rsCpuCore.cpp revision 64c682b65cd04ac83b51251b40dca14423df351a 
1/*
2 * Copyright (C) 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 "rsCpuScriptGroup.h"
20#include "rsCpuScriptGroup2.h"
21
22#include <malloc.h>
23#include "rsContext.h"
24
25#include <sys/types.h>
26#include <sys/resource.h>
27#include <sched.h>
28#include <sys/syscall.h>
29#include <string.h>
30#include <unistd.h>
31
32#include <stdio.h>
33#include <stdlib.h>
34#include <fcntl.h>
35
36#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB)
37#include <cutils/properties.h>
38#include "utils/StopWatch.h"
39#endif
40
41#ifdef RS_SERVER
42// Android exposes gettid(), standard Linux does not
43static pid_t gettid() {
44    return syscall(SYS_gettid);
45}
46#endif
47
48using namespace android;
49using namespace android::renderscript;
50
51typedef void (*outer_foreach_t)(
52    const android::renderscript::RsExpandKernelParams *,
53    uint32_t x1, uint32_t x2, uint32_t outstep);
54
55
56static pthread_key_t gThreadTLSKey = 0;
57static uint32_t gThreadTLSKeyCount = 0;
58static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER;
59
60bool android::renderscript::gArchUseSIMD = false;
61
62RsdCpuReference::~RsdCpuReference() {
63}
64
65RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major,
66        uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn
67        , bcc::RSLinkRuntimeCallback pLinkRuntimeCallback,
68        RSSelectRTCallback pSelectRTCallback,
69        const char *pBccPluginName
70        ) {
71
72    RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc);
73    if (!cpu) {
74        return nullptr;
75    }
76    if (!cpu->init(version_major, version_minor, lfn, slfn)) {
77        delete cpu;
78        return nullptr;
79    }
80
81    cpu->setLinkRuntimeCallback(pLinkRuntimeCallback);
82    cpu->setSelectRTCallback(pSelectRTCallback);
83    if (pBccPluginName) {
84        cpu->setBccPluginName(pBccPluginName);
85    }
86
87    return cpu;
88}
89
90
91Context * RsdCpuReference::getTlsContext() {
92    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
93    return tls->mContext;
94}
95
96const Script * RsdCpuReference::getTlsScript() {
97    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
98    return tls->mScript;
99}
100
101pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; }
102
103////////////////////////////////////////////////////////////
104///
105
106RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) {
107    mRSC = rsc;
108
109    version_major = 0;
110    version_minor = 0;
111    mInForEach = false;
112    memset(&mWorkers, 0, sizeof(mWorkers));
113    memset(&mTlsStruct, 0, sizeof(mTlsStruct));
114    mExit = false;
115    mLinkRuntimeCallback = nullptr;
116    mSelectRTCallback = nullptr;
117    mSetupCompilerCallback = nullptr;
118}
119
120
121void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) {
122    RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc;
123
124    uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1);
125
126    //ALOGV("RS helperThread starting %p idx=%i", dc, idx);
127
128    dc->mWorkers.mLaunchSignals[idx].init();
129    dc->mWorkers.mNativeThreadId[idx] = gettid();
130
131    memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct));
132    int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct);
133    if (status) {
134        ALOGE("pthread_setspecific %i", status);
135    }
136
137#if 0
138    typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
139    cpu_set_t cpuset;
140    memset(&cpuset, 0, sizeof(cpuset));
141    cpuset.bits[idx / 64] |= 1ULL << (idx % 64);
142    int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
143              sizeof(cpuset), &cpuset);
144    ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
145#endif
146
147    while (!dc->mExit) {
148        dc->mWorkers.mLaunchSignals[idx].wait();
149        if (dc->mWorkers.mLaunchCallback) {
150           // idx +1 is used because the calling thread is always worker 0.
151           dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1);
152        }
153        __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1);
154        dc->mWorkers.mCompleteSignal.set();
155    }
156
157    //ALOGV("RS helperThread exited %p idx=%i", dc, idx);
158    return nullptr;
159}
160
161void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) {
162    mWorkers.mLaunchData = data;
163    mWorkers.mLaunchCallback = cbk;
164
165    // fast path for very small launches
166    MTLaunchStruct *mtls = (MTLaunchStruct *)data;
167    if (mtls && mtls->fep.dim.y <= 1 && mtls->end.x <= mtls->start.x + mtls->mSliceSize) {
168        if (mWorkers.mLaunchCallback) {
169            mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
170        }
171        return;
172    }
173
174    mWorkers.mRunningCount = mWorkers.mCount;
175    __sync_synchronize();
176
177    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
178        mWorkers.mLaunchSignals[ct].set();
179    }
180
181    // We use the calling thread as one of the workers so we can start without
182    // the delay of the thread wakeup.
