rsCpuScript.cpp revision 68173de001c5d41c74828a426651e07cfd9e8710 
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#define THREADABLE_STR "isThreadable: "
440
441// Copy up to a newline or size chars from str -> s, updating str
442// Returns s when successful and nullptr when '\0' is finally reached.
443static char* strgets(char *s, int size, const char **ppstr) {
444    if (!ppstr || !*ppstr || **ppstr == '\0' || size < 1) {
445        return nullptr;
446    }
447
448    int i;
449    for (i = 0; i < (size - 1); i++) {
450        s[i] = **ppstr;
451        (*ppstr)++;
452        if (s[i] == '\0') {
453            return s;
454        } else if (s[i] == '\n') {
455            s[i+1] = '\0';
456            return s;
457        }
458    }
459
460    // size has been exceeded.
461    s[i] = '\0';
462
463    return s;
464}
465
466RsdCpuScriptImpl::RsdCpuScriptImpl(RsdCpuReferenceImpl *ctx, const Script *s) {
467    mCtx = ctx;
468    mScript = s;
469
470    mScriptSO = nullptr;
471
472#ifndef RS_COMPATIBILITY_LIB
473    mCompilerDriver = nullptr;
474#endif
475
476
477    mRoot = nullptr;
478    mRootExpand = nullptr;
479    mInit = nullptr;
480    mFreeChildren = nullptr;
481    mScriptExec = nullptr;
482
483    mBoundAllocs = nullptr;
484    mIntrinsicData = nullptr;
485    mIsThreadable = true;
486}
487
488bool RsdCpuScriptImpl::storeRSInfoFromSO() {
489    mRoot = (RootFunc_t) dlsym(mScriptSO, "root");
490    if (mRoot) {
491        //ALOGE("Found root(): %p", mRoot);
492    }
493    mRootExpand = (RootFunc_t) dlsym(mScriptSO, "root.expand");
494    if (mRootExpand) {
495        //ALOGE("Found root.expand(): %p", mRootExpand);
496    }
497    mInit = (InvokeFunc_t) dlsym(mScriptSO, "init");
498    if (mInit) {
499        //ALOGE("Found init(): %p", mInit);
500    }
501    mFreeChildren = (InvokeFunc_t) dlsym(mScriptSO, ".rs.dtor");
502    if (mFreeChildren) {
503        //ALOGE("Found .rs.dtor(): %p", mFreeChildren);
504    }
505
506    mScriptExec = ScriptExecutable::createFromSharedObject(
507            mCtx->getContext(), mScriptSO);
508
509    if (mScriptExec == nullptr) {
510        return false;
511    }
512
513    size_t varCount = mScriptExec->getExportedVariableCount();
514    if (varCount > 0) {
515        mBoundAllocs = new Allocation *[varCount];
516        memset(mBoundAllocs, 0, varCount * sizeof(*mBoundAllocs));
517    }
518
519    mIsThreadable = mScriptExec->getThreadable();
520    //ALOGE("Script isThreadable? %d", mIsThreadable);
521
522    return true;
523}
524
525ScriptExecutable* ScriptExecutable::createFromSharedObject(
526    Context* RSContext, void* sharedObj) {
527    char line[MAXLINE];
528
529    size_t varCount = 0;
530    size_t funcCount = 0;
531    size_t forEachCount = 0;
532    size_t objectSlotCount = 0;
533    size_t pragmaCount = 0;
534    bool isThreadable = true;
535
536    const char *rsInfo = (const char *) dlsym(sharedObj, ".rs.info");
537
538    if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
539        return nullptr;
540    }
541    if (sscanf(line, EXPORT_VAR_STR "%zu", &varCount) != 1) {
542        ALOGE("Invalid export var count!: %s", line);
543        return nullptr;
544    }
545
546    std::vector<void*> fieldAddress;
547
548    for (size_t i = 0; i < varCount; ++i) {
549        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
550            return nullptr;
551        }
552        char *c = strrchr(line, '\n');
553        if (c) {
554            *c = '\0';
555        }
556        void* addr = dlsym(sharedObj, line);
557        if (addr == nullptr) {
558            ALOGE("Failed to find variable address for %s: %s",
559                  line, dlerror());
560            // Not a critical error if we don't find a global variable.
