rsdBcc.cpp revision 1415ca46b289604fd727310e4f6ae3c8c68276c9 
1/*
2 * Copyright (C) 2011 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
18#include "rsdCore.h"
19#include "rsdBcc.h"
20#include "rsdRuntime.h"
21
22#include <bcinfo/MetadataExtractor.h>
23
24#include "rsContext.h"
25#include "rsScriptC.h"
26
27#include "utils/Timers.h"
28#include "utils/StopWatch.h"
29extern "C" {
30#include "libdex/ZipArchive.h"
31}
32
33
34using namespace android;
35using namespace android::renderscript;
36
37struct DrvScript {
38    int (*mRoot)();
39    void (*mInit)();
40    void (*mFreeChildren)();
41
42    BCCScriptRef mBccScript;
43
44    bcinfo::MetadataExtractor *ME;
45
46    InvokeFunc_t *mInvokeFunctions;
47    void ** mFieldAddress;
48    bool * mFieldIsObject;
49    const uint32_t *mExportForEachSignatureList;
50
51    const uint8_t * mScriptText;
52    uint32_t mScriptTextLength;
53};
54
55
56static Script * setTLS(Script *sc) {
57    ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(rsdgThreadTLSKey);
58    rsAssert(tls);
59    Script *old = tls->mScript;
60    tls->mScript = sc;
61    return old;
62}
63
64
65bool rsdScriptInit(const Context *rsc,
66                     ScriptC *script,
67                     char const *resName,
68                     char const *cacheDir,
69                     uint8_t const *bitcode,
70                     size_t bitcodeSize,
71                     uint32_t flags) {
72    //LOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir, bitcode, bitcodeSize, flags, lookupFunc);
73
74    pthread_mutex_lock(&rsdgInitMutex);
75
76    size_t exportFuncCount = 0;
77    size_t exportVarCount = 0;
78    size_t objectSlotCount = 0;
79    size_t exportForEachSignatureCount = 0;
80
81    DrvScript *drv = (DrvScript *)calloc(1, sizeof(DrvScript));
82    if (drv == NULL) {
83        goto error;
84    }
85    script->mHal.drv = drv;
86
87    drv->mBccScript = bccCreateScript();
88    script->mHal.info.isThreadable = true;
89    drv->mScriptText = bitcode;
90    drv->mScriptTextLength = bitcodeSize;
91
92
93    drv->ME = new bcinfo::MetadataExtractor((const char*)drv->mScriptText,
94                                            drv->mScriptTextLength);
95    if (!drv->ME->extract()) {
96      LOGE("bcinfo: failed to read script metadata");
97      goto error;
98    }
99
100    //LOGE("mBccScript %p", script->mBccScript);
101
102    if (bccRegisterSymbolCallback(drv->mBccScript, &rsdLookupRuntimeStub, script) != 0) {
103        LOGE("bcc: FAILS to register symbol callback");
104        goto error;
105    }
106
107    if (bccReadBC(drv->mBccScript,
108                  resName,
109                  (char const *)drv->mScriptText,
110                  drv->mScriptTextLength, 0) != 0) {
111        LOGE("bcc: FAILS to read bitcode");
112        goto error;
113    }
114
115    if (bccLinkFile(drv->mBccScript, "/system/lib/libclcore.bc", 0) != 0) {
116        LOGE("bcc: FAILS to link bitcode");
117        goto error;
118    }
119
120    if (bccPrepareExecutable(drv->mBccScript, cacheDir, resName, 0) != 0) {
121        LOGE("bcc: FAILS to prepare executable");
