rsScriptC.cpp revision a914f340ae5b267dc3ab36c1156c795b8fa18f5d 
1/*
2 * Copyright (C) 2009 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 "rsContext.h"
18#include "rsScriptC.h"
19#include "rsMatrix.h"
20#include "../../compile/libbcc/include/bcc/bcc.h"
21#include "utils/Timers.h"
22#include "utils/StopWatch.h"
23
24#include <GLES/gl.h>
25#include <GLES/glext.h>
26
27using namespace android;
28using namespace android::renderscript;
29
30#define GET_TLS()  Context::ScriptTLSStruct * tls = \
31    (Context::ScriptTLSStruct *)pthread_getspecific(Context::gThreadTLSKey); \
32    Context * rsc = tls->mContext; \
33    ScriptC * sc = (ScriptC *) tls->mScript
34
35
36ScriptC::ScriptC(Context *rsc) : Script(rsc) {
37    mBccScript = NULL;
38    memset(&mProgram, 0, sizeof(mProgram));
39}
40
41ScriptC::~ScriptC() {
42    if (mBccScript) {
43        bccDeleteScript(mBccScript);
44    }
45    free(mEnviroment.mScriptText);
46    mEnviroment.mScriptText = NULL;
47}
48
49void ScriptC::setupScript(Context *rsc) {
50    setupGLState(rsc);
51    mEnviroment.mStartTimeMillis
52                = nanoseconds_to_milliseconds(systemTime(SYSTEM_TIME_MONOTONIC));
53
54    for (uint32_t ct=0; ct < mEnviroment.mFieldCount; ct++) {
55        if (mSlots[ct].get() && !mTypes[ct].get()) {
56            mTypes[ct].set(mSlots[ct]->getType());
57        }
58
59        if (!mTypes[ct].get())
60            continue;
61        void *ptr = NULL;
62        if (mSlots[ct].get()) {
63            ptr = mSlots[ct]->getPtr();
64        }
65        void **dest = ((void ***)mEnviroment.mFieldAddress)[ct];
66
67        if (rsc->props.mLogScripts) {
68            if (mSlots[ct].get() != NULL) {
69                LOGV("%p ScriptC::setupScript slot=%i  dst=%p  src=%p  type=%p", rsc, ct, dest, ptr, mSlots[ct]->getType());
70            } else {
71                LOGV("%p ScriptC::setupScript slot=%i  dst=%p  src=%p  type=null", rsc, ct, dest, ptr);
72            }
73        }
74
75        if (dest) {
76            *dest = ptr;
77        }
78    }
79}
80
81const Allocation *ScriptC::ptrToAllocation(const void *ptr) const {
82    if (!ptr) {
83        return NULL;
84    }
85    for (uint32_t ct=0; ct < mEnviroment.mFieldCount; ct++) {
86        if (!mSlots[ct].get())
87            continue;
88        if (mSlots[ct]->getPtr() == ptr) {
89            return mSlots[ct].get();
90        }
91    }
92    LOGE("ScriptC::ptrToAllocation, failed to find %p", ptr);
93    return NULL;
94}
95
96Script * ScriptC::setTLS(Script *sc) {
97    Context::ScriptTLSStruct * tls = (Context::ScriptTLSStruct *)
98                                  pthread_getspecific(Context::gThreadTLSKey);
99    rsAssert(tls);
100    Script *old = tls->mScript;
101    tls->mScript = sc;
102    return old;
103}
104
105void ScriptC::setupGLState(Context *rsc) {
106    if (mEnviroment.mFragmentStore.get()) {
107        rsc->setProgramStore(mEnviroment.mFragmentStore.get());
108    }
109    if (mEnviroment.mFragment.get()) {
110        rsc->setProgramFragment(mEnviroment.mFragment.get());
111    }
112    if (mEnviroment.mVertex.get()) {
113        rsc->setProgramVertex(mEnviroment.mVertex.get());
114    }
115    if (mEnviroment.mRaster.get()) {
116        rsc->setProgramRaster(mEnviroment.mRaster.get());
117    }
118}
119
120uint32_t ScriptC::run(Context *rsc) {
121    if (mProgram.mRoot == NULL) {
122        rsc->setError(RS_ERROR_BAD_SCRIPT, "Attempted to run bad script");
123        return 0;
124    }
125
126    setupScript(rsc);
127
128    uint32_t ret = 0;
129    Script * oldTLS = setTLS(this);
130
131    if (rsc->props.mLogScripts) {
132        LOGV("%p ScriptC::run invoking root,  ptr %p", rsc, mProgram.mRoot);
