rsCpuScript.cpp revision 14ce007a633b10e3b9a3fae29d8f53a7e8c9b59f 
1/*
2 * Copyright (C) 2011-2012 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include "rsCpuCore.h"
18#include "rsCpuScript.h"
19#include "rsCpuExecutable.h"
20
21#ifdef RS_COMPATIBILITY_LIB
22    #include <stdio.h>
23    #include <sys/stat.h>
24    #include <unistd.h>
25#else
26    #include "rsCppUtils.h"
27
28    #include <bcc/Config/Config.h>
29    #include <bcinfo/MetadataExtractor.h>
30    #include <cutils/properties.h>
31
32    #include <zlib.h>
33    #include <sys/file.h>
34    #include <sys/types.h>
35    #include <unistd.h>
36
37    #include <string>
38    #include <vector>
39#endif
40
41#include <set>
42#include <string>
43#include <dlfcn.h>
44#include <stdlib.h>
45#include <string.h>
46#include <iostream>
47#include <sstream>
48
49namespace {
50
51static const bool kDebugGlobalVariables = false;
52
53static bool allocationLODIsNull(const android::renderscript::Allocation *alloc) {
54  // Even if alloc != nullptr, mallocPtr could be null if
55  // IO_OUTPUT/IO_INPUT with no bound surface.
56  return alloc && alloc->mHal.drvState.lod[0].mallocPtr == nullptr;
57}
58
59#ifndef RS_COMPATIBILITY_LIB
60
61static bool is_force_recompile() {
62#ifdef RS_SERVER
63  return false;
64#else
65  char buf[PROPERTY_VALUE_MAX];
66
67  // Re-compile if floating point precision has been overridden.
68  property_get("debug.rs.precision", buf, "");
69  if (buf[0] != '\0') {
70    return true;
71  }
72
73  // Re-compile if debug.rs.forcerecompile is set.
74  property_get("debug.rs.forcerecompile", buf, "0");
75  if ((::strcmp(buf, "1") == 0) || (::strcmp(buf, "true") == 0)) {
76    return true;
77  } else {
78    return false;
79  }
80#endif  // RS_SERVER
81}
82
83static void setCompileArguments(std::vector<const char*>* args,
84                                const std::string& bcFileName,
85                                const char* cacheDir, const char* resName,
86                                const char* core_lib, bool useRSDebugContext,
87                                const char* bccPluginName, bool emitGlobalInfo,
88                                int optLevel, bool emitGlobalInfoSkipConstant) {
89    rsAssert(cacheDir && resName && core_lib);
90    args->push_back(android::renderscript::RsdCpuScriptImpl::BCC_EXE_PATH);
91    args->push_back("-unroll-runtime");
92    args->push_back("-scalarize-load-store");
93    if (emitGlobalInfo) {
94        args->push_back("-rs-global-info");
95        if (emitGlobalInfoSkipConstant) {
96            args->push_back("-rs-global-info-skip-constant");
97        }
98    }
99    args->push_back("-o");
100    args->push_back(resName);
101    args->push_back("-output_path");
102    args->push_back(cacheDir);
103    args->push_back("-bclib");
104    args->push_back(core_lib);
105    args->push_back("-mtriple");
106    args->push_back(DEFAULT_TARGET_TRIPLE_STRING);
107    args->push_back("-O");
108
109    switch (optLevel) {
110    case (0):
111        args->push_back("0");
112        break;
113    case (3):
114        args->push_back("3");
115        break;
116    default:
117        ALOGW("Expected optimization level of 0 or 3. Received %d", optLevel);
118        args->push_back("3");
119        break;
120    }
121
122    // Enable workaround for A53 codegen by default.
123#if defined(__aarch64__) && !defined(DISABLE_A53_WORKAROUND)
124    args->push_back("-aarch64-fix-cortex-a53-835769");
125#endif
126
127    // Execute the bcc compiler.
128    if (useRSDebugContext) {
129        args->push_back("-rs-debug-ctx");
130    } else {
131        // Only load additional libraries for compiles that don't use
132        // the debug context.