183    if (mWorkers.mLaunchCallback) {
184        mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
185    }
186
187    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
188        mWorkers.mCompleteSignal.wait();
189    }
190}
191
192
193void RsdCpuReferenceImpl::lockMutex() {
194    pthread_mutex_lock(&gInitMutex);
195}
196
197void RsdCpuReferenceImpl::unlockMutex() {
198    pthread_mutex_unlock(&gInitMutex);
199}
200
201static int
202read_file(const char*  pathname, char*  buffer, size_t  buffsize)
203{
204    int  fd, len;
205
206    fd = open(pathname, O_RDONLY);
207    if (fd < 0)
208        return -1;
209
210    do {
211        len = read(fd, buffer, buffsize);
212    } while (len < 0 && errno == EINTR);
213
214    close(fd);
215
216    return len;
217}
218
219static void GetCpuInfo() {
220    char cpuinfo[4096];
221    int  cpuinfo_len;
222
223    cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo);
224    if (cpuinfo_len < 0)  /* should not happen */ {
225        return;
226    }
227
228#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS)
229    gArchUseSIMD = (!!strstr(cpuinfo, " neon")) ||
230                   (!!strstr(cpuinfo, " asimd"));
231#elif defined(ARCH_X86_HAVE_SSSE3)
232    gArchUseSIMD = !!strstr(cpuinfo, " ssse3");
233#endif
234}
235
236bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor,
237                               sym_lookup_t lfn, script_lookup_t slfn) {
238
239    mSymLookupFn = lfn;
240    mScriptLookupFn = slfn;
241
242    lockMutex();
243    if (!gThreadTLSKeyCount) {
244        int status = pthread_key_create(&gThreadTLSKey, nullptr);
245        if (status) {
246            ALOGE("Failed to init thread tls key.");
247            unlockMutex();
248            return false;
249        }
250    }
251    gThreadTLSKeyCount++;
252    unlockMutex();
253
254    mTlsStruct.mContext = mRSC;
255    mTlsStruct.mScript = nullptr;
256    int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct);
257    if (status) {
258        ALOGE("pthread_setspecific %i", status);
259    }
260
261    GetCpuInfo();
262
263    int cpu = sysconf(_SC_NPROCESSORS_CONF);
264    if(mRSC->props.mDebugMaxThreads) {
265        cpu = mRSC->props.mDebugMaxThreads;
266    }
267    if (cpu < 2) {
268        mWorkers.mCount = 0;
269        return true;
270    }
271
272    // Subtract one from the cpu count because we also use the command thread as a worker.
273    mWorkers.mCount = (uint32_t)(cpu - 1);
274
275    ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1);
276
277    mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t));
278    mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t));
279    mWorkers.mLaunchSignals = new Signal[mWorkers.mCount];
280    mWorkers.mLaunchCallback = nullptr;
281
282    mWorkers.mCompleteSignal.init();
283
284    mWorkers.mRunningCount = mWorkers.mCount;
285    mWorkers.mLaunchCount = 0;
286    __sync_synchronize();
287
288    pthread_attr_t threadAttr;
289    status = pthread_attr_init(&threadAttr);
290    if (status) {
291        ALOGE("Failed to init thread attribute.");
292        return false;
293    }
294
295    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
296        status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this);
297        if (status) {
298            mWorkers.mCount = ct;
299            ALOGE("Created fewer than expected number of RS threads.");
300            break;
301        }
302    }
303    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
304        usleep(100);
305    }
306
307    pthread_attr_destroy(&threadAttr);
308    return true;
309}
310
311
312void RsdCpuReferenceImpl::setPriority(int32_t priority) {
313    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
314        setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority);
315    }
316}
317
318RsdCpuReferenceImpl::~RsdCpuReferenceImpl() {
319    mExit = true;
320    mWorkers.mLaunchData = nullptr;
321    mWorkers.mLaunchCallback = nullptr;
322    mWorkers.mRunningCount = mWorkers.mCount;
323    __sync_synchronize();
324    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
325        mWorkers.mLaunchSignals[ct].set();
326    }
327    void *res;
328    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
329        pthread_join(mWorkers.mThreadId[ct], &res);
330    }
331    rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0);
332    free(mWorkers.mThreadId);
333    free(mWorkers.mNativeThreadId);
334    delete[] mWorkers.mLaunchSignals;
335
336    // Global structure cleanup.
