rsCpuCore.cpp revision 415a249300cb9ca3f560814bfde52c6796fed1f7 
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 <stdio.h>
30#include <string.h>
31#include <unistd.h>
32
33#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB)
34#include <cutils/properties.h>
35#include "utils/StopWatch.h"
36#endif
37
38#ifdef RS_SERVER
39// Android exposes gettid(), standard Linux does not
40static pid_t gettid() {
41    return syscall(SYS_gettid);
42}
43#endif
44
45using namespace android;
46using namespace android::renderscript;
47
48#define REDUCE_NEW_ALOGV(mtls, level, ...) do { if ((mtls)->logReduce >= (level)) ALOGV(__VA_ARGS__); } while(0)
49
50static pthread_key_t gThreadTLSKey = 0;
51static uint32_t gThreadTLSKeyCount = 0;
52static pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER;
53
54bool android::renderscript::gArchUseSIMD = false;
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        , RSSelectRTCallback pSelectRTCallback,
62        const char *pBccPluginName
63        ) {
64
65    RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc);
66    if (!cpu) {
67        return nullptr;
68    }
69    if (!cpu->init(version_major, version_minor, lfn, slfn)) {
70        delete cpu;
71        return nullptr;
72    }
73
74    cpu->setSelectRTCallback(pSelectRTCallback);
75    if (pBccPluginName) {
76        cpu->setBccPluginName(pBccPluginName);
77    }
78
79    return cpu;
80}
81
82
83Context * RsdCpuReference::getTlsContext() {
84    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
85    return tls->mContext;
86}
87
88const Script * RsdCpuReference::getTlsScript() {
89    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
90    return tls->mScript;
91}
92
93pthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; }
94
95////////////////////////////////////////////////////////////
96///
97
98RsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) {
99    mRSC = rsc;
100
101    version_major = 0;
102    version_minor = 0;
103    mInKernel = false;
104    memset(&mWorkers, 0, sizeof(mWorkers));
105    memset(&mTlsStruct, 0, sizeof(mTlsStruct));
106    mExit = false;
107    mSelectRTCallback = nullptr;
108    mEmbedGlobalInfo = true;
109    mEmbedGlobalInfoSkipConstant = true;
110}
111
112
113void * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) {
114    RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc;
115
116    uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1);
117
118    //ALOGV("RS helperThread starting %p idx=%i", dc, idx);
119
120    dc->mWorkers.mLaunchSignals[idx].init();
121    dc->mWorkers.mNativeThreadId[idx] = gettid();
122
123    memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct));
124    int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct);
125    if (status) {
126        ALOGE("pthread_setspecific %i", status);
127    }
128
129#if 0
130    typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
131    cpu_set_t cpuset;
132    memset(&cpuset, 0, sizeof(cpuset));
133    cpuset.bits[idx / 64] |= 1ULL << (idx % 64);
134    int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
135              sizeof(cpuset), &cpuset);
136    ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
137#endif
138
139    while (!dc->mExit) {
140        dc->mWorkers.mLaunchSignals[idx].wait();
141        if (dc->mWorkers.mLaunchCallback) {
142           // idx +1 is used because the calling thread is always worker 0.
143           dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1);
144        }
145        __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1);
146        dc->mWorkers.mCompleteSignal.set();
147    }
148
149    //ALOGV("RS helperThread exited %p idx=%i", dc, idx);
150    return nullptr;
151}
152
153// Launch a kernel.
154// The callback function is called to execute the kernel.
155void RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) {
156    mWorkers.mLaunchData = data;
157    mWorkers.mLaunchCallback = cbk;
158
159    // fast path for very small launches
160    MTLaunchStructCommon *mtls = (MTLaunchStructCommon *)data;
161    if (mtls && mtls->dimPtr->y <= 1 && mtls->end.x <= mtls->start.x + 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
187void RsdCpuReferenceImpl::lockMutex() {
188    pthread_mutex_lock(&gInitMutex);
189}
190
191void RsdCpuReferenceImpl::unlockMutex() {
192    pthread_mutex_unlock(&gInitMutex);
193}
194
195// Determine if the CPU we're running on supports SIMD instructions.
196static void GetCpuInfo() {
197    // Read the CPU flags from /proc/cpuinfo.
198    FILE *cpuinfo = fopen("/proc/cpuinfo", "r");
199
200    if (!cpuinfo) {
201        return;
202    }
203
204    char cpuinfostr[4096];
205    // fgets() ends with newline or EOF, need to check the whole
206    // "cpuinfo" file to make sure we can use SIMD or not.
