rsCpuScript.cpp revision 358ffb84f56929cd2d61f6429a790a1606eaf865 
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 <set>
22    #include <string>
23    #include <dlfcn.h>
24    #include <stdio.h>
25    #include <stdlib.h>
26    #include <string.h>
27    #include <sys/stat.h>
28    #include <unistd.h>
29    #include <fstream>
30    #include <iostream>
31#else
32    #include <bcc/BCCContext.h>
33    #include <bcc/Config/Config.h>
34    #include <bcc/Renderscript/RSCompilerDriver.h>
35    #include <bcc/Renderscript/RSExecutable.h>
36    #include <bcc/Renderscript/RSInfo.h>
37    #include <bcinfo/MetadataExtractor.h>
38    #include <cutils/properties.h>
39
40    #include <sys/types.h>
41    #include <sys/wait.h>
42    #include <unistd.h>
43
44    #include <string>
45    #include <vector>
46#endif
47
48namespace {
49#ifdef RS_COMPATIBILITY_LIB
50
51// Create a len length string containing random characters from [A-Za-z0-9].
52static std::string getRandomString(size_t len) {
53    char buf[len + 1];
54    for (size_t i = 0; i < len; i++) {
55        uint32_t r = arc4random() & 0xffff;
56        r %= 62;
57        if (r < 26) {
58            // lowercase
59            buf[i] = 'a' + r;
60        } else if (r < 52) {
61            // uppercase
62            buf[i] = 'A' + (r - 26);
63        } else {
64            // Use a number
65            buf[i] = '0' + (r - 52);
66        }
67    }
68    buf[len] = '\0';
69    return std::string(buf);
70}
71
72// Check if a path exists and attempt to create it if it doesn't.
73static bool ensureCacheDirExists(const char *path) {
74    if (access(path, R_OK | W_OK | X_OK) == 0) {
75        // Done if we can rwx the directory
76        return true;
77    }
78    if (mkdir(path, 0700) == 0) {
79        return true;
80    }
81    return false;
82}
83
84// Copy the file named \p srcFile to \p dstFile.
85// Return 0 on success and -1 if anything wasn't copied.
86static int copyFile(const char *dstFile, const char *srcFile) {
87    std::ifstream srcStream(srcFile);
88    if (!srcStream) {
89        ALOGE("Could not verify or read source file: %s", srcFile);
90        return -1;
91    }
92    std::ofstream dstStream(dstFile);
93    if (!dstStream) {
94        ALOGE("Could not verify or write destination file: %s", dstFile);
95        return -1;
96    }
97    dstStream << srcStream.rdbuf();
98    if (!dstStream) {
99        ALOGE("Could not write destination file: %s", dstFile);
100        return -1;
101    }
102
103    srcStream.close();
104    dstStream.close();
105
106    return 0;
107}
108
109// Attempt to load the shared library from origName, but then fall back to
110// creating a copy of the shared library if necessary (to ensure instancing).
111// This function returns the dlopen()-ed handle if successful.
112static void *loadSOHelper(const char *origName, const char *cacheDir,
113                          const char *resName) {
114    // Keep track of which .so libraries have been loaded. Once a library is
115    // in the set (per-process granularity), we must instead make a copy of
116    // the original shared object (randomly named .so file) and load that one
117    // instead. If we don't do this, we end up aliasing global data between
118    // the various Script instances (which are supposed to be completely
119    // independent).
120    static std::set<std::string> LoadedLibraries;
121
122    void *loaded = nullptr;
123
124    // Skip everything if we don't even have the original library available.
125    if (access(origName, F_OK) != 0) {
126        return nullptr;
127    }
128
129    // Common path is that we have not loaded this Script/library before.
130    if (LoadedLibraries.find(origName) == LoadedLibraries.end()) {
131        loaded = dlopen(origName, RTLD_NOW | RTLD_LOCAL);
132        if (loaded) {
133            LoadedLibraries.insert(origName);
134        }
135        return loaded;
136    }
137
138    std::string newName(cacheDir);
139    newName.append("/com.android.renderscript.cache/");
140
141    if (!ensureCacheDirExists(newName.c_str())) {
142        ALOGE("Could not verify or create cache dir: %s", cacheDir);
143        return nullptr;
144    }
145
146    // Construct an appropriately randomized filename for the copy.
147    newName.append("librs.");
148    newName.append(resName);
149    newName.append("#");
150    newName.append(getRandomString(6));  // 62^6 potential filename variants.
151    newName.append(".so");
152
153    int r = copyFile(newName.c_str(), origName);
154    if (r != 0) {
155        ALOGE("Could not create copy %s -> %s", origName, newName.c_str());
156        return nullptr;
157    }
158    loaded = dlopen(newName.c_str(), RTLD_NOW | RTLD_LOCAL);
159    r = unlink(newName.c_str());
160    if (r != 0) {
161        ALOGE("Could not unlink copy %s", newName.c_str());
162    }
163    if (loaded) {
164        LoadedLibraries.insert(newName.c_str());
165    }
166
167    return loaded;
168}
169
170// Load the shared library referred to by cacheDir and resName. If we have
171// already loaded this library, we instead create a new copy (in the
172// cache dir) and then load that. We then immediately destroy the copy.
173// This is required behavior to implement script instancing for the support
174// library, since shared objects are loaded and de-duped by name only.
175static void *loadSharedLibrary(const char *cacheDir, const char *resName) {
176    void *loaded = nullptr;
177#ifndef RS_SERVER
178    std::string scriptSOName(cacheDir);
179    size_t cutPos = scriptSOName.rfind("cache");
180    if (cutPos != std::string::npos) {
181        scriptSOName.erase(cutPos);
182    } else {
183        ALOGE("Found peculiar cacheDir (missing \"cache\"): %s", cacheDir);
184    }
185    scriptSOName.append("/lib/librs.");
186#else
187    std::string scriptSOName("lib");
188#endif
189    scriptSOName.append(resName);
190    scriptSOName.append(".so");
191
192    // We should check if we can load the library from the standard app
193    // location for shared libraries first.
