 346 347 if (!batch->mClosures.front()->mClosure->mIsKernel) { 348 rsAssert(batch->size() == 1); 349 generateSourceSlot(mCpuRefImpl, *batch->mClosures.front()->mClosure, inputs, ss); 350 invokeBatches.push_back(ss.str()); 351 } else { 352 for (const auto& cpuClosure : batch->mClosures) { 353 generateSourceSlot(mCpuRefImpl, *cpuClosure->mClosure, inputs, ss); 354 } 355 kernelBatches.push_back(ss.str()); 356 } 357 } 358 359 rsAssert(cacheDir != nullptr); 360 string objFilePath(cacheDir); 361 objFilePath.append("/"); 362 objFilePath.append(mGroup->mName); 363 objFilePath.append(".o"); 364 365 const char* resName = mGroup->mName; 366 string coreLibRelaxedPath; 367 const string& coreLibPath = getCoreLibPath(getCpuRefImpl()->getContext(), 368 &coreLibRelaxedPath); 369 370 vector<const char*> arguments; 371 bool emitGlobalInfo = getCpuRefImpl()->getEmbedGlobalInfo(); 372 bool emitGlobalInfoSkipConstant = getCpuRefImpl()->getEmbedGlobalInfoSkipConstant(); 373 setupCompileArguments(inputs, kernelBatches, invokeBatches, cacheDir, 374 resName, coreLibPath.c_str(), coreLibRelaxedPath.c_str(), 375 emitGlobalInfo, emitGlobalInfoSkipConstant, 376 &arguments); 377 378 std::unique_ptr<const char> cmdLine(rsuJoinStrings(arguments.size() - 1, 379 arguments.data())); 380 381 inputs.push_back(coreLibPath.c_str()); 382 inputs.push_back(coreLibRelaxedPath.c_str()); 383 384 uint32_t checksum = constructBuildChecksum(nullptr, 0, cmdLine.get(), 385 inputs.data(), inputs.size()); 386 387 if (checksum == 0) { 388 return; 389 } 390 391 std::stringstream ss; 392 ss << std::hex << checksum; 393 const char* checksumStr = ss.str().c_str(); 394 395 //===--------------------------------------------------------------------===// 396 // Try to load a shared lib from code cache matching filename and checksum 397 //===--------------------------------------------------------------------===// 398 399 mScriptObj = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 400 if (mScriptObj != nullptr) { 401 mExecutable = ScriptExecutable::createFromSharedObject( 402 getCpuRefImpl()->getContext(), mScriptObj, checksum); 403 if (mExecutable != nullptr) { 404 return; 405 } else { 406 ALOGE("Failed to create an executable object from so file"); 407 } 408 dlclose(mScriptObj); 409 mScriptObj = nullptr; 410 } 411 412 //===--------------------------------------------------------------------===// 413 // Fuse the input kernels and generate native code in an object file 414 //===--------------------------------------------------------------------===// 415 416 arguments.push_back("-build-checksum"); 417 arguments.push_back(checksumStr); 418 arguments.push_back(nullptr); 419 420 bool compiled = rsuExecuteCommand(RsdCpuScriptImpl::BCC_EXE_PATH, 421 arguments.size()-1, 422 arguments.data()); 423 if (!compiled) { 424 return; 425 } 426 427 //===--------------------------------------------------------------------===// 428 // Create and load the shared lib 429 //===--------------------------------------------------------------------===// 430 431 if (!SharedLibraryUtils::createSharedLibrary( 432 getCpuRefImpl()->getContext()->getDriverName(), cacheDir, resName)) { 433 ALOGE("Failed to link object file '%s'", resName); 434 unlink(objFilePath.c_str()); 435 return; 436 } 437 438 unlink(objFilePath.c_str()); 439 440 mScriptObj = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 441 if (mScriptObj == nullptr) { 442 ALOGE("Unable to load '%s'", resName); 443 return; 444 } 445 446 mExecutable = ScriptExecutable::createFromSharedObject( 447 getCpuRefImpl()->getContext(), 448 mScriptObj); 449 450#endif // RS_COMPATIBILITY_LIB 451} 452 453void CpuScriptGroup2Impl::execute() { 454 for (auto batch : mBatches) { 455 batch->setGlobalsForBatch(); 456 batch->run(); 457 } 458} 459 460void Batch::setGlobalsForBatch() { 461 for (CPUClosure* cpuClosure : mClosures) { 462 const Closure* closure = cpuClosure->mClosure; 463 const IDBase* funcID = closure->mFunctionID.get(); 464 Script* s = funcID->mScript;; 465 for (const auto& p : closure->mGlobals) { 466 const void* value = p.second.first; 467 int size = p.second.second; 468 if (value == nullptr && size == 0) { 469 // This indicates the current closure depends on another closure for a 470 // global in their shared module (script). In this case we don't need to 471 // copy the value. For example, an invoke intializes a global variable 472 // which a kernel later reads. 