133        if (bccPluginName && strlen(bccPluginName) > 0) {
134            args->push_back("-load");
135            args->push_back(bccPluginName);
136        }
137    }
138
139    args->push_back("-fPIC");
140    args->push_back("-embedRSInfo");
141
142    args->push_back(bcFileName.c_str());
143    args->push_back(nullptr);
144}
145
146static bool compileBitcode(const std::string &bcFileName,
147                           const char *bitcode,
148                           size_t bitcodeSize,
149                           std::vector<const char *> &compileArguments) {
150    rsAssert(bitcode && bitcodeSize);
151
152    FILE *bcfile = fopen(bcFileName.c_str(), "w");
153    if (!bcfile) {
154        ALOGE("Could not write to %s", bcFileName.c_str());
155        return false;
156    }
157    size_t nwritten = fwrite(bitcode, 1, bitcodeSize, bcfile);
158    fclose(bcfile);
159    if (nwritten != bitcodeSize) {
160        ALOGE("Could not write %zu bytes to %s", bitcodeSize,
161              bcFileName.c_str());
162        return false;
163    }
164
165    return android::renderscript::rsuExecuteCommand(
166                   android::renderscript::RsdCpuScriptImpl::BCC_EXE_PATH,
167                   compileArguments.size()-1, compileArguments.data());
168}
169
170// The checksum is unnecessary under a few conditions, since the primary
171// use-case for it is debugging. If we are loading something from the
172// system partition (read-only), we know that it was precompiled as part of
173// application ahead of time (and thus the checksum is completely
174// unnecessary). The checksum is also unnecessary on release (non-debug)
175// builds, as the only way to get a shared object is to have compiled the
176// script once already. On a release build, there is no way to adjust the
177// other libraries/dependencies, and so the only reason to recompile would
178// be for a source APK change or an OTA. In either case, the APK would be
179// reinstalled, which would already clear the code_cache/ directory.
180bool isChecksumNeeded(const char *cacheDir) {
181    if ((::strcmp(SYSLIBPATH, cacheDir) == 0) ||
182        (::strcmp(SYSLIBPATH_VENDOR, cacheDir) == 0))
183        return false;
184    char buf[PROPERTY_VALUE_MAX];
185    property_get("ro.debuggable", buf, "");
186    return (buf[0] == '1');
187}
188
189bool addFileToChecksum(const char *fileName, uint32_t &checksum) {
190    int FD = open(fileName, O_RDONLY);
191    if (FD == -1) {
192        ALOGE("Cannot open file \'%s\' to compute checksum", fileName);
193        return false;
194    }
195
196    char buf[256];
197    while (true) {
198        ssize_t nread = read(FD, buf, sizeof(buf));
199        if (nread < 0) { // bail out on failed read
200            ALOGE("Error while computing checksum for file \'%s\'", fileName);
201            return false;
202        }
203
204        checksum = adler32(checksum, (const unsigned char *) buf, nread);
205        if (static_cast<size_t>(nread) < sizeof(buf)) // EOF
206            break;
207    }
208
209    if (close(FD) != 0) {
210        ALOGE("Cannot close file \'%s\' after computing checksum", fileName);
211        return false;
212    }
213    return true;
214}
215
216#endif  // !defined(RS_COMPATIBILITY_LIB)
217}  // namespace
218
219namespace android {
220namespace renderscript {
221
222#ifndef RS_COMPATIBILITY_LIB
223
224uint32_t constructBuildChecksum(uint8_t const *bitcode, size_t bitcodeSize,
225                                const char *commandLine,
226                                const char** bccFiles, size_t numFiles) {
227    uint32_t checksum = adler32(0L, Z_NULL, 0);
228
229    // include checksum of bitcode
230    if (bitcode != nullptr && bitcodeSize > 0) {
231        checksum = adler32(checksum, bitcode, bitcodeSize);
232    }
233
234    // include checksum of command line arguments
235    checksum = adler32(checksum, (const unsigned char *) commandLine,
236                       strlen(commandLine));
237
238    // include checksum of bccFiles
239    for (size_t i = 0; i < numFiles; i++) {
240        const char* bccFile = bccFiles[i];
241        if (bccFile[0] != 0 && !addFileToChecksum(bccFile, checksum)) {
242            // return empty checksum instead of something partial/corrupt
243            return 0;
244        }
245    }
246
247    return checksum;
248}
249
250#endif  // !RS_COMPATIBILITY_LIB
251
252RsdCpuScriptImpl::RsdCpuScriptImpl(RsdCpuReferenceImpl *ctx, const Script *s) {
253    mCtx = ctx;
254    mScript = s;
255
256    mScriptSO = nullptr;
257
258    mRoot = nullptr;
259    mRootExpand = nullptr;
260    mInit = nullptr;
261    mFreeChildren = nullptr;
262    mScriptExec = nullptr;
263
264    mBoundAllocs = nullptr;
265    mIntrinsicData = nullptr;
266    mIsThreadable = true;
267
268    mBuildChecksum = 0;
269    mChecksumNeeded = false;
270}
271
272bool RsdCpuScriptImpl::storeRSInfoFromSO() {
273    // The shared object may have an invalid build checksum.