561        }
562        fieldAddress.push_back(addr);
563    }
564
565    if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
566        return nullptr;
567    }
568    if (sscanf(line, EXPORT_FUNC_STR "%zu", &funcCount) != 1) {
569        ALOGE("Invalid export func count!: %s", line);
570        return nullptr;
571    }
572
573    std::vector<InvokeFunc_t> invokeFunctions(funcCount);
574
575    for (size_t i = 0; i < funcCount; ++i) {
576        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
577            return nullptr ;
578        }
579        char *c = strrchr(line, '\n');
580        if (c) {
581            *c = '\0';
582        }
583
584        invokeFunctions[i] = (InvokeFunc_t) dlsym(sharedObj, line);
585        if (invokeFunctions[i] == nullptr) {
586            ALOGE("Failed to get function address for %s(): %s",
587                  line, dlerror());
588            return nullptr;
589        }
590    }
591
592    if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
593        return nullptr;
594    }
595    if (sscanf(line, EXPORT_FOREACH_STR "%zu", &forEachCount) != 1) {
596        ALOGE("Invalid export forEach count!: %s", line);
597        return nullptr;
598    }
599
600    std::vector<ForEachFunc_t> forEachFunctions(forEachCount);
601    std::vector<uint32_t> forEachSignatures(forEachCount);
602
603    for (size_t i = 0; i < forEachCount; ++i) {
604        unsigned int tmpSig = 0;
605        char tmpName[MAXLINE];
606
607        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
608            return nullptr;
609        }
610        if (sscanf(line, "%u - %" MAKE_STR(MAXLINE) "s",
611                   &tmpSig, tmpName) != 2) {
612          ALOGE("Invalid export forEach!: %s", line);
613          return nullptr;
614        }
615
616        // Lookup the expanded ForEach kernel.
617        strncat(tmpName, ".expand", MAXLINE-1-strlen(tmpName));
618        forEachSignatures[i] = tmpSig;
619        forEachFunctions[i] =
620            (ForEachFunc_t) dlsym(sharedObj, tmpName);
621        if (i != 0 && forEachFunctions[i] == nullptr) {
622            // Ignore missing root.expand functions.
623            // root() is always specified at location 0.
624            ALOGE("Failed to find forEach function address for %s: %s",
625                  tmpName, dlerror());
626            return nullptr;
627        }
628    }
629
630    if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
631        return nullptr;
632    }
633    if (sscanf(line, OBJECT_SLOT_STR "%zu", &objectSlotCount) != 1) {
634        ALOGE("Invalid object slot count!: %s", line);
635        return nullptr;
636    }
637
638    std::vector<bool> fieldIsObject(varCount, false);
639
640    rsAssert(varCount > 0);
641    for (size_t i = 0; i < objectSlotCount; ++i) {
642        uint32_t varNum = 0;
643        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
644            return nullptr;
645        }
646        if (sscanf(line, "%u", &varNum) != 1) {
647            ALOGE("Invalid object slot!: %s", line);
648            return nullptr;
649        }
650
651        if (varNum < varCount) {
652            fieldIsObject[varNum] = true;
653        }
654    }
655
656#ifdef RS_COMPATIBILITY_LIB
657    // Do not attempt to read pragmas or isThreadable flag in compat lib path.
