rsCpuCore.cpp revision 77d57a305f4134e78ebc91869011c4009988104e 
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
21#include <malloc.h>
22#include "rsContext.h"
23
24#include <sys/types.h>
25#include <sys/resource.h>
26#include <sched.h>
27#include <sys/syscall.h>
28#include <string.h>
29#include <unistd.h>
30
31#include <stdio.h>
32#include <stdlib.h>
33#include <fcntl.h>
34
35#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB)
36#include <cutils/properties.h>
37#include "utils/StopWatch.h"
38#endif
39
40#ifdef RS_SERVER
41// Android exposes gettid(), standard Linux does not
42static pid_t gettid() {
43    return syscall(SYS_gettid);
44}
45#endif
46
47using namespace android;
48using namespace android::renderscript;
49
50typedef void (*outer_foreach_t)(
51    const android::renderscript::RsExpandKernelParams *,
52    uint32_t x1, uint32_t x2, uint32_t outstep);
53
54
55static pthread_key_t gThreadTLSKey = 0;
56static uint32_t gThreadTLSKeyCount = 0;
57static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER;
58
59bool android::renderscript::gArchUseSIMD = false;
60
61RsdCpuReference::~RsdCpuReference() {
62}
63
64RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major,
65        uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn
66#ifndef RS_COMPATIBILITY_LIB
67        , bcc::RSLinkRuntimeCallback pLinkRuntimeCallback,
68        RSSelectRTCallback pSelectRTCallback,
69        const char *pBccPluginName
70#endif
71        ) {
72
73    RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc);
74    if (!cpu) {
75        return nullptr;
76    }
77    if (!cpu->init(version_major, version_minor, lfn, slfn)) {
78        delete cpu;
79        return nullptr;
80    }
81
82#ifndef RS_COMPATIBILITY_LIB
83    cpu->setLinkRuntimeCallback(pLinkRuntimeCallback);
84    cpu->setSelectRTCallback(pSelectRTCallback);
85    if (pBccPluginName) {
86        cpu->setBccPluginName(pBccPluginName);
87    }
88#endif
89
90    return cpu;
91}
92
93
94Context * RsdCpuReference::getTlsContext() {
95    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
96    return tls->mContext;
97}
98
99const Script * RsdCpuReference::getTlsScript() {
100    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
101    return tls->mScript;
102}
103
104pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; }
105
106////////////////////////////////////////////////////////////
107///
108
109RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) {
110    mRSC = rsc;
111
112    version_major = 0;
113    version_minor = 0;
114    mInForEach = false;
115    memset(&mWorkers, 0, sizeof(mWorkers));
116    memset(&mTlsStruct, 0, sizeof(mTlsStruct));
117    mExit = false;
118#ifndef RS_COMPATIBILITY_LIB
119    mLinkRuntimeCallback = nullptr;
120    mSelectRTCallback = nullptr;
121    mSetupCompilerCallback = nullptr;
122#endif
123}
124
125
126void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) {
127    RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc;
128
129    uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1);
130
131    //ALOGV("RS helperThread starting %p idx=%i", dc, idx);
132
133    dc->mWorkers.mLaunchSignals[idx].init();
134    dc->mWorkers.mNativeThreadId[idx] = gettid();
135
136    memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct));
137    int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct);
138    if (status) {
139        ALOGE("pthread_setspecific %i", status);
140    }
141
142#if 0
143    typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
144    cpu_set_t cpuset;
145    memset(&cpuset, 0, sizeof(cpuset));
146    cpuset.bits[idx / 64] |= 1ULL << (idx % 64);
147    int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
148              sizeof(cpuset), &cpuset);
149    ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
150#endif
151
152    while (!dc->mExit) {
153        dc->mWorkers.mLaunchSignals[idx].wait();
154        if (dc->mWorkers.mLaunchCallback) {
155           // idx +1 is used because the calling thread is always worker 0.
