rsCpuScript.cpp revision bf2111d3b3de310932099514f06924e48fa1d7b2
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 20#ifdef RS_COMPATIBILITY_LIB 21 #include <stdio.h> 22 #include <sys/stat.h> 23 #include <unistd.h> 24#else 25 #include <bcc/BCCContext.h> 26 #include <bcc/Config/Config.h> 27 #include <bcc/Renderscript/RSCompilerDriver.h> 28 #include <bcc/Renderscript/RSInfo.h> 29 #include <bcinfo/MetadataExtractor.h> 30 #include <cutils/properties.h> 31 32 #include <sys/types.h> 33 #include <sys/wait.h> 34 #include <unistd.h> 35 36 #include <string> 37 #include <vector> 38#endif 39 40#include <set> 41#include <string> 42#include <dlfcn.h> 43#include <stdlib.h> 44#include <string.h> 45#include <fstream> 46#include <iostream> 47 48#ifdef __LP64__ 49#define SYSLIBPATH "/system/lib64" 50#else 51#define SYSLIBPATH "/system/lib" 52#endif 53 54namespace { 55 56// Create a len length string containing random characters from [A-Za-z0-9]. 57static std::string getRandomString(size_t len) { 58 char buf[len + 1]; 59 for (size_t i = 0; i < len; i++) { 60 uint32_t r = arc4random() & 0xffff; 61 r %= 62; 62 if (r < 26) { 63 // lowercase 64 buf[i] = 'a' + r; 65 } else if (r < 52) { 66 // uppercase 67 buf[i] = 'A' + (r - 26); 68 } else { 69 // Use a number 70 buf[i] = '0' + (r - 52); 71 } 72 } 73 buf[len] = '\0'; 74 return std::string(buf); 75} 76 77// Check if a path exists and attempt to create it if it doesn't. 78static bool ensureCacheDirExists(const char *path) { 79 if (access(path, R_OK | W_OK | X_OK) == 0) { 80 // Done if we can rwx the directory 81 return true; 82 } 83 if (mkdir(path, 0700) == 0) { 84 return true; 85 } 86 return false; 87} 88 89// Copy the file named \p srcFile to \p dstFile. 90// Return 0 on success and -1 if anything wasn't copied. 91static int copyFile(const char *dstFile, const char *srcFile) { 92 std::ifstream srcStream(srcFile); 93 if (!srcStream) { 94 ALOGE("Could not verify or read source file: %s", srcFile); 95 return -1; 96 } 97 std::ofstream dstStream(dstFile); 98 if (!dstStream) { 99 ALOGE("Could not verify or write destination file: %s", dstFile); 100 return -1; 101 } 102 dstStream << srcStream.rdbuf(); 103 if (!dstStream) { 104 ALOGE("Could not write destination file: %s", dstFile); 105 return -1; 106 } 107 108 srcStream.close(); 109 dstStream.close(); 110 111 return 0; 112} 113 114static std::string findSharedObjectName(const char *cacheDir, 115 const char *resName) { 116 117#ifndef RS_SERVER 118 std::string scriptSOName(cacheDir); 119#ifdef RS_COMPATIBILITY_LIB 120 size_t cutPos = scriptSOName.rfind("cache"); 121 if (cutPos != std::string::npos) { 122 scriptSOName.erase(cutPos); 123 } else { 124 ALOGE("Found peculiar cacheDir (missing \"cache\"): %s", cacheDir); 125 } 126 scriptSOName.append("/lib/librs."); 127#else 128 scriptSOName.append("/librs."); 129#endif 130 131#else 132 std::string scriptSOName("lib"); 133#endif 134 scriptSOName.append(resName); 135 scriptSOName.append(".so"); 136 137 return scriptSOName; 138} 139 140#ifndef RS_COMPATIBILITY_LIB 141 142static bool is_force_recompile() { 143#ifdef RS_SERVER 144 return false; 145#else 146 char buf[PROPERTY_VALUE_MAX]; 147 148 // Re-compile if floating point precision has been overridden. 149 property_get("debug.rs.precision", buf, ""); 150 if (buf[0] != '\0') { 151 return true; 152 } 153 154 // Re-compile if debug.rs.forcerecompile is set. 155 property_get("debug.rs.forcerecompile", buf, "0"); 156 if ((::strcmp(buf, "1") == 0) || (::strcmp(buf, "true") == 0)) { 157 return true; 158 } else { 159 return false; 160 } 161#endif // RS_SERVER 162} 163 164const static char *BCC_EXE_PATH = "/system/bin/bcc"; 165 166static void setCompileArguments(std::vector<const char*>* args, 167 const std::string& bcFileName, 168 const char* cacheDir, const char* resName, 169 const char* core_lib, bool useRSDebugContext, 170 const char* bccPluginName) { 171 rsAssert(cacheDir && resName && core_lib); 172 args->push_back(BCC_EXE_PATH); 173 args->push_back("-unroll-runtime"); 174 args->push_back("-scalarize-load-store"); 175 args->push_back("-o"); 176 args->push_back(resName); 177 args->push_back("-output_path"); 178 args->push_back(cacheDir); 179 args->push_back("-bclib"); 180 args->push_back(core_lib); 181 args->push_back("-mtriple"); 182 args->push_back(DEFAULT_TARGET_TRIPLE_STRING); 183 184 // Enable workaround for A53 codegen by default. 185#if defined(__aarch64__) && !defined(DISABLE_A53_WORKAROUND) 186 args->push_back("-aarch64-fix-cortex-a53-835769"); 187#endif 188 189 // Execute the bcc compiler. 