207    while (fgets(cpuinfostr, sizeof(cpuinfostr), cpuinfo)) {
208#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS)
209        gArchUseSIMD = strstr(cpuinfostr, " neon") || strstr(cpuinfostr, " asimd");
210#elif defined(ARCH_X86_HAVE_SSSE3)
211        gArchUseSIMD = strstr(cpuinfostr, " ssse3");
212#endif
213        if (gArchUseSIMD) {
214            break;
215        }
216    }
217    fclose(cpuinfo);
218}
219
220bool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor,
221                               sym_lookup_t lfn, script_lookup_t slfn) {
222    mSymLookupFn = lfn;
223    mScriptLookupFn = slfn;
224
225    lockMutex();
226    if (!gThreadTLSKeyCount) {
227        int status = pthread_key_create(&gThreadTLSKey, nullptr);
228        if (status) {
229            ALOGE("Failed to init thread tls key.");
230            unlockMutex();
231            return false;
232        }
233    }
234    gThreadTLSKeyCount++;
235    unlockMutex();
236
237    mTlsStruct.mContext = mRSC;
238    mTlsStruct.mScript = nullptr;
239    int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct);
240    if (status) {
241        ALOGE("pthread_setspecific %i", status);
242    }
243
244    mPageSize = sysconf(_SC_PAGE_SIZE);
245    // ALOGV("page size = %ld", mPageSize);
246
247    GetCpuInfo();
248
249    int cpu = sysconf(_SC_NPROCESSORS_CONF);
250    if(mRSC->props.mDebugMaxThreads) {
251        cpu = mRSC->props.mDebugMaxThreads;
252    }
253    if (cpu < 2) {
254        mWorkers.mCount = 0;
255        return true;
256    }
257
258    // Subtract one from the cpu count because we also use the command thread as a worker.
259    mWorkers.mCount = (uint32_t)(cpu - 1);
260
261    ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1);
262
263    mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t));
264    mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t));
265    mWorkers.mLaunchSignals = new Signal[mWorkers.mCount];
266    mWorkers.mLaunchCallback = nullptr;
267
268    mWorkers.mCompleteSignal.init();
269
270    mWorkers.mRunningCount = mWorkers.mCount;
271    mWorkers.mLaunchCount = 0;
272    __sync_synchronize();
273
274    pthread_attr_t threadAttr;
275    status = pthread_attr_init(&threadAttr);
276    if (status) {
277        ALOGE("Failed to init thread attribute.");
278        return false;
279    }
280
281    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
282        status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this);
283        if (status) {
284            mWorkers.mCount = ct;
285            ALOGE("Created fewer than expected number of RS threads.");
286            break;
287        }
288    }
289    while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) {
290        usleep(100);
291    }
292
293    pthread_attr_destroy(&threadAttr);
294    return true;
295}
296
297
298void RsdCpuReferenceImpl::setPriority(int32_t priority) {
299    for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
300        setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority);
301    }
302}
303
304RsdCpuReferenceImpl::~RsdCpuReferenceImpl() {
305    mExit = true;
306    mWorkers.mLaunchData = nullptr;
307    mWorkers.mLaunchCallback = nullptr;
308    mWorkers.mRunningCount = mWorkers.mCount;
309    __sync_synchronize();
310    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
311        mWorkers.mLaunchSignals[ct].set();
312    }
313    void *res;
314    for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
315        pthread_join(mWorkers.mThreadId[ct], &res);
316    }
317    rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0);
318    free(mWorkers.mThreadId);
319    free(mWorkers.mNativeThreadId);
320    delete[] mWorkers.mLaunchSignals;
321
322    // Global structure cleanup.
323    lockMutex();
324    --gThreadTLSKeyCount;
325    if (!gThreadTLSKeyCount) {
326        pthread_key_delete(gThreadTLSKey);
327    }
328    unlockMutex();
329
330}
331
332// Set up the appropriate input and output pointers to the kernel driver info structure.
333// Inputs:
334//   mtls - The MTLaunchStruct holding information about the kernel launch
335//   fep - The forEach parameters (driver info structure)
336//   x, y, z, lod, face, a1, a2, a3, a4 - The start offsets into each dimension
337static inline void FepPtrSetup(const MTLaunchStructForEach *mtls, RsExpandKernelDriverInfo *fep,
338                               uint32_t x, uint32_t y,
339                               uint32_t z = 0, uint32_t lod = 0,
340                               RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X,
341                               uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) {
342    for (uint32_t i = 0; i < fep->inLen; i++) {
343        fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4);
344    }
345    if (mtls->aout[0] != nullptr) {
346        fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4);
347    }
348}
349
350// Set up the appropriate input and output pointers to the kernel driver info structure.
