rsCpuCore.cpp revision b0934b67b95cc27e2358c2aa4db5f7c1067c8f9b 
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#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB)
32#include <cutils/properties.h>
33#include "utils/StopWatch.h"
34#endif
35
36#ifdef RS_SERVER
37// Android exposes gettid(), standard Linux does not
38static pid_t gettid() {
39    return syscall(SYS_gettid);
40}
41#endif
42
43using namespace android;
44using namespace android::renderscript;
45
46typedef void (*outer_foreach_t)(
47    const android::renderscript::RsForEachStubParamStruct *,
48    uint32_t x1, uint32_t x2,
49    uint32_t instep, uint32_t outstep);
50
51
52static pthread_key_t gThreadTLSKey = 0;
53static uint32_t gThreadTLSKeyCount = 0;
54static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER;
55
56RsdCpuReference::~RsdCpuReference() {
57}
58
59RsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major,
60        uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn
61#ifndef RS_COMPATIBILITY_LIB
62        , bcc::RSLinkRuntimeCallback pLinkRuntimeCallback,
63        RSSelectRTCallback pSelectRTCallback
64#endif
65        ) {
66
67    RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc);
68    if (!cpu) {
69        return NULL;
70    }
71    if (!cpu->init(version_major, version_minor, lfn, slfn)) {
72        delete cpu;
73        return NULL;
74    }
75
76#ifndef RS_COMPATIBILITY_LIB
77    cpu->setLinkRuntimeCallback(pLinkRuntimeCallback);
78    cpu->setSelectRTCallback(pSelectRTCallback);
79#endif
80
81    return cpu;
82}
83
84
85Context * RsdCpuReference::getTlsContext() {
86    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
87    return tls->mContext;
88}
89
90const Script * RsdCpuReference::getTlsScript() {
91    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
92    return tls->mScript;
93}
94
95pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; }
96
97////////////////////////////////////////////////////////////
98///
99
100RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) {
101    mRSC = rsc;
102
103    version_major = 0;
104    version_minor = 0;
105    mInForEach = false;
106    memset(&mWorkers, 0, sizeof(mWorkers));
107    memset(&mTlsStruct, 0, sizeof(mTlsStruct));
108    mExit = false;
109#ifndef RS_COMPATIBILITY_LIB
110    mLinkRuntimeCallback = NULL;
111    mSelectRTCallback = NULL;
112#endif
113}
114
115
116void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) {
117    RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc;
118
119    uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1);
120
121    //ALOGV("RS helperThread starting %p idx=%i", dc, idx);
122
123    dc->mWorkers.mLaunchSignals[idx].init();
124    dc->mWorkers.mNativeThreadId[idx] = gettid();
125
126    memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct));
127    int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct);
128    if (status) {
129        ALOGE("pthread_setspecific %i", status);
130    }
131
132#if 0
133    typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
134    cpu_set_t cpuset;
135    memset(&cpuset, 0, sizeof(cpuset));
136    cpuset.bits[idx / 64] |= 1ULL << (idx % 64);
137    int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
138              sizeof(cpuset), &cpuset);
139    ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
140#endif
141
142    while (!dc->mExit) {
143        dc->mWorkers.mLaunchSignals[idx].wait();
144        if (dc->mWorkers.mLaunchCallback) {
145           // idx +1 is used because the calling thread is always worker 0.
146           dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1);
147        }
148        __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1);
149        dc->mWorkers.mCompleteSignal.set();
150    }
151
152    //ALOGV("RS helperThread exited %p idx=%i", dc, idx);
153    return NULL;
154}
155
156void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) {
157    mWorkers.mLaunchData = data;
158    mWorkers.mLaunchCallback = cbk;
159
160    // fast path for very small launches
161    MTLaunchStruct *mtls = (MTLaunchStruct *)data;
162    if (mtls && mtls->fep.dimY <= 1 && mtls->xEnd <= mtls->xStart + mtls->mSliceSize) {
163        if (mWorkers.mLaunchCallback) {
164            mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
165        }
166        return;
167    }
168
169    mWorkers.mRunningCount = mWorkers.mCount;
170    __sync_synchronize();
171
172    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
173        mWorkers.mLaunchSignals[ct].set();
174    }
175
176    // We use the calling thread as one of the workers so we can start without
177    // the delay of the thread wakeup.