190 if (useRSDebugContext) { 191 args->push_back("-rs-debug-ctx"); 192 } else { 193 // Only load additional libraries for compiles that don't use 194 // the debug context. 195 if (bccPluginName && strlen(bccPluginName) > 0) { 196 args->push_back("-load"); 197 args->push_back(bccPluginName); 198 } 199 } 200 201 args->push_back("-fPIC"); 202 args->push_back("-embedRSInfo"); 203 204 args->push_back(bcFileName.c_str()); 205 args->push_back(nullptr); 206} 207 208static bool compileBitcode(const std::string &bcFileName, 209 const char *bitcode, 210 size_t bitcodeSize, 211 const char **compileArguments, 212 const std::string &compileCommandLine) { 213 rsAssert(bitcode && bitcodeSize); 214 215 FILE *bcfile = fopen(bcFileName.c_str(), "w"); 216 if (!bcfile) { 217 ALOGE("Could not write to %s", bcFileName.c_str()); 218 return false; 219 } 220 size_t nwritten = fwrite(bitcode, 1, bitcodeSize, bcfile); 221 fclose(bcfile); 222 if (nwritten != bitcodeSize) { 223 ALOGE("Could not write %zu bytes to %s", bitcodeSize, 224 bcFileName.c_str()); 225 return false; 226 } 227 228 pid_t pid = fork(); 229 230 switch (pid) { 231 case -1: { // Error occurred (we attempt no recovery) 232 ALOGE("Couldn't fork for bcc compiler execution"); 233 return false; 234 } 235 case 0: { // Child process 236 ALOGV("Invoking BCC with: %s", compileCommandLine.c_str()); 237 execv(BCC_EXE_PATH, (char* const*)compileArguments); 238 239 ALOGE("execv() failed: %s", strerror(errno)); 240 abort(); 241 return false; 242 } 243 default: { // Parent process (actual driver) 244 // Wait on child process to finish compiling the source. 245 int status = 0; 246 pid_t w = waitpid(pid, &status, 0); 247 if (w == -1) { 248 ALOGE("Could not wait for bcc compiler"); 249 return false; 250 } 251 252 if (WIFEXITED(status) && WEXITSTATUS(status) == 0) { 253 return true; 254 } 255 256 ALOGE("bcc compiler terminated unexpectedly"); 257 return false; 258 } 259 } 260} 261 262#endif // !defined(RS_COMPATIBILITY_LIB) 263} // namespace 264 265namespace android { 266namespace renderscript { 267 268const char* SharedLibraryUtils::LD_EXE_PATH = "/system/bin/ld.mc"; 269const char* SharedLibraryUtils::RS_CACHE_DIR = "com.android.renderscript.cache"; 270 271#ifndef RS_COMPATIBILITY_LIB 272 273bool SharedLibraryUtils::createSharedLibrary(const char *cacheDir, const char *resName) { 274 std::string sharedLibName = findSharedObjectName(cacheDir, resName); 275 std::string objFileName = cacheDir; 276 objFileName.append("/"); 277 objFileName.append(resName); 278 objFileName.append(".o"); 279 280 const char *compiler_rt = SYSLIBPATH"/libcompiler_rt.so"; 281 std::vector<const char *> args = { 282 LD_EXE_PATH, 283 "-shared", 284 "-nostdlib", 285 compiler_rt, 286 "-mtriple", DEFAULT_TARGET_TRIPLE_STRING, 287 "-L", SYSLIBPATH, 288 "-lRSDriver", "-lm", "-lc", 289 objFileName.c_str(), 290 "-o", sharedLibName.c_str(), 291 nullptr 292 }; 293 294 std::string cmdLineStr = bcc::getCommandLine(args.size()-1, args.data()); 295 296 pid_t pid = fork(); 297 298 switch (pid) { 299 case -1: { // Error occurred (we attempt no recovery) 300 ALOGE("Couldn't fork for linker (%s) execution", LD_EXE_PATH); 301 return false; 302 } 303 case 0: { // Child process 304 ALOGV("Invoking ld.mc with args '%s'", cmdLineStr.c_str()); 305 execv(LD_EXE_PATH, (char* const*) args.data()); 306 307 ALOGE("execv() failed: %s", strerror(errno)); 308 abort(); 309 return false; 310 } 311 default: { // Parent process (actual driver) 312 // Wait on child process to finish compiling the source. 313 int status = 0; 314 pid_t w = waitpid(pid, &status, 0); 315 if (w == -1) { 316 ALOGE("Could not wait for linker (%s)", LD_EXE_PATH); 317 return false; 318 } 319 320 if (WIFEXITED(status) && WEXITSTATUS(status) == 0) { 321 return true; 322 } 323 324 ALOGE("Linker (%s) terminated unexpectedly", LD_EXE_PATH); 325 return false; 326 } 327 } 328} 329 330#endif // RS_COMPATIBILITY_LIB 331 332void* SharedLibraryUtils::loadSharedLibrary(const char *cacheDir, const char *resName) { 333 void *loaded = nullptr; 334 335 std::string scriptSOName = findSharedObjectName(cacheDir, resName); 336 337 // We should check if we can load the library from the standard app 338 // location for shared libraries first. 339 loaded = loadSOHelper(scriptSOName.c_str(), cacheDir, resName); 340 341 if (loaded == nullptr) { 342 ALOGE("Unable to open shared library (%s): %s", 343 scriptSOName.c_str(), dlerror()); 344 345#ifdef RS_COMPATIBILITY_LIB 346 // One final attempt to find the library in "/system/lib". 347 // We do this to allow bundled applications to use the compatibility 348 // library fallback path. Those applications don't have a private 349 // library path, so they need to install to the system directly. 