122        goto error;
123    }
124
125    drv->mRoot = reinterpret_cast<int (*)()>(bccGetFuncAddr(drv->mBccScript, "root"));
126    drv->mInit = reinterpret_cast<void (*)()>(bccGetFuncAddr(drv->mBccScript, "init"));
127    drv->mFreeChildren = reinterpret_cast<void (*)()>(bccGetFuncAddr(drv->mBccScript, ".rs.dtor"));
128
129    exportFuncCount = drv->ME->getExportFuncCount();
130    if (exportFuncCount > 0) {
131        drv->mInvokeFunctions = (InvokeFunc_t*) calloc(exportFuncCount,
132                                                       sizeof(InvokeFunc_t));
133        bccGetExportFuncList(drv->mBccScript, exportFuncCount,
134                             (void **) drv->mInvokeFunctions);
135    } else {
136        drv->mInvokeFunctions = NULL;
137    }
138
139    exportVarCount = drv->ME->getExportVarCount();
140    if (exportVarCount > 0) {
141        drv->mFieldAddress = (void **) calloc(exportVarCount, sizeof(void*));
142        drv->mFieldIsObject = (bool *) calloc(exportVarCount, sizeof(bool));
143        bccGetExportVarList(drv->mBccScript, exportVarCount,
144                            (void **) drv->mFieldAddress);
145    } else {
146        drv->mFieldAddress = NULL;
147        drv->mFieldIsObject = NULL;
148    }
149
150    objectSlotCount = drv->ME->getObjectSlotCount();
151    if (objectSlotCount > 0) {
152        const uint32_t *objectSlotList = drv->ME->getObjectSlotList();
153        for (uint32_t ct=0; ct < objectSlotCount; ct++) {
154            drv->mFieldIsObject[objectSlotList[ct]] = true;
155        }
156    }
157
158    exportForEachSignatureCount = drv->ME->getExportForEachSignatureCount();
159    rsAssert(exportForEachSignatureCount <= 1);
160    drv->mExportForEachSignatureList = drv->ME->getExportForEachSignatureList();
161
162    // Copy info over to runtime
163    script->mHal.info.exportedFunctionCount = drv->ME->getExportFuncCount();
164    script->mHal.info.exportedVariableCount = drv->ME->getExportVarCount();
165    script->mHal.info.exportedPragmaCount = drv->ME->getPragmaCount();
166    script->mHal.info.exportedPragmaKeyList = drv->ME->getPragmaKeyList();
167    script->mHal.info.exportedPragmaValueList = drv->ME->getPragmaValueList();
168    script->mHal.info.root = drv->mRoot;
169
170    pthread_mutex_unlock(&rsdgInitMutex);
171    return true;
172
173error:
174
175    pthread_mutex_unlock(&rsdgInitMutex);
176    if (drv->ME) {
177        delete drv->ME;
178        drv->ME = NULL;
179    }
180    free(drv);
181    return false;
182
183}
184
185typedef struct {
186    Context *rsc;
187    Script *script;
188    uint32_t sig;
189    const Allocation * ain;
190    Allocation * aout;
191    const void * usr;
192    size_t usrLen;
193
194    uint32_t mSliceSize;
195    volatile int mSliceNum;
196
197    const uint8_t *ptrIn;
198    uint32_t eStrideIn;
199    uint8_t *ptrOut;
200    uint32_t eStrideOut;
201
202    uint32_t xStart;
203    uint32_t xEnd;
204    uint32_t yStart;
205    uint32_t yEnd;
206    uint32_t zStart;
207    uint32_t zEnd;
208    uint32_t arrayStart;
209    uint32_t arrayEnd;
210
211    uint32_t dimX;
212    uint32_t dimY;
