rsScriptC.cpp revision 15849d9a926fbd83587b9da30b8888574998a0fd 
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    mEnviroment.mStartTimeMillis
51                = nanoseconds_to_milliseconds(systemTime(SYSTEM_TIME_MONOTONIC));
52
53    for (uint32_t ct=0; ct < mEnviroment.mFieldCount; ct++) {
54        if (mSlots[ct].get() && !mTypes[ct].get()) {
55            mTypes[ct].set(mSlots[ct]->getType());
56        }
57
58        if (!mTypes[ct].get())
59            continue;
60        void *ptr = NULL;
61        if (mSlots[ct].get()) {
62            ptr = mSlots[ct]->getPtr();
63        }
64        void **dest = ((void ***)mEnviroment.mFieldAddress)[ct];
65
66        if (rsc->props.mLogScripts) {
67            if (mSlots[ct].get() != NULL) {
68                LOGV("%p ScriptC::setupScript slot=%i  dst=%p  src=%p  type=%p", rsc, ct, dest, ptr, mSlots[ct]->getType());
69            } else {
70                LOGV("%p ScriptC::setupScript slot=%i  dst=%p  src=%p  type=null", rsc, ct, dest, ptr);
71            }
72        }
73
74        if (dest) {
75            *dest = ptr;
76        }
77    }
78}
79
80const Allocation *ScriptC::ptrToAllocation(const void *ptr) const {
81    if (!ptr) {
82        return NULL;
83    }
84    for (uint32_t ct=0; ct < mEnviroment.mFieldCount; ct++) {
85        if (!mSlots[ct].get())
86            continue;
87        if (mSlots[ct]->getPtr() == ptr) {
88            return mSlots[ct].get();
89        }
90    }
91    LOGE("ScriptC::ptrToAllocation, failed to find %p", ptr);
92    return NULL;
93}
94
95Script * ScriptC::setTLS(Script *sc) {
96    Context::ScriptTLSStruct * tls = (Context::ScriptTLSStruct *)
97                                  pthread_getspecific(Context::gThreadTLSKey);
98    rsAssert(tls);
99    Script *old = tls->mScript;
100    tls->mScript = sc;
101    return old;
102}
103
104void ScriptC::setupGLState(Context *rsc) {
105    if (mEnviroment.mFragmentStore.get()) {
106        rsc->setProgramStore(mEnviroment.mFragmentStore.get());
107    }
108    if (mEnviroment.mFragment.get()) {
109        rsc->setProgramFragment(mEnviroment.mFragment.get());
110    }
111    if (mEnviroment.mVertex.get()) {
112        rsc->setProgramVertex(mEnviroment.mVertex.get());
113    }
114    if (mEnviroment.mRaster.get()) {
115        rsc->setProgramRaster(mEnviroment.mRaster.get());
116    }
117}
118
119uint32_t ScriptC::run(Context *rsc) {
120    if (mProgram.mRoot == NULL) {
121        rsc->setError(RS_ERROR_BAD_SCRIPT, "Attempted to run bad script");
122        return 0;
123    }
124
125    setupGLState(rsc);
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    setupGLState(rsc);
282    setupScript(rsc);
283    Script * oldTLS = setTLS(this);
284
285    mtls.rsc = rsc;
286    mtls.ain = ain;
287    mtls.aout = aout;
288    mtls.script = this;
289    mtls.usr = usr;
290    mtls.mSliceSize = 10;
291    mtls.mSliceNum = 0;
292
293    mtls.ptrIn = NULL;
294    mtls.eStrideIn = 0;
295    if (ain) {
296        mtls.ptrIn = (const uint8_t *)ain->getPtr();
297        mtls.eStrideIn = ain->getType()->getElementSizeBytes();
298    }
299
300    mtls.ptrOut = NULL;
301    mtls.eStrideOut = 0;
302    if (aout) {
303        mtls.ptrOut = (uint8_t *)aout->getPtr();
304        mtls.eStrideOut = aout->getType()->getElementSizeBytes();
305    }
306
307    if ((rsc->getWorkerPoolSize() > 1) && mEnviroment.mIsThreadable) {
308        if (mtls.dimY > 1) {
309            rsc->launchThreads(wc_xy, &mtls);
310        } else {
311            rsc->launchThreads(wc_x, &mtls);
312        }
313
314        //LOGE("launch 1");
315    } else {
316        //LOGE("launch 3");
317        for (uint32_t ar = mtls.arrayStart; ar < mtls.arrayEnd; ar++) {
318            for (uint32_t z = mtls.zStart; z < mtls.zEnd; z++) {