350 // Note that this is really just a testing path. 351 std::string scriptSONameSystem("/system/lib/librs."); 352 scriptSONameSystem.append(resName); 353 scriptSONameSystem.append(".so"); 354 loaded = loadSOHelper(scriptSONameSystem.c_str(), cacheDir, 355 resName); 356 if (loaded == nullptr) { 357 ALOGE("Unable to open system shared library (%s): %s", 358 scriptSONameSystem.c_str(), dlerror()); 359 } 360#endif 361 } 362 363 return loaded; 364} 365 366void* SharedLibraryUtils::loadSOHelper(const char *origName, const char *cacheDir, 367 const char *resName) { 368 // Keep track of which .so libraries have been loaded. Once a library is 369 // in the set (per-process granularity), we must instead make a copy of 370 // the original shared object (randomly named .so file) and load that one 371 // instead. If we don't do this, we end up aliasing global data between 372 // the various Script instances (which are supposed to be completely 373 // independent). 374 static std::set<std::string> LoadedLibraries; 375 376 void *loaded = nullptr; 377 378 // Skip everything if we don't even have the original library available. 379 if (access(origName, F_OK) != 0) { 380 return nullptr; 381 } 382 383 // Common path is that we have not loaded this Script/library before. 384 if (LoadedLibraries.find(origName) == LoadedLibraries.end()) { 385 loaded = dlopen(origName, RTLD_NOW | RTLD_LOCAL); 386 if (loaded) { 387 LoadedLibraries.insert(origName); 388 } 389 return loaded; 390 } 391 392 std::string newName(cacheDir); 393 394 // Append RS_CACHE_DIR only if it is not found in cacheDir 395 // In driver mode, RS_CACHE_DIR is already appended to cacheDir. 396 if (newName.find(RS_CACHE_DIR) == std::string::npos) { 397 newName.append("/"); 398 newName.append(RS_CACHE_DIR); 399 newName.append("/"); 400 } 401 402 if (!ensureCacheDirExists(newName.c_str())) { 403 ALOGE("Could not verify or create cache dir: %s", cacheDir); 404 return nullptr; 405 } 406 407 // Construct an appropriately randomized filename for the copy. 408 newName.append("librs."); 409 newName.append(resName); 410 newName.append("#"); 411 newName.append(getRandomString(6)); // 62^6 potential filename variants. 412 newName.append(".so"); 413 414 int r = copyFile(newName.c_str(), origName); 415 if (r != 0) { 416 ALOGE("Could not create copy %s -> %s", origName, newName.c_str()); 417 return nullptr; 418 } 419 loaded = dlopen(newName.c_str(), RTLD_NOW | RTLD_LOCAL); 420 r = unlink(newName.c_str()); 421 if (r != 0) { 422 ALOGE("Could not unlink copy %s", newName.c_str()); 423 } 424 if (loaded) { 425 LoadedLibraries.insert(newName.c_str()); 426 } 427 428 return loaded; 429} 430 431#define MAXLINE 500 432#define MAKE_STR_HELPER(S) #S 433#define MAKE_STR(S) MAKE_STR_HELPER(S) 434#define EXPORT_VAR_STR "exportVarCount: " 435#define EXPORT_FUNC_STR "exportFuncCount: " 436#define EXPORT_FOREACH_STR "exportForEachCount: " 437#define OBJECT_SLOT_STR "objectSlotCount: " 438#define PRAGMA_STR "pragmaCount: " 439 440// Copy up to a newline or size chars from str -> s, updating str 441// Returns s when successful and nullptr when '\0' is finally reached. 442static char* strgets(char *s, int size, const char **ppstr) { 443 if (!ppstr || !*ppstr || **ppstr == '\0' || size < 1) { 444 return nullptr; 445 } 446 447 int i; 448 for (i = 0; i < (size - 1); i++) { 449 s[i] = **ppstr; 450 (*ppstr)++; 451 if (s[i] == '\0') { 452 return s; 453 } else if (s[i] == '\n') { 454 s[i+1] = '\0'; 455 return s; 456 } 457 } 458 459 // size has been exceeded. 460 s[i] = '\0'; 461 462 return s; 463} 464 465RsdCpuScriptImpl::RsdCpuScriptImpl(RsdCpuReferenceImpl *ctx, const Script *s) { 466 mCtx = ctx; 467 mScript = s; 468 469 mScriptSO = nullptr; 470 471#ifndef RS_COMPATIBILITY_LIB 472 mCompilerDriver = nullptr; 473#endif 474 475 476 mRoot = nullptr; 477 mRootExpand = nullptr; 478 mInit = nullptr; 479 mFreeChildren = nullptr; 480 mScriptExec = nullptr; 481 482 mBoundAllocs = nullptr; 483 mIntrinsicData = nullptr; 484 mIsThreadable = true; 485} 486 487bool RsdCpuScriptImpl::storeRSInfoFromSO() { 488 mRoot = (RootFunc_t) dlsym(mScriptSO, "root"); 489 if (mRoot) { 490 //ALOGE("Found root(): %p", mRoot); 491 } 492 mRootExpand = (RootFunc_t) dlsym(mScriptSO, "root.expand"); 493 if (mRootExpand) { 494 //ALOGE("Found root.expand(): %p", mRootExpand); 495 } 496 mInit = (InvokeFunc_t) dlsym(mScriptSO, "init"); 497 if (mInit) { 498 //ALOGE("Found init(): %p", mInit); 499 } 500 mFreeChildren = (InvokeFunc_t) dlsym(mScriptSO, ".rs.dtor"); 501 if (mFreeChildren) { 502 //ALOGE("Found .rs.dtor(): %p", mFreeChildren); 