351// Inputs:
352//   mtls - The MTLaunchStruct holding information about the kernel launch
353//   redp - The reduce parameters (driver info structure)
354//   x, y, z - The start offsets into each dimension
355static inline void RedpPtrSetup(const MTLaunchStructReduceNew *mtls, RsExpandKernelDriverInfo *redp,
356                                uint32_t x, uint32_t y, uint32_t z) {
357    for (uint32_t i = 0; i < redp->inLen; i++) {
358        redp->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z);
359    }
360}
361
362static uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) {
363    if (start >= end) {
364        *p = start;
365        return val;
366    }
367
368    uint32_t div = end - start;
369
370    uint32_t n = val / div;
371    *p = (val - (n * div)) + start;
372    return n;
373}
374
375static bool SelectOuterSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) {
376    uint32_t r = sliceNum;
377    r = sliceInt(&info->current.z, r, mtls->start.z, mtls->end.z);
378    r = sliceInt(&info->current.lod, r, mtls->start.lod, mtls->end.lod);
379    r = sliceInt(&info->current.face, r, mtls->start.face, mtls->end.face);
380    r = sliceInt(&info->current.array[0], r, mtls->start.array[0], mtls->end.array[0]);
381    r = sliceInt(&info->current.array[1], r, mtls->start.array[1], mtls->end.array[1]);
382    r = sliceInt(&info->current.array[2], r, mtls->start.array[2], mtls->end.array[2]);
383    r = sliceInt(&info->current.array[3], r, mtls->start.array[3], mtls->end.array[3]);
384    return r == 0;
385}
386
387static bool SelectZSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) {
388    return sliceInt(&info->current.z, sliceNum, mtls->start.z, mtls->end.z) == 0;
389}
390
391static void walk_general_foreach(void *usr, uint32_t idx) {
392    MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr;
393    RsExpandKernelDriverInfo fep = mtls->fep;
394    fep.lid = idx;
395    ForEachFunc_t fn = mtls->kernel;
396
397    while(1) {
398        uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
399
400        if (!SelectOuterSlice(mtls, &fep, slice)) {
401            return;
402        }
403
404        for (fep.current.y = mtls->start.y; fep.current.y < mtls->end.y;
405             fep.current.y++) {
406
407            FepPtrSetup(mtls, &fep, mtls->start.x,
408                        fep.current.y, fep.current.z, fep.current.lod,
409                        (RsAllocationCubemapFace)fep.current.face,
410                        fep.current.array[0], fep.current.array[1],
411                        fep.current.array[2], fep.current.array[3]);
412
413            fn(&fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]);
414        }
415    }
416}
417
418static void walk_2d_foreach(void *usr, uint32_t idx) {
419    MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr;
420    RsExpandKernelDriverInfo fep = mtls->fep;
421    fep.lid = idx;
422    ForEachFunc_t fn = mtls->kernel;
423
424    while (1) {
425        uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
426        uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize;
427        uint32_t yEnd   = yStart + mtls->mSliceSize;
428
429        yEnd = rsMin(yEnd, mtls->end.y);
430
431        if (yEnd <= yStart) {
432            return;
433        }
434
435        for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) {
436            FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y);
437
438            fn(&fep, mtls->start.x, mtls->end.x, fep.outStride[0]);
439        }
440    }
441}
442
443static void walk_1d_foreach(void *usr, uint32_t idx) {
444    MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr;
445    RsExpandKernelDriverInfo fep = mtls->fep;
446    fep.lid = idx;
447    ForEachFunc_t fn = mtls->kernel;
448
449    while (1) {
450        uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
451        uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize;
452        uint32_t xEnd   = xStart + mtls->mSliceSize;
453
454        xEnd = rsMin(xEnd, mtls->end.x);
455
456        if (xEnd <= xStart) {
457            return;
458        }
459
460        FepPtrSetup(mtls, &fep, xStart, 0);
461
462        fn(&fep, xStart, xEnd, fep.outStride[0]);
463    }
464}
465
466// The function format_bytes() is an auxiliary function to assist in logging.
467//
468// Bytes are read from an input (inBuf) and written (as pairs of hex digits)
469// to an output (outBuf).
470//
471// Output format:
472// - starts with ": "
473// - each input byte is translated to a pair of hex digits
474// - bytes are separated by "." except that every fourth separator is "|"
475// - if the input is sufficiently long, the output is truncated and terminated with "..."