274    // Validate and fail early.
275    mScriptExec = ScriptExecutable::createFromSharedObject(
276            mCtx->getContext(), mScriptSO,
277            mChecksumNeeded ? mBuildChecksum : 0);
278
279    if (mScriptExec == nullptr) {
280        return false;
281    }
282
283    mRoot = (RootFunc_t) dlsym(mScriptSO, "root");
284    if (mRoot) {
285        //ALOGE("Found root(): %p", mRoot);
286    }
287    mRootExpand = (RootFunc_t) dlsym(mScriptSO, "root.expand");
288    if (mRootExpand) {
289        //ALOGE("Found root.expand(): %p", mRootExpand);
290    }
291    mInit = (InitOrDtorFunc_t) dlsym(mScriptSO, "init");
292    if (mInit) {
293        //ALOGE("Found init(): %p", mInit);
294    }
295    mFreeChildren = (InitOrDtorFunc_t) dlsym(mScriptSO, ".rs.dtor");
296    if (mFreeChildren) {
297        //ALOGE("Found .rs.dtor(): %p", mFreeChildren);
298    }
299
300    size_t varCount = mScriptExec->getExportedVariableCount();
301    if (varCount > 0) {
302        mBoundAllocs = new Allocation *[varCount];
303        memset(mBoundAllocs, 0, varCount * sizeof(*mBoundAllocs));
304    }
305
306    mIsThreadable = mScriptExec->getThreadable();
307    //ALOGE("Script isThreadable? %d", mIsThreadable);
308
309    if (kDebugGlobalVariables) {
310        mScriptExec->dumpGlobalInfo();
311    }
312
313    return true;
314}
315
316bool RsdCpuScriptImpl::init(char const *resName, char const *cacheDir,
317                            uint8_t const *bitcode, size_t bitcodeSize,
318                            uint32_t flags, char const *bccPluginName) {
319    //ALOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir,
320    // bitcode, bitcodeSize, flags, lookupFunc);
321    //ALOGE("rsdScriptInit %p %p", rsc, script);
322
323    mCtx->lockMutex();
324#ifndef RS_COMPATIBILITY_LIB
325    bool useRSDebugContext = false;
326
327    bcinfo::MetadataExtractor bitcodeMetadata((const char *) bitcode, bitcodeSize);
328    if (!bitcodeMetadata.extract()) {
329        ALOGE("Could not extract metadata from bitcode");
330        mCtx->unlockMutex();
331        return false;
332    }
333
334    const char* core_lib = findCoreLib(bitcodeMetadata, (const char*)bitcode, bitcodeSize);
335
336    if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) {
337        useRSDebugContext = true;
338    }
339
340    int optLevel = mCtx->getContext()->getOptLevel();
341
342    std::string bcFileName(cacheDir);
343    bcFileName.append("/");
344    bcFileName.append(resName);
345    bcFileName.append(".bc");
346
347    std::vector<const char*> compileArguments;
348    bool emitGlobalInfo = mCtx->getEmbedGlobalInfo();
349    bool emitGlobalInfoSkipConstant = mCtx->getEmbedGlobalInfoSkipConstant();
350    setCompileArguments(&compileArguments, bcFileName, cacheDir, resName, core_lib,
351                        useRSDebugContext, bccPluginName, emitGlobalInfo,
352                        optLevel, emitGlobalInfoSkipConstant);
353
354    mChecksumNeeded = isChecksumNeeded(cacheDir);
355    if (mChecksumNeeded) {
356        std::vector<const char *> bccFiles = { BCC_EXE_PATH,
357                                               core_lib,
358                                             };
359
360        // The last argument of compileArguments is a nullptr, so remove 1 from
361        // the size.