658    // Neither is applicable for compat lib
659    std::vector<const char *> pragmaKeys(pragmaCount);
660    std::vector<const char *> pragmaValues(pragmaCount);
661
662    isThreadable = true;
663
664#else
665    if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
666        return nullptr;
667    }
668
669    if (sscanf(line, PRAGMA_STR "%zu", &pragmaCount) != 1) {
670        ALOGE("Invalid pragma count!: %s", line);
671        return nullptr;
672    }
673
674    std::vector<const char *> pragmaKeys(pragmaCount);
675    std::vector<const char *> pragmaValues(pragmaCount);
676
677    for (size_t i = 0; i < pragmaCount; ++i) {
678        if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
679            ALOGE("Unable to read pragma at index %zu!", i);
680            return nullptr;
681        }
682
683        char key[MAXLINE];
684        char value[MAXLINE] = ""; // initialize in case value is empty
685
686        // pragmas can just have a key and no value.  Only check to make sure
687        // that the key is not empty
688        if (sscanf(line, "%" MAKE_STR(MAXLINE) "s - %" MAKE_STR(MAXLINE) "s",
689                   key, value) == 0 ||
690            strlen(key) == 0)
691        {
692            ALOGE("Invalid pragma value!: %s", line);
693
694            // free previously allocated keys and values
695            for (size_t idx = 0; idx < i; ++idx) {
696                delete [] pragmaKeys[idx];
697                delete [] pragmaValues[idx];
698            }
699            return nullptr;
700        }
701
702        char *pKey = new char[strlen(key)+1];
703        strcpy(pKey, key);
704        pragmaKeys[i] = pKey;
705
706        char *pValue = new char[strlen(value)+1];
707        strcpy(pValue, value);
708        pragmaValues[i] = pValue;
709        //ALOGE("Pragma %zu: Key: '%s' Value: '%s'", i, pKey, pValue);
710    }
711
712    if (strgets(line, MAXLINE, &rsInfo) == nullptr) {
713        return nullptr;
714    }
715
716    char tmpFlag[4];
717    if (sscanf(line, THREADABLE_STR "%4s", tmpFlag) != 1) {
718        ALOGE("Invalid threadable flag!: %s", line);
719        return nullptr;
720    }
721    if (strcmp(tmpFlag, "yes") == 0)
722        isThreadable = true;
723    else if (strcmp(tmpFlag, "no") == 0)
724        isThreadable = false;
725    else {
726        ALOGE("Invalid threadable flag!: %s", tmpFlag);
727        return nullptr;
728    }
729
730#endif
731
732    return new ScriptExecutable(
733        RSContext, fieldAddress, fieldIsObject, invokeFunctions,
734        forEachFunctions, forEachSignatures, pragmaKeys, pragmaValues,
735        isThreadable);
736}
737
738bool RsdCpuScriptImpl::init(char const *resName, char const *cacheDir,
739                            uint8_t const *bitcode, size_t bitcodeSize,
740                            uint32_t flags, char const *bccPluginName) {
741    //ALOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir, bitcode, bitcodeSize, flags, lookupFunc);
742    //ALOGE("rsdScriptInit %p %p", rsc, script);
743
744    mCtx->lockMutex();
745#ifndef RS_COMPATIBILITY_LIB
746    bool useRSDebugContext = false;
747
748    mCompilerDriver = nullptr;
749
750    mCompilerDriver = new bcc::RSCompilerDriver();
751    if (mCompilerDriver == nullptr) {
752        ALOGE("bcc: FAILS to create compiler driver (out of memory)");
753        mCtx->unlockMutex();
754        return false;
755    }
756
757    // Run any compiler setup functions we have been provided with.
758    RSSetupCompilerCallback setupCompilerCallback =
759            mCtx->getSetupCompilerCallback();
760    if (setupCompilerCallback != nullptr) {
761        setupCompilerCallback(mCompilerDriver);
762    }
763
764    bcinfo::MetadataExtractor bitcodeMetadata((const char *) bitcode, bitcodeSize);
765    if (!bitcodeMetadata.extract()) {
766        ALOGE("Could not extract metadata from bitcode");
767        mCtx->unlockMutex();
768        return false;
769    }
770
771    const char* core_lib = findCoreLib(bitcodeMetadata, (const char*)bitcode, bitcodeSize);
772
773    if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) {
774        mCompilerDriver->setDebugContext(true);
775        useRSDebugContext = true;
776    }
777
778    std::string bcFileName(cacheDir);
779    bcFileName.append("/");
780    bcFileName.append(resName);
781    bcFileName.append(".bc");
782
783    std::vector<const char*> compileArguments;
784    setCompileArguments(&compileArguments, bcFileName, cacheDir, resName, core_lib,
785                        useRSDebugContext, bccPluginName);
786    // The last argument of compileArguments ia a nullptr, so remove 1 from the size.