337    lockMutex();
338    --gThreadTLSKeyCount;
339    if (!gThreadTLSKeyCount) {
340        pthread_key_delete(gThreadTLSKey);
341    }
342    unlockMutex();
343
344}
345
346typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
347typedef void (*walk_loop_t)(const MTLaunchStruct*,
348                            RsExpandKernelDriverInfo,
349                            outer_foreach_t);
350
351static void kparamSetup(RsExpandKernelParams *kparams, const RsExpandKernelDriverInfo *fep) {
352    //ALOGE("kp  usr %p", fep->usr);
353    //ALOGE("kp  slot %i", fep->slot);
354    //ALOGE("kp  dim %i %i %i", fep->dim.x, fep->dim.y, fep->dim.z);
355    //ALOGE("kp  lid %i", fep->lid);
356    //ALOGE("kp  in[0] stide %i  ptr %p", fep->inStride[0], fep->inPtr[0]);
357    //ALOGE("kp  out[0] ptr %p", fep->outPtr[0]);
358    //ALOGE("kp  loc %i %i %i", fep->current.x, fep->current.y, fep->current.z);
359
360    kparams->usr  = fep->usr;
361    kparams->slot = fep->slot;
362    kparams->dimX = fep->dim.x;
363    kparams->dimY = fep->dim.y;
364    kparams->dimZ = fep->dim.z;
365    kparams->lid = fep->lid;
366    kparams->inEStrides = (uint32_t *)&fep->inStride[0];
367    kparams->ins = (const void **)&fep->inPtr[0];
368    kparams->out = fep->outPtr[0];
369    kparams->y = fep->current.y;
370    kparams->z = fep->current.z;
371}
372
373static inline void FepPtrSetup(const MTLaunchStruct *mtls, RsExpandKernelDriverInfo *fep,
374                               uint32_t x, uint32_t y,
375                               uint32_t z = 0, uint32_t lod = 0,
376                               RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X,
377                               uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) {
378
379    for (uint32_t i = 0; i < fep->inLen; i++) {
380        fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4);
381    }
382
383    if (mtls->aout[0] != nullptr) {
384        fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4);
385    }
386}
387
388static uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) {
389    if (start >= end) {
390        *p = start;
391        return val;
392    }
393
394    uint32_t div = end - start;
395
396    uint32_t n = val / div;
397    *p = (val - (n * div)) + start;
398    return n;
399}
400
401static bool SelectOuterSlice(MTLaunchStruct* mtls, uint32_t sliceNum) {
402
403    uint32_t r = sliceNum;
404    r = sliceInt(&mtls->fep.current.z, r, mtls->start.z, mtls->end.z);
405    r = sliceInt(&mtls->fep.current.lod, r, mtls->start.lod, mtls->end.lod);
406    r = sliceInt(&mtls->fep.current.face, r, mtls->start.face, mtls->end.face);
407    r = sliceInt(&mtls->fep.current.array[0], r, mtls->start.array[0], mtls->end.array[0]);
408    r = sliceInt(&mtls->fep.current.array[1], r, mtls->start.array[1], mtls->end.array[1]);
409    r = sliceInt(&mtls->fep.current.array[2], r, mtls->start.array[2], mtls->end.array[2]);
410    r = sliceInt(&mtls->fep.current.array[3], r, mtls->start.array[3], mtls->end.array[3]);
411    return r == 0;
412}
413
414
415static void walk_general(void *usr, uint32_t idx) {
416    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
417    RsExpandKernelDriverInfo fep = mtls->fep;
418    fep.lid = idx;
419    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
420
421
422    while(1) {
423        uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
424
425        if (!SelectOuterSlice(mtls, slice)) {
426            return;
427        }
428
429        for (mtls->fep.current.y = mtls->start.y;
430             mtls->fep.current.y < mtls->end.y;
431             mtls->fep.current.y++) {
432
433            FepPtrSetup(mtls, &mtls->fep, mtls->start.x,
434                        mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod,
435                        (RsAllocationCubemapFace)mtls->fep.current.face,
436                        mtls->fep.current.array[0], mtls->fep.current.array[1],
437                        mtls->fep.current.array[2], mtls->fep.current.array[3]);
438
439            RsExpandKernelParams kparams;
440            kparamSetup(&kparams, &mtls->fep);
441            fn(&kparams, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]);
442        }
443    }
444
445}