476//
477// Arguments:
478// - outBuf  -- Pointer to buffer of type "FormatBuf" into which output is written
479// - inBuf   -- Pointer to bytes which are to be formatted into outBuf
480// - inBytes -- Number of bytes in inBuf
481//
482// Constant:
483// - kFormatInBytesMax -- Only min(kFormatInBytesMax, inBytes) bytes will be read
484//                        from inBuf
485//
486// Return value:
487// - pointer (const char *) to output (which is part of outBuf)
488//
489static const int kFormatInBytesMax = 16;
490// ": " + 2 digits per byte + 1 separator between bytes + "..." + null
491typedef char FormatBuf[2 + kFormatInBytesMax*2 + (kFormatInBytesMax - 1) + 3 + 1];
492static const char *format_bytes(FormatBuf *outBuf, const uint8_t *inBuf, const int inBytes) {
493  strcpy(*outBuf, ": ");
494  int pos = 2;
495  const int lim = std::min(kFormatInBytesMax, inBytes);
496  for (int i = 0; i < lim; ++i) {
497    if (i) {
498      sprintf(*outBuf + pos, (i % 4 ? "." : "|"));
499      ++pos;
500    }
501    sprintf(*outBuf + pos, "%02x", inBuf[i]);
502    pos += 2;
503  }
504  if (kFormatInBytesMax < inBytes)
505    strcpy(*outBuf + pos, "...");
506  return *outBuf;
507}
508
509static void reduce_new_get_accumulator(uint8_t *&accumPtr, const MTLaunchStructReduceNew *mtls,
510                                       const char *walkerName, uint32_t threadIdx) {
511  rsAssert(!accumPtr);
512
513  uint32_t accumIdx = (uint32_t)__sync_fetch_and_add(&mtls->accumCount, 1);
514  if (mtls->outFunc) {
515    accumPtr = mtls->accumAlloc + mtls->accumStride * accumIdx;
516  } else {
517    if (accumIdx == 0) {
518      accumPtr = mtls->redp.outPtr[0];
519    } else {
520      accumPtr = mtls->accumAlloc + mtls->accumStride * (accumIdx - 1);
521    }
522  }
523  REDUCE_NEW_ALOGV(mtls, 2, "%s(%p): idx = %u got accumCount %u and accumPtr %p",
524                   walkerName, mtls->accumFunc, threadIdx, accumIdx, accumPtr);
525  // initialize accumulator
526  if (mtls->initFunc) {
527    mtls->initFunc(accumPtr);
528  } else {
529    memset(accumPtr, 0, mtls->accumSize);
530  }
531}
532
533static void walk_1d_reduce_new(void *usr, uint32_t idx) {
534  const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr;
535  RsExpandKernelDriverInfo redp = mtls->redp;
536
537  // find accumulator
538  uint8_t *&accumPtr = mtls->accumPtr[idx];
539  if (!accumPtr) {
540    reduce_new_get_accumulator(accumPtr, mtls, __func__, idx);
541  }
542
543  // accumulate
544  const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc;
545  while (1) {
546    uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
547    uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize;
548    uint32_t xEnd   = xStart + mtls->mSliceSize;
549
550    xEnd = rsMin(xEnd, mtls->end.x);
551
552    if (xEnd <= xStart) {
553      return;
554    }
555
556    RedpPtrSetup(mtls, &redp, xStart, 0, 0);
557    fn(&redp, xStart, xEnd, accumPtr);
558
559    // Emit log line after slice has been run, so that we can include
560    // the results of the run on that line.