194    loaded = loadSOHelper(scriptSOName.c_str(), cacheDir, resName);
195
196    if (loaded == nullptr) {
197        ALOGE("Unable to open shared library (%s): %s",
198              scriptSOName.c_str(), dlerror());
199
200        // One final attempt to find the library in "/system/lib".
201        // We do this to allow bundled applications to use the compatibility
202        // library fallback path. Those applications don't have a private
203        // library path, so they need to install to the system directly.
204        // Note that this is really just a testing path.
205        std::string scriptSONameSystem("/system/lib/librs.");
206        scriptSONameSystem.append(resName);
207        scriptSONameSystem.append(".so");
208        loaded = loadSOHelper(scriptSONameSystem.c_str(), cacheDir,
209                              resName);
210        if (loaded == nullptr) {
211            ALOGE("Unable to open system shared library (%s): %s",
212                  scriptSONameSystem.c_str(), dlerror());
213        }
214    }
215
216    return loaded;
217}
218
219#else  // RS_COMPATIBILITY_LIB is not defined
220
221static bool is_force_recompile() {
222#ifdef RS_SERVER
223  return false;
224#else
225  char buf[PROPERTY_VALUE_MAX];
226
227  // Re-compile if floating point precision has been overridden.
228  property_get("debug.rs.precision", buf, "");
229  if (buf[0] != '\0') {
230    return true;
231  }
232
233  // Re-compile if debug.rs.forcerecompile is set.
234  property_get("debug.rs.forcerecompile", buf, "0");
235  if ((::strcmp(buf, "1") == 0) || (::strcmp(buf, "true") == 0)) {
236    return true;
237  } else {
238    return false;
239  }
240#endif  // RS_SERVER
241}
242
243const static char *BCC_EXE_PATH = "/system/bin/bcc";
244
245static void setCompileArguments(std::vector<const char*>* args,
246                                const std::string& bcFileName,
247                                const char* cacheDir, const char* resName,
248                                const char* core_lib, bool useRSDebugContext,
249                                const char* bccPluginName) {
250    rsAssert(cacheDir && resName && core_lib);
251    args->push_back(BCC_EXE_PATH);
252    args->push_back("-o");
253    args->push_back(resName);
254    args->push_back("-output_path");
255    args->push_back(cacheDir);
256    args->push_back("-bclib");
257    args->push_back(core_lib);
258    args->push_back("-mtriple");
259    args->push_back(DEFAULT_TARGET_TRIPLE_STRING);
260
261    // Enable workaround for A53 codegen by default.
262#if defined(__aarch64__) && !defined(DISABLE_A53_WORKAROUND)
263    args->push_back("-aarch64-fix-cortex-a53-835769");
264#endif
265
266    // Execute the bcc compiler.
267    if (useRSDebugContext) {
268        args->push_back("-rs-debug-ctx");
269    } else {
270        // Only load additional libraries for compiles that don't use
271        // the debug context.
272        if (bccPluginName && strlen(bccPluginName) > 0) {
273            args->push_back("-load");
274            args->push_back(bccPluginName);
275        }
276    }
277
278    args->push_back(bcFileName.c_str());
279    args->push_back(nullptr);
280}
281
282static bool compileBitcode(const std::string &bcFileName,
283                           const char *bitcode,
284                           size_t bitcodeSize,
285                           const char **compileArguments,
286                           const std::string &compileCommandLine) {
287    rsAssert(bitcode && bitcodeSize);
288
289    FILE *bcfile = fopen(bcFileName.c_str(), "w");
290    if (!bcfile) {
291        ALOGE("Could not write to %s", bcFileName.c_str());
292        return false;
293    }
294    size_t nwritten = fwrite(bitcode, 1, bitcodeSize, bcfile);
295    fclose(bcfile);
296    if (nwritten != bitcodeSize) {
297        ALOGE("Could not write %zu bytes to %s", bitcodeSize,
298              bcFileName.c_str());
299        return false;
300    }
301
302    pid_t pid = fork();
303
304    switch (pid) {
305    case -1: {  // Error occurred (we attempt no recovery)
306        ALOGE("Couldn't fork for bcc compiler execution");
307        return false;
308    }
309    case 0: {  // Child process
310        ALOGV("Invoking BCC with: %s", compileCommandLine.c_str());
311        execv(BCC_EXE_PATH, (char* const*)compileArguments);
312
313        ALOGE("execv() failed: %s", strerror(errno));
314        abort();
315        return false;
316    }
317    default: {  // Parent process (actual driver)
318        // Wait on child process to finish compiling the source.
319        int status = 0;
320        pid_t w = waitpid(pid, &status, 0);
321        if (w == -1) {
322            ALOGE("Could not wait for bcc compiler");
323            return false;
324        }
325
326        if (WIFEXITED(status) && WEXITSTATUS(status) == 0) {
327            return true;
328        }
329
330        ALOGE("bcc compiler terminated unexpectedly");
331        return false;
332    }
333    }
334}
335
336#endif  // !defined(RS_COMPATIBILITY_LIB)
337}  // namespace
338
339namespace android {
340namespace renderscript {
341
342#ifdef RS_COMPATIBILITY_LIB
343#define MAXLINE 500
344#define MAKE_STR_HELPER(S) #S
345#define MAKE_STR(S) MAKE_STR_HELPER(S)
346#define EXPORT_VAR_STR "exportVarCount: "
347#define EXPORT_FUNC_STR "exportFuncCount: "
348#define EXPORT_FOREACH_STR "exportForEachCount: "
349#define OBJECT_SLOT_STR "objectSlotCount: "
350
351// Copy up to a newline or size chars from str -> s, updating str
352// Returns s when successful and nullptr when '\0' is finally reached.
