rsCpuCore.cpp revision 962e720b3d1c27bcfec90374ff393584b99577b3 
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
30#ifndef RS_SERVER
31#include <cutils/properties.h>
32#include "utils/StopWatch.h"
33#endif
34
35#ifdef RS_SERVER
36// Android exposes gettid(), standard Linux does not
37static pid_t gettid() {
38    return syscall(SYS_gettid);
39}
40#endif
41
42using namespace android;
43using namespace android::renderscript;
44
45typedef void (*outer_foreach_t)(
46    const android::renderscript::RsForEachStubParamStruct *,
47    uint32_t x1, uint32_t x2,
48    uint32_t instep, uint32_t outstep);
49
50
51static pthread_key_t gThreadTLSKey = 0;
52static uint32_t gThreadTLSKeyCount = 0;
53static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER;
54
55RsdCpuReference::~RsdCpuReference() {
56}
57
58RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major,
59        uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn
60#ifndef RS_COMPATIBILITY_LIB
61        , bcc::RSLinkRuntimeCallback pLinkRuntimeCallback
62#endif
63        ) {
64
65    RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc);
66    if (!cpu) {
67        return NULL;
68    }
69    if (!cpu->init(version_major, version_minor, lfn, slfn)) {
70        delete cpu;
71        return NULL;
72    }
73
74#ifndef RS_COMPATIBILITY_LIB
75    cpu->setLinkRuntimeCallback(pLinkRuntimeCallback);
76#endif
77
78    return cpu;
79}
80
81
82Context * RsdCpuReference::getTlsContext() {
83    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
84    return tls->mContext;
85}
86
87const Script * RsdCpuReference::getTlsScript() {
88    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
89    return tls->mScript;
90}
91
92pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; }
93
94////////////////////////////////////////////////////////////
95///
96
97RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) {
98    mRSC = rsc;
99
100    version_major = 0;
101    version_minor = 0;
102    mInForEach = false;
103    memset(&mWorkers, 0, sizeof(mWorkers));
104    memset(&mTlsStruct, 0, sizeof(mTlsStruct));
105    mExit = false;
106#ifndef RS_COMPATIBILITY_LIB
107    mLinkRuntimeCallback = NULL;
108#endif
109}
110
111
112void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) {
113    RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc;
114
115    uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1);
116
117    //ALOGV("RS helperThread starting %p idx=%i", dc, idx);
118
119    dc->mWorkers.mLaunchSignals[idx].init();
120    dc->mWorkers.mNativeThreadId[idx] = gettid();
121
122    memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct));
123    int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct);
124    if (status) {
125        ALOGE("pthread_setspecific %i", status);
126    }
127
128#if 0
129    typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
130    cpu_set_t cpuset;
131    memset(&cpuset, 0, sizeof(cpuset));
132    cpuset.bits[idx / 64] |= 1ULL << (idx % 64);
133    int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
134              sizeof(cpuset), &cpuset);
135    ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
136#endif
137
138    while (!dc->mExit) {
139        dc->mWorkers.mLaunchSignals[idx].wait();
140        if (dc->mWorkers.mLaunchCallback) {
141           // idx +1 is used because the calling thread is always worker 0.
142           dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1);
143        }
144        __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1);
145        dc->mWorkers.mCompleteSignal.set();
146    }
147
148    //ALOGV("RS helperThread exited %p idx=%i", dc, idx);
149    return NULL;
150}
151
152void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) {
153    mWorkers.mLaunchData = data;
154    mWorkers.mLaunchCallback = cbk;
155
156    // fast path for very small launches
157    MTLaunchStruct *mtls = (MTLaunchStruct *)data;
158    if (mtls && mtls->fep.dimY <= 1 && mtls->xEnd <= mtls->xStart + mtls->mSliceSize) {
159        if (mWorkers.mLaunchCallback) {
160            mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
161        }
162        return;
163    }
164
165    mWorkers.mRunningCount = mWorkers.mCount;
166    __sync_synchronize();
167
168    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
169        mWorkers.mLaunchSignals[ct].set();
170    }
171
172    // We use the calling thread as one of the workers so we can start without
173    // the delay of the thread wakeup.
174    if (mWorkers.mLaunchCallback) {
175        mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
176    }
177
178    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
179        mWorkers.mCompleteSignal.wait();
180    }
181}
182
183
184void RsdCpuReferenceImpl::lockMutex() {
185    pthread_mutex_lock(&gInitMutex);
186}
187
188void RsdCpuReferenceImpl::unlockMutex() {
189    pthread_mutex_unlock(&gInitMutex);
190}
191
192bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor,
193                               sym_lookup_t lfn, script_lookup_t slfn) {
194
195    mSymLookupFn = lfn;
196    mScriptLookupFn = slfn;
197
198    lockMutex();
199    if (!gThreadTLSKeyCount) {
200        int status = pthread_key_create(&gThreadTLSKey, NULL);
201        if (status) {
202            ALOGE("Failed to init thread tls key.");
203            unlockMutex();
204            return false;
205        }
206    }
207    gThreadTLSKeyCount++;
208    unlockMutex();
209
210    mTlsStruct.mContext = mRSC;
211    mTlsStruct.mScript = NULL;
212    int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct);
213    if (status) {
214        ALOGE("pthread_setspecific %i", status);
215    }
216
217    int cpu = sysconf(_SC_NPROCESSORS_ONLN);
218    if(mRSC->props.mDebugMaxThreads) {
219        cpu = mRSC->props.mDebugMaxThreads;
220    }
221    if (cpu < 2) {
222        mWorkers.mCount = 0;
223        return true;
224    }
225
226    // Subtract one from the cpu count because we also use the command thread as a worker.