133    }
134
135    ret = mProgram.mRoot();
136
137    if (rsc->props.mLogScripts) {
138        LOGV("%p ScriptC::run invoking complete, ret=%i", rsc, ret);
139    }
140
141    setTLS(oldTLS);
142    return ret;
143}
144
145typedef struct {
146    Context *rsc;
147    ScriptC *script;
148    const Allocation * ain;
149    Allocation * aout;
150    const void * usr;
151
152    uint32_t mSliceSize;
153    volatile int mSliceNum;
154
155    const uint8_t *ptrIn;
156    uint32_t eStrideIn;
157    uint8_t *ptrOut;
158    uint32_t eStrideOut;
159
160    uint32_t xStart;
161    uint32_t xEnd;
162    uint32_t yStart;
163    uint32_t yEnd;
164    uint32_t zStart;
165    uint32_t zEnd;
166    uint32_t arrayStart;
167    uint32_t arrayEnd;
168
169    uint32_t dimX;
170    uint32_t dimY;
171    uint32_t dimZ;
172    uint32_t dimArray;
173} MTLaunchStruct;
174typedef int (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
175
176static void wc_xy(void *usr, uint32_t idx) {
177    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
178
179    while (1) {
180        uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum);
181        uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize;
182        uint32_t yEnd = yStart + mtls->mSliceSize;
183        yEnd = rsMin(yEnd, mtls->yEnd);
184        if (yEnd <= yStart) {
185            return;
186        }
187
188        //LOGE("usr idx %i, x %i,%i  y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
189        //LOGE("usr ptr in %p,  out %p", mtls->ptrIn, mtls->ptrOut);
190        for (uint32_t y = yStart; y < yEnd; y++) {
191            uint32_t offset = mtls->dimX * y;
192            uint8_t *xPtrOut = mtls->ptrOut + (mtls->eStrideOut * offset);
193            const uint8_t *xPtrIn = mtls->ptrIn + (mtls->eStrideIn * offset);
194
195            for (uint32_t x = mtls->xStart; x < mtls->xEnd; x++) {
196                ((rs_t)mtls->script->mProgram.mRoot) (xPtrIn, xPtrOut, mtls->usr, x, y, 0, 0);
197                xPtrIn += mtls->eStrideIn;
198                xPtrOut += mtls->eStrideOut;
199            }
200        }
201    }
202}
203
204static void wc_x(void *usr, uint32_t idx) {
205    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
206
207    while (1) {
208        uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum);
209        uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize;
210        uint32_t xEnd = xStart + mtls->mSliceSize;
211        xEnd = rsMin(xEnd, mtls->xEnd);
212        if (xEnd <= xStart) {
213            return;
214        }
215
216        //LOGE("usr idx %i, x %i,%i  y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
217        //LOGE("usr ptr in %p,  out %p", mtls->ptrIn, mtls->ptrOut);
218        uint8_t *xPtrOut = mtls->ptrOut + (mtls->eStrideOut * xStart);
219        const uint8_t *xPtrIn = mtls->ptrIn + (mtls->eStrideIn * xStart);
220        for (uint32_t x = xStart; x < xEnd; x++) {
221            ((rs_t)mtls->script->mProgram.mRoot) (xPtrIn, xPtrOut, mtls->usr, x, 0, 0, 0);
222            xPtrIn += mtls->eStrideIn;
223            xPtrOut += mtls->eStrideOut;
224        }
225    }
226}
227
228void ScriptC::runForEach(Context *rsc,
229                         const Allocation * ain,
230                         Allocation * aout,
231                         const void * usr,
232                         const RsScriptCall *sc) {
233    MTLaunchStruct mtls;
234    memset(&mtls, 0, sizeof(mtls));
235    Context::PushState ps(rsc);
236
237    if (ain) {
238        mtls.dimX = ain->getType()->getDimX();
239        mtls.dimY = ain->getType()->getDimY();
240        mtls.dimZ = ain->getType()->getDimZ();