353static char* strgets(char *s, int size, const char **ppstr) {
354    if (!ppstr || !*ppstr || **ppstr == '\0' || size < 1) {
355        return nullptr;
356    }
357
358    int i;
359    for (i = 0; i < (size - 1); i++) {
360        s[i] = **ppstr;
361        (*ppstr)++;
362        if (s[i] == '\0') {
363            return s;
364        } else if (s[i] == '\n') {
365            s[i+1] = '\0';
366            return s;
367        }
368    }
369
370    // size has been exceeded.
371    s[i] = '\0';
372
373    return s;
374}
375#endif
376
377RsdCpuScriptImpl::RsdCpuScriptImpl(RsdCpuReferenceImpl *ctx, const Script *s) {
378    mCtx = ctx;
379    mScript = s;
380
381#ifdef RS_COMPATIBILITY_LIB
382    mScriptSO = nullptr;
383    mInvokeFunctions = nullptr;
384    mForEachFunctions = nullptr;
385    mFieldAddress = nullptr;
386    mFieldIsObject = nullptr;
387    mForEachSignatures = nullptr;
388#else
389    mCompilerContext = nullptr;
390    mCompilerDriver = nullptr;
391    mExecutable = nullptr;
392#endif
393
394
395    mRoot = nullptr;
396    mRootExpand = nullptr;
397    mInit = nullptr;
398    mFreeChildren = nullptr;
399
400
401    mBoundAllocs = nullptr;
402    mIntrinsicData = nullptr;
403    mIsThreadable = true;
404}
405
406
407bool RsdCpuScriptImpl::init(char const *resName, char const *cacheDir,
408                            uint8_t const *bitcode, size_t bitcodeSize,
409                            uint32_t flags, char const *bccPluginName) {
410    //ALOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir, bitcode, bitcodeSize, flags, lookupFunc);
411    //ALOGE("rsdScriptInit %p %p", rsc, script);
412
413    mCtx->lockMutex();
414#ifndef RS_COMPATIBILITY_LIB
415    bool useRSDebugContext = false;
416
417    mCompilerContext = nullptr;
418    mCompilerDriver = nullptr;
419    mExecutable = nullptr;
420
421    mCompilerContext = new bcc::BCCContext();
422    if (mCompilerContext == nullptr) {
423        ALOGE("bcc: FAILS to create compiler context (out of memory)");
424        mCtx->unlockMutex();
425        return false;
426    }
427
428    mCompilerDriver = new bcc::RSCompilerDriver();
429    if (mCompilerDriver == nullptr) {
430        ALOGE("bcc: FAILS to create compiler driver (out of memory)");
431        mCtx->unlockMutex();
432        return false;
433    }
434
435    // Configure symbol resolvers (via compiler-rt and the RS runtime).
436    mRSRuntime.setLookupFunction(lookupRuntimeStub);
437    mRSRuntime.setContext(this);
438    mResolver.chainResolver(mCompilerRuntime);
439    mResolver.chainResolver(mRSRuntime);
440
441    // Run any compiler setup functions we have been provided with.
442    RSSetupCompilerCallback setupCompilerCallback =
443            mCtx->getSetupCompilerCallback();
444    if (setupCompilerCallback != nullptr) {
445        setupCompilerCallback(mCompilerDriver);
446    }
447
448    bcinfo::MetadataExtractor bitcodeMetadata((const char *) bitcode, bitcodeSize);
449    if (!bitcodeMetadata.extract()) {
450        ALOGE("Could not extract metadata from bitcode");
451        mCtx->unlockMutex();
452        return false;
453    }
454
455    const char* core_lib = findCoreLib(bitcodeMetadata, (const char*)bitcode, bitcodeSize);
456
457    if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) {
458        mCompilerDriver->setDebugContext(true);
459        useRSDebugContext = true;
460    }
461
462    std::string bcFileName(cacheDir);
463    bcFileName.append("/");
464    bcFileName.append(resName);
465    bcFileName.append(".bc");
466
467    std::vector<const char*> compileArguments;
468    setCompileArguments(&compileArguments, bcFileName, cacheDir, resName, core_lib,
469                        useRSDebugContext, bccPluginName);
470    // The last argument of compileArguments ia a nullptr, so remove 1 from the size.
471    std::string compileCommandLine =
472                bcc::getCommandLine(compileArguments.size() - 1, compileArguments.data());
473
474    if (!is_force_recompile()) {
475        // Load the compiled script that's in the cache, if any.
476        mExecutable = bcc::RSCompilerDriver::loadScript(cacheDir, resName, (const char*)bitcode,
477                                                        bitcodeSize, compileCommandLine.c_str(),
478                                                        mResolver);
479    }
480
481    // If we can't, it's either not there or out of date.  We compile the bit code and try loading
482    // again.