319                for (uint32_t y = mtls.yStart; y < mtls.yEnd; y++) {
320                    uint32_t offset = mtls.dimX * mtls.dimY * mtls.dimZ * ar +
321                                      mtls.dimX * mtls.dimY * z +
322                                      mtls.dimX * y;
323                    uint8_t *xPtrOut = mtls.ptrOut + (mtls.eStrideOut * offset);
324                    const uint8_t *xPtrIn = mtls.ptrIn + (mtls.eStrideIn * offset);
325
326                    for (uint32_t x = mtls.xStart; x < mtls.xEnd; x++) {
327                        ((rs_t)mProgram.mRoot) (xPtrIn, xPtrOut, usr, x, y, z, ar);
328                        xPtrIn += mtls.eStrideIn;
329                        xPtrOut += mtls.eStrideOut;
330                    }
331                }
332            }
333        }
334    }
335
336    setTLS(oldTLS);
337}
338
339void ScriptC::Invoke(Context *rsc, uint32_t slot, const void *data, uint32_t len) {
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    if (!strcmp(name, "__isThreadable")) {
382      return (BCCvoid*) s->mEnviroment.mIsThreadable;
383    } else if (!strcmp(name, "__clearThreadable")) {
384      s->mEnviroment.mIsThreadable = false;
385      return NULL;
386    }
387    sym = ScriptCState::lookupSymbol(name);
388    if (!sym) {
389        sym = ScriptCState::lookupSymbolCL(name);
390    }
391    if (!sym) {
392        sym = ScriptCState::lookupSymbolGL(name);
393    }
394    if (sym) {
395        s->mEnviroment.mIsThreadable &= sym->threadable;
396        return sym->mPtr;
397    }
398    LOGE("ScriptC sym lookup failed for %s", name);
399    return NULL;
400}
401
402extern const char rs_runtime_lib_bc[];
403extern unsigned rs_runtime_lib_bc_size;
404
405void ScriptCState::runCompiler(Context *rsc, ScriptC *s, const char *resName, const char *cacheDir) {
406    {
407        s->mBccScript = bccCreateScript();
408        s->mEnviroment.mIsThreadable = true;
409        bccRegisterSymbolCallback(s->mBccScript, symbolLookup, s);
410        // bccReadBC() reads in the BitCode, if no cache file corresponding to
411        // the resName is found. Otherwise, bccReadBC() returns a negative value
412        // and the "else" branch will be taken.
413        if (bccReadBC(s->mBccScript,
414                      s->mEnviroment.mScriptText,
415                      s->mEnviroment.mScriptTextLength,
416                      resName,
417                      cacheDir) >= 0) {
418          //bccLinkBC(s->mBccScript, rs_runtime_lib_bc, rs_runtime_lib_bc_size);
419          bccCompileBC(s->mBccScript);
420        } else {
421          // bccReadBC returns a neagative value: Didn't read any script,
422          // So, use cached binary instead
423          bccLoadBinary(s->mBccScript);
424        }
425        bccGetScriptLabel(s->mBccScript, "root", (BCCvoid**) &s->mProgram.mRoot);
426        bccGetScriptLabel(s->mBccScript, "init", (BCCvoid**) &s->mProgram.mInit);
427    }
428    LOGV("%p ScriptCState::runCompiler root %p,  init %p", rsc, s->mProgram.mRoot, s->mProgram.mInit);
429
430    if (s->mProgram.mInit) {
431        s->mProgram.mInit();
432    }
433
434    bccGetExportFuncs(s->mBccScript, (BCCsizei*) &s->mEnviroment.mInvokeFunctionCount, 0, NULL);
435    if (s->mEnviroment.mInvokeFunctionCount <= 0)
436        s->mEnviroment.mInvokeFunctions = NULL;
437    else {
438        s->mEnviroment.mInvokeFunctions = (Script::InvokeFunc_t*) calloc(s->mEnviroment.mInvokeFunctionCount, sizeof(Script::InvokeFunc_t));
439        bccGetExportFuncs(s->mBccScript, NULL, s->mEnviroment.mInvokeFunctionCount, (BCCvoid **) s->mEnviroment.mInvokeFunctions);
440    }
441
442    bccGetExportVars(s->mBccScript, (BCCsizei*) &s->mEnviroment.mFieldCount, 0, NULL);