241        //mtls.dimArray = ain->getType()->getDimArray();
242    } else if (aout) {
243        mtls.dimX = aout->getType()->getDimX();
244        mtls.dimY = aout->getType()->getDimY();
245        mtls.dimZ = aout->getType()->getDimZ();
246        //mtls.dimArray = aout->getType()->getDimArray();
247    } else {
248        rsc->setError(RS_ERROR_BAD_SCRIPT, "rsForEach called with null allocations");
249        return;
250    }
251
252    if (!sc || (sc->xEnd == 0)) {
253        mtls.xEnd = mtls.dimX;
254    } else {
255        rsAssert(sc->xStart < mtls.dimX);
256        rsAssert(sc->xEnd <= mtls.dimX);
257        rsAssert(sc->xStart < sc->xEnd);
258        mtls.xStart = rsMin(mtls.dimX, sc->xStart);
259        mtls.xEnd = rsMin(mtls.dimX, sc->xEnd);
260        if (mtls.xStart >= mtls.xEnd) return;
261    }
262
263    if (!sc || (sc->yEnd == 0)) {
264        mtls.yEnd = mtls.dimY;
265    } else {
266        rsAssert(sc->yStart < mtls.dimY);
267        rsAssert(sc->yEnd <= mtls.dimY);
268        rsAssert(sc->yStart < sc->yEnd);
269        mtls.yStart = rsMin(mtls.dimY, sc->yStart);
270        mtls.yEnd = rsMin(mtls.dimY, sc->yEnd);
271        if (mtls.yStart >= mtls.yEnd) return;
272    }
273
274    mtls.xEnd = rsMax((uint32_t)1, mtls.xEnd);
275    mtls.yEnd = rsMax((uint32_t)1, mtls.yEnd);
276    mtls.zEnd = rsMax((uint32_t)1, mtls.zEnd);
277    mtls.arrayEnd = rsMax((uint32_t)1, mtls.arrayEnd);
278
279    rsAssert(ain->getType()->getDimZ() == 0);
280
281    setupScript(rsc);
282    Script * oldTLS = setTLS(this);
283
284    mtls.rsc = rsc;
285    mtls.ain = ain;
286    mtls.aout = aout;
287    mtls.script = this;
288    mtls.usr = usr;
289    mtls.mSliceSize = 10;
290    mtls.mSliceNum = 0;
291
292    mtls.ptrIn = NULL;
293    mtls.eStrideIn = 0;
294    if (ain) {
295        mtls.ptrIn = (const uint8_t *)ain->getPtr();
296        mtls.eStrideIn = ain->getType()->getElementSizeBytes();
297    }
298
299    mtls.ptrOut = NULL;
300    mtls.eStrideOut = 0;
301    if (aout) {
302        mtls.ptrOut = (uint8_t *)aout->getPtr();
303        mtls.eStrideOut = aout->getType()->getElementSizeBytes();
304    }
305
306    if ((rsc->getWorkerPoolSize() > 1) && mEnviroment.mIsThreadable) {
307        if (mtls.dimY > 1) {
308            rsc->launchThreads(wc_xy, &mtls);
309        } else {
310            rsc->launchThreads(wc_x, &mtls);
311        }
312
313        //LOGE("launch 1");
314    } else {
315        //LOGE("launch 3");
316        for (uint32_t ar = mtls.arrayStart; ar < mtls.arrayEnd; ar++) {
317            for (uint32_t z = mtls.zStart; z < mtls.zEnd; z++) {
318                for (uint32_t y = mtls.yStart; y < mtls.yEnd; y++) {
319                    uint32_t offset = mtls.dimX * mtls.dimY * mtls.dimZ * ar +
320                                      mtls.dimX * mtls.dimY * z +
321                                      mtls.dimX * y;
322                    uint8_t *xPtrOut = mtls.ptrOut + (mtls.eStrideOut * offset);
323                    const uint8_t *xPtrIn = mtls.ptrIn + (mtls.eStrideIn * offset);
324
325                    for (uint32_t x = mtls.xStart; x < mtls.xEnd; x++) {
326                        ((rs_t)mProgram.mRoot) (xPtrIn, xPtrOut, usr, x, y, z, ar);
327                        xPtrIn += mtls.eStrideIn;
328                        xPtrOut += mtls.eStrideOut;
329                    }
330                }
331            }
332        }
333    }
334
335    setTLS(oldTLS);
336}
337
338void ScriptC::Invoke(Context *rsc, uint32_t slot, const void *data, uint32_t len) {
339    //LOGE("rsi_ScriptInvoke %i", slot);
340    if ((slot >= mEnviroment.mInvokeFunctionCount) ||