156           dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1);
157        }
158        __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1);
159        dc->mWorkers.mCompleteSignal.set();
160    }
161
162    //ALOGV("RS helperThread exited %p idx=%i", dc, idx);
163    return nullptr;
164}
165
166void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) {
167    mWorkers.mLaunchData = data;
168    mWorkers.mLaunchCallback = cbk;
169
170    // fast path for very small launches
171    MTLaunchStruct *mtls = (MTLaunchStruct *)data;
172    if (mtls && mtls->fep.dimY <= 1 && mtls->xEnd <= mtls->xStart + mtls->mSliceSize) {
173        if (mWorkers.mLaunchCallback) {
174            mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
175        }
176        return;
177    }
178
179    mWorkers.mRunningCount = mWorkers.mCount;
180    __sync_synchronize();
181
182    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
183        mWorkers.mLaunchSignals[ct].set();
184    }
185
186    // We use the calling thread as one of the workers so we can start without
187    // the delay of the thread wakeup.
188    if (mWorkers.mLaunchCallback) {
189        mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
190    }
191
192    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
193        mWorkers.mCompleteSignal.wait();
194    }
195}
196
197
198void RsdCpuReferenceImpl::lockMutex() {
199    pthread_mutex_lock(&gInitMutex);
200}
201
202void RsdCpuReferenceImpl::unlockMutex() {
203    pthread_mutex_unlock(&gInitMutex);
204}
205
206static int
207read_file(const char*  pathname, char*  buffer, size_t  buffsize)
208{
209    int  fd, len;
210
211    fd = open(pathname, O_RDONLY);
212    if (fd < 0)
213        return -1;
214
215    do {
216        len = read(fd, buffer, buffsize);
217    } while (len < 0 && errno == EINTR);
218
219    close(fd);
220
221    return len;
222}
223
224static void GetCpuInfo() {
225    char cpuinfo[4096];
226    int  cpuinfo_len;
227
228    cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo);
229    if (cpuinfo_len < 0)  /* should not happen */ {
230        return;
231    }
232
233#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS)
234    gArchUseSIMD = (!!strstr(cpuinfo, " neon")) ||
235                   (!!strstr(cpuinfo, " asimd"));
236#elif defined(ARCH_X86_HAVE_SSSE3)
237    gArchUseSIMD = !!strstr(cpuinfo, " ssse3");
238#endif
239}
240
241bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor,
242                               sym_lookup_t lfn, script_lookup_t slfn) {
243
244    mSymLookupFn = lfn;
245    mScriptLookupFn = slfn;
246
247    lockMutex();
248    if (!gThreadTLSKeyCount) {
249        int status = pthread_key_create(&gThreadTLSKey, nullptr);
250        if (status) {
251            ALOGE("Failed to init thread tls key.");
252            unlockMutex();
253            return false;
254        }
255    }
256    gThreadTLSKeyCount++;
257    unlockMutex();
258
259    mTlsStruct.mContext = mRSC;
260    mTlsStruct.mScript = nullptr;
261    int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct);
262    if (status) {
263        ALOGE("pthread_setspecific %i", status);
264    }
265
266    GetCpuInfo();
267
268    int cpu = sysconf(_SC_NPROCESSORS_CONF);
269    if(mRSC->props.mDebugMaxThreads) {
270        cpu = mRSC->props.mDebugMaxThreads;
271    }
272    if (cpu < 2) {
273        mWorkers.mCount = 0;
274        return true;
275    }
276
277    // Subtract one from the cpu count because we also use the command thread as a worker.