561    FormatBuf fmt;
562    if (mtls->logReduce >= 3) {
563      format_bytes(&fmt, accumPtr, mtls->accumSize);
564    } else {
565      fmt[0] = 0;
566    }
567    REDUCE_NEW_ALOGV(mtls, 2, "walk_1d_reduce_new(%p): idx = %u, x in [%u, %u)%s",
568                     mtls->accumFunc, idx, xStart, xEnd, fmt);
569  }
570}
571
572static void walk_2d_reduce_new(void *usr, uint32_t idx) {
573  const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr;
574  RsExpandKernelDriverInfo redp = mtls->redp;
575
576  // find accumulator
577  uint8_t *&accumPtr = mtls->accumPtr[idx];
578  if (!accumPtr) {
579    reduce_new_get_accumulator(accumPtr, mtls, __func__, idx);
580  }
581
582  // accumulate
583  const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc;
584  while (1) {
585    uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
586    uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize;
587    uint32_t yEnd   = yStart + mtls->mSliceSize;
588
589    yEnd = rsMin(yEnd, mtls->end.y);
590
591    if (yEnd <= yStart) {
592      return;
593    }
594
595    for (redp.current.y = yStart; redp.current.y < yEnd; redp.current.y++) {
596      RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, 0);
597      fn(&redp, mtls->start.x, mtls->end.x, accumPtr);
598    }
599
600    FormatBuf fmt;
601    if (mtls->logReduce >= 3) {
602      format_bytes(&fmt, accumPtr, mtls->accumSize);
603    } else {
604      fmt[0] = 0;
605    }
606    REDUCE_NEW_ALOGV(mtls, 2, "walk_2d_reduce_new(%p): idx = %u, y in [%u, %u)%s",
607                     mtls->accumFunc, idx, yStart, yEnd, fmt);
608  }
609}
610
611static void walk_3d_reduce_new(void *usr, uint32_t idx) {
612  const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr;
613  RsExpandKernelDriverInfo redp = mtls->redp;
614
615  // find accumulator
616  uint8_t *&accumPtr = mtls->accumPtr[idx];
617  if (!accumPtr) {
618    reduce_new_get_accumulator(accumPtr, mtls, __func__, idx);
619  }
620
621  // accumulate
622  const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc;
623  while (1) {
624    uint32_t slice  = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1);
625
626    if (!SelectZSlice(mtls, &redp, slice)) {
627      return;
628    }
629
630    for (redp.current.y = mtls->start.y; redp.current.y < mtls->end.y; redp.current.y++) {
631      RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, redp.current.z);
632      fn(&redp, mtls->start.x, mtls->end.x, accumPtr);
633    }
634
635    FormatBuf fmt;
636    if (mtls->logReduce >= 3) {
637      format_bytes(&fmt, accumPtr, mtls->accumSize);
638    } else {
639      fmt[0] = 0;
640    }
641    REDUCE_NEW_ALOGV(mtls, 2, "walk_3d_reduce_new(%p): idx = %u, z = %u%s",
642                     mtls->accumFunc, idx, redp.current.z, fmt);
643  }
644}
645
646// Launch a simple reduce-style kernel.
647// Inputs:
648//  ain:  The allocation that contains the input
649//  aout: The allocation that will hold the output
650//  mtls: Holds launch parameters
651void RsdCpuReferenceImpl::launchReduce(const Allocation *ain,
652                                       Allocation *aout,
653                                       MTLaunchStructReduce *mtls) {
654    const uint32_t xStart = mtls->start.x;
655    const uint32_t xEnd = mtls->end.x;
656
657    if (xStart >= xEnd) {
658      return;
659    }
660
661    const uint32_t startOffset = ain->getType()->getElementSizeBytes() * xStart;
662    mtls->kernel(&mtls->inBuf[startOffset], mtls->outBuf, xEnd - xStart);
663}
664
665// Launch a general reduce-style kernel.
666// Inputs:
667//   ains[0..inLen-1]: Array of allocations that contain the inputs
668//   aout:             The allocation that will hold the output
669//   mtls:             Holds launch parameters
670void RsdCpuReferenceImpl::launchReduceNew(const Allocation ** ains,
671                                          uint32_t inLen,
672                                          Allocation * aout,
673                                          MTLaunchStructReduceNew *mtls) {
674  mtls->logReduce = mRSC->props.mLogReduce;
675  if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) {
676    launchReduceNewParallel(ains, inLen, aout, mtls);
677  } else {
678    launchReduceNewSerial(ains, inLen, aout, mtls);
679  }
680}
681
682// Launch a general reduce-style kernel, single-threaded.
683// Inputs:
684//   ains[0..inLen-1]: Array of allocations that contain the inputs
685//   aout:             The allocation that will hold the output
686//   mtls:             Holds launch parameters
687void RsdCpuReferenceImpl::launchReduceNewSerial(const Allocation ** ains,
688                                                uint32_t inLen,
689                                                Allocation * aout,
690                                                MTLaunchStructReduceNew *mtls) {
691  REDUCE_NEW_ALOGV(mtls, 1, "launchReduceNewSerial(%p): %u x %u x %u", mtls->accumFunc,
692                   mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z);
693
694  // In the presence of outconverter, we allocate temporary memory for
695  // the accumulator.
696  //
697  // In the absence of outconverter, we use the output allocation as the
698  // accumulator.