787    std::string compileCommandLine =
788                bcc::getCommandLine(compileArguments.size() - 1, compileArguments.data());
789
790    if (!is_force_recompile() && !useRSDebugContext) {
791        mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName);
792    }
793
794    // If we can't, it's either not there or out of date.  We compile the bit code and try loading
795    // again.
796    if (mScriptSO == nullptr) {
797        if (!compileBitcode(bcFileName, (const char*)bitcode, bitcodeSize,
798                            compileArguments.data(), compileCommandLine))
799        {
800            ALOGE("bcc: FAILS to compile '%s'", resName);
801            mCtx->unlockMutex();
802            return false;
803        }
804
805        if (!SharedLibraryUtils::createSharedLibrary(cacheDir, resName)) {
806            ALOGE("Linker: Failed to link object file '%s'", resName);
807            mCtx->unlockMutex();
808            return false;
809        }
810
811        mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName);
812        if (mScriptSO == nullptr) {
813            ALOGE("Unable to load '%s'", resName);
814            mCtx->unlockMutex();
815            return false;
816        }
817    }
818
819    // Read RS symbol information from the .so.
820    if ( !mScriptSO) {
821        goto error;
822    }
823
824    if ( !storeRSInfoFromSO()) {
825      goto error;
826    }
827#else  // RS_COMPATIBILITY_LIB is defined
828
829    mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName);
830
831    if (!mScriptSO) {
832        goto error;
833    }
834
835    if (!storeRSInfoFromSO()) {
836        goto error;
837    }
838#endif
839    mCtx->unlockMutex();
840    return true;
841
842error:
843
844    mCtx->unlockMutex();
845    if (mScriptSO) {
846        dlclose(mScriptSO);
847    }
848    return false;
849}
850
851#ifndef RS_COMPATIBILITY_LIB
852
853const char* RsdCpuScriptImpl::findCoreLib(const bcinfo::MetadataExtractor& ME, const char* bitcode,
854                                          size_t bitcodeSize) {
855    const char* defaultLib = SYSLIBPATH"/libclcore.bc";
856
857    // If we're debugging, use the debug library.
858    if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) {
859        return SYSLIBPATH"/libclcore_debug.bc";
860    }
861
862    // If a callback has been registered to specify a library, use that.
863    RSSelectRTCallback selectRTCallback = mCtx->getSelectRTCallback();
864    if (selectRTCallback != nullptr) {
865        return selectRTCallback((const char*)bitcode, bitcodeSize);
866    }
867
868    // Check for a platform specific library
869#if defined(ARCH_ARM_HAVE_NEON) && !defined(DISABLE_CLCORE_NEON)
870    enum bcinfo::RSFloatPrecision prec = ME.getRSFloatPrecision();
871    if (prec == bcinfo::RS_FP_Relaxed) {
872        // NEON-capable ARMv7a devices can use an accelerated math library
873        // for all reduced precision scripts.
874        // ARMv8 does not use NEON, as ASIMD can be used with all precision
875        // levels.
876        return SYSLIBPATH"/libclcore_neon.bc";
877    } else {
878        return defaultLib;
879    }
880#elif defined(__i386__) || defined(__x86_64__)
881    // x86 devices will use an optimized library.