278    mWorkers.mCount = (uint32_t)(cpu - 1);
279
280    ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1);
281
282    mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t));
283    mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t));
284    mWorkers.mLaunchSignals = new Signal[mWorkers.mCount];
285    mWorkers.mLaunchCallback = nullptr;
286
287    mWorkers.mCompleteSignal.init();
288
289    mWorkers.mRunningCount = mWorkers.mCount;
290    mWorkers.mLaunchCount = 0;
291    __sync_synchronize();
292
293    pthread_attr_t threadAttr;
294    status = pthread_attr_init(&threadAttr);
295    if (status) {
296        ALOGE("Failed to init thread attribute.");
297        return false;
298    }
299
300    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
301        status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this);
302        if (status) {
303            mWorkers.mCount = ct;
304            ALOGE("Created fewer than expected number of RS threads.");
305            break;
306        }
307    }
308    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
309        usleep(100);
310    }
311
312    pthread_attr_destroy(&threadAttr);
313    return true;
314}
315
316
317void RsdCpuReferenceImpl::setPriority(int32_t priority) {
318    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
319        setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority);
320    }
321}
322
323RsdCpuReferenceImpl::~RsdCpuReferenceImpl() {
324    mExit = true;
325    mWorkers.mLaunchData = nullptr;
326    mWorkers.mLaunchCallback = nullptr;
327    mWorkers.mRunningCount = mWorkers.mCount;
328    __sync_synchronize();
329    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
330        mWorkers.mLaunchSignals[ct].set();
331    }
332    void *res;
333    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
334        pthread_join(mWorkers.mThreadId[ct], &res);
335    }
336    rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0);
337    free(mWorkers.mThreadId);
338    free(mWorkers.mNativeThreadId);
339    delete[] mWorkers.mLaunchSignals;
340
341    // Global structure cleanup.
342    lockMutex();
343    --gThreadTLSKeyCount;
344    if (!gThreadTLSKeyCount) {
345        pthread_key_delete(gThreadTLSKey);
346    }
347    unlockMutex();
348
349}
350
351typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
352typedef void (*walk_loop_t)(MTLaunchStruct*,
353                            RsExpandKernelParams&,
354                            outer_foreach_t);
355
356
357static void walk_wrapper(void* usr, uint32_t idx, walk_loop_t walk_loop) {
358    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
359
360    uint32_t inLen = mtls->fep.inLen;
361
362    RsExpandKernelParams kparams;
363    kparams.takeFields(mtls->fep);
364
365    // Used by CpuScriptGroup, IntrinsicBlur, and IntrinsicHistogram
366    kparams.lid = idx;
367
368    if (inLen > 0) {
369        // Allocate space for our input base pointers.
370        kparams.ins = (const void**)alloca(inLen * sizeof(void*));
371
372        // Allocate space for our input stride information.
373        kparams.inEStrides = (uint32_t*)alloca(inLen * sizeof(uint32_t));
374
375        // Fill our stride information.
376        for (int inIndex = inLen; --inIndex >= 0;) {
377          kparams.inEStrides[inIndex] = mtls->fep.inStrides[inIndex].eStride;
378        }
379    }
380
381    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
382
383    walk_loop(mtls, kparams, fn);
384}
385
386static void walk_2d(void *usr, uint32_t idx) {
387    walk_wrapper(usr, idx, [](MTLaunchStruct *mtls,
388                              RsExpandKernelParams &kparams,
389                              outer_foreach_t fn) {
390
391        while (1) {
392            uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
393            uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize;
394            uint32_t yEnd   = yStart + mtls->mSliceSize;
395
396            yEnd = rsMin(yEnd, mtls->yEnd);
397
398            if (yEnd <= yStart) {
399                return;
400            }
401
402            for (kparams.y = yStart; kparams.y < yEnd; kparams.y++) {
403                kparams.out = mtls->fep.outPtr +
404                              (mtls->fep.outStride.yStride * kparams.y) +