699  uint8_t *const accumPtr = (mtls->outFunc
700                             ? static_cast<uint8_t *>(malloc(mtls->accumSize))
701                             : mtls->redp.outPtr[0]);
702
703  // initialize
704  if (mtls->initFunc) {
705    mtls->initFunc(accumPtr);
706  } else {
707    memset(accumPtr, 0, mtls->accumSize);
708  }
709
710  // accumulate
711  const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc;
712  uint32_t slice = 0;
713  while (SelectOuterSlice(mtls, &mtls->redp, slice++)) {
714    for (mtls->redp.current.y = mtls->start.y;
715         mtls->redp.current.y < mtls->end.y;
716         mtls->redp.current.y++) {
717      RedpPtrSetup(mtls, &mtls->redp, mtls->start.x, mtls->redp.current.y, mtls->redp.current.z);
718      fn(&mtls->redp, mtls->start.x, mtls->end.x, accumPtr);
719    }
720  }
721
722  // outconvert
723  if (mtls->outFunc) {
724    mtls->outFunc(mtls->redp.outPtr[0], accumPtr);
725    free(accumPtr);
726  }
727}
728
729// Launch a general reduce-style kernel, multi-threaded.
730// Inputs:
731//   ains[0..inLen-1]: Array of allocations that contain the inputs
732//   aout:             The allocation that will hold the output
733//   mtls:             Holds launch parameters
734void RsdCpuReferenceImpl::launchReduceNewParallel(const Allocation ** ains,
735                                                  uint32_t inLen,
736                                                  Allocation * aout,
737                                                  MTLaunchStructReduceNew *mtls) {
738  // For now, we don't know how to go parallel in the absence of a combiner.
739  if (!mtls->combFunc) {
740    launchReduceNewSerial(ains, inLen, aout, mtls);
741    return;
742  }
743
744  // Number of threads = "main thread" + number of other (worker) threads
745  const uint32_t numThreads = mWorkers.mCount + 1;
746
747  // In the absence of outconverter, we use the output allocation as
748  // an accumulator, and therefore need to allocate one fewer accumulator.
749  const uint32_t numAllocAccum = numThreads - (mtls->outFunc == nullptr);
750
751  // If mDebugReduceSplitAccum, then we want each accumulator to start
752  // on a page boundary.  (TODO: Would some unit smaller than a page
753  // be sufficient to avoid false sharing?)
754  if (mRSC->props.mDebugReduceSplitAccum) {
755    // Round up accumulator size to an integral number of pages
756    mtls->accumStride =
757        (unsigned(mtls->accumSize) + unsigned(mPageSize)-1) &
758        ~(unsigned(mPageSize)-1);
759    // Each accumulator gets its own page.  Alternatively, if we just
760    // wanted to make sure no two accumulators are on the same page,
761    // we could instead do
762    //   allocSize = mtls->accumStride * (numAllocation - 1) + mtls->accumSize
763    const size_t allocSize = mtls->accumStride * numAllocAccum;
764    mtls->accumAlloc = static_cast<uint8_t *>(memalign(mPageSize, allocSize));
765  } else {
766    mtls->accumStride = mtls->accumSize;
767    mtls->accumAlloc = static_cast<uint8_t *>(malloc(mtls->accumStride * numAllocAccum));
768  }
769
770  const size_t accumPtrArrayBytes = sizeof(uint8_t *) * numThreads;
771  mtls->accumPtr = static_cast<uint8_t **>(malloc(accumPtrArrayBytes));
772  memset(mtls->accumPtr, 0, accumPtrArrayBytes);
773
774  mtls->accumCount = 0;
775
776  rsAssert(!mInKernel);
777  mInKernel = true;
778  REDUCE_NEW_ALOGV(mtls, 1, "launchReduceNewParallel(%p): %u x %u x %u, %u threads, accumAlloc = %p",
779                   mtls->accumFunc,
780                   mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z,
781                   numThreads, mtls->accumAlloc);
782  if (mtls->redp.dim.z > 1) {
783    mtls->mSliceSize = 1;
784    launchThreads(walk_3d_reduce_new, mtls);
785  } else if (mtls->redp.dim.y > 1) {
786    mtls->mSliceSize = rsMax(1U, mtls->redp.dim.y / (numThreads * 4));
787    launchThreads(walk_2d_reduce_new, mtls);
788  } else {
789    mtls->mSliceSize = rsMax(1U, mtls->redp.dim.x / (numThreads * 4));
790    launchThreads(walk_1d_reduce_new, mtls);
791  }
792  mInKernel = false;
793
794  // Combine accumulators and identify final accumulator
795  uint8_t *finalAccumPtr = (mtls->outFunc ? nullptr : mtls->redp.outPtr[0]);
796  //   Loop over accumulators, combining into finalAccumPtr.  If finalAccumPtr
797  //   is null, then the first accumulator I find becomes finalAccumPtr.