483    if (mExecutable == nullptr) {
484        if (!compileBitcode(bcFileName, (const char*)bitcode, bitcodeSize, compileArguments.data(),
485                            compileCommandLine)) {
486            ALOGE("bcc: FAILS to compile '%s'", resName);
487            mCtx->unlockMutex();
488            return false;
489        }
490        mExecutable = bcc::RSCompilerDriver::loadScript(cacheDir, resName, (const char*)bitcode,
491                                                        bitcodeSize, compileCommandLine.c_str(),
492                                                        mResolver);
493        if (mExecutable == nullptr) {
494            ALOGE("bcc: FAILS to load freshly compiled executable for '%s'", resName);
495            mCtx->unlockMutex();
496            return false;
497        }
498    }
499
500    mExecutable->setThreadable(mIsThreadable);
501    if (!mExecutable->syncInfo()) {
502        ALOGW("bcc: FAILS to synchronize the RS info file to the disk");
503    }
504
505    mRoot = reinterpret_cast<int (*)()>(mExecutable->getSymbolAddress("root"));
506    mRootExpand =
507        reinterpret_cast<int (*)()>(mExecutable->getSymbolAddress("root.expand"));
508    mInit = reinterpret_cast<void (*)()>(mExecutable->getSymbolAddress("init"));
509    mFreeChildren =
510        reinterpret_cast<void (*)()>(mExecutable->getSymbolAddress(".rs.dtor"));
511
512
513    if (bitcodeMetadata.getExportVarCount()) {
514        mBoundAllocs = new Allocation *[bitcodeMetadata.getExportVarCount()];
515        memset(mBoundAllocs, 0, sizeof(void *) * bitcodeMetadata.getExportVarCount());
516    }
517
518    for (size_t i = 0; i < bitcodeMetadata.getExportForEachSignatureCount(); i++) {
519        char* name = new char[strlen(bitcodeMetadata.getExportForEachNameList()[i]) + 1];
520        mExportedForEachFuncList.push_back(
521                    std::make_pair(name, bitcodeMetadata.getExportForEachSignatureList()[i]));
522    }
523
524#else  // RS_COMPATIBILITY_LIB is defined
525
526    mScriptSO = loadSharedLibrary(cacheDir, resName);
527
528    if (mScriptSO) {
529        char line[MAXLINE];
530        mRoot = (RootFunc_t) dlsym(mScriptSO, "root");
531        if (mRoot) {
532            //ALOGE("Found root(): %p", mRoot);
533        }
534        mRootExpand = (RootFunc_t) dlsym(mScriptSO, "root.expand");
535        if (mRootExpand) {
536            //ALOGE("Found root.expand(): %p", mRootExpand);
537        }
538        mInit = (InvokeFunc_t) dlsym(mScriptSO, "init");
539        if (mInit) {
540            //ALOGE("Found init(): %p", mInit);
541        }
542        mFreeChildren = (InvokeFunc_t) dlsym(mScriptSO, ".rs.dtor");
543        if (mFreeChildren) {
544            //ALOGE("Found .rs.dtor(): %p", mFreeChildren);
545        }
546
547        const char *rsInfo = (const char *) dlsym(mScriptSO, ".rs.info");
548        if (rsInfo) {
549            //ALOGE("Found .rs.info(): %p - %s", rsInfo, rsInfo);
550        }
551
552        size_t varCount = 0;
553        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
554            goto error;
555        }
556        if (sscanf(line, EXPORT_VAR_STR "%zu", &varCount) != 1) {
557            ALOGE("Invalid export var count!: %s", line);
558            goto error;
559        }
560
561        mExportedVariableCount = varCount;
562        //ALOGE("varCount: %zu", varCount);
563        if (varCount > 0) {
564            // Start by creating/zeroing this member, since we don't want to
565            // accidentally clean up invalid pointers later (if we error out).
566            mFieldIsObject = new bool[varCount];
567            if (mFieldIsObject == nullptr) {
568                goto error;
569            }
570            memset(mFieldIsObject, 0, varCount * sizeof(*mFieldIsObject));
571            mFieldAddress = new void*[varCount];
572            if (mFieldAddress == nullptr) {
573                goto error;
574            }
575            for (size_t i = 0; i < varCount; ++i) {
576                if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
577                    goto error;
578                }
579                char *c = strrchr(line, '\n');
580                if (c) {
581                    *c = '\0';
582                }
583                mFieldAddress[i] = dlsym(mScriptSO, line);
584                if (mFieldAddress[i] == nullptr) {
585                    ALOGE("Failed to find variable address for %s: %s",
586                          line, dlerror());
587                    // Not a critical error if we don't find a global variable.
588                }
589                else {
590                    //ALOGE("Found variable %s at %p", line,
591                    //mFieldAddress[i]);
592                }
593            }
594        }
595
596        size_t funcCount = 0;
597        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
598            goto error;
599        }
600        if (sscanf(line, EXPORT_FUNC_STR "%zu", &funcCount) != 1) {
601            ALOGE("Invalid export func count!: %s", line);
602            goto error;
603        }
604
605        mExportedFunctionCount = funcCount;
606        //ALOGE("funcCount: %zu", funcCount);
607
608        if (funcCount > 0) {
609            mInvokeFunctions = new InvokeFunc_t[funcCount];
610            if (mInvokeFunctions == nullptr) {
611                goto error;
612            }
613            for (size_t i = 0; i < funcCount; ++i) {
614                if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
615                    goto error;
616                }
617                char *c = strrchr(line, '\n');
618                if (c) {
619                    *c = '\0';
620                }
621
622                mInvokeFunctions[i] = (InvokeFunc_t) dlsym(mScriptSO, line);
623                if (mInvokeFunctions[i] == nullptr) {
624                    ALOGE("Failed to get function address for %s(): %s",
625                          line, dlerror());
626                    goto error;
627                }
628                else {
629                    //ALOGE("Found InvokeFunc_t %s at %p", line, mInvokeFunctions[i]);
630                }
631            }
632        }
633
634        size_t forEachCount = 0;
635        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
636            goto error;
637        }
638        if (sscanf(line, EXPORT_FOREACH_STR "%zu", &forEachCount) != 1) {
639            ALOGE("Invalid export forEach count!: %s", line);
640            goto error;
641        }
642
643        if (forEachCount > 0) {
644
645            mForEachSignatures = new uint32_t[forEachCount];
646            if (mForEachSignatures == nullptr) {
647                goto error;
648            }
649            mForEachFunctions = new ForEachFunc_t[forEachCount];
650            if (mForEachFunctions == nullptr) {
651                goto error;
652            }
653            for (size_t i = 0; i < forEachCount; ++i) {
654                unsigned int tmpSig = 0;
655                char tmpName[MAXLINE];
656
657                if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
658                    goto error;
659                }
660                if (sscanf(line, "%u - %" MAKE_STR(MAXLINE) "s",
661                           &tmpSig, tmpName) != 2) {
662                    ALOGE("Invalid export forEach!: %s", line);
663                    goto error;
664                }
665
666                // Lookup the expanded ForEach kernel.
667                strncat(tmpName, ".expand", MAXLINE-1-strlen(tmpName));
668                mForEachSignatures[i] = tmpSig;
669                mForEachFunctions[i] =
670                        (ForEachFunc_t) dlsym(mScriptSO, tmpName);
671                if (i != 0 && mForEachFunctions[i] == nullptr) {
672                    // Ignore missing root.expand functions.
673                    // root() is always specified at location 0.