882    return SYSLIBPATH"/libclcore_x86.bc";
883#else
884    return defaultLib;
885#endif
886}
887
888#endif
889
890void RsdCpuScriptImpl::populateScript(Script *script) {
891    // Copy info over to runtime
892    script->mHal.info.exportedFunctionCount = mScriptExec->getExportedFunctionCount();
893    script->mHal.info.exportedVariableCount = mScriptExec->getExportedVariableCount();
894    script->mHal.info.exportedPragmaCount = mScriptExec->getPragmaCount();;
895    script->mHal.info.exportedPragmaKeyList =
896        const_cast<const char**>(&mScriptExec->getPragmaKeys().front());
897    script->mHal.info.exportedPragmaValueList =
898        const_cast<const char**>(&mScriptExec->getPragmaValues().front());
899
900    // Bug, need to stash in metadata
901    if (mRootExpand) {
902        script->mHal.info.root = mRootExpand;
903    } else {
904        script->mHal.info.root = mRoot;
905    }
906}
907
908
909typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
910
911bool RsdCpuScriptImpl::forEachMtlsSetup(const Allocation ** ains,
912                                        uint32_t inLen,
913                                        Allocation * aout,
914                                        const void * usr, uint32_t usrLen,
915                                        const RsScriptCall *sc,
916                                        MTLaunchStruct *mtls) {
917
918    memset(mtls, 0, sizeof(MTLaunchStruct));
919
920    for (int index = inLen; --index >= 0;) {
921        const Allocation* ain = ains[index];
922
923        // possible for this to occur if IO_OUTPUT/IO_INPUT with no bound surface
924        if (ain != nullptr &&
925            (const uint8_t *)ain->mHal.drvState.lod[0].mallocPtr == nullptr) {
926
927            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
928                                         "rsForEach called with null in allocations");
929            return false;
930        }
931    }
932
933    if (aout &&
934        (const uint8_t *)aout->mHal.drvState.lod[0].mallocPtr == nullptr) {
935
936        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
937                                     "rsForEach called with null out allocations");
938        return false;
939    }
940
941    if (inLen > 0) {
942        const Allocation *ain0   = ains[0];
943        const Type       *inType = ain0->getType();
944
945        mtls->fep.dim.x = inType->getDimX();
946        mtls->fep.dim.y = inType->getDimY();
947        mtls->fep.dim.z = inType->getDimZ();
948
949        for (int Index = inLen; --Index >= 1;) {
950            if (!ain0->hasSameDims(ains[Index])) {
951                mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
952                  "Failed to launch kernel; dimensions of input and output allocations do not match.");
953
954                return false;
955            }
956        }
957
958    } else if (aout != nullptr) {
959        const Type *outType = aout->getType();
960
961        mtls->fep.dim.x = outType->getDimX();
962        mtls->fep.dim.y = outType->getDimY();
963        mtls->fep.dim.z = outType->getDimZ();
964
965    } else {
966        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
967                                     "rsForEach called with null allocations");
968        return false;
969    }
970
971    if (inLen > 0 && aout != nullptr) {
972        if (!ains[0]->hasSameDims(aout)) {
973            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
974              "Failed to launch kernel; dimensions of input and output allocations do not match.");
975
976            return false;
977        }
978    }
979
980    if (!sc || (sc->xEnd == 0)) {
981        mtls->end.x = mtls->fep.dim.x;
982    } else {
983        mtls->start.x = rsMin(mtls->fep.dim.x, sc->xStart);
984        mtls->end.x = rsMin(mtls->fep.dim.x, sc->xEnd);