362        std::unique_ptr<const char> compileCommandLine(
363            rsuJoinStrings(compileArguments.size()-1, compileArguments.data()));
364
365        mBuildChecksum = constructBuildChecksum(bitcode, bitcodeSize,
366                                                compileCommandLine.get(),
367                                                bccFiles.data(), bccFiles.size());
368
369        if (mBuildChecksum == 0) {
370            // cannot compute checksum but verification is enabled
371            mCtx->unlockMutex();
372            return false;
373        }
374    }
375    else {
376        // add a dummy/constant as a checksum if verification is disabled
377        mBuildChecksum = 0xabadcafe;
378    }
379
380    // Append build checksum to commandline
381    // Handle the terminal nullptr in compileArguments
382    compileArguments.pop_back();
383    compileArguments.push_back("-build-checksum");
384    std::stringstream ss;
385    ss << std::hex << mBuildChecksum;
386    compileArguments.push_back(ss.str().c_str());
387    compileArguments.push_back(nullptr);
388
389    if (!is_force_recompile() && !useRSDebugContext) {
390        mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName);
391
392        // Read RS info from the shared object to detect checksum mismatch
393        if (mScriptSO != nullptr && !storeRSInfoFromSO()) {
394            dlclose(mScriptSO);
395            mScriptSO = nullptr;
396        }
397    }
398
399    // If we can't, it's either not there or out of date.  We compile the bit code and try loading
400    // again.
401    if (mScriptSO == nullptr) {
402        if (!compileBitcode(bcFileName, (const char*)bitcode, bitcodeSize,
403                            compileArguments))
404        {
405            ALOGE("bcc: FAILS to compile '%s'", resName);
406            mCtx->unlockMutex();
407            return false;
408        }
409
410        if (!SharedLibraryUtils::createSharedLibrary(mCtx->getContext()->getDriverName(),
411                                                     cacheDir, resName)) {
412            ALOGE("Linker: Failed to link object file '%s'", resName);
413            mCtx->unlockMutex();
414            return false;
415        }
416
417        mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName);
418        if (mScriptSO == nullptr) {
419            ALOGE("Unable to load '%s'", resName);
420            mCtx->unlockMutex();
421            return false;
422        }
423
424        // Read RS symbol information from the .so.
425        if (!storeRSInfoFromSO()) {
426            goto error;
427        }
428    }
429
430    mBitcodeFilePath.setTo(bcFileName.c_str());
431
432#else  // RS_COMPATIBILITY_LIB is defined
433    const char *nativeLibDir = mCtx->getContext()->getNativeLibDir();
434    mScriptSO = SharedLibraryUtils::loadSharedLibrary(cacheDir, resName, nativeLibDir);
435
436    if (!mScriptSO) {
437        goto error;
438    }
439
440    if (!storeRSInfoFromSO()) {
441        goto error;
442    }
443#endif
444    mCtx->unlockMutex();
445    return true;
446
447error:
448
449    mCtx->unlockMutex();
450    if (mScriptSO) {
451        dlclose(mScriptSO);
452        mScriptSO = nullptr;
453    }
454    return false;
455}
456
457#ifndef RS_COMPATIBILITY_LIB
458
459const char* RsdCpuScriptImpl::findCoreLib(const bcinfo::MetadataExtractor& ME, const char* bitcode,
460                                          size_t bitcodeSize) {
461    const char* defaultLib = SYSLIBPATH"/libclcore.bc";
462
463    // If we're debugging, use the debug library.
464    if (mCtx->getContext()->getContextType() == RS_CONTEXT_TYPE_DEBUG) {
465        return SYSLIBPATH"/libclcore_debug.bc";
466    }
467
468    // If a callback has been registered to specify a library, use that.
469    RSSelectRTCallback selectRTCallback = mCtx->getSelectRTCallback();
470    if (selectRTCallback != nullptr) {
471        return selectRTCallback((const char*)bitcode, bitcodeSize);
472    }
473
474    // Check for a platform specific library
475#if defined(ARCH_ARM_HAVE_NEON) && !defined(DISABLE_CLCORE_NEON)
476    enum bcinfo::RSFloatPrecision prec = ME.getRSFloatPrecision();
477    if (prec == bcinfo::RS_FP_Relaxed) {
478        // NEON-capable ARMv7a devices can use an accelerated math library
479        // for all reduced precision scripts.
480        // ARMv8 does not use NEON, as ASIMD can be used with all precision
481        // levels.
482        return SYSLIBPATH"/libclcore_neon.bc";
483    } else {
484        return defaultLib;
485    }
486#elif defined(__i386__) || defined(__x86_64__)
487    // x86 devices will use an optimized library.