405                              (mtls->fep.outStride.eStride * mtls->xStart);
406
407                for (int inIndex = mtls->fep.inLen; --inIndex >= 0;) {
408                    StridePair &strides = mtls->fep.inStrides[inIndex];
409
410                    kparams.ins[inIndex] =
411                      mtls->fep.inPtrs[inIndex] +
412                      (strides.yStride * kparams.y) +
413                      (strides.eStride * mtls->xStart);
414                }
415
416                fn(&kparams, mtls->xStart, mtls->xEnd,
417                   mtls->fep.outStride.eStride);
418            }
419        }
420    });
421}
422
423static void walk_1d(void *usr, uint32_t idx) {
424    walk_wrapper(usr, idx, [](MTLaunchStruct *mtls,
425                              RsExpandKernelParams &kparams,
426                              outer_foreach_t fn) {
427
428        while (1) {
429            uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
430            uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize;
431            uint32_t xEnd   = xStart + mtls->mSliceSize;
432
433            xEnd = rsMin(xEnd, mtls->xEnd);
434
435            if (xEnd <= xStart) {
436                return;
437            }
438
439            kparams.out = mtls->fep.outPtr +
440                          (mtls->fep.outStride.eStride * xStart);
441
442            for (int inIndex = mtls->fep.inLen; --inIndex >= 0;) {
443                StridePair &strides = mtls->fep.inStrides[inIndex];
444
445                kparams.ins[inIndex] =
446                  mtls->fep.inPtrs[inIndex] + (strides.eStride * xStart);
447            }
448
449            fn(&kparams, xStart, xEnd, mtls->fep.outStride.eStride);
450        }
451    });
452}
453
454
455void RsdCpuReferenceImpl::launchThreads(const Allocation ** ains,
456                                        uint32_t inLen,
457                                        Allocation* aout,
458                                        const RsScriptCall* sc,
459                                        MTLaunchStruct* mtls) {
460
461    //android::StopWatch kernel_time("kernel time");
462
463    if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) {
464        const size_t targetByteChunk = 16 * 1024;
465        mInForEach = true;
466
467        if (mtls->fep.dimY > 1) {
468            uint32_t s1 = mtls->fep.dimY / ((mWorkers.mCount + 1) * 4);
469            uint32_t s2 = 0;
470
471            // This chooses our slice size to rate limit atomic ops to
472            // one per 16k bytes of reads/writes.
473            if (mtls->fep.outStride.yStride) {
474                s2 = targetByteChunk / mtls->fep.outStride.yStride;
475            } else {
476                // We know that there is either an output or an input.
477                s2 = targetByteChunk / mtls->fep.inStrides[0].yStride;
478            }
479            mtls->mSliceSize = rsMin(s1, s2);
480
481            if(mtls->mSliceSize < 1) {
482                mtls->mSliceSize = 1;
483            }
484
485            launchThreads(walk_2d, mtls);
486        } else {
487            uint32_t s1 = mtls->fep.dimX / ((mWorkers.mCount + 1) * 4);
488            uint32_t s2 = 0;
489
490            // This chooses our slice size to rate limit atomic ops to
491            // one per 16k bytes of reads/writes.
492            if (mtls->fep.outStride.eStride) {
493                s2 = targetByteChunk / mtls->fep.outStride.eStride;
494            } else {
495                // We know that there is either an output or an input.
496                s2 = targetByteChunk / mtls->fep.inStrides[0].eStride;
497            }
498            mtls->mSliceSize = rsMin(s1, s2);
499
500            if (mtls->mSliceSize < 1) {
501                mtls->mSliceSize = 1;
502            }
503
504            launchThreads(walk_1d, mtls);
505        }
506        mInForEach = false;
507
508    } else {
509        RsExpandKernelParams kparams;
510        kparams.takeFields(mtls->fep);
511
512        if (inLen > 0) {
513            // Allocate space for our input base pointers.
514            kparams.ins = (const void**)alloca(inLen * sizeof(void*));
515
516            // Allocate space for our input stride information.
517            kparams.inEStrides = (uint32_t*)alloca(inLen * sizeof(uint32_t));
518
519            // Fill our stride information.