341        (mEnviroment.mInvokeFunctions[slot] == NULL)) {
342        rsc->setError(RS_ERROR_BAD_SCRIPT, "Calling invoke on bad script");
343        return;
344    }
345    setupScript(rsc);
346    Script * oldTLS = setTLS(this);
347
348    if (rsc->props.mLogScripts) {
349        LOGV("%p ScriptC::Invoke invoking slot %i,  ptr %p", rsc, slot, mEnviroment.mInvokeFunctions[slot]);
350    }
351    ((void (*)(const void *, uint32_t))
352        mEnviroment.mInvokeFunctions[slot])(data, len);
353    if (rsc->props.mLogScripts) {
354        LOGV("%p ScriptC::Invoke complete", rsc);
355    }
356
357    setTLS(oldTLS);
358}
359
360ScriptCState::ScriptCState() {
361    mScript.clear();
362}
363
364ScriptCState::~ScriptCState() {
365    mScript.clear();
366}
367
368void ScriptCState::init(Context *rsc) {
369    clear(rsc);
370}
371
372void ScriptCState::clear(Context *rsc) {
373    rsAssert(rsc);
374    mScript.clear();
375    mScript.set(new ScriptC(rsc));
376}
377
378static BCCvoid* symbolLookup(BCCvoid* pContext, const BCCchar* name) {
379    const ScriptCState::SymbolTable_t *sym;
380    ScriptC *s = (ScriptC *)pContext;
381    sym = ScriptCState::lookupSymbol(name);
382    if (!sym) {
383        sym = ScriptCState::lookupSymbolCL(name);
384    }
385    if (!sym) {
386        sym = ScriptCState::lookupSymbolGL(name);
387    }
388    if (sym) {
389        s->mEnviroment.mIsThreadable &= sym->threadable;
390        return sym->mPtr;
391    }
392    LOGE("ScriptC sym lookup failed for %s", name);
393    return NULL;
394}
395
396extern const char rs_runtime_lib_bc[];
397extern unsigned rs_runtime_lib_bc_size;
398
399void ScriptCState::runCompiler(Context *rsc, ScriptC *s, const char *resName) {
400    {
401        s->mBccScript = bccCreateScript();
402        s->mEnviroment.mIsThreadable = true;
403        bccRegisterSymbolCallback(s->mBccScript, symbolLookup, s);
404        // bccReadBC() reads in the BitCode, if no cache file corresponding to
405        // the resName is found. Otherwise, bccReadBC() returns a negative value
406        // and the "else" branch will be taken.
407        if (bccReadBC(s->mBccScript,
408                      s->mEnviroment.mScriptText,
409                      s->mEnviroment.mScriptTextLength,
410                      resName) >= 0) {
411          //bccLinkBC(s->mBccScript, rs_runtime_lib_bc, rs_runtime_lib_bc_size);
412          bccCompileBC(s->mBccScript);
413        } else {
414          // bccReadBC returns a neagative value: Didn't read any script,
415          // So, use cached binary instead
416          bccLoadBinary(s->mBccScript);
417        }
418        bccGetScriptLabel(s->mBccScript, "root", (BCCvoid**) &s->mProgram.mRoot);
419        bccGetScriptLabel(s->mBccScript, "init", (BCCvoid**) &s->mProgram.mInit);
420    }
421    LOGV("%p ScriptCState::runCompiler root %p,  init %p", rsc, s->mProgram.mRoot, s->mProgram.mInit);
422
423    if (s->mProgram.mInit) {
424        s->mProgram.mInit();
425    }
426
427    bccGetExportFuncs(s->mBccScript, (BCCsizei*) &s->mEnviroment.mInvokeFunctionCount, 0, NULL);
428    if (s->mEnviroment.mInvokeFunctionCount <= 0)
429        s->mEnviroment.mInvokeFunctions = NULL;
430    else {
431        s->mEnviroment.mInvokeFunctions = (Script::InvokeFunc_t*) calloc(s->mEnviroment.mInvokeFunctionCount, sizeof(Script::InvokeFunc_t));
432        bccGetExportFuncs(s->mBccScript, NULL, s->mEnviroment.mInvokeFunctionCount, (BCCvoid **) s->mEnviroment.mInvokeFunctions);
433    }
434
435    bccGetExportVars(s->mBccScript, (BCCsizei*) &s->mEnviroment.mFieldCount, 0, NULL);
436    if (s->mEnviroment.mFieldCount <= 0)
437        s->mEnviroment.mFieldAddress = NULL;
