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