473 continue; 474 } 475 rsAssert(p.first != nullptr); 476 Script* script = p.first->mScript; 477 RsdCpuReferenceImpl* ctxt = mGroup->getCpuRefImpl(); 478 const RsdCpuScriptImpl *cpuScript = 479 (const RsdCpuScriptImpl *)ctxt->lookupScript(script); 480 int slot = p.first->mSlot; 481 ScriptExecutable* exec = mGroup->getExecutable(); 482 if (exec != nullptr) { 483 const char* varName = cpuScript->getFieldName(slot); 484 void* addr = exec->getFieldAddress(varName); 485 if (size < 0) { 486 rsrSetObject(mGroup->getCpuRefImpl()->getContext(), 487 (rs_object_base*)addr, (ObjectBase*)value); 488 } else { 489 memcpy(addr, (const void*)&value, size); 490 } 491 } else { 492 // We use -1 size to indicate an ObjectBase rather than a primitive type 493 if (size < 0) { 494 s->setVarObj(slot, (ObjectBase*)value); 495 } else { 496 s->setVar(slot, (const void*)&value, size); 497 } 498 } 499 } 500 } 501} 502 503void Batch::run() { 504 if (!mClosures.front()->mClosure->mIsKernel) { 505 rsAssert(mClosures.size() == 1); 506 507 // This batch contains a single closure for an invoke function 508 CPUClosure* cc = mClosures.front(); 509 const Closure* c = cc->mClosure; 510 511 if (mFunc != nullptr) { 512 // TODO: Need align pointers for x86_64. 513 // See RsdCpuScriptImpl::invokeFunction in rsCpuScript.cpp 514 ((InvokeFuncTy)mFunc)(c->mParams, c->mParamLength); 515 } else { 516 const ScriptInvokeID* invokeID = (const ScriptInvokeID*)c->mFunctionID.get(); 517 rsAssert(invokeID != nullptr); 518 cc->mSi->invokeFunction(invokeID->mSlot, c->mParams, c->mParamLength); 519 } 520 521 return; 522 } 523 524 if (mFunc != nullptr) { 525 MTLaunchStruct mtls; 526 const CPUClosure* firstCpuClosure = mClosures.front(); 527 const CPUClosure* lastCpuClosure = mClosures.back(); 528 529 firstCpuClosure->mSi->forEachMtlsSetup( 530 (const Allocation**)firstCpuClosure->mClosure->mArgs, 531 firstCpuClosure->mClosure->mNumArg, 532 lastCpuClosure->mClosure->mReturnValue, 533 nullptr, 0, nullptr, &mtls); 534 535 mtls.script = nullptr; 536 mtls.fep.usr = nullptr; 537 mtls.kernel = (ForEachFunc_t)mFunc; 538 539 mGroup->getCpuRefImpl()->launchThreads( 540 (const Allocation**)firstCpuClosure->mClosure->mArgs, 541 firstCpuClosure->mClosure->mNumArg, 542 lastCpuClosure->mClosure->mReturnValue, 543 nullptr, &mtls); 544 545 return; 546 } 547 548 for (CPUClosure* cpuClosure : mClosures) { 549 const Closure* closure = cpuClosure->mClosure; 550 const ScriptKernelID* kernelID = 551 (const ScriptKernelID*)closure->mFunctionID.get(); 552 cpuClosure->mSi->preLaunch(kernelID->mSlot, 553 (const Allocation**)closure->mArgs, 554 closure->mNumArg, closure->mReturnValue, 555 nullptr, 0, nullptr); 556 } 557 558 const CPUClosure* cpuClosure = mClosures.front(); 559 const Closure* closure = cpuClosure->mClosure; 560 MTLaunchStruct mtls; 561 562 if (cpuClosure->mSi->forEachMtlsSetup((const Allocation**)closure->mArgs, 563 closure->mNumArg, 564 closure->mReturnValue, 565 nullptr, 0, nullptr, &mtls)) { 566 567 mtls.script = nullptr; 568 mtls.kernel = (void (*)())&groupRoot; 569 mtls.fep.usr = &mClosures; 570 571 mGroup->getCpuRefImpl()->launchThreads(nullptr, 0, nullptr, nullptr, &mtls); 572 } 573 574 for (CPUClosure* cpuClosure : mClosures) { 575 const Closure* closure = cpuClosure->mClosure; 576 const ScriptKernelID* kernelID = 577 (const ScriptKernelID*)closure->mFunctionID.get(); 578 cpuClosure->mSi->postLaunch(kernelID->mSlot, 579 (const Allocation**)closure->mArgs, 580 closure->mNumArg, closure->mReturnValue, 581 nullptr, 0, nullptr); 582 } 583} 584 585} // namespace renderscript 586} // namespace android 587