674                    ALOGE("Failed to find forEach function address for %s: %s",
675                          tmpName, dlerror());
676                    goto error;
677                }
678                else {
679                    //ALOGE("Found forEach %s at %p", tmpName, mForEachFunctions[i]);
680                }
681            }
682        }
683
684        size_t objectSlotCount = 0;
685        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
686            goto error;
687        }
688        if (sscanf(line, OBJECT_SLOT_STR "%zu", &objectSlotCount) != 1) {
689            ALOGE("Invalid object slot count!: %s", line);
690            goto error;
691        }
692
693        if (objectSlotCount > 0) {
694            rsAssert(varCount > 0);
695            for (size_t i = 0; i < objectSlotCount; ++i) {
696                uint32_t varNum = 0;
697                if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
698                    goto error;
699                }
700                if (sscanf(line, "%u", &varNum) != 1) {
701                    ALOGE("Invalid object slot!: %s", line);
702                    goto error;
703                }
704
705                if (varNum < varCount) {
706                    mFieldIsObject[varNum] = true;
707                }
708            }
709        }
710
711        if (varCount > 0) {
712            mBoundAllocs = new Allocation *[varCount];
713            memset(mBoundAllocs, 0, varCount * sizeof(*mBoundAllocs));
714        }
715
716        if (mScriptSO == (void*)1) {
717            //rsdLookupRuntimeStub(script, "acos");
718        }
719    } else {
720        goto error;
721    }
722#endif
723    mCtx->unlockMutex();
724    return true;
725
726#ifdef RS_COMPATIBILITY_LIB
727error:
728
729    mCtx->unlockMutex();
730    delete[] mInvokeFunctions;
731    delete[] mForEachFunctions;
732    delete[] mFieldAddress;
733    delete[] mFieldIsObject;
734    delete[] mForEachSignatures;
735    delete[] mBoundAllocs;
736    if (mScriptSO) {
737        dlclose(mScriptSO);
738    }
739    return false;
740#endif
741}
742
743#ifndef RS_COMPATIBILITY_LIB
744
745#ifdef __LP64__
746#define SYSLIBPATH "/system/lib64"
747#else
748#define SYSLIBPATH "/system/lib"
749#endif
750
751const char* RsdCpuScriptImpl::findCoreLib(const bcinfo::MetadataExtractor& ME, const char* bitcode,
752                                          size_t bitcodeSize) {
753    const char* defaultLib = SYSLIBPATH"/libclcore.bc";
754
755    // If we're debugging, use the debug library.
756    if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) {
757        return SYSLIBPATH"/libclcore_debug.bc";
758    }
759
760    // If a callback has been registered to specify a library, use that.
761    RSSelectRTCallback selectRTCallback = mCtx->getSelectRTCallback();
762    if (selectRTCallback != nullptr) {
763        return selectRTCallback((const char*)bitcode, bitcodeSize);
764    }
765
766    // Check for a platform specific library
767#if defined(ARCH_ARM_HAVE_NEON) && !defined(DISABLE_CLCORE_NEON)
768    enum bcinfo::RSFloatPrecision prec = ME.getRSFloatPrecision();
769    if (prec == bcinfo::RS_FP_Relaxed) {
770        // NEON-capable ARMv7a devices can use an accelerated math library
771        // for all reduced precision scripts.
772        // ARMv8 does not use NEON, as ASIMD can be used with all precision
773        // levels.
774        return SYSLIBPATH"/libclcore_neon.bc";
775    } else {
776        return defaultLib;
777    }
778#elif defined(__i386__) || defined(__x86_64__)
779    // x86 devices will use an optimized library.
780    return SYSLIBPATH"/libclcore_x86.bc";
781#else
782    return defaultLib;
783#endif
784}
785
786#endif
787
788void RsdCpuScriptImpl::populateScript(Script *script) {
789#ifndef RS_COMPATIBILITY_LIB
790    // Copy info over to runtime
791    script->mHal.info.exportedFunctionCount = mExecutable->getExportFuncAddrs().size();
792    script->mHal.info.exportedVariableCount = mExecutable->getExportVarAddrs().size();
793    script->mHal.info.exportedForeachFuncList = &mExportedForEachFuncList[0];
794    script->mHal.info.exportedPragmaCount = mExecutable->getPragmaKeys().size();
795    script->mHal.info.exportedPragmaKeyList =
796        const_cast<const char**>(&mExecutable->getPragmaKeys().front());
797    script->mHal.info.exportedPragmaValueList =
798        const_cast<const char**>(&mExecutable->getPragmaValues().front());
799
800    if (mRootExpand) {
801        script->mHal.info.root = mRootExpand;
802    } else {
803        script->mHal.info.root = mRoot;
804    }
805#else
806    // Copy info over to runtime
807    script->mHal.info.exportedFunctionCount = mExportedFunctionCount;
808    script->mHal.info.exportedVariableCount = mExportedVariableCount;
809    script->mHal.info.exportedPragmaCount = 0;
810    script->mHal.info.exportedPragmaKeyList = 0;
811    script->mHal.info.exportedPragmaValueList = 0;
812
813    // Bug, need to stash in metadata
814    if (mRootExpand) {
815        script->mHal.info.root = mRootExpand;
816    } else {
817        script->mHal.info.root = mRoot;
818    }
819#endif
820}
821
822
823typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
824
825void RsdCpuScriptImpl::forEachMtlsSetup(const Allocation ** ains,
826                                        uint32_t inLen,
827                                        Allocation * aout,