503 } 504 505 mScriptExec = ScriptExecutable::createFromSharedObject( 506 mCtx->getContext(), mScriptSO); 507 508 if (mScriptExec == nullptr) { 509 return false; 510 } 511 512 size_t varCount = mScriptExec->getExportedVariableCount(); 513 if (varCount > 0) { 514 mBoundAllocs = new Allocation *[varCount]; 515 memset(mBoundAllocs, 0, varCount * sizeof(*mBoundAllocs)); 516 } 517 518 return true; 519} 520 521ScriptExecutable* ScriptExecutable::createFromSharedObject( 522 Context* RSContext, void* sharedObj) { 523 char line[MAXLINE]; 524 525 size_t varCount = 0; 526 size_t funcCount = 0; 527 size_t forEachCount = 0; 528 size_t objectSlotCount = 0; 529 size_t pragmaCount = 0; 530 531 const char *rsInfo = (const char *) dlsym(sharedObj, ".rs.info"); 532 533 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 534 return nullptr; 535 } 536 if (sscanf(line, EXPORT_VAR_STR "%zu", &varCount) != 1) { 537 ALOGE("Invalid export var count!: %s", line); 538 return nullptr; 539 } 540 541 std::vector<void*> fieldAddress; 542 543 for (size_t i = 0; i < varCount; ++i) { 544 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 545 return nullptr; 546 } 547 char *c = strrchr(line, '\n'); 548 if (c) { 549 *c = '\0'; 550 } 551 void* addr = dlsym(sharedObj, line); 552 if (addr == nullptr) { 553 ALOGE("Failed to find variable address for %s: %s", 554 line, dlerror()); 555 // Not a critical error if we don't find a global variable. 556 } 557 fieldAddress.push_back(addr); 558 } 559 560 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 561 return nullptr; 562 } 563 if (sscanf(line, EXPORT_FUNC_STR "%zu", &funcCount) != 1) { 564 ALOGE("Invalid export func count!: %s", line); 565 return nullptr; 566 } 567 568 std::vector<InvokeFunc_t> invokeFunctions(funcCount); 569 570 for (size_t i = 0; i < funcCount; ++i) { 571 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 572 return nullptr ; 573 } 574 char *c = strrchr(line, '\n'); 575 if (c) { 576 *c = '\0'; 577 } 578 579 invokeFunctions[i] = (InvokeFunc_t) dlsym(sharedObj, line); 580 if (invokeFunctions[i] == nullptr) { 581 ALOGE("Failed to get function address for %s(): %s", 582 line, dlerror()); 583 return nullptr; 584 } 585 } 586 587 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 588 return nullptr; 589 } 590 if (sscanf(line, EXPORT_FOREACH_STR "%zu", &forEachCount) != 1) { 591 ALOGE("Invalid export forEach count!: %s", line); 592 return nullptr; 593 } 594 595 std::vector<ForEachFunc_t> forEachFunctions(forEachCount); 596 std::vector<uint32_t> forEachSignatures(forEachCount); 597 598 for (size_t i = 0; i < forEachCount; ++i) { 599 unsigned int tmpSig = 0; 600 char tmpName[MAXLINE]; 601 602 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 603 return nullptr; 604 } 605 if (sscanf(line, "%u - %" MAKE_STR(MAXLINE) "s", 606 &tmpSig, tmpName) != 2) { 607 ALOGE("Invalid export forEach!: %s", line); 608 return nullptr; 609 } 610 611 // Lookup the expanded ForEach kernel. 612 strncat(tmpName, ".expand", MAXLINE-1-strlen(tmpName)); 613 forEachSignatures[i] = tmpSig; 614 forEachFunctions[i] = 615 (ForEachFunc_t) dlsym(sharedObj, tmpName); 616 if (i != 0 && forEachFunctions[i] == nullptr) { 617 // Ignore missing root.expand functions. 618 // root() is always specified at location 0. 619 ALOGE("Failed to find forEach function address for %s: %s", 620 tmpName, dlerror()); 621 return nullptr; 622 } 623 } 624 625 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 626 return nullptr; 627 } 628 if (sscanf(line, OBJECT_SLOT_STR "%zu", &objectSlotCount) != 1) { 629 ALOGE("Invalid object slot count!: %s", line); 630 return nullptr; 631 } 632 633 std::vector<bool> fieldIsObject(varCount, false); 634 635 rsAssert(varCount > 0); 636 for (size_t i = 0; i < objectSlotCount; ++i) { 637 uint32_t varNum = 0; 638 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 639 return nullptr; 640 } 641 if (sscanf(line, "%u", &varNum) != 1) { 642 ALOGE("Invalid object slot!: %s", line); 643 return nullptr; 644 } 645 646 if (varNum < varCount) { 647 fieldIsObject[varNum] = true; 648 } 649 } 650 651#ifdef RS_COMPATIBILITY_LIB 652 // Do not attempt to read pragmas in compat lib path 653 std::vector<const char *> pragmaKeys(pragmaCount); 654 std::vector<const char *> pragmaValues(pragmaCount); 655 656#else 657 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 658 return nullptr; 659 } 660 661 if (sscanf(line, PRAGMA_STR "%zu", &pragmaCount) != 1) { 662 ALOGE("Invalid pragma count!: %s", line); 663 return nullptr; 664 } 665 666 std::vector<const char *> pragmaKeys(pragmaCount); 667 std::vector<const char *> pragmaValues(pragmaCount); 668 669 for (size_t i = 0; i < pragmaCount; ++i) { 670 if (strgets(line, MAXLINE, &rsInfo) == nullptr) { 671 ALOGE("Unable to read pragma at index %zu!", i); 672 return nullptr; 673 } 674 675 char key[MAXLINE]; 676 char value[MAXLINE] = ""; // initialize in case value is empty 677 678 // pragmas can just have a key and no value. Only check to make sure 679 // that the key is not empty 680 if (sscanf(line, "%" MAKE_STR(MAXLINE) "s - %" MAKE_STR(MAXLINE) "s", 681 key, value) == 0 || 682 strlen(key) == 0) 683 { 684 ALOGE("Invalid pragma value!: %s", line); 685 686 // free previously allocated keys and values 687 for (size_t idx = 0; idx < i; ++idx) { 688 delete [] pragmaKeys[idx]; 689 delete [] pragmaValues[idx]; 690 } 691 return nullptr; 692 } 693 694 char *pKey = new char[strlen(key)+1]; 695 strcpy(pKey, key); 696 pragmaKeys[i] = pKey; 697 698 char *pValue = new char[strlen(value)+1]; 699 strcpy(pValue, value); 700 pragmaValues[i] = pValue; 701 //ALOGE("Pragma %zu: Key: '%s' Value: '%s'", i, pKey, pValue); 702 } 703#endif 704 705 return new ScriptExecutable( 706 RSContext, fieldAddress, fieldIsObject, invokeFunctions, 707 forEachFunctions, forEachSignatures, pragmaKeys, pragmaValues); 708} 709 710bool RsdCpuScriptImpl::init(char const *resName, char const *cacheDir, 711 uint8_t const *bitcode, size_t bitcodeSize, 712 uint32_t flags, char const *bccPluginName) { 713 //ALOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir, bitcode, bitcodeSize, flags, lookupFunc); 714 //ALOGE("rsdScriptInit %p %p", rsc, script); 715 716 mCtx->lockMutex(); 717#ifndef RS_COMPATIBILITY_LIB 718 bool useRSDebugContext = false; 719 720 mCompilerDriver = nullptr; 721 722 mCompilerDriver = new bcc::RSCompilerDriver(); 723 if (mCompilerDriver == nullptr) { 724 ALOGE("bcc: FAILS to create compiler driver (out of memory)"); 725 mCtx->unlockMutex(); 726 return false; 727 } 728 729 // Run any compiler setup functions we have been provided with. 730 RSSetupCompilerCallback setupCompilerCallback = 731 mCtx->getSetupCompilerCallback(); 732 if (setupCompilerCallback != nullptr) { 733 setupCompilerCallback(mCompilerDriver); 734 } 735 736 bcinfo::MetadataExtractor bitcodeMetadata((const char *) bitcode, bitcodeSize); 737 if (!bitcodeMetadata.extract()) { 738 ALOGE("Could not extract metadata from bitcode"); 739 mCtx->unlockMutex(); 740 return false; 741 } 742 743 const char* core_lib = findCoreLib(bitcodeMetadata, (const char*)bitcode, bitcodeSize); 744 745 if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) { 746 mCompilerDriver->setDebugContext(true); 747 useRSDebugContext = true; 748 } 749 750 std::string bcFileName(cacheDir); 751 bcFileName.append("/"); 752 bcFileName.append(resName); 753 bcFileName.append(".bc"); 754 755 std::vector<const char*> compileArguments; 756 setCompileArguments(&compileArguments, bcFileName, cacheDir, resName, core_lib, 757 useRSDebugContext, bccPluginName); 758 // The last argument of compileArguments ia a nullptr, so remove 1 from the size. 759 std::string compileCommandLine = 760 bcc::getCommandLine(compileArguments.size() - 1, compileArguments.data()); 761 762 if (!is_force_recompile() && !useRSDebugContext) { 763 mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 764 } 765 766 // If we can't, it's either not there or out of date. We compile the bit code and try loading 767 // again. 768 if (mScriptSO == nullptr) { 769 if (!compileBitcode(bcFileName, (const char*)bitcode, bitcodeSize, 770 compileArguments.data(), compileCommandLine)) 771 { 772 ALOGE("bcc: FAILS to compile '%s'", resName); 773 mCtx->unlockMutex(); 774 return false; 775 } 776 777 if (!SharedLibraryUtils::createSharedLibrary(cacheDir, resName)) { 778 ALOGE("Linker: Failed to link object file '%s'", resName); 779 mCtx->unlockMutex(); 780 return false; 781 } 782 783 mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 784 if (mScriptSO == nullptr) { 785 ALOGE("Unable to load '%s'", resName); 786 mCtx->unlockMutex(); 787 return false; 788 } 789 } 790 791 // Read RS symbol information from the .so. 792 if ( !mScriptSO) { 793 goto error; 794 } 795 796 if ( !storeRSInfoFromSO()) { 797 goto error; 798 } 799#else // RS_COMPATIBILITY_LIB is defined 800 801 mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName); 802 803 if (!mScriptSO) { 804 goto error; 805 } 806 807 if (!storeRSInfoFromSO()) { 808 goto error; 809 } 810#endif 811 mCtx->unlockMutex(); 812 return true; 813 814error: 815 816 mCtx->unlockMutex(); 817 if (mScriptSO) { 818 dlclose(mScriptSO); 819 } 820 return false; 821} 822 823#ifndef RS_COMPATIBILITY_LIB 824 825const char* RsdCpuScriptImpl::findCoreLib(const bcinfo::MetadataExtractor& ME, const char* bitcode, 826 size_t bitcodeSize) { 827 const char* defaultLib = SYSLIBPATH"/libclcore.bc"; 828 829 // If we're debugging, use the debug library. 