488    return SYSLIBPATH"/libclcore_x86.bc";
489#else
490    return defaultLib;
491#endif
492}
493
494#endif
495
496void RsdCpuScriptImpl::populateScript(Script *script) {
497    // Copy info over to runtime
498    script->mHal.info.exportedFunctionCount = mScriptExec->getExportedFunctionCount();
499    script->mHal.info.exportedReduceCount = mScriptExec->getExportedReduceCount();
500    script->mHal.info.exportedForEachCount = mScriptExec->getExportedForEachCount();
501    script->mHal.info.exportedVariableCount = mScriptExec->getExportedVariableCount();
502    script->mHal.info.exportedPragmaCount = mScriptExec->getPragmaCount();;
503    script->mHal.info.exportedPragmaKeyList = mScriptExec->getPragmaKeys();
504    script->mHal.info.exportedPragmaValueList = mScriptExec->getPragmaValues();
505
506    // Bug, need to stash in metadata
507    if (mRootExpand) {
508        script->mHal.info.root = mRootExpand;
509    } else {
510        script->mHal.info.root = mRoot;
511    }
512}
513
514// Set up the launch dimensions, and write the values of the launch
515// dimensions into the mtls start/end fields.
516//
517// Inputs:
518//    baseDim - base shape of the input
519//         sc - used to constrain the launch dimensions
520//
521// Returns:
522//   True on success, false on failure to set up
523bool RsdCpuScriptImpl::setUpMtlsDimensions(MTLaunchStructCommon *mtls,
524                                           const RsLaunchDimensions &baseDim,
525                                           const RsScriptCall *sc) {
526    rsAssert(mtls);
527
528#define SET_UP_DIMENSION(DIM_FIELD, SC_FIELD) do {            \
529    if (!sc || (sc->SC_FIELD##End == 0)) {                    \
530        mtls->end.DIM_FIELD = baseDim.DIM_FIELD;              \
531    } else {                                                  \
532        mtls->start.DIM_FIELD =                               \
533            rsMin(baseDim.DIM_FIELD, sc->SC_FIELD##Start);    \
534        mtls->end.DIM_FIELD =                                 \
535            rsMin(baseDim.DIM_FIELD, sc->SC_FIELD##End);      \
536        if (mtls->start.DIM_FIELD >= mtls->end.DIM_FIELD) {   \
537            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT, \
538                "Failed to launch kernel; Invalid "           \
539                #SC_FIELD "Start or " #SC_FIELD "End.");      \
540            return false;                                     \
541        }                                                     \
542    }} while(0)
543
544    SET_UP_DIMENSION(x, x);
545    SET_UP_DIMENSION(y, y);
546    SET_UP_DIMENSION(z, z);
547    SET_UP_DIMENSION(array[0], array);
548    SET_UP_DIMENSION(array[1], array2);
549    SET_UP_DIMENSION(array[2], array3);
550    SET_UP_DIMENSION(array[3], array4);
551#undef SET_UP_DIMENSION
552
553    return true;
554}
555
556// Preliminary work to prepare a reduce-style kernel for launch.
557bool RsdCpuScriptImpl::reduceMtlsSetup(const Allocation *ain,
558                                       const Allocation *aout,
559                                       const RsScriptCall *sc,
560                                       MTLaunchStructReduce *mtls) {
561    rsAssert(ain && aout);
562    memset(mtls, 0, sizeof(MTLaunchStructReduce));
563    mtls->dimPtr = &mtls->inputDim;
564
565    if (allocationLODIsNull(ain) || allocationLODIsNull(aout)) {
566        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
567                                     "reduce called with a null allocation");
568        return false;
569    }
570
571    // Set up the dimensions of the input.
572    const Type *inType = ain->getType();
573    mtls->inputDim.x = inType->getDimX();
574    rsAssert(inType->getDimY() == 0);
575
576    if (!setUpMtlsDimensions(mtls, mtls->inputDim, sc)) {
577        return false;
578    }
579
580    mtls->rs = mCtx;
581    // Currently not threaded.
582    mtls->isThreadable = false;
583    mtls->mSliceNum = -1;
584
585    // Set up input and output.
586    mtls->inBuf = static_cast<uint8_t *>(ain->getPointerUnchecked(0, 0));
587    mtls->outBuf = static_cast<uint8_t *>(aout->getPointerUnchecked(0, 0));
588
589    rsAssert(mtls->inBuf && mtls->outBuf);
590
591    return true;
592}
593
594// Preliminary work to prepare a forEach-style kernel for launch.