828                                        const void * usr, uint32_t usrLen,
829                                        const RsScriptCall *sc,
830                                        MTLaunchStruct *mtls) {
831
832    memset(mtls, 0, sizeof(MTLaunchStruct));
833
834    for (int index = inLen; --index >= 0;) {
835        const Allocation* ain = ains[index];
836
837        // possible for this to occur if IO_OUTPUT/IO_INPUT with no bound surface
838        if (ain != nullptr &&
839            (const uint8_t *)ain->mHal.drvState.lod[0].mallocPtr == nullptr) {
840
841            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
842                                         "rsForEach called with null in allocations");
843            return;
844        }
845    }
846
847    if (aout &&
848        (const uint8_t *)aout->mHal.drvState.lod[0].mallocPtr == nullptr) {
849
850        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
851                                     "rsForEach called with null out allocations");
852        return;
853    }
854
855    if (inLen > 0) {
856        const Allocation *ain0   = ains[0];
857        const Type       *inType = ain0->getType();
858
859        mtls->fep.dimX = inType->getDimX();
860        mtls->fep.dimY = inType->getDimY();
861        mtls->fep.dimZ = inType->getDimZ();
862
863        for (int Index = inLen; --Index >= 1;) {
864            if (!ain0->hasSameDims(ains[Index])) {
865                mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
866                  "Failed to launch kernel; dimensions of input and output allocations do not match.");
867
868                return;
869            }
870        }
871
872    } else if (aout != nullptr) {
873        const Type *outType = aout->getType();
874
875        mtls->fep.dimX = outType->getDimX();
876        mtls->fep.dimY = outType->getDimY();
877        mtls->fep.dimZ = outType->getDimZ();
878
879    } else {
880        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
881                                     "rsForEach called with null allocations");
882        return;
883    }
884
885    if (inLen > 0 && aout != nullptr) {
886        if (!ains[0]->hasSameDims(aout)) {
887            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
888              "Failed to launch kernel; dimensions of input and output allocations do not match.");
889
890            return;
891        }
892    }
893
894    if (!sc || (sc->xEnd == 0)) {
895        mtls->xEnd = mtls->fep.dimX;
896    } else {
897        rsAssert(sc->xStart < mtls->fep.dimX);
898        rsAssert(sc->xEnd <= mtls->fep.dimX);
899        rsAssert(sc->xStart < sc->xEnd);
900        mtls->xStart = rsMin(mtls->fep.dimX, sc->xStart);
901        mtls->xEnd = rsMin(mtls->fep.dimX, sc->xEnd);
902        if (mtls->xStart >= mtls->xEnd) return;
903    }
904
905    if (!sc || (sc->yEnd == 0)) {
906        mtls->yEnd = mtls->fep.dimY;
907    } else {
908        rsAssert(sc->yStart < mtls->fep.dimY);
909        rsAssert(sc->yEnd <= mtls->fep.dimY);
910        rsAssert(sc->yStart < sc->yEnd);
911        mtls->yStart = rsMin(mtls->fep.dimY, sc->yStart);
912        mtls->yEnd = rsMin(mtls->fep.dimY, sc->yEnd);
913        if (mtls->yStart >= mtls->yEnd) return;
914    }
915
916    if (!sc || (sc->zEnd == 0)) {
917        mtls->zEnd = mtls->fep.dimZ;
918    } else {
919        rsAssert(sc->zStart < mtls->fep.dimZ);
920        rsAssert(sc->zEnd <= mtls->fep.dimZ);
921        rsAssert(sc->zStart < sc->zEnd);
922        mtls->zStart = rsMin(mtls->fep.dimZ, sc->zStart);
923        mtls->zEnd = rsMin(mtls->fep.dimZ, sc->zEnd);
924        if (mtls->zStart >= mtls->zEnd) return;
925    }
926
927    mtls->xEnd     = rsMax((uint32_t)1, mtls->xEnd);
928    mtls->yEnd     = rsMax((uint32_t)1, mtls->yEnd);
929    mtls->zEnd     = rsMax((uint32_t)1, mtls->zEnd);
930    mtls->arrayEnd = rsMax((uint32_t)1, mtls->arrayEnd);
931
932    rsAssert(inLen == 0 || (ains[0]->getType()->getDimZ() == 0));
933
934    mtls->rsc        = mCtx;
935    mtls->ains       = ains;
936    mtls->aout       = aout;
937    mtls->fep.usr    = usr;
938    mtls->fep.usrLen = usrLen;
939    mtls->mSliceSize = 1;
940    mtls->mSliceNum  = 0;
941
942    mtls->fep.inPtrs    = nullptr;
943    mtls->fep.inStrides = nullptr;
944    mtls->isThreadable  = mIsThreadable;
945
946    if (inLen > 0) {
947
948        if (inLen <= RS_KERNEL_INPUT_THRESHOLD) {
949            mtls->fep.inPtrs    = (const uint8_t**)mtls->inPtrsBuff;
950            mtls->fep.inStrides = mtls->inStridesBuff;
951        } else {
952            mtls->fep.heapAllocatedArrays = true;
953
954            mtls->fep.inPtrs    = new const uint8_t*[inLen];
955            mtls->fep.inStrides = new StridePair[inLen];
956        }
957
958        mtls->fep.inLen = inLen;
959
960        for (int index = inLen; --index >= 0;) {
961            const Allocation *ain = ains[index];
962
963            mtls->fep.inPtrs[index] =
964              (const uint8_t*)ain->mHal.drvState.lod[0].mallocPtr;
965
966            mtls->fep.inStrides[index].eStride =
967              ain->getType()->getElementSizeBytes();
968            mtls->fep.inStrides[index].yStride =
969              ain->mHal.drvState.lod[0].stride;
970        }
971    }
972
973    mtls->fep.outPtr            = nullptr;
974    mtls->fep.outStride.eStride = 0;
975    mtls->fep.outStride.yStride = 0;
976    if (aout != nullptr) {
977        mtls->fep.outPtr = (uint8_t *)aout->mHal.drvState.lod[0].mallocPtr;