985        if (mtls->start.x >= mtls->end.x) return false;
986    }
987
988    if (!sc || (sc->yEnd == 0)) {
989        mtls->end.y = mtls->fep.dim.y;
990    } else {
991        mtls->start.y = rsMin(mtls->fep.dim.y, sc->yStart);
992        mtls->end.y = rsMin(mtls->fep.dim.y, sc->yEnd);
993        if (mtls->start.y >= mtls->end.y) return false;
994    }
995
996    if (!sc || (sc->zEnd == 0)) {
997        mtls->end.z = mtls->fep.dim.z;
998    } else {
999        mtls->start.z = rsMin(mtls->fep.dim.z, sc->zStart);
1000        mtls->end.z = rsMin(mtls->fep.dim.z, sc->zEnd);
1001        if (mtls->start.z >= mtls->end.z) return false;
1002    }
1003
1004    if (!sc || (sc->arrayEnd == 0)) {
1005        mtls->end.array[0] = mtls->fep.dim.array[0];
1006    } else {
1007        mtls->start.array[0] = rsMin(mtls->fep.dim.array[0], sc->arrayStart);
1008        mtls->end.array[0] = rsMin(mtls->fep.dim.array[0], sc->arrayEnd);
1009        if (mtls->start.array[0] >= mtls->end.array[0]) return false;
1010    }
1011
1012    if (!sc || (sc->array2End == 0)) {
1013        mtls->end.array[1] = mtls->fep.dim.array[1];
1014    } else {
1015        mtls->start.array[1] = rsMin(mtls->fep.dim.array[1], sc->array2Start);
1016        mtls->end.array[1] = rsMin(mtls->fep.dim.array[1], sc->array2End);
1017        if (mtls->start.array[1] >= mtls->end.array[1]) return false;
1018    }
1019
1020    if (!sc || (sc->array3End == 0)) {
1021        mtls->end.array[2] = mtls->fep.dim.array[2];
1022    } else {
1023        mtls->start.array[2] = rsMin(mtls->fep.dim.array[2], sc->array3Start);
1024        mtls->end.array[2] = rsMin(mtls->fep.dim.array[2], sc->array3End);
1025        if (mtls->start.array[2] >= mtls->end.array[2]) return false;
1026    }
1027
1028    if (!sc || (sc->array4End == 0)) {
1029        mtls->end.array[3] = mtls->fep.dim.array[3];
1030    } else {
1031        mtls->start.array[3] = rsMin(mtls->fep.dim.array[3], sc->array4Start);
1032        mtls->end.array[3] = rsMin(mtls->fep.dim.array[3], sc->array4End);
1033        if (mtls->start.array[3] >= mtls->end.array[3]) return false;
1034    }
1035
1036
1037    // The X & Y walkers always want 0-1 min even if dim is not present
1038    mtls->end.x    = rsMax((uint32_t)1, mtls->end.x);
1039    mtls->end.y    = rsMax((uint32_t)1, mtls->end.y);
1040
1041    mtls->rsc        = mCtx;
1042    if (ains) {
1043        memcpy(mtls->ains, ains, inLen * sizeof(ains[0]));
1044    }
1045    mtls->aout[0]    = aout;
1046    mtls->fep.usr    = usr;
1047    mtls->fep.usrLen = usrLen;
1048    mtls->mSliceSize = 1;
1049    mtls->mSliceNum  = 0;
1050
1051    mtls->isThreadable  = mIsThreadable;
1052
1053    if (inLen > 0) {
1054        mtls->fep.inLen = inLen;
1055        for (int index = inLen; --index >= 0;) {
1056            mtls->fep.inPtr[index] = (const uint8_t*)ains[index]->mHal.drvState.lod[0].mallocPtr;
1057            mtls->fep.inStride[index] = ains[index]->getType()->getElementSizeBytes();
1058        }
1059    }
1060
1061    if (aout != nullptr) {
1062        mtls->fep.outPtr[0] = (uint8_t *)aout->mHal.drvState.lod[0].mallocPtr;
1063        mtls->fep.outStride[0] = aout->getType()->getElementSizeBytes();
1064    }
1065
1066    // All validation passed, ok to launch threads
1067    return true;
1068}
1069
1070
1071void RsdCpuScriptImpl::invokeForEach(uint32_t slot,
1072                                     const Allocation ** ains,
1073                                     uint32_t inLen,
1074                                     Allocation * aout,
1075                                     const void * usr,
1076                                     uint32_t usrLen,
1077                                     const RsScriptCall *sc) {
1078
1079    MTLaunchStruct mtls;
1080
1081    if (forEachMtlsSetup(ains, inLen, aout, usr, usrLen, sc, &mtls)) {