595bool RsdCpuScriptImpl::forEachMtlsSetup(const Allocation ** ains,
596                                        uint32_t inLen,
597                                        Allocation * aout,
598                                        const void * usr, uint32_t usrLen,
599                                        const RsScriptCall *sc,
600                                        MTLaunchStructForEach *mtls) {
601    memset(mtls, 0, sizeof(MTLaunchStructForEach));
602    mtls->dimPtr = &mtls->fep.dim;
603
604    for (int index = inLen; --index >= 0;) {
605        if (allocationLODIsNull(ains[index])) {
606            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
607                                         "rsForEach called with null in allocations");
608            return false;
609        }
610    }
611
612    if (allocationLODIsNull(aout)) {
613        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
614                                     "rsForEach called with null out allocations");
615        return false;
616    }
617
618    if (inLen > 0) {
619        const Allocation *ain0   = ains[0];
620        const Type       *inType = ain0->getType();
621
622        mtls->fep.dim.x = inType->getDimX();
623        mtls->fep.dim.y = inType->getDimY();
624        mtls->fep.dim.z = inType->getDimZ();
625
626        for (int Index = inLen; --Index >= 1;) {
627            if (!ain0->hasSameDims(ains[Index])) {
628                mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
629                  "Failed to launch kernel; dimensions of input and output"
630                  "allocations do not match.");
631
632                return false;
633            }
634        }
635
636    } else if (aout != nullptr) {
637        const Type *outType = aout->getType();
638
639        mtls->fep.dim.x = outType->getDimX();
640        mtls->fep.dim.y = outType->getDimY();
641        mtls->fep.dim.z = outType->getDimZ();
642
643    } else if (sc != nullptr) {
644        mtls->fep.dim.x = sc->xEnd;
645        mtls->fep.dim.y = sc->yEnd;
646        mtls->fep.dim.z = 0;
647    } else {
648        mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
649                                     "rsForEach called with null allocations");
650        return false;
651    }
652
653    if (inLen > 0 && aout != nullptr) {
654        if (!ains[0]->hasSameDims(aout)) {
655            mCtx->getContext()->setError(RS_ERROR_BAD_SCRIPT,
656              "Failed to launch kernel; dimensions of input and output allocations do not match.");
657
658            return false;
659        }
660    }
661
662    if (!setUpMtlsDimensions(mtls, mtls->fep.dim, sc)) {
663        return false;
664    }
665
666    // The X & Y walkers always want 0-1 min even if dim is not present
667    mtls->end.x    = rsMax((uint32_t)1, mtls->end.x);
668    mtls->end.y    = rsMax((uint32_t)1, mtls->end.y);
669    mtls->rs       = mCtx;
670    if (ains) {
671        memcpy(mtls->ains, ains, inLen * sizeof(ains[0]));
672    }
673    mtls->aout[0]    = aout;
674    mtls->fep.usr    = usr;
675    mtls->fep.usrLen = usrLen;
676    mtls->mSliceSize = 1;
677    mtls->mSliceNum  = 0;
678
679    mtls->isThreadable  = mIsThreadable;
680
681    if (inLen > 0) {
682        mtls->fep.inLen = inLen;
683        for (int index = inLen; --index >= 0;) {
684            mtls->fep.inPtr[index] = (const uint8_t*)ains[index]->mHal.drvState.lod[0].mallocPtr;
685            mtls->fep.inStride[index] = ains[index]->getType()->getElementSizeBytes();
686        }
687    }
688
689    if (aout != nullptr) {
690        mtls->fep.outPtr[0] = (uint8_t *)aout->mHal.drvState.lod[0].mallocPtr;
691        mtls->fep.outStride[0] = aout->getType()->getElementSizeBytes();
692    }
693
694    // All validation passed, ok to launch threads
695    return true;
696}
697
698
699void RsdCpuScriptImpl::invokeForEach(uint32_t slot,
700                                     const Allocation ** ains,
701                                     uint32_t inLen,
702                                     Allocation * aout,
703                                     const void * usr,
704                                     uint32_t usrLen,
705                                     const RsScriptCall *sc) {
706
707    MTLaunchStructForEach mtls;
708
709    if (forEachMtlsSetup(ains, inLen, aout, usr, usrLen, sc, &mtls)) {
710        forEachKernelSetup(slot, &mtls);
711
712        RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