443    if (s->mEnviroment.mFieldCount <= 0)
444        s->mEnviroment.mFieldAddress = NULL;
445    else {
446        s->mEnviroment.mFieldAddress = (void **) calloc(s->mEnviroment.mFieldCount, sizeof(void *));
447        bccGetExportVars(s->mBccScript, NULL, s->mEnviroment.mFieldCount, (BCCvoid **) s->mEnviroment.mFieldAddress);
448        s->initSlots();
449    }
450
451    s->mEnviroment.mFragment.set(rsc->getDefaultProgramFragment());
452    s->mEnviroment.mVertex.set(rsc->getDefaultProgramVertex());
453    s->mEnviroment.mFragmentStore.set(rsc->getDefaultProgramStore());
454    s->mEnviroment.mRaster.set(rsc->getDefaultProgramRaster());
455
456    if (s->mProgram.mRoot) {
457        const static int pragmaMax = 16;
458        BCCsizei pragmaCount;
459        BCCchar * str[pragmaMax];
460        bccGetPragmas(s->mBccScript, &pragmaCount, pragmaMax, &str[0]);
461
462        for (int ct=0; ct < pragmaCount; ct+=2) {
463            //LOGE("pragme %s %s", str[ct], str[ct+1]);
464            if (!strcmp(str[ct], "version")) {
465                continue;
466            }
467
468            if (!strcmp(str[ct], "stateVertex")) {
469                if (!strcmp(str[ct+1], "default")) {
470                    continue;
471                }
472                if (!strcmp(str[ct+1], "parent")) {
473                    s->mEnviroment.mVertex.clear();
474                    continue;
475                }
476                LOGE("Unreconized value %s passed to stateVertex", str[ct+1]);
477            }
478
479            if (!strcmp(str[ct], "stateRaster")) {
480                if (!strcmp(str[ct+1], "default")) {
481                    continue;
482                }
483                if (!strcmp(str[ct+1], "parent")) {
484                    s->mEnviroment.mRaster.clear();
485                    continue;
486                }
487                LOGE("Unreconized value %s passed to stateRaster", str[ct+1]);
488            }
489
490            if (!strcmp(str[ct], "stateFragment")) {
491                if (!strcmp(str[ct+1], "default")) {
492                    continue;
493                }
494                if (!strcmp(str[ct+1], "parent")) {
495                    s->mEnviroment.mFragment.clear();
496                    continue;
497                }
498                LOGE("Unreconized value %s passed to stateFragment", str[ct+1]);
499            }
500
501            if (!strcmp(str[ct], "stateStore")) {
502                if (!strcmp(str[ct+1], "default")) {
503                    continue;
504                }
505                if (!strcmp(str[ct+1], "parent")) {
506                    s->mEnviroment.mFragmentStore.clear();
507                    continue;
508                }
509                LOGE("Unreconized value %s passed to stateStore", str[ct+1]);
510            }
511
512        }
513
514
515    } else {
516        // Deal with an error.
517    }
518}
519
520namespace android {
521namespace renderscript {
522
523void rsi_ScriptCBegin(Context * rsc) {
524    ScriptCState *ss = &rsc->mScriptC;
525    ss->clear(rsc);
526}
527
528void rsi_ScriptCSetText(Context *rsc, const char *text, uint32_t len) {
529    ScriptCState *ss = &rsc->mScriptC;
530
531    char *t = (char *)malloc(len + 1);
532    memcpy(t, text, len);
533    t[len] = 0;
534    ss->mScript->mEnviroment.mScriptText = t;
535    ss->mScript->mEnviroment.mScriptTextLength = len;
536}
537
538RsScript rsi_ScriptCCreate(Context * rsc, const char *resName, const char *cacheDir)
539{
540    ScriptCState *ss = &rsc->mScriptC;
541
542    ObjectBaseRef<ScriptC> s(ss->mScript);
543    ss->mScript.clear();
544    s->incUserRef();
545
546    ss->runCompiler(rsc, s.get(), resName, cacheDir);
547    ss->clear(rsc);
548    return s.get();
549}
550
551}
552}
553