1082        forEachKernelSetup(slot, &mtls);
1083
1084        RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
1085        mCtx->launchThreads(ains, inLen, aout, sc, &mtls);
1086        mCtx->setTLS(oldTLS);
1087    }
1088}
1089
1090void RsdCpuScriptImpl::forEachKernelSetup(uint32_t slot, MTLaunchStruct *mtls) {
1091    mtls->script = this;
1092    mtls->fep.slot = slot;
1093    mtls->kernel = mScriptExec->getForEachFunction(slot);
1094    rsAssert(mtls->kernel != nullptr);
1095    mtls->sig = mScriptExec->getForEachSignature(slot);
1096}
1097
1098int RsdCpuScriptImpl::invokeRoot() {
1099    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
1100    int ret = mRoot();
1101    mCtx->setTLS(oldTLS);
1102    return ret;
1103}
1104
1105void RsdCpuScriptImpl::invokeInit() {
1106    if (mInit) {
1107        mInit();
1108    }
1109}
1110
1111void RsdCpuScriptImpl::invokeFreeChildren() {
1112    if (mFreeChildren) {
1113        mFreeChildren();
1114    }
1115}
1116
1117void RsdCpuScriptImpl::invokeFunction(uint32_t slot, const void *params,
1118                                      size_t paramLength) {
1119    //ALOGE("invoke %i %p %zu", slot, params, paramLength);
1120    void * ap = nullptr;
1121
1122#if defined(__x86_64__)
1123    // The invoked function could have input parameter of vector type for example float4 which
1124    // requires void* params to be 16 bytes aligned when using SSE instructions for x86_64 platform.
1125    // So try to align void* params before passing them into RS exported function.
1126
1127    if ((uint8_t)(uint64_t)params & 0x0F) {
1128        if ((ap = (void*)memalign(16, paramLength)) != nullptr) {
1129            memcpy(ap, params, paramLength);
1130        } else {
1131            ALOGE("x86_64: invokeFunction memalign error, still use params which is not 16 bytes aligned.");
1132        }
1133    }
1134#endif
1135
1136    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
1137    reinterpret_cast<void (*)(const void *, uint32_t)>(
1138        mScriptExec->getInvokeFunction(slot))(ap? (const void *) ap: params, paramLength);
1139
1140    mCtx->setTLS(oldTLS);
1141}
1142
1143void RsdCpuScriptImpl::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
1144    //rsAssert(!script->mFieldIsObject[slot]);
1145    //ALOGE("setGlobalVar %i %p %zu", slot, data, dataLength);
1146
1147    //if (mIntrinsicID) {
1148        //mIntrinsicFuncs.setVar(dc, script, drv->mIntrinsicData, slot, data, dataLength);
1149        //return;
1150    //}
1151
1152    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
1153    if (!destPtr) {
1154        //ALOGV("Calling setVar on slot = %i which is null", slot);
1155        return;
1156    }
1157
1158    memcpy(destPtr, data, dataLength);
1159}
1160
1161void RsdCpuScriptImpl::getGlobalVar(uint32_t slot, void *data, size_t dataLength) {
1162    //rsAssert(!script->mFieldIsObject[slot]);
1163    //ALOGE("getGlobalVar %i %p %zu", slot, data, dataLength);
1164
1165    int32_t *srcPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
1166    if (!srcPtr) {
1167        //ALOGV("Calling setVar on slot = %i which is null", slot);
1168        return;
1169    }
1170    memcpy(data, srcPtr, dataLength);
1171}
1172
1173
1174void RsdCpuScriptImpl::setGlobalVarWithElemDims(uint32_t slot, const void *data, size_t dataLength,
1175                                                const Element *elem,
1176                                                const uint32_t *dims, size_t dimLength) {
1177    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
1178    if (!destPtr) {
1179        //ALOGV("Calling setVar on slot = %i which is null", slot);
1180        return;
1181    }
1182
1183    // We want to look at dimension in terms of integer components,
1184    // but dimLength is given in terms of bytes.