713        mCtx->launchForEach(ains, inLen, aout, sc, &mtls);
714        mCtx->setTLS(oldTLS);
715    }
716}
717
718void RsdCpuScriptImpl::invokeReduce(uint32_t slot,
719                                    const Allocation *ain,
720                                    Allocation *aout,
721                                    const RsScriptCall *sc) {
722    MTLaunchStructReduce mtls;
723
724    if (reduceMtlsSetup(ain, aout, sc, &mtls)) {
725        reduceKernelSetup(slot, &mtls);
726        RsdCpuScriptImpl *oldTLS = mCtx->setTLS(this);
727        mCtx->launchReduce(ain, aout, &mtls);
728        mCtx->setTLS(oldTLS);
729    }
730}
731
732void RsdCpuScriptImpl::forEachKernelSetup(uint32_t slot, MTLaunchStructForEach *mtls) {
733    mtls->script = this;
734    mtls->fep.slot = slot;
735    mtls->kernel = mScriptExec->getForEachFunction(slot);
736    rsAssert(mtls->kernel != nullptr);
737    mtls->sig = mScriptExec->getForEachSignature(slot);
738}
739
740void RsdCpuScriptImpl::reduceKernelSetup(uint32_t slot, MTLaunchStructReduce *mtls) {
741    mtls->script = this;
742    mtls->kernel = mScriptExec->getReduceFunction(slot);
743    rsAssert(mtls->kernel != nullptr);
744}
745
746int RsdCpuScriptImpl::invokeRoot() {
747    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
748    int ret = mRoot();
749    mCtx->setTLS(oldTLS);
750    return ret;
751}
752
753void RsdCpuScriptImpl::invokeInit() {
754    if (mInit) {
755        mInit();
756    }
757}
758
759void RsdCpuScriptImpl::invokeFreeChildren() {
760    if (mFreeChildren) {
761        mFreeChildren();
762    }
763}
764
765void RsdCpuScriptImpl::invokeFunction(uint32_t slot, const void *params,
766                                      size_t paramLength) {
767    //ALOGE("invoke %i %p %zu", slot, params, paramLength);
768    void * ap = nullptr;
769
770#if defined(__x86_64__)
771    // The invoked function could have input parameter of vector type for example float4 which
772    // requires void* params to be 16 bytes aligned when using SSE instructions for x86_64 platform.
773    // So try to align void* params before passing them into RS exported function.
774
775    if ((uint8_t)(uint64_t)params & 0x0F) {
776        if ((ap = (void*)memalign(16, paramLength)) != nullptr) {
777            memcpy(ap, params, paramLength);
778        } else {
779            ALOGE("x86_64: invokeFunction memalign error, still use params which"
780                  " is not 16 bytes aligned.");
781        }
782    }
783#endif
784
785    RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
786    reinterpret_cast<void (*)(const void *, uint32_t)>(
787        mScriptExec->getInvokeFunction(slot))(ap? (const void *) ap: params, paramLength);
788
789#if defined(__x86_64__)
790    free(ap);
791#endif
792
793    mCtx->setTLS(oldTLS);
794}
795
796void RsdCpuScriptImpl::setGlobalVar(uint32_t slot, const void *data, size_t dataLength) {
797    //rsAssert(!script->mFieldIsObject[slot]);
798    //ALOGE("setGlobalVar %i %p %zu", slot, data, dataLength);
799
800    //if (mIntrinsicID) {
801        //mIntrinsicFuncs.setVar(dc, script, drv->mIntrinsicData, slot, data, dataLength);
802        //return;
803    //}
804
805    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
806    if (!destPtr) {
807        //ALOGV("Calling setVar on slot = %i which is null", slot);
808        return;
809    }
810
811    memcpy(destPtr, data, dataLength);
812}
813
814void RsdCpuScriptImpl::getGlobalVar(uint32_t slot, void *data, size_t dataLength) {
815    //rsAssert(!script->mFieldIsObject[slot]);
816    //ALOGE("getGlobalVar %i %p %zu", slot, data, dataLength);
817
818    int32_t *srcPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
819    if (!srcPtr) {
820        //ALOGV("Calling setVar on slot = %i which is null", slot);
821        return;
822    }
823    memcpy(data, srcPtr, dataLength);
824}
825
826
827void RsdCpuScriptImpl::setGlobalVarWithElemDims(uint32_t slot, const void *data, size_t dataLength,
828                                                const Element *elem,
829                                                const uint32_t *dims, size_t dimLength) {
830    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
831    if (!destPtr) {
832        //ALOGV("Calling setVar on slot = %i which is null", slot);
833        return;
834    }
835
836    // We want to look at dimension in terms of integer components,
837    // but dimLength is given in terms of bytes.