213    uint32_t dimZ;
214    uint32_t dimArray;
215} MTLaunchStruct;
216typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t);
217
218static void wc_xy(void *usr, uint32_t idx) {
219    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
220    RsForEachStubParamStruct p;
221    memset(&p, 0, sizeof(p));
222    p.usr = mtls->usr;
223    p.usr_len = mtls->usrLen;
224    RsdHal * dc = (RsdHal *)mtls->rsc->mHal.drv;
225    uint32_t sig = mtls->sig;
226
227    outer_foreach_t fn = dc->mForEachLaunch[sig];
228    while (1) {
229        uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum);
230        uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize;
231        uint32_t yEnd = yStart + mtls->mSliceSize;
232        yEnd = rsMin(yEnd, mtls->yEnd);
233        if (yEnd <= yStart) {
234            return;
235        }
236
237        //LOGE("usr idx %i, x %i,%i  y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
238        //LOGE("usr ptr in %p,  out %p", mtls->ptrIn, mtls->ptrOut);
239        for (p.y = yStart; p.y < yEnd; p.y++) {
240            uint32_t offset = mtls->dimX * p.y;
241            p.out = mtls->ptrOut + (mtls->eStrideOut * offset);
242            p.in = mtls->ptrIn + (mtls->eStrideIn * offset);
243            fn(&mtls->script->mHal.info.root, &p, mtls->xStart, mtls->xEnd,
244               mtls->eStrideIn, mtls->eStrideOut);
245        }
246    }
247}
248
249static void wc_x(void *usr, uint32_t idx) {
250    MTLaunchStruct *mtls = (MTLaunchStruct *)usr;
251    RsForEachStubParamStruct p;
252    memset(&p, 0, sizeof(p));
253    p.usr = mtls->usr;
254    p.usr_len = mtls->usrLen;
255    RsdHal * dc = (RsdHal *)mtls->rsc->mHal.drv;
256    uint32_t sig = mtls->sig;
257
258    outer_foreach_t fn = dc->mForEachLaunch[sig];
259    while (1) {
260        uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum);
261        uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize;
262        uint32_t xEnd = xStart + mtls->mSliceSize;
263        xEnd = rsMin(xEnd, mtls->xEnd);
264        if (xEnd <= xStart) {
265            return;
266        }
267
268        //LOGE("usr slice %i idx %i, x %i,%i", slice, idx, xStart, xEnd);
269        //LOGE("usr ptr in %p,  out %p", mtls->ptrIn, mtls->ptrOut);
270
271        p.out = mtls->ptrOut + (mtls->eStrideOut * xStart);
272        p.in = mtls->ptrIn + (mtls->eStrideIn * xStart);
273        fn(&mtls->script->mHal.info.root, &p, xStart, xEnd, mtls->eStrideIn, mtls->eStrideOut);
274    }
275}
276
277void rsdScriptInvokeForEach(const Context *rsc,
278                            Script *s,
279                            uint32_t slot,
280                            const Allocation * ain,
281                            Allocation * aout,
282                            const void * usr,
283                            uint32_t usrLen,
284                            const RsScriptCall *sc) {
285
286    RsdHal * dc = (RsdHal *)rsc->mHal.drv;
287
288    MTLaunchStruct mtls;
289    memset(&mtls, 0, sizeof(mtls));
290
291    DrvScript *drv = (DrvScript *)s->mHal.drv;
292    // We only support slot 0 (root) at this point in time.