798  for (unsigned idx = 0; idx < mtls->accumCount; ++idx) {
799    uint8_t *const thisAccumPtr = mtls->accumPtr[idx];
800    if (finalAccumPtr) {
801      if (finalAccumPtr != thisAccumPtr) {
802        if (mtls->combFunc) {
803          if (mtls->logReduce >= 3) {
804            FormatBuf fmt;
805            REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): accumulating into%s",
806                             mtls->accumFunc,
807                             format_bytes(&fmt, finalAccumPtr, mtls->accumSize));
808            REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p):    accumulator[%d]%s",
809                             mtls->accumFunc, idx,
810                             format_bytes(&fmt, thisAccumPtr, mtls->accumSize));
811          }
812          mtls->combFunc(finalAccumPtr, thisAccumPtr);
813        } else {
814          rsAssert(!"expected combiner");
815        }
816      }
817    } else {
818      finalAccumPtr = thisAccumPtr;
819    }
820  }
821  rsAssert(finalAccumPtr != nullptr);
822  if (mtls->logReduce >= 3) {
823    FormatBuf fmt;
824    REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): final accumulator%s",
825                     mtls->accumFunc, format_bytes(&fmt, finalAccumPtr, mtls->accumSize));
826  }
827
828  // Outconvert
829  if (mtls->outFunc) {
830    mtls->outFunc(mtls->redp.outPtr[0], finalAccumPtr);
831    if (mtls->logReduce >= 3) {
832      FormatBuf fmt;
833      REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): final outconverted result%s",
834                       mtls->accumFunc,
835                       format_bytes(&fmt, mtls->redp.outPtr[0], mtls->redp.outStride[0]));
836    }
837  }
838
839  // Clean up
840  free(mtls->accumPtr);
841  free(mtls->accumAlloc);
842}
843
844
845void RsdCpuReferenceImpl::launchForEach(const Allocation ** ains,
846                                        uint32_t inLen,
847                                        Allocation* aout,
848                                        const RsScriptCall* sc,
849                                        MTLaunchStructForEach* mtls) {
850
851    //android::StopWatch kernel_time("kernel time");
852
853    bool outerDims = (mtls->start.z != mtls->end.z) ||
854                     (mtls->start.face != mtls->end.face) ||
855                     (mtls->start.lod != mtls->end.lod) ||
856                     (mtls->start.array[0] != mtls->end.array[0]) ||
857                     (mtls->start.array[1] != mtls->end.array[1]) ||
858                     (mtls->start.array[2] != mtls->end.array[2]) ||
859                     (mtls->start.array[3] != mtls->end.array[3]);
860
861    if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) {
862        const size_t targetByteChunk = 16 * 1024;
863        mInKernel = true;  // NOTE: The guard immediately above ensures this was !mInKernel
864
865        if (outerDims) {
866            // No fancy logic for chunk size
867            mtls->mSliceSize = 1;
868            launchThreads(walk_general_foreach, mtls);
869        } else if (mtls->fep.dim.y > 1) {
870            uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4);
871            uint32_t s2 = 0;
872
873            // This chooses our slice size to rate limit atomic ops to
874            // one per 16k bytes of reads/writes.
875            if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) {
876                s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride;
877            } else if (mtls->ains[0]) {
878                s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride;
879            } else {
880                // Launch option only case
881                // Use s1 based only on the dimensions
882                s2 = s1;
883            }
884            mtls->mSliceSize = rsMin(s1, s2);
885
886            if(mtls->mSliceSize < 1) {
887                mtls->mSliceSize = 1;
888            }
889
890            launchThreads(walk_2d_foreach, mtls);
891        } else {
892            uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4);
893            uint32_t s2 = 0;
894
895            // This chooses our slice size to rate limit atomic ops to
896            // one per 16k bytes of reads/writes.