178    if (mWorkers.mLaunchCallback) {
179        mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0);
180    }
181
182    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
183        mWorkers.mCompleteSignal.wait();
184    }
185}
186
187
188void RsdCpuReferenceImpl::lockMutex() {
189    pthread_mutex_lock(&gInitMutex);
190}
191
192void RsdCpuReferenceImpl::unlockMutex() {
193    pthread_mutex_unlock(&gInitMutex);
194}
195
196bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor,
197                               sym_lookup_t lfn, script_lookup_t slfn) {
198
199    mSymLookupFn = lfn;
200    mScriptLookupFn = slfn;
201
202    lockMutex();
203    if (!gThreadTLSKeyCount) {
204        int status = pthread_key_create(&gThreadTLSKey, NULL);
205        if (status) {
206            ALOGE("Failed to init thread tls key.");
207            unlockMutex();
208            return false;
209        }
210    }
211    gThreadTLSKeyCount++;
212    unlockMutex();
213
214    mTlsStruct.mContext = mRSC;
215    mTlsStruct.mScript = NULL;
216    int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct);
217    if (status) {
218        ALOGE("pthread_setspecific %i", status);
219    }
220
221    int cpu = sysconf(_SC_NPROCESSORS_ONLN);
222    if(mRSC->props.mDebugMaxThreads) {
223        cpu = mRSC->props.mDebugMaxThreads;
224    }
225    if (cpu < 2) {
226        mWorkers.mCount = 0;
227        return true;
228    }
229
230    // Subtract one from the cpu count because we also use the command thread as a worker.
231    mWorkers.mCount = (uint32_t)(cpu - 1);
232
233    ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1);
234
235    mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t));
236    mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t));
237    mWorkers.mLaunchSignals = new Signal[mWorkers.mCount];
238    mWorkers.mLaunchCallback = NULL;
239
240    mWorkers.mCompleteSignal.init();
241
242    mWorkers.mRunningCount = mWorkers.mCount;
243    mWorkers.mLaunchCount = 0;
244    __sync_synchronize();
245
246    pthread_attr_t threadAttr;
247    status = pthread_attr_init(&threadAttr);
248    if (status) {
249        ALOGE("Failed to init thread attribute.");
250        return false;
251    }
252
253    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
254        status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this);
255        if (status) {
256            mWorkers.mCount = ct;
257            ALOGE("Created fewer than expected number of RS threads.");
258            break;
259        }
260    }
261    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
262        usleep(100);
263    }
264
265    pthread_attr_destroy(&threadAttr);
266    return true;
267}
268
269
270void RsdCpuReferenceImpl::setPriority(int32_t priority) {
271    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
272        setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority);
273    }
274}
275
276RsdCpuReferenceImpl::~RsdCpuReferenceImpl() {
277    mExit = true;
278    mWorkers.mLaunchData = NULL;
279    mWorkers.mLaunchCallback = NULL;
280    mWorkers.mRunningCount = mWorkers.mCount;
281    __sync_synchronize();
282    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
283        mWorkers.mLaunchSignals[ct].set();
284    }
285    void *res;
286    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
287        pthread_join(mWorkers.mThreadId[ct], &res);
288    }
289    rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0);
290
291    // Global structure cleanup.