830 if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) { 831 return SYSLIBPATH"/libclcore_debug.bc"; 832 } 833 834 // If a callback has been registered to specify a library, use that. 835 RSSelectRTCallback selectRTCallback = mCtx->getSelectRTCallback(); 836 if (selectRTCallback != nullptr) { 837 return selectRTCallback((const char*)bitcode, bitcodeSize); 838 } 839 840 // Check for a platform specific library 841#if defined(ARCH_ARM_HAVE_NEON) && !defined(DISABLE_CLCORE_NEON) 842 enum bcinfo::RSFloatPrecision prec = ME.getRSFloatPrecision(); 843 if (prec == bcinfo::RS_FP_Relaxed) { 844 // NEON-capable ARMv7a devices can use an accelerated math library 845 // for all reduced precision scripts. 846 // ARMv8 does not use NEON, as ASIMD can be used with all precision 847 // levels. 848 return SYSLIBPATH"/libclcore_neon.bc"; 849 } else { 850 return defaultLib; 851 } 852#elif defined(__i386__) || defined(__x86_64__) 853 // x86 devices will use an optimized library. 854 return SYSLIBPATH"/libclcore_x86.bc"; 855#else 856 return defaultLib; 857#endif 858} 859 860#endif 861 862void RsdCpuScriptImpl::populateScript(Script *script) { 863 // Copy info over to runtime 864 script->mHal.info.exportedFunctionCount = mScriptExec->getExportedFunctionCount(); 865 script->mHal.info.exportedVariableCount = mScriptExec->getExportedVariableCount(); 866 script->mHal.info.exportedPragmaCount = mScriptExec->getPragmaCount();; 867 script->mHal.info.exportedPragmaKeyList = 868 const_cast<const char**>(&mScriptExec->getPragmaKeys().front()); 869 script->mHal.info.exportedPragmaValueList = 870 const_cast<const char**>(&mScriptExec->getPragmaValues().front()); 871 872 // Bug, need to stash in metadata 873 if (mRootExpand) { 874 script->mHal.info.root = mRootExpand; 875 } else { 876 script->mHal.info.root = mRoot; 877 } 878} 879 880 881typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t); 882 883bool RsdCpuScriptImpl::forEachMtlsSetup(const Allocation ** ains, 884 uint32_t inLen, 885 Allocation * aout, 886 const void * usr, uint32_t usrLen, 887 const RsScriptCall *sc, 888 MTLaunchStruct *mtls) { 889 890 memset(mtls, 0, sizeof(MTLaunchStruct)); 891 892 for (int index = inLen; --index >= 0;) { 893 const Allocation* ain = ains[index]; 894 895 // possible for this to occur if IO_OUTPUT/IO_INPUT with no bound surface 896 if (ain != nullptr && 897 (const uint8_t *)ain->mHal.drvState.lod[0].mallocPtr == nullptr) { 898 899 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 900 "rsForEach called with null in allocations"); 901 return false; 902 } 903 } 904 905 if (aout && 906 (const uint8_t *)aout->mHal.drvState.lod[0].mallocPtr == nullptr) { 907 908 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 909 "rsForEach called with null out allocations"); 910 return false; 911 } 912 913 if (inLen > 0) { 914 const Allocation *ain0 = ains[0]; 915 const Type *inType = ain0->getType(); 916 917 mtls->fep.dim.x = inType->getDimX(); 918 mtls->fep.dim.y = inType->getDimY(); 919 mtls->fep.dim.z = inType->getDimZ(); 920 921 for (int Index = inLen; --Index >= 1;) { 922 if (!ain0->hasSameDims(ains[Index])) { 923 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 924 "Failed to launch kernel; dimensions of input and output allocations do not match."); 925 926 return false; 927 } 928 } 929 930 } else if (aout != nullptr) { 931 const Type *outType = aout->getType(); 932 933 mtls->fep.dim.x = outType->getDimX(); 934 mtls->fep.dim.y = outType->getDimY(); 935 mtls->fep.dim.z = outType->getDimZ(); 936 937 } else { 938 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 939 "rsForEach called with null allocations"); 940 return false; 941 } 942 943 if (inLen > 0 && aout != nullptr) { 944 if (!ains[0]->hasSameDims(aout)) { 945 mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, 946 "Failed to launch kernel; dimensions of input and output allocations do not match."); 947 948 return false; 949 } 950 } 951 952 if (!sc || (sc->xEnd == 0)) { 953 mtls->end.x = mtls->fep.dim.x; 954 } else { 955 mtls->start.x = rsMin(mtls->fep.dim.x, sc->xStart); 956 mtls->end.x = rsMin(mtls->fep.dim.x, sc->xEnd); 957 if (mtls->start.x >= mtls->end.x) return false; 958 } 959 960 if (!sc || (sc->yEnd == 0)) { 961 mtls->end.y = mtls->fep.dim.y; 962 } else { 963 mtls->start.y = rsMin(mtls->fep.dim.y, sc->yStart); 964 mtls->end.y = rsMin(mtls->fep.dim.y, sc->yEnd); 965 if (mtls->start.y >= mtls->end.y) return false; 966 } 967 968 if (!sc || (sc->zEnd == 0)) { 969 mtls->end.z = mtls->fep.dim.z; 970 } else { 971 mtls->start.z = rsMin(mtls->fep.dim.z, sc->zStart); 972 mtls->end.z = rsMin(mtls->fep.dim.z, sc->zEnd); 973 if (mtls->start.z >= mtls->end.z) return false; 974 } 975 976 if (!sc || (sc->arrayEnd == 0)) { 977 mtls->end.array[0] = mtls->fep.dim.array[0]; 978 } else { 979 mtls->start.array[0] = rsMin(mtls->fep.dim.array[0], sc->arrayStart); 980 mtls->end.array[0] = rsMin(mtls->fep.dim.array[0], sc->arrayEnd); 981 if (mtls->start.array[0] >= mtls->end.array[0]) return false; 982 } 983 984 if (!sc || (sc->array2End == 0)) { 985 mtls->end.array[1] = mtls->fep.dim.array[1]; 986 } else { 987 mtls->start.array[1] = rsMin(mtls->fep.dim.array[1], sc->array2Start); 988 mtls->end.array[1] = rsMin(mtls->fep.dim.array[1], sc->array2End); 989 if (mtls->start.array[1] >= mtls->end.array[1]) return false; 990 } 991 992 if (!sc || (sc->array3End == 0)) { 993 mtls->end.array[2] = mtls->fep.dim.array[2]; 994 } else { 995 mtls->start.array[2] = rsMin(mtls->fep.dim.array[2], sc->array3Start); 996 mtls->end.array[2] = rsMin(mtls->fep.dim.array[2], sc->array3End); 997 if (mtls->start.array[2] >= mtls->end.array[2]) return false; 998 } 999 1000 if (!sc || (sc->array4End == 0)) { 1001 mtls->end.array[3] = mtls->fep.dim.array[3]; 1002 } else { 1003 mtls->start.array[3] = rsMin(mtls->fep.dim.array[3], sc->array4Start); 1004 mtls->end.array[3] = rsMin(mtls->fep.dim.array[3], sc->array4End); 1005 if (mtls->start.array[3] >= mtls->end.array[3]) return false; 1006 } 1007 1008 1009 // The X & Y walkers always want 0-1 min even if dim is not present 1010 mtls->end.x = rsMax((uint32_t)1, mtls->end.x); 1011 mtls->end.y = rsMax((uint32_t)1, mtls->end.y); 1012 1013 mtls->rsc = mCtx; 1014 if (ains) { 1015 memcpy(mtls->ains, ains, inLen * sizeof(ains[0])); 1016 } 1017 mtls->aout[0] = aout; 1018 mtls->fep.usr = usr; 1019 mtls->fep.usrLen = usrLen; 1020 mtls->mSliceSize = 1; 1021 mtls->mSliceNum = 0; 1022 1023 mtls->isThreadable = mIsThreadable; 1024 1025 if (inLen > 0) { 1026 mtls->fep.inLen = inLen; 1027 for (int index = inLen; --index >= 0;) { 1028 mtls->fep.inPtr[index] = (const uint8_t*)ains[index]->mHal.drvState.lod[0].mallocPtr; 1029 mtls->fep.inStride[index] = ains[index]->getType()->getElementSizeBytes(); 1030 } 1031 } 1032 1033 if (aout != nullptr) { 1034 mtls->fep.outPtr[0] = (uint8_t *)aout->mHal.drvState.lod[0].mallocPtr; 1035 mtls->fep.outStride[0] = aout->getType()->getElementSizeBytes(); 1036 } 1037 1038 // All validation passed, ok to launch threads 1039 return true; 1040} 1041 1042 1043void RsdCpuScriptImpl::invokeForEach(uint32_t slot, 1044 const Allocation ** ains, 1045 uint32_t inLen, 1046 Allocation * aout, 1047 const void * usr, 1048 uint32_t usrLen, 1049 const RsScriptCall *sc) { 1050 1051 MTLaunchStruct mtls; 1052 1053 if (forEachMtlsSetup(ains, inLen, aout, usr, usrLen, sc, &mtls)) { 1054 forEachKernelSetup(slot, &mtls); 1055 1056 RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this); 1057 mCtx->launchThreads(ains, inLen, aout, sc, &mtls); 1058 mCtx->setTLS(oldTLS); 1059 } 1060} 1061 1062void RsdCpuScriptImpl::forEachKernelSetup(uint32_t slot, MTLaunchStruct *mtls) { 1063 mtls->script = this; 1064 mtls->fep.slot = slot; 1065 mtls->kernel = mScriptExec->getForEachFunction(slot); 1066 rsAssert(mtls->kernel != nullptr); 1067 mtls->sig = mScriptExec->getForEachSignature(slot); 1068} 1069 1070int RsdCpuScriptImpl::invokeRoot() { 1071 RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this); 1072 int ret = mRoot(); 1073 mCtx->setTLS(oldTLS); 1074 return ret; 1075} 1076 1077void RsdCpuScriptImpl::invokeInit() { 1078 if (mInit) { 1079 mInit(); 1080 } 1081} 1082 1083void RsdCpuScriptImpl::invokeFreeChildren() { 1084 if (mFreeChildren) { 1085 mFreeChildren(); 1086 } 1087} 1088 1089void RsdCpuScriptImpl::invokeFunction(uint32_t slot, const void *params, 1090 size_t paramLength) { 1091 //ALOGE("invoke %i %p %zu", slot, params, paramLength); 1092 void * ap = nullptr; 1093 1094#if defined(__x86_64__) 1095 // The invoked function could have input parameter of vector type for example float4 which 1096 // requires void* params to be 16 bytes aligned when using SSE instructions for x86_64 platform. 1097 // So try to align void* params before passing them into RS exported function. 