438    else {
439        s->mEnviroment.mFieldAddress = (void **) calloc(s->mEnviroment.mFieldCount, sizeof(void *));
440        bccGetExportVars(s->mBccScript, NULL, s->mEnviroment.mFieldCount, (BCCvoid **) s->mEnviroment.mFieldAddress);
441        s->initSlots();
442    }
443
444    s->mEnviroment.mFragment.set(rsc->getDefaultProgramFragment());
445    s->mEnviroment.mVertex.set(rsc->getDefaultProgramVertex());
446    s->mEnviroment.mFragmentStore.set(rsc->getDefaultProgramStore());
447    s->mEnviroment.mRaster.set(rsc->getDefaultProgramRaster());
448
449    if (s->mProgram.mRoot) {
450        const static int pragmaMax = 16;
451        BCCsizei pragmaCount;
452        BCCchar * str[pragmaMax];
453        bccGetPragmas(s->mBccScript, &pragmaCount, pragmaMax, &str[0]);
454
455        for (int ct=0; ct < pragmaCount; ct+=2) {
456            //LOGE("pragme %s %s", str[ct], str[ct+1]);
457            if (!strcmp(str[ct], "version")) {
458                continue;
459            }
460
461            if (!strcmp(str[ct], "stateVertex")) {
462                if (!strcmp(str[ct+1], "default")) {
463                    continue;
464                }
465                if (!strcmp(str[ct+1], "parent")) {
466                    s->mEnviroment.mVertex.clear();
467                    continue;
468                }
469                LOGE("Unreconized value %s passed to stateVertex", str[ct+1]);
470            }
471
472            if (!strcmp(str[ct], "stateRaster")) {
473                if (!strcmp(str[ct+1], "default")) {
474                    continue;
475                }
476                if (!strcmp(str[ct+1], "parent")) {
477                    s->mEnviroment.mRaster.clear();
478                    continue;
479                }
480                LOGE("Unreconized value %s passed to stateRaster", str[ct+1]);
481            }
482
483            if (!strcmp(str[ct], "stateFragment")) {
484                if (!strcmp(str[ct+1], "default")) {
485                    continue;
486                }
487                if (!strcmp(str[ct+1], "parent")) {
488                    s->mEnviroment.mFragment.clear();
489                    continue;
490                }
491                LOGE("Unreconized value %s passed to stateFragment", str[ct+1]);
492            }
493
494            if (!strcmp(str[ct], "stateStore")) {
495                if (!strcmp(str[ct+1], "default")) {
496                    continue;
497                }
498                if (!strcmp(str[ct+1], "parent")) {
499                    s->mEnviroment.mFragmentStore.clear();
500                    continue;
501                }
502                LOGE("Unreconized value %s passed to stateStore", str[ct+1]);
503            }
504
505        }
506
507
508    } else {
509        // Deal with an error.
510    }
511}
512
513namespace android {
514namespace renderscript {
515
516void rsi_ScriptCBegin(Context * rsc) {
517    ScriptCState *ss = &rsc->mScriptC;
518    ss->clear(rsc);
519}
520
521void rsi_ScriptCSetText(Context *rsc, const char *text, uint32_t len) {
522    ScriptCState *ss = &rsc->mScriptC;
523
524    char *t = (char *)malloc(len + 1);
525    memcpy(t, text, len);
526    t[len] = 0;
527    ss->mScript->mEnviroment.mScriptText = t;
528    ss->mScript->mEnviroment.mScriptTextLength = len;
529}
530
531RsScript rsi_ScriptCCreate(Context * rsc, const char *resName)
532{
533    ScriptCState *ss = &rsc->mScriptC;
534
535    ObjectBaseRef<ScriptC> s(ss->mScript);
536    ss->mScript.clear();
537    s->incUserRef();
538
539    ss->runCompiler(rsc, s.get(), resName);
540    ss->clear(rsc);
541    return s.get();
542}
543
544}
545}
546