292    lockMutex();
293    --gThreadTLSKeyCount;
294    if (!gThreadTLSKeyCount) {
295        pthread_key_delete(gThreadTLSKey);
296    }
297    unlockMutex();
298
299}
300
301typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
302
303static void wc_xy(void *usr, uint32_t idx) {
304    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
305    RsForEachStubParamStruct p;
306    memcpy(&p, &mtls->fep, sizeof(p));
307    p.lid = idx;
308    uint32_t sig = mtls->sig;
309
310    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
311    while (1) {
312        uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
313        uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize;
314        uint32_t yEnd = yStart + mtls->mSliceSize;
315        yEnd = rsMin(yEnd, mtls->yEnd);
316        if (yEnd <= yStart) {
317            return;
318        }
319
320        //ALOGE("usr idx %i, x %i,%i  y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
321        //ALOGE("usr ptr in %p,  out %p", mtls->fep.ptrIn, mtls->fep.ptrOut);
322
323        for (p.y = yStart; p.y < yEnd; p.y++) {
324            p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * p.y) +
325                    (mtls->fep.eStrideOut * mtls->xStart);
326            p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * p.y) +
327                   (mtls->fep.eStrideIn * mtls->xStart);
328            fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut);
329        }
330    }
331}
332
333static void wc_x(void *usr, uint32_t idx) {
334    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
335    RsForEachStubParamStruct p;
336    memcpy(&p, &mtls->fep, sizeof(p));
337    p.lid = idx;
338    uint32_t sig = mtls->sig;
339
340    outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
341    while (1) {
342        uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
343        uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize;
344        uint32_t xEnd = xStart + mtls->mSliceSize;
345        xEnd = rsMin(xEnd, mtls->xEnd);
346        if (xEnd <= xStart) {
347            return;
348        }
349
350        //ALOGE("usr slice %i idx %i, x %i,%i", slice, idx, xStart, xEnd);
351        //ALOGE("usr ptr in %p,  out %p", mtls->fep.ptrIn, mtls->fep.ptrOut);
352
353        p.out = mtls->fep.ptrOut + (mtls->fep.eStrideOut * xStart);
354        p.in = mtls->fep.ptrIn + (mtls->fep.eStrideIn * xStart);
355        fn(&p, xStart, xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut);
356    }
357}
358
359void RsdCpuReferenceImpl::launchThreads(const Allocation * ain, Allocation * aout,
360                                     const RsScriptCall *sc, MTLaunchStruct *mtls) {
361
362    //android::StopWatch kernel_time("kernel time");
363
364    if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) {
365        const size_t targetByteChunk = 16 * 1024;
366        mInForEach = true;
367        if (mtls->fep.dimY > 1) {
368            uint32_t s1 = mtls->fep.dimY / ((mWorkers.mCount + 1) * 4);
369            uint32_t s2 = 0;
370
371            // This chooses our slice size to rate limit atomic ops to
372            // one per 16k bytes of reads/writes.
373            if (mtls->fep.yStrideOut) {
374                s2 = targetByteChunk / mtls->fep.yStrideOut;
375            } else {
376                s2 = targetByteChunk / mtls->fep.yStrideIn;
377            }
378            mtls->mSliceSize = rsMin(s1, s2);
379
380            if(mtls->mSliceSize < 1) {
381                mtls->mSliceSize = 1;
382            }
383
384         //   mtls->mSliceSize = 2;
385            launchThreads(wc_xy, mtls);
386        } else {
387            uint32_t s1 = mtls->fep.dimX / ((mWorkers.mCount + 1) * 4);
388            uint32_t s2 = 0;
389
390            // This chooses our slice size to rate limit atomic ops to
391            // one per 16k bytes of reads/writes.