227    mWorkers.mCount = (uint32_t)(cpu - 1);
228
229    ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1);
230
231    mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t));
232    mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t));
233    mWorkers.mLaunchSignals = new Signal[mWorkers.mCount];
234    mWorkers.mLaunchCallback = NULL;
235
236    mWorkers.mCompleteSignal.init();
237
238    mWorkers.mRunningCount = mWorkers.mCount;
239    mWorkers.mLaunchCount = 0;
240    __sync_synchronize();
241
242    pthread_attr_t threadAttr;
243    status = pthread_attr_init(&threadAttr);
244    if (status) {
245        ALOGE("Failed to init thread attribute.");
246        return false;
247    }
248
249    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
250        status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this);
251        if (status) {
252            mWorkers.mCount = ct;
253            ALOGE("Created fewer than expected number of RS threads.");
254            break;
255        }
256    }
257    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
258        usleep(100);
259    }
260
261    pthread_attr_destroy(&threadAttr);
262    return true;
263}
264
265
266void RsdCpuReferenceImpl::setPriority(int32_t priority) {
267    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
268        setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority);
269    }
270}
271
272RsdCpuReferenceImpl::~RsdCpuReferenceImpl() {
273    mExit = true;
274    mWorkers.mLaunchData = NULL;
275    mWorkers.mLaunchCallback = NULL;
276    mWorkers.mRunningCount = mWorkers.mCount;
277    __sync_synchronize();
278    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
279        mWorkers.mLaunchSignals[ct].set();
280    }
281    void *res;
282    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
283        pthread_join(mWorkers.mThreadId[ct], &res);
284    }
285    rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0);
286
287    // Global structure cleanup.
288    lockMutex();
289    --gThreadTLSKeyCount;
290    if (!gThreadTLSKeyCount) {
291        pthread_key_delete(gThreadTLSKey);
292    }
293    unlockMutex();
294
295}
296
297typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
298
299static void wc_xy(void *usr, uint32_t idx) {
300    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
301    RsForEachStubParamStruct p;
302    memcpy(&p, &mtls->fep, sizeof(p));
303    p.lid = idx;
304    uint32_t sig = mtls->sig;
305
306    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
307    while (1) {
308        uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
309        uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize;
310        uint32_t yEnd = yStart + mtls->mSliceSize;
311        yEnd = rsMin(yEnd, mtls->yEnd);
312        if (yEnd <= yStart) {
313            return;
314        }
315
316        //ALOGE("usr idx %i, x %i,%i  y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
317        //ALOGE("usr ptr in %p,  out %p", mtls->fep.ptrIn, mtls->fep.ptrOut);
318
319        for (p.y = yStart; p.y < yEnd; p.y++) {
320            p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * p.y) +
321                    (mtls->fep.eStrideOut * mtls->xStart);
322            p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * p.y) +
323                   (mtls->fep.eStrideIn * mtls->xStart);
324            fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut);
325        }
326    }
327}
328
329static void wc_x(void *usr, uint32_t idx) {
330    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
331    RsForEachStubParamStruct p;
332    memcpy(&p, &mtls->fep, sizeof(p));
333    p.lid = idx;
334    uint32_t sig = mtls->sig;
335
336    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
337    while (1) {
338        uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
339        uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize;
340        uint32_t xEnd = xStart + mtls->mSliceSize;
341        xEnd = rsMin(xEnd, mtls->xEnd);
342        if (xEnd <= xStart) {
343            return;
344        }
345
346        //ALOGE("usr slice %i idx %i, x %i,%i", slice, idx, xStart, xEnd);
347        //ALOGE("usr ptr in %p,  out %p", mtls->fep.ptrIn, mtls->fep.ptrOut);
348
349        p.out = mtls->fep.ptrOut + (mtls->fep.eStrideOut * xStart);
350        p.in = mtls->fep.ptrIn + (mtls->fep.eStrideIn * xStart);
351        fn(&p, xStart, xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut);
352    }
353}
354
355void RsdCpuReferenceImpl::launchThreads(const Allocation * ain, Allocation * aout,
356                                     const RsScriptCall *sc, MTLaunchStruct *mtls) {
357
358    //android::StopWatch kernel_time("kernel time");
359
360    if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) {
361        const size_t targetByteChunk = 16 * 1024;
362        mInForEach = true;
363        if (mtls->fep.dimY > 1) {
364            uint32_t s1 = mtls->fep.dimY / ((mWorkers.mCount + 1) * 4);
365            uint32_t s2 = 0;
366
367            // This chooses our slice size to rate limit atomic ops to
368            // one per 16k bytes of reads/writes.