520            for (int inIndex = inLen; --inIndex >= 0;) {
521                kparams.inEStrides[inIndex] =
522                    mtls->fep.inStrides[inIndex].eStride;
523            }
524        }
525
526        //ALOGE("launch 3");
527        outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
528        for (uint32_t arrayIndex = mtls->arrayStart;
529             arrayIndex < mtls->arrayEnd; arrayIndex++) {
530
531            for (kparams.z = mtls->zStart; kparams.z < mtls->zEnd;
532                 kparams.z++) {
533
534                for (kparams.y = mtls->yStart; kparams.y < mtls->yEnd;
535                     kparams.y++) {
536
537                    uint32_t offset =
538                      mtls->fep.dimY * mtls->fep.dimZ * arrayIndex +
539                      mtls->fep.dimY * kparams.z + kparams.y;
540
541                    kparams.out = mtls->fep.outPtr +
542                                  (mtls->fep.outStride.yStride * offset) +
543                                  (mtls->fep.outStride.eStride * mtls->xStart);
544
545                    for (int inIndex = inLen; --inIndex >= 0;) {
546                        StridePair &strides = mtls->fep.inStrides[inIndex];
547
548                        kparams.ins[inIndex] =
549                          mtls->fep.inPtrs[inIndex] +
550                          (strides.yStride * offset) +
551                          (strides.eStride * mtls->xStart);
552                    }
553
554                    fn(&kparams, mtls->xStart, mtls->xEnd,
555                       mtls->fep.outStride.eStride);
556                }
557            }
558        }
559    }
560}
561
562RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) {
563    //ALOGE("setTls %p", sc);
564    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
565    rsAssert(tls);
566    RsdCpuScriptImpl *old = tls->mImpl;
567    tls->mImpl = sc;
568    tls->mContext = mRSC;
569    if (sc) {
570        tls->mScript = sc->getScript();
571    } else {
572        tls->mScript = nullptr;
573    }
574    return old;
575}
576
577const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) {
578    return mSymLookupFn(mRSC, name);
579}
580
581
582RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s,
583                                    char const *resName, char const *cacheDir,
584                                    uint8_t const *bitcode, size_t bitcodeSize,
585                                    uint32_t flags) {
586
587    RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s);
588    if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags
589#ifndef RS_COMPATIBILITY_LIB
590        , getBccPluginName()
591#endif
592        )) {
593        delete i;
594        return nullptr;
595    }
596    return i;
597}
598
599extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx,
600                                             const Script *s, const Element *e);
601extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx,
602                                                   const Script *s, const Element *e);
603extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx,
604                                                   const Script *s, const Element *e);
605extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx,
606                                           const Script *s, const Element *e);
607extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx,
608                                                   const Script *s, const Element *e);
609extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx,
610                                            const Script *s, const Element *e);
611extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx,
612                                                const Script *s, const Element *e);
613extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx,
614                                             const Script *s, const Element *e);
615extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx,
616                                                 const Script *s, const Element *e);
617extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx,
618                                              const Script *s, const Element *e);
619
620RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s,
621                                    RsScriptIntrinsicID iid, Element *e) {
622
623    RsdCpuScriptImpl *i = nullptr;
624    switch (iid) {
625    case RS_SCRIPT_INTRINSIC_ID_3DLUT:
626        i = rsdIntrinsic_3DLUT(this, s, e);
627        break;
628    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3:
629        i = rsdIntrinsic_Convolve3x3(this, s, e);
630        break;
631    case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX:
632        i = rsdIntrinsic_ColorMatrix(this, s, e);
633        break;
634    case RS_SCRIPT_INTRINSIC_ID_LUT:
635        i = rsdIntrinsic_LUT(this, s, e);
636        break;
637    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5:
638        i = rsdIntrinsic_Convolve5x5(this, s, e);
639        break;
640    case RS_SCRIPT_INTRINSIC_ID_BLUR:
641        i = rsdIntrinsic_Blur(this, s, e);
642        break;
643    case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB:
644        i = rsdIntrinsic_YuvToRGB(this, s, e);
645        break;
646    case RS_SCRIPT_INTRINSIC_ID_BLEND:
647        i = rsdIntrinsic_Blend(this, s, e);
648        break;
649    case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM:
650        i = rsdIntrinsic_Histogram(this, s, e);
651        break;
652    case RS_SCRIPT_INTRINSIC_ID_RESIZE:
653        i = rsdIntrinsic_Resize(this, s, e);
654        break;
655
656    default:
657        rsAssert(0);
658    }
659
660    return i;
661}
662
663RsdCpuReference::CpuScriptGroup * RsdCpuReferenceImpl::createScriptGroup(const ScriptGroup *sg) {
664    CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg);
665    if (!sgi->init()) {
666        delete sgi;
667        return nullptr;
668    }
669    return sgi;
670}
671