1098 1099 if ((uint8_t)(uint64_t)params & 0x0F) { 1100 if ((ap = (void*)memalign(16, paramLength)) != nullptr) { 1101 memcpy(ap, params, paramLength); 1102 } else { 1103 ALOGE("x86_64: invokeFunction memalign error, still use params which is not 16 bytes aligned."); 1104 } 1105 } 1106#endif 1107 1108 RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this); 1109 reinterpret_cast<void (*)(const void *, uint32_t)>( 1110 mScriptExec->getInvokeFunction(slot))(ap? (const void *) ap: params, paramLength); 1111 1112 mCtx->setTLS(oldTLS); 1113} 1114 1115void RsdCpuScriptImpl::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) { 1116 //rsAssert(!script->mFieldIsObject[slot]); 1117 //ALOGE("setGlobalVar %i %p %zu", slot, data, dataLength); 1118 1119 //if (mIntrinsicID) { 1120 //mIntrinsicFuncs.setVar(dc, script, drv->mIntrinsicData, slot, data, dataLength); 1121 //return; 1122 //} 1123 1124 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 1125 if (!destPtr) { 1126 //ALOGV("Calling setVar on slot = %i which is null", slot); 1127 return; 1128 } 1129 1130 memcpy(destPtr, data, dataLength); 1131} 1132 1133void RsdCpuScriptImpl::getGlobalVar(uint32_t slot, void *data, size_t dataLength) { 1134 //rsAssert(!script->mFieldIsObject[slot]); 1135 //ALOGE("getGlobalVar %i %p %zu", slot, data, dataLength); 1136 1137 int32_t *srcPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 1138 if (!srcPtr) { 1139 //ALOGV("Calling setVar on slot = %i which is null", slot); 1140 return; 1141 } 1142 memcpy(data, srcPtr, dataLength); 1143} 1144 1145 1146void RsdCpuScriptImpl::setGlobalVarWithElemDims(uint32_t slot, const void *data, size_t dataLength, 1147 const Element *elem, 1148 const uint32_t *dims, size_t dimLength) { 1149 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 1150 if (!destPtr) { 1151 //ALOGV("Calling setVar on slot = %i which is null", slot); 1152 return; 1153 } 1154 1155 // We want to look at dimension in terms of integer components, 1156 // but dimLength is given in terms of bytes. 1157 dimLength /= sizeof(int); 1158 1159 // Only a single dimension is currently supported. 1160 rsAssert(dimLength == 1); 1161 if (dimLength == 1) { 1162 // First do the increment loop. 1163 size_t stride = elem->getSizeBytes(); 1164 const char *cVal = reinterpret_cast<const char *>(data); 1165 for (uint32_t i = 0; i < dims[0]; i++) { 1166 elem->incRefs(cVal); 1167 cVal += stride; 1168 } 1169 1170 // Decrement loop comes after (to prevent race conditions). 1171 char *oldVal = reinterpret_cast<char *>(destPtr); 1172 for (uint32_t i = 0; i < dims[0]; i++) { 1173 elem->decRefs(oldVal); 1174 oldVal += stride; 1175 } 1176 } 1177 1178 memcpy(destPtr, data, dataLength); 1179} 1180 1181void RsdCpuScriptImpl::setGlobalBind(uint32_t slot, Allocation *data) { 1182 1183 //rsAssert(!script->mFieldIsObject[slot]); 1184 //ALOGE("setGlobalBind %i %p", slot, data); 1185 1186 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 1187 if (!destPtr) { 1188 //ALOGV("Calling setVar on slot = %i which is null", slot); 1189 return; 1190 } 1191 1192 void *ptr = nullptr; 1193 mBoundAllocs[slot] = data; 1194 if (data) { 1195 ptr = data->mHal.drvState.lod[0].mallocPtr; 1196 } 1197 memcpy(destPtr, &ptr, sizeof(void *)); 1198} 1199 1200void RsdCpuScriptImpl::setGlobalObj(uint32_t slot, ObjectBase *data) { 1201 1202 //rsAssert(script->mFieldIsObject[slot]); 1203 //ALOGE("setGlobalObj %i %p", slot, data); 1204 1205 int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot)); 1206 if (!destPtr) { 1207 //ALOGV("Calling setVar on slot = %i which is null", slot); 1208 return; 1209 } 1210 1211 rsrSetObject(mCtx->getContext(), (rs_object_base *)destPtr, data); 1212} 1213 1214RsdCpuScriptImpl::~RsdCpuScriptImpl() { 1215#ifndef RS_COMPATIBILITY_LIB 1216 if (mCompilerDriver) { 1217 delete mCompilerDriver; 1218 } 1219#endif 1220 1221 if (mScriptExec != nullptr) { 1222 delete mScriptExec; 1223 } 1224 if (mBoundAllocs) delete[] mBoundAllocs; 1225 if (mScriptSO) { 1226 dlclose(mScriptSO); 1227 } 1228} 1229 1230Allocation * RsdCpuScriptImpl::getAllocationForPointer(const void *ptr) const { 1231 if (!ptr) { 1232 return nullptr; 1233 } 1234 1235 for (uint32_t ct=0; ct < mScript->mHal.info.exportedVariableCount; ct++) { 1236 Allocation *a = mBoundAllocs[ct]; 1237 if (!a) continue; 1238 if (a->mHal.drvState.lod[0].mallocPtr == ptr) { 1239 return a; 1240 } 1241 } 1242 ALOGE("rsGetAllocation, failed to find %p", ptr); 1243 return nullptr; 1244} 1245 1246void RsdCpuScriptImpl::preLaunch(uint32_t slot, const Allocation ** ains, 1247 uint32_t inLen, Allocation * aout, 1248 const void * usr, uint32_t usrLen, 1249 const RsScriptCall *sc) {} 1250 1251void RsdCpuScriptImpl::postLaunch(uint32_t slot, const Allocation ** ains, 1252 uint32_t inLen, Allocation * aout, 1253 const void * usr, uint32_t usrLen, 1254 const RsScriptCall *sc) {} 1255 1256 1257} 1258} 1259