446
447static void walk_2d(void *usr, uint32_t idx) {
448    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
449    RsExpandKernelDriverInfo fep = mtls->fep;
450    fep.lid = idx;
451    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
452
453    while (1) {
454        uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
455        uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize;
456        uint32_t yEnd   = yStart + mtls->mSliceSize;
457
458        yEnd = rsMin(yEnd, mtls->end.y);
459
460        if (yEnd <= yStart) {
461            return;
462        }
463
464        for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) {
465            FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y);
466
467            RsExpandKernelParams kparams;
468            kparamSetup(&kparams, &fep);
469
470            fn(&kparams, mtls->start.x, mtls->end.x, fep.outStride[0]);
471        }
472    }
473}
474
475static void walk_1d(void *usr, uint32_t idx) {
476    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
477    RsExpandKernelDriverInfo fep = mtls->fep;
478    fep.lid = idx;
479    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
480
481    while (1) {
482        uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
483        uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize;
484        uint32_t xEnd   = xStart + mtls->mSliceSize;
485
486        xEnd = rsMin(xEnd, mtls->end.x);
487
488        if (xEnd <= xStart) {
489            return;
490        }
491
492        FepPtrSetup(mtls, &fep, xStart, 0);
493
494        RsExpandKernelParams kparams;
495        kparamSetup(&kparams, &fep);
496
497        fn(&kparams, xStart, xEnd, fep.outStride[0]);
498    }
499}
500
501void RsdCpuReferenceImpl::launchThreads(const Allocation ** ains,
502                                        uint32_t inLen,
503                                        Allocation* aout,
504                                        const RsScriptCall* sc,
505                                        MTLaunchStruct* mtls) {
506
507    //android::StopWatch kernel_time("kernel time");
508
509    bool outerDims = (mtls->start.z != mtls->end.z) ||
510                     (mtls->start.face != mtls->end.face) ||
511                     (mtls->start.lod != mtls->end.lod) ||
512                     (mtls->start.array[0] != mtls->end.array[0]) ||
513                     (mtls->start.array[1] != mtls->end.array[1]) ||
514                     (mtls->start.array[2] != mtls->end.array[2]) ||
515                     (mtls->start.array[3] != mtls->end.array[3]);
516
517    if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) {
518        const size_t targetByteChunk = 16 * 1024;
519        mInForEach = true;
520
521        if (outerDims) {
522            // No fancy logic for chunk size
523            mtls->mSliceSize = 1;
524            launchThreads(walk_general, mtls);
525        } else if (mtls->fep.dim.y > 1) {
526            uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4);
527            uint32_t s2 = 0;
528
529            // This chooses our slice size to rate limit atomic ops to
530            // one per 16k bytes of reads/writes.
531            if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) {
532                s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride;
533            } else {
534                // We know that there is either an output or an input.
535                s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride;
536            }
537            mtls->mSliceSize = rsMin(s1, s2);
538
539            if(mtls->mSliceSize < 1) {
540                mtls->mSliceSize = 1;
541            }
542
543            launchThreads(walk_2d, mtls);
544        } else {
545            uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4);
546            uint32_t s2 = 0;
547
548            // This chooses our slice size to rate limit atomic ops to
549            // one per 16k bytes of reads/writes.
550            if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) {
551                s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes();
552            } else {
553                // We know that there is either an output or an input.