978
979        mtls->fep.outStride.eStride = aout->getType()->getElementSizeBytes();
980        mtls->fep.outStride.yStride = aout->mHal.drvState.lod[0].stride;
981    }
982}
983
984
985void RsdCpuScriptImpl::invokeForEach(uint32_t slot,
986                                     const Allocation ** ains,
987                                     uint32_t inLen,
988                                     Allocation * aout,
989                                     const void * usr,
990                                     uint32_t usrLen,
991                                     const RsScriptCall *sc) {
992
993    MTLaunchStruct mtls;
994
995    forEachMtlsSetup(ains, inLen, aout, usr, usrLen, sc, &mtls);
996    forEachKernelSetup(slot, &mtls);
997
998    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
999    mCtx->launchThreads(ains, inLen, aout, sc, &mtls);
1000    mCtx->setTLS(oldTLS);
1001}
1002
1003void RsdCpuScriptImpl::forEachKernelSetup(uint32_t slot, MTLaunchStruct *mtls) {
1004    mtls->script = this;
1005    mtls->fep.slot = slot;
1006#ifndef RS_COMPATIBILITY_LIB
1007    rsAssert(slot < mExecutable->getExportForeachFuncAddrs().size());
1008    mtls->kernel = reinterpret_cast<ForEachFunc_t>(
1009                      mExecutable->getExportForeachFuncAddrs()[slot]);
1010    rsAssert(mtls->kernel != nullptr);
1011    mtls->sig = mExecutable->getInfo().getExportForeachFuncs()[slot].second;
1012#else
1013    mtls->kernel = reinterpret_cast<ForEachFunc_t>(mForEachFunctions[slot]);
1014    rsAssert(mtls->kernel != nullptr);
1015    mtls->sig = mForEachSignatures[slot];
1016#endif
1017}
1018
1019int RsdCpuScriptImpl::invokeRoot() {
1020    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
1021    int ret = mRoot();
1022    mCtx->setTLS(oldTLS);
1023    return ret;
1024}
1025
1026void RsdCpuScriptImpl::invokeInit() {
1027    if (mInit) {
1028        mInit();
1029    }
1030}
1031
1032void RsdCpuScriptImpl::invokeFreeChildren() {
1033    if (mFreeChildren) {
1034        mFreeChildren();
1035    }
1036}
1037
1038void RsdCpuScriptImpl::invokeFunction(uint32_t slot, const void *params,
1039                                      size_t paramLength) {
1040    //ALOGE("invoke %p %p %i %p %i", dc, script, slot, params, paramLength);
1041
1042    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
1043    reinterpret_cast<void (*)(const void *, uint32_t)>(
1044#ifndef RS_COMPATIBILITY_LIB
1045        mExecutable->getExportFuncAddrs()[slot])(params, paramLength);
1046#else
1047        mInvokeFunctions[slot])(params, paramLength);
1048#endif
1049    mCtx->setTLS(oldTLS);
1050}
1051
1052void RsdCpuScriptImpl::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
1053    //rsAssert(!script->mFieldIsObject[slot]);
1054    //ALOGE("setGlobalVar %p %p %i %p %i", dc, script, slot, data, dataLength);
1055
1056    //if (mIntrinsicID) {
1057        //mIntrinsicFuncs.setVar(dc, script, drv->mIntrinsicData, slot, data, dataLength);
1058        //return;
1059    //}
1060
1061#ifndef RS_COMPATIBILITY_LIB
1062    int32_t *destPtr = reinterpret_cast<int32_t *>(
1063                          mExecutable->getExportVarAddrs()[slot]);
1064#else
1065    int32_t *destPtr = reinterpret_cast<int32_t *>(mFieldAddress[slot]);
1066#endif
1067    if (!destPtr) {
1068        //ALOGV("Calling setVar on slot = %i which is null", slot);
1069        return;
1070    }
1071
1072    memcpy(destPtr, data, dataLength);
1073}
1074
1075void RsdCpuScriptImpl::getGlobalVar(uint32_t slot, void *data, size_t dataLength) {
1076    //rsAssert(!script->mFieldIsObject[slot]);
1077    //ALOGE("getGlobalVar %p %p %i %p %i", dc, script, slot, data, dataLength);
1078
1079#ifndef RS_COMPATIBILITY_LIB
1080    int32_t *srcPtr = reinterpret_cast<int32_t *>(
1081                          mExecutable->getExportVarAddrs()[slot]);
1082#else
1083    int32_t *srcPtr = reinterpret_cast<int32_t *>(mFieldAddress[slot]);
1084#endif
1085    if (!srcPtr) {
1086        //ALOGV("Calling setVar on slot = %i which is null", slot);
1087        return;
1088    }
1089    memcpy(data, srcPtr, dataLength);
1090}
1091
1092
1093void RsdCpuScriptImpl::setGlobalVarWithElemDims(uint32_t slot, const void *data, size_t dataLength,
1094                                                const Element *elem,
1095                                                const uint32_t *dims, size_t dimLength) {
1096
1097#ifndef RS_COMPATIBILITY_LIB
1098    int32_t *destPtr = reinterpret_cast<int32_t *>(
1099        mExecutable->getExportVarAddrs()[slot]);
1100#else
1101    int32_t *destPtr = reinterpret_cast<int32_t *>(mFieldAddress[slot]);
1102#endif
1103    if (!destPtr) {
1104        //ALOGV("Calling setVar on slot = %i which is null", slot);
1105        return;
1106    }
1107
1108    // We want to look at dimension in terms of integer components,
1109    // but dimLength is given in terms of bytes.
1110    dimLength /= sizeof(int);
1111
1112    // Only a single dimension is currently supported.
1113    rsAssert(dimLength == 1);
1114    if (dimLength == 1) {
1115        // First do the increment loop.
1116        size_t stride = elem->getSizeBytes();
1117        const char *cVal = reinterpret_cast<const char *>(data);
1118        for (uint32_t i = 0; i < dims[0]; i++) {
1119            elem->incRefs(cVal);
1120            cVal += stride;
1121        }
1122
1123        // Decrement loop comes after (to prevent race conditions).