369            if (mtls->fep.yStrideOut) {
370                s2 = targetByteChunk / mtls->fep.yStrideOut;
371            } else {
372                s2 = targetByteChunk / mtls->fep.yStrideIn;
373            }
374            mtls->mSliceSize = rsMin(s1, s2);
375
376            if(mtls->mSliceSize < 1) {
377                mtls->mSliceSize = 1;
378            }
379
380         //   mtls->mSliceSize = 2;
381            launchThreads(wc_xy, mtls);
382        } else {
383            uint32_t s1 = mtls->fep.dimX / ((mWorkers.mCount + 1) * 4);
384            uint32_t s2 = 0;
385
386            // This chooses our slice size to rate limit atomic ops to
387            // one per 16k bytes of reads/writes.
388            if (mtls->fep.eStrideOut) {
389                s2 = targetByteChunk / mtls->fep.eStrideOut;
390            } else {
391                s2 = targetByteChunk / mtls->fep.eStrideIn;
392            }
393            mtls->mSliceSize = rsMin(s1, s2);
394
395            if(mtls->mSliceSize < 1) {
396                mtls->mSliceSize = 1;
397            }
398
399            launchThreads(wc_x, mtls);
400        }
401        mInForEach = false;
402
403        //ALOGE("launch 1");
404    } else {
405        RsForEachStubParamStruct p;
406        memcpy(&p, &mtls->fep, sizeof(p));
407        uint32_t sig = mtls->sig;
408
409        //ALOGE("launch 3");
410        outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
411        for (p.ar[0] = mtls->arrayStart; p.ar[0] < mtls->arrayEnd; p.ar[0]++) {
412            for (p.z = mtls->zStart; p.z < mtls->zEnd; p.z++) {
413                for (p.y = mtls->yStart; p.y < mtls->yEnd; p.y++) {
414                    uint32_t offset = mtls->fep.dimY * mtls->fep.dimZ * p.ar[0] +
415                                      mtls->fep.dimY * p.z + p.y;
416                    p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * offset) +
417                            (mtls->fep.eStrideOut * mtls->xStart);
418                    p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * offset) +
419                           (mtls->fep.eStrideIn * mtls->xStart);
420                    fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut);
421                }
422            }
423        }
424    }
425}
426
427RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) {
428    //ALOGE("setTls %p", sc);
429    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
430    rsAssert(tls);
431    RsdCpuScriptImpl *old = tls->mImpl;
432    tls->mImpl = sc;
433    tls->mContext = mRSC;
434    if (sc) {
435        tls->mScript = sc->getScript();
436    } else {
437        tls->mScript = NULL;
438    }
439    return old;
440}
441
442const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) {
443    return mSymLookupFn(mRSC, name);
444}
445
446
447RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s,
448                                    char const *resName, char const *cacheDir,
449                                    uint8_t const *bitcode, size_t bitcodeSize,
450                                    uint32_t flags) {
451
452    RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s);
453    if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags)) {
454        delete i;
455        return NULL;
456    }
457    return i;
458}
459
460extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx,
461                                             const Script *s, const Element *e);
462extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx,
463                                                   const Script *s, const Element *e);
464extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx,
465                                                   const Script *s, const Element *e);
466extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx,
467                                           const Script *s, const Element *e);
468extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx,
469                                                   const Script *s, const Element *e);
470extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx,
471                                            const Script *s, const Element *e);
472extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx,
473                                                const Script *s, const Element *e);
474extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx,
475                                             const Script *s, const Element *e);
476
477RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s,
478                                    RsScriptIntrinsicID iid, Element *e) {
479
480    RsdCpuScriptImpl *i = NULL;
481    switch (iid) {
482    case RS_SCRIPT_INTRINSIC_ID_3DLUT:
483        i = rsdIntrinsic_3DLUT(this, s, e);
484        break;
485    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3:
486        i = rsdIntrinsic_Convolve3x3(this, s, e);
487        break;
488    case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX:
489        i = rsdIntrinsic_ColorMatrix(this, s, e);
490        break;
491    case RS_SCRIPT_INTRINSIC_ID_LUT:
492        i = rsdIntrinsic_LUT(this, s, e);
493        break;
494    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5:
495        i = rsdIntrinsic_Convolve5x5(this, s, e);
496        break;
497    case RS_SCRIPT_INTRINSIC_ID_BLUR:
498        i = rsdIntrinsic_Blur(this, s, e);
499        break;
500    case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB:
501        i = rsdIntrinsic_YuvToRGB(this, s, e);
502        break;
503    case RS_SCRIPT_INTRINSIC_ID_BLEND:
504        i = rsdIntrinsic_Blend(this, s, e);
505        break;
506
507    default:
508        rsAssert(0);
509    }
510
511    return i;
512}
513
514RsdCpuReference::CpuScriptGroup * RsdCpuReferenceImpl::createScriptGroup(const ScriptGroup *sg) {
515    CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg);
516    if (!sgi->init()) {
517        delete sgi;
518        return NULL;
519    }
520    return sgi;
521}
522
523
524