554                s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes();
555            }
556            mtls->mSliceSize = rsMin(s1, s2);
557
558            if (mtls->mSliceSize < 1) {
559                mtls->mSliceSize = 1;
560            }
561
562            launchThreads(walk_1d, mtls);
563        }
564        mInForEach = false;
565
566    } else {
567        outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
568        uint32_t slice = 0;
569
570
571        while(SelectOuterSlice(mtls, slice++)) {
572            for (mtls->fep.current.y = mtls->start.y;
573                 mtls->fep.current.y < mtls->end.y;
574                 mtls->fep.current.y++) {
575
576                FepPtrSetup(mtls, &mtls->fep, mtls->start.x,
577                            mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod,
578                            (RsAllocationCubemapFace) mtls->fep.current.face,
579                            mtls->fep.current.array[0], mtls->fep.current.array[1],
580                            mtls->fep.current.array[2], mtls->fep.current.array[3]);
581
582                RsExpandKernelParams kparams;
583                kparamSetup(&kparams, &mtls->fep);
584                fn(&kparams, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]);
585            }
586        }
587    }
588}
589
590RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) {
591    //ALOGE("setTls %p", sc);
592    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
593    rsAssert(tls);
594    RsdCpuScriptImpl *old = tls->mImpl;
595    tls->mImpl = sc;
596    tls->mContext = mRSC;
597    if (sc) {
598        tls->mScript = sc->getScript();
599    } else {
600        tls->mScript = nullptr;
601    }
602    return old;
603}
604
605const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) {
606    return mSymLookupFn(mRSC, name);
607}
608
609
610RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s,
611                                    char const *resName, char const *cacheDir,
612                                    uint8_t const *bitcode, size_t bitcodeSize,
613                                    uint32_t flags) {
614
615    RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s);
616    if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags
617        , getBccPluginName()
618        )) {
619        delete i;
620        return nullptr;
621    }
622    return i;
623}
624
625extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx,
626                                             const Script *s, const Element *e);
627extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx,
628                                                   const Script *s, const Element *e);
629extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx,
630                                                   const Script *s, const Element *e);
631extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx,
632                                           const Script *s, const Element *e);
633extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx,
634                                                   const Script *s, const Element *e);
635extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx,
636                                            const Script *s, const Element *e);
637extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx,
638                                                const Script *s, const Element *e);
639extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx,
640                                             const Script *s, const Element *e);
641extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx,
642                                                 const Script *s, const Element *e);
643extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx,
644                                              const Script *s, const Element *e);
645extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx,
646                                              const Script *s, const Element *e);
647
648RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s,
649                                    RsScriptIntrinsicID iid, Element *e) {
650
651    RsdCpuScriptImpl *i = nullptr;
652    switch (iid) {
653    case RS_SCRIPT_INTRINSIC_ID_3DLUT:
654        i = rsdIntrinsic_3DLUT(this, s, e);
655        break;
656    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3:
657        i = rsdIntrinsic_Convolve3x3(this, s, e);
658        break;
659    case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX:
660        i = rsdIntrinsic_ColorMatrix(this, s, e);
661        break;
662    case RS_SCRIPT_INTRINSIC_ID_LUT:
663        i = rsdIntrinsic_LUT(this, s, e);
664        break;
665    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5:
666        i = rsdIntrinsic_Convolve5x5(this, s, e);
667        break;
668    case RS_SCRIPT_INTRINSIC_ID_BLUR:
669        i = rsdIntrinsic_Blur(this, s, e);
670        break;
671    case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB:
672        i = rsdIntrinsic_YuvToRGB(this, s, e);
673        break;
674    case RS_SCRIPT_INTRINSIC_ID_BLEND:
675        i = rsdIntrinsic_Blend(this, s, e);
676        break;
677    case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM:
678        i = rsdIntrinsic_Histogram(this, s, e);
679        break;
680    case RS_SCRIPT_INTRINSIC_ID_RESIZE:
681        i = rsdIntrinsic_Resize(this, s, e);
682        break;
683#if !defined(RS_COMPATIBILITY_LIB)
684    case RS_SCRIPT_INTRINSIC_ID_BLAS:
685        i = rsdIntrinsic_BLAS(this, s, e);
686        break;
687#endif
688
689    default:
690        rsAssert(0);
691    }
692
693    return i;
694}
695
696void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) {
697  switch (sg->getApiVersion()) {
698    case ScriptGroupBase::SG_V1: {
699      CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg);
700      if (!sgi->init()) {
701        delete sgi;
702        return nullptr;
703      }
704      return sgi;
705    }
706    case ScriptGroupBase::SG_V2: {
707      return new CpuScriptGroup2Impl(this, sg);
708    }
709  }
710  return nullptr;
711}
712