897            if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) {
898                s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes();
899            } else if (mtls->ains[0]) {
900                s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes();
901            } else {
902                // Launch option only case
903                // Use s1 based only on the dimensions
904                s2 = s1;
905            }
906            mtls->mSliceSize = rsMin(s1, s2);
907
908            if (mtls->mSliceSize < 1) {
909                mtls->mSliceSize = 1;
910            }
911
912            launchThreads(walk_1d_foreach, mtls);
913        }
914        mInKernel = false;
915
916    } else {
917        ForEachFunc_t fn = mtls->kernel;
918        uint32_t slice = 0;
919
920
921        while(SelectOuterSlice(mtls, &mtls->fep, slice++)) {
922            for (mtls->fep.current.y = mtls->start.y;
923                 mtls->fep.current.y < mtls->end.y;
924                 mtls->fep.current.y++) {
925
926                FepPtrSetup(mtls, &mtls->fep, mtls->start.x,
927                            mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod,
928                            (RsAllocationCubemapFace) mtls->fep.current.face,
929                            mtls->fep.current.array[0], mtls->fep.current.array[1],
930                            mtls->fep.current.array[2], mtls->fep.current.array[3]);
931
932                fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]);
933            }
934        }
935    }
936}
937
938RsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) {
939    //ALOGE("setTls %p", sc);
940    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey);
941    rsAssert(tls);
942    RsdCpuScriptImpl *old = tls->mImpl;
943    tls->mImpl = sc;
944    tls->mContext = mRSC;
945    if (sc) {
946        tls->mScript = sc->getScript();
947    } else {
948        tls->mScript = nullptr;
949    }
950    return old;
951}
952
953const RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) {
954    return mSymLookupFn(mRSC, name);
955}
956
957
958RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s,
959                                    char const *resName, char const *cacheDir,
960                                    uint8_t const *bitcode, size_t bitcodeSize,
961                                    uint32_t flags) {
962
963    RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s);
964    if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags
965        , getBccPluginName()
966        )) {
967        delete i;
968        return nullptr;
969    }
970    return i;
971}
972
973extern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx,
974                                             const Script *s, const Element *e);
975extern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx,
976                                                   const Script *s, const Element *e);
977extern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx,
978                                                   const Script *s, const Element *e);
979extern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx,
980                                           const Script *s, const Element *e);
981extern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx,
982                                                   const Script *s, const Element *e);
983extern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx,
984                                            const Script *s, const Element *e);
985extern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx,
986                                                const Script *s, const Element *e);
987extern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx,
988                                             const Script *s, const Element *e);
989extern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx,
990                                                 const Script *s, const Element *e);
991extern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx,
992                                              const Script *s, const Element *e);
993extern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx,
994                                              const Script *s, const Element *e);
995
996RsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s,
997                                    RsScriptIntrinsicID iid, Element *e) {
998
999    RsdCpuScriptImpl *i = nullptr;
1000    switch (iid) {
1001    case RS_SCRIPT_INTRINSIC_ID_3DLUT:
1002        i = rsdIntrinsic_3DLUT(this, s, e);
1003        break;
1004    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3:
1005        i = rsdIntrinsic_Convolve3x3(this, s, e);
1006        break;
1007    case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX:
1008        i = rsdIntrinsic_ColorMatrix(this, s, e);
1009        break;
1010    case RS_SCRIPT_INTRINSIC_ID_LUT:
1011        i = rsdIntrinsic_LUT(this, s, e);
1012        break;
1013    case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5:
1014        i = rsdIntrinsic_Convolve5x5(this, s, e);
1015        break;
1016    case RS_SCRIPT_INTRINSIC_ID_BLUR:
1017        i = rsdIntrinsic_Blur(this, s, e);
1018        break;
1019    case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB:
1020        i = rsdIntrinsic_YuvToRGB(this, s, e);
1021        break;
1022    case RS_SCRIPT_INTRINSIC_ID_BLEND:
1023        i = rsdIntrinsic_Blend(this, s, e);
1024        break;
1025    case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM:
1026        i = rsdIntrinsic_Histogram(this, s, e);
1027        break;
1028    case RS_SCRIPT_INTRINSIC_ID_RESIZE:
1029        i = rsdIntrinsic_Resize(this, s, e);
1030        break;
1031    case RS_SCRIPT_INTRINSIC_ID_BLAS:
1032        i = rsdIntrinsic_BLAS(this, s, e);
1033        break;
1034
1035    default:
1036        rsAssert(0);
1037    }
1038
1039    return i;
1040}
1041
1042void* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) {
1043  switch (sg->getApiVersion()) {
1044    case ScriptGroupBase::SG_V1: {
1045      CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg);
1046      if (!sgi->init()) {
1047        delete sgi;
1048        return nullptr;
1049      }
1050      return sgi;
1051    }
1052    case ScriptGroupBase::SG_V2: {
1053      return new CpuScriptGroup2Impl(this, sg);
1054    }
1055  }
1056  return nullptr;
1057}
1058