392            if (mtls->fep.eStrideOut) {
393                s2 = targetByteChunk / mtls->fep.eStrideOut;
394            } else {
395                s2 = targetByteChunk / mtls->fep.eStrideIn;
396            }
397            mtls->mSliceSize = rsMin(s1, s2);
398
399            if(mtls->mSliceSize < 1) {
400                mtls->mSliceSize = 1;
401            }
402
403            launchThreads(wc_x, mtls);
404        }
405        mInForEach = false;
406
407        //ALOGE("launch 1");
408    } else {
409        RsForEachStubParamStruct p;
410        memcpy(&p, &mtls->fep, sizeof(p));
411        uint32_t sig = mtls->sig;
412
413        //ALOGE("launch 3");
414        outer_foreach_t fn = (outer_foreach_t) mtls->kernel;
415        for (p.ar[0] = mtls->arrayStart; p.ar[0] < mtls->arrayEnd; p.ar[0]++) {
416            for (p.z = mtls->zStart; p.z < mtls->zEnd; p.z++) {
417                for (p.y = mtls->yStart; p.y < mtls->yEnd; p.y++) {
418                    uint32_t offset = mtls->fep.dimY * mtls->fep.dimZ * p.ar[0] +
419                                      mtls->fep.dimY * p.z + p.y;
420                    p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * offset) +
421                            (mtls->fep.eStrideOut * mtls->xStart);
422                    p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * offset) +
423                           (mtls->fep.eStrideIn * mtls->xStart);
424                    fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut);
425                }
426            }
427        }
428    }
429}
430
431RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) {
432    //ALOGE("setTls %p", sc);
433    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
434    rsAssert(tls);
435    RsdCpuScriptImpl *old = tls->mImpl;
436    tls->mImpl = sc;
437    tls->mContext = mRSC;
438    if (sc) {
439        tls->mScript = sc->getScript();
440    } else {
441        tls->mScript = NULL;
442    }
443    return old;
444}
445
446const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) {
447    return mSymLookupFn(mRSC, name);
448}
449
450
451RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s,
452                                    char const *resName, char const *cacheDir,
453                                    uint8_t const *bitcode, size_t bitcodeSize,
454                                    uint32_t flags) {
455
456    RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s);
457    if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags)) {
458        delete i;
459        return NULL;
460    }
461    return i;
462}
463
464extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx,
465                                             const Script *s, const Element *e);
466extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx,
467                                                   const Script *s, const Element *e);
468extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx,
469                                                   const Script *s, const Element *e);
470extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx,
471                                           const Script *s, const Element *e);
472extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx,
473                                                   const Script *s, const Element *e);
474extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx,
475                                            const Script *s, const Element *e);
476extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx,
477                                                const Script *s, const Element *e);
478extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx,
479                                             const Script *s, const Element *e);
480
481RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s,
482                                    RsScriptIntrinsicID iid, Element *e) {
483
484    RsdCpuScriptImpl *i = NULL;
485    switch (iid) {
486    case RS_SCRIPT_INTRINSIC_ID_3DLUT:
487        i = rsdIntrinsic_3DLUT(this, s, e);
488        break;
489    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3:
490        i = rsdIntrinsic_Convolve3x3(this, s, e);
491        break;
492    case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX:
493        i = rsdIntrinsic_ColorMatrix(this, s, e);
494        break;
495    case RS_SCRIPT_INTRINSIC_ID_LUT:
496        i = rsdIntrinsic_LUT(this, s, e);
497        break;
498    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5:
499        i = rsdIntrinsic_Convolve5x5(this, s, e);
500        break;
501    case RS_SCRIPT_INTRINSIC_ID_BLUR:
502        i = rsdIntrinsic_Blur(this, s, e);
503        break;
504    case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB:
505        i = rsdIntrinsic_YuvToRGB(this, s, e);
506        break;
507    case RS_SCRIPT_INTRINSIC_ID_BLEND:
508        i = rsdIntrinsic_Blend(this, s, e);
509        break;
510
511    default:
512        rsAssert(0);
513    }
514
515    return i;
516}
517
518RsdCpuReference::CpuScriptGroup * RsdCpuReferenceImpl::createScriptGroup(const ScriptGroup *sg) {
519    CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg);
520    if (!sgi->init()) {
521        delete sgi;
522        return NULL;
523    }
524    return sgi;
525}
526
527
528