1185    dimLength /= sizeof(int);
1186
1187    // Only a single dimension is currently supported.
1188    rsAssert(dimLength == 1);
1189    if (dimLength == 1) {
1190        // First do the increment loop.
1191        size_t stride = elem->getSizeBytes();
1192        const char *cVal = reinterpret_cast<const char *>(data);
1193        for (uint32_t i = 0; i < dims[0]; i++) {
1194            elem->incRefs(cVal);
1195            cVal += stride;
1196        }
1197
1198        // Decrement loop comes after (to prevent race conditions).
1199        char *oldVal = reinterpret_cast<char *>(destPtr);
1200        for (uint32_t i = 0; i < dims[0]; i++) {
1201            elem->decRefs(oldVal);
1202            oldVal += stride;
1203        }
1204    }
1205
1206    memcpy(destPtr, data, dataLength);
1207}
1208
1209void RsdCpuScriptImpl::setGlobalBind(uint32_t slot, Allocation *data) {
1210
1211    //rsAssert(!script->mFieldIsObject[slot]);
1212    //ALOGE("setGlobalBind %i %p", slot, data);
1213
1214    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
1215    if (!destPtr) {
1216        //ALOGV("Calling setVar on slot = %i which is null", slot);
1217        return;
1218    }
1219
1220    void *ptr = nullptr;
1221    mBoundAllocs[slot] = data;
1222    if (data) {
1223        ptr = data->mHal.drvState.lod[0].mallocPtr;
1224    }
1225    memcpy(destPtr, &ptr, sizeof(void *));
1226}
1227
1228void RsdCpuScriptImpl::setGlobalObj(uint32_t slot, ObjectBase *data) {
1229
1230    //rsAssert(script->mFieldIsObject[slot]);
1231    //ALOGE("setGlobalObj %i %p", slot, data);
1232
1233    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
1234    if (!destPtr) {
1235        //ALOGV("Calling setVar on slot = %i which is null", slot);
1236        return;
1237    }
1238
1239    rsrSetObject(mCtx->getContext(), (rs_object_base *)destPtr, data);
1240}
1241
1242RsdCpuScriptImpl::~RsdCpuScriptImpl() {
1243#ifndef RS_COMPATIBILITY_LIB
1244    if (mCompilerDriver) {
1245        delete mCompilerDriver;
1246    }
1247#endif
1248
1249    if (mScriptExec != nullptr) {
1250        delete mScriptExec;
1251    }
1252    if (mBoundAllocs) delete[] mBoundAllocs;
1253    if (mScriptSO) {
1254        dlclose(mScriptSO);
1255    }
1256}
1257
1258Allocation * RsdCpuScriptImpl::getAllocationForPointer(const void *ptr) const {
1259    if (!ptr) {
1260        return nullptr;
1261    }
1262
1263    for (uint32_t ct=0; ct < mScript->mHal.info.exportedVariableCount; ct++) {
1264        Allocation *a = mBoundAllocs[ct];
1265        if (!a) continue;
1266        if (a->mHal.drvState.lod[0].mallocPtr == ptr) {
1267            return a;
1268        }
1269    }
1270    ALOGE("rsGetAllocation, failed to find %p", ptr);
1271    return nullptr;
1272}
1273
1274void RsdCpuScriptImpl::preLaunch(uint32_t slot, const Allocation ** ains,
1275                                 uint32_t inLen, Allocation * aout,
1276                                 const void * usr, uint32_t usrLen,
1277                                 const RsScriptCall *sc) {}
1278
1279void RsdCpuScriptImpl::postLaunch(uint32_t slot, const Allocation ** ains,
1280                                  uint32_t inLen, Allocation * aout,
1281                                  const void * usr, uint32_t usrLen,
1282                                  const RsScriptCall *sc) {}
1283
1284
1285}
1286}
1287