838    dimLength /= sizeof(int);
839
840    // Only a single dimension is currently supported.
841    rsAssert(dimLength == 1);
842    if (dimLength == 1) {
843        // First do the increment loop.
844        size_t stride = elem->getSizeBytes();
845        const char *cVal = reinterpret_cast<const char *>(data);
846        for (uint32_t i = 0; i < dims[0]; i++) {
847            elem->incRefs(cVal);
848            cVal += stride;
849        }
850
851        // Decrement loop comes after (to prevent race conditions).
852        char *oldVal = reinterpret_cast<char *>(destPtr);
853        for (uint32_t i = 0; i < dims[0]; i++) {
854            elem->decRefs(oldVal);
855            oldVal += stride;
856        }
857    }
858
859    memcpy(destPtr, data, dataLength);
860}
861
862void RsdCpuScriptImpl::setGlobalBind(uint32_t slot, Allocation *data) {
863
864    //rsAssert(!script->mFieldIsObject[slot]);
865    //ALOGE("setGlobalBind %i %p", slot, data);
866
867    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
868    if (!destPtr) {
869        //ALOGV("Calling setVar on slot = %i which is null", slot);
870        return;
871    }
872
873    void *ptr = nullptr;
874    mBoundAllocs[slot] = data;
875    if (data) {
876        ptr = data->mHal.drvState.lod[0].mallocPtr;
877    }
878    memcpy(destPtr, &ptr, sizeof(void *));
879}
880
881void RsdCpuScriptImpl::setGlobalObj(uint32_t slot, ObjectBase *data) {
882
883    //rsAssert(script->mFieldIsObject[slot]);
884    //ALOGE("setGlobalObj %i %p", slot, data);
885
886    int32_t *destPtr = reinterpret_cast<int32_t *>(mScriptExec->getFieldAddress(slot));
887    if (!destPtr) {
888        //ALOGV("Calling setVar on slot = %i which is null", slot);
889        return;
890    }
891
892    rsrSetObject(mCtx->getContext(), (rs_object_base *)destPtr, data);
893}
894
895const char* RsdCpuScriptImpl::getFieldName(uint32_t slot) const {
896    return mScriptExec->getFieldName(slot);
897}
898
899RsdCpuScriptImpl::~RsdCpuScriptImpl() {
900    delete mScriptExec;
901    delete[] mBoundAllocs;
902    if (mScriptSO) {
903        dlclose(mScriptSO);
904    }
905}
906
907Allocation * RsdCpuScriptImpl::getAllocationForPointer(const void *ptr) const {
908    if (!ptr) {
909        return nullptr;
910    }
911
912    for (uint32_t ct=0; ct < mScript->mHal.info.exportedVariableCount; ct++) {
913        Allocation *a = mBoundAllocs[ct];
914        if (!a) continue;
915        if (a->mHal.drvState.lod[0].mallocPtr == ptr) {
916            return a;
917        }
918    }
919    ALOGE("rsGetAllocation, failed to find %p", ptr);
920    return nullptr;
921}
922
923int RsdCpuScriptImpl::getGlobalEntries() const {
924    return mScriptExec->getGlobalEntries();
925}
926
927const char * RsdCpuScriptImpl::getGlobalName(int i) const {
928    return mScriptExec->getGlobalName(i);
929}
930
931const void * RsdCpuScriptImpl::getGlobalAddress(int i) const {
932    return mScriptExec->getGlobalAddress(i);
933}
934
935size_t RsdCpuScriptImpl::getGlobalSize(int i) const {
936    return mScriptExec->getGlobalSize(i);
937}
938
939uint32_t RsdCpuScriptImpl::getGlobalProperties(int i) const {
940    return mScriptExec->getGlobalProperties(i);
941}
942
943void RsdCpuScriptImpl::preLaunch(uint32_t slot, const Allocation ** ains,
944                                 uint32_t inLen, Allocation * aout,
945                                 const void * usr, uint32_t usrLen,
946                                 const RsScriptCall *sc) {}
947
948void RsdCpuScriptImpl::postLaunch(uint32_t slot, const Allocation ** ains,
949                                  uint32_t inLen, Allocation * aout,
950                                  const void * usr, uint32_t usrLen,
951                                  const RsScriptCall *sc) {}
952
953
954}
955}
956