1124        char *oldVal = reinterpret_cast<char *>(destPtr);
1125        for (uint32_t i = 0; i < dims[0]; i++) {
1126            elem->decRefs(oldVal);
1127            oldVal += stride;
1128        }
1129    }
1130
1131    memcpy(destPtr, data, dataLength);
1132}
1133
1134void RsdCpuScriptImpl::setGlobalBind(uint32_t slot, Allocation *data) {
1135
1136    //rsAssert(!script->mFieldIsObject[slot]);
1137    //ALOGE("setGlobalBind %p %p %i %p", dc, script, slot, data);
1138
1139#ifndef RS_COMPATIBILITY_LIB
1140    int32_t *destPtr = reinterpret_cast<int32_t *>(
1141                          mExecutable->getExportVarAddrs()[slot]);
1142#else
1143    int32_t *destPtr = reinterpret_cast<int32_t *>(mFieldAddress[slot]);
1144#endif
1145    if (!destPtr) {
1146        //ALOGV("Calling setVar on slot = %i which is null", slot);
1147        return;
1148    }
1149
1150    void *ptr = nullptr;
1151    mBoundAllocs[slot] = data;
1152    if(data) {
1153        ptr = data->mHal.drvState.lod[0].mallocPtr;
1154    }
1155    memcpy(destPtr, &ptr, sizeof(void *));
1156}
1157
1158void RsdCpuScriptImpl::setGlobalObj(uint32_t slot, ObjectBase *data) {
1159
1160    //rsAssert(script->mFieldIsObject[slot]);
1161    //ALOGE("setGlobalObj %p %p %i %p", dc, script, slot, data);
1162
1163#ifndef RS_COMPATIBILITY_LIB
1164    int32_t *destPtr = reinterpret_cast<int32_t *>(
1165                          mExecutable->getExportVarAddrs()[slot]);
1166#else
1167    int32_t *destPtr = reinterpret_cast<int32_t *>(mFieldAddress[slot]);
1168#endif
1169
1170    if (!destPtr) {
1171        //ALOGV("Calling setVar on slot = %i which is null", slot);
1172        return;
1173    }
1174
1175    rsrSetObject(mCtx->getContext(), (rs_object_base *)destPtr, data);
1176}
1177
1178RsdCpuScriptImpl::~RsdCpuScriptImpl() {
1179#ifndef RS_COMPATIBILITY_LIB
1180    if (mExecutable) {
1181        std::vector<void *>::const_iterator var_addr_iter =
1182            mExecutable->getExportVarAddrs().begin();
1183        std::vector<void *>::const_iterator var_addr_end =
1184            mExecutable->getExportVarAddrs().end();
1185
1186        bcc::RSInfo::ObjectSlotListTy::const_iterator is_object_iter =
1187            mExecutable->getInfo().getObjectSlots().begin();
1188        bcc::RSInfo::ObjectSlotListTy::const_iterator is_object_end =
1189            mExecutable->getInfo().getObjectSlots().end();
1190
1191        while ((var_addr_iter != var_addr_end) &&
1192               (is_object_iter != is_object_end)) {
1193            // The field address can be nullptr if the script-side has optimized
1194            // the corresponding global variable away.
1195            rs_object_base *obj_addr =
1196                reinterpret_cast<rs_object_base *>(*var_addr_iter);
1197            if (*is_object_iter) {
1198                if (*var_addr_iter != nullptr && mCtx->getContext() != nullptr) {
1199                    rsrClearObject(mCtx->getContext(), obj_addr);
1200                }
1201            }
1202            var_addr_iter++;
1203            is_object_iter++;
1204        }
1205    }
1206
1207    if (mCompilerContext) {
1208        delete mCompilerContext;
1209    }
1210    if (mCompilerDriver) {
1211        delete mCompilerDriver;
1212    }
1213    if (mExecutable) {
1214        delete mExecutable;
1215    }
1216    if (mBoundAllocs) {
1217        delete[] mBoundAllocs;
1218    }
1219
1220    for (size_t i = 0; i < mExportedForEachFuncList.size(); i++) {
1221        delete[] mExportedForEachFuncList[i].first;
1222    }
1223#else
1224    if (mFieldIsObject) {
1225        for (size_t i = 0; i < mExportedVariableCount; ++i) {
1226            if (mFieldIsObject[i]) {
1227                if (mFieldAddress[i] != nullptr) {
1228                    rs_object_base *obj_addr =
1229                        reinterpret_cast<rs_object_base *>(mFieldAddress[i]);
1230                    rsrClearObject(mCtx->getContext(), obj_addr);
1231                }
1232            }
1233        }
1234    }
1235
1236    if (mInvokeFunctions) delete[] mInvokeFunctions;
1237    if (mForEachFunctions) delete[] mForEachFunctions;
1238    if (mFieldAddress) delete[] mFieldAddress;
1239    if (mFieldIsObject) delete[] mFieldIsObject;
1240    if (mForEachSignatures) delete[] mForEachSignatures;
1241    if (mBoundAllocs) delete[] mBoundAllocs;
1242    if (mScriptSO) {
1243        dlclose(mScriptSO);
1244    }
1245#endif
1246}
1247
1248Allocation * RsdCpuScriptImpl::getAllocationForPointer(const void *ptr) const {
1249    if (!ptr) {
1250        return nullptr;
1251    }
1252
1253    for (uint32_t ct=0; ct < mScript->mHal.info.exportedVariableCount; ct++) {
1254        Allocation *a = mBoundAllocs[ct];
1255        if (!a) continue;
1256        if (a->mHal.drvState.lod[0].mallocPtr == ptr) {
1257            return a;
1258        }
1259    }
1260    ALOGE("rsGetAllocation, failed to find %p", ptr);
1261    return nullptr;
1262}
1263
1264void RsdCpuScriptImpl::preLaunch(uint32_t slot, const Allocation ** ains,
1265                                 uint32_t inLen, Allocation * aout,
1266                                 const void * usr, uint32_t usrLen,
1267                                 const RsScriptCall *sc) {}
1268
1269void RsdCpuScriptImpl::postLaunch(uint32_t slot, const Allocation ** ains,
1270                                  uint32_t inLen, Allocation * aout,
1271                                  const void * usr, uint32_t usrLen,
1272                                  const RsScriptCall *sc) {}
1273
1274
1275}
1276}
1277