293    rsAssert(slot == 0);
294    mtls.sig = 0x1f;  // temp fix for old apps, full table in slang_rs_export_foreach.cpp
295    if (drv->mExportForEachSignatureList) {
296        mtls.sig = drv->mExportForEachSignatureList[slot];
297    }
298    if (ain) {
299        mtls.dimX = ain->getType()->getDimX();
300        mtls.dimY = ain->getType()->getDimY();
301        mtls.dimZ = ain->getType()->getDimZ();
302        //mtls.dimArray = ain->getType()->getDimArray();
303    } else if (aout) {
304        mtls.dimX = aout->getType()->getDimX();
305        mtls.dimY = aout->getType()->getDimY();
306        mtls.dimZ = aout->getType()->getDimZ();
307        //mtls.dimArray = aout->getType()->getDimArray();
308    } else {
309        rsc->setError(RS_ERROR_BAD_SCRIPT, "rsForEach called with null allocations");
310        return;
311    }
312
313    if (!sc || (sc->xEnd == 0)) {
314        mtls.xEnd = mtls.dimX;
315    } else {
316        rsAssert(sc->xStart < mtls.dimX);
317        rsAssert(sc->xEnd <= mtls.dimX);
318        rsAssert(sc->xStart < sc->xEnd);
319        mtls.xStart = rsMin(mtls.dimX, sc->xStart);
320        mtls.xEnd = rsMin(mtls.dimX, sc->xEnd);
321        if (mtls.xStart >= mtls.xEnd) return;
322    }
323
324    if (!sc || (sc->yEnd == 0)) {
325        mtls.yEnd = mtls.dimY;
326    } else {
327        rsAssert(sc->yStart < mtls.dimY);
328        rsAssert(sc->yEnd <= mtls.dimY);
329        rsAssert(sc->yStart < sc->yEnd);
330        mtls.yStart = rsMin(mtls.dimY, sc->yStart);
331        mtls.yEnd = rsMin(mtls.dimY, sc->yEnd);
332        if (mtls.yStart >= mtls.yEnd) return;
333    }
334
335    mtls.xEnd = rsMax((uint32_t)1, mtls.xEnd);
336    mtls.yEnd = rsMax((uint32_t)1, mtls.yEnd);
337    mtls.zEnd = rsMax((uint32_t)1, mtls.zEnd);
338    mtls.arrayEnd = rsMax((uint32_t)1, mtls.arrayEnd);
339
340    rsAssert(!ain || (ain->getType()->getDimZ() == 0));
341
342    Context *mrsc = (Context *)rsc;
343    Script * oldTLS = setTLS(s);
344
345    mtls.rsc = mrsc;
346    mtls.ain = ain;
347    mtls.aout = aout;
348    mtls.script = s;
349    mtls.usr = usr;
350    mtls.usrLen = usrLen;
351    mtls.mSliceSize = 10;
352    mtls.mSliceNum = 0;
353
354    mtls.ptrIn = NULL;
355    mtls.eStrideIn = 0;
356    if (ain) {
357        mtls.ptrIn = (const uint8_t *)ain->getPtr();
358        mtls.eStrideIn = ain->getType()->getElementSizeBytes();
359    }
360
361    mtls.ptrOut = NULL;
362    mtls.eStrideOut = 0;
363    if (aout) {
364        mtls.ptrOut = (uint8_t *)aout->getPtr();
365        mtls.eStrideOut = aout->getType()->getElementSizeBytes();
366    }
367
368    if ((dc->mWorkers.mCount > 1) && s->mHal.info.isThreadable) {
369        if (mtls.dimY > 1) {
370            rsdLaunchThreads(mrsc, wc_xy, &mtls);
371        } else {
372            rsdLaunchThreads(mrsc, wc_x, &mtls);
373        }
374
375        //LOGE("launch 1");
376    } else {
377        RsForEachStubParamStruct p;
378        memset(&p, 0, sizeof(p));
379        p.usr = mtls.usr;
380        p.usr_len = mtls.usrLen;
381        uint32_t sig = mtls.sig;
382
383        //LOGE("launch 3");
384        outer_foreach_t fn = dc->mForEachLaunch[sig];
385        for (p.ar[0] = mtls.arrayStart; p.ar[0] < mtls.arrayEnd; p.ar[0]++) {
386            for (p.z = mtls.zStart; p.z < mtls.zEnd; p.z++) {
387                for (p.y = mtls.yStart; p.y < mtls.yEnd; p.y++) {
388                    uint32_t offset = mtls.dimX * mtls.dimY * mtls.dimZ * p.ar[0] +
389                                      mtls.dimX * mtls.dimY * p.z +
390                                      mtls.dimX * p.y;
391                    p.out = mtls.ptrOut + (mtls.eStrideOut * offset);
392                    p.in = mtls.ptrIn + (mtls.eStrideIn * offset);
393                    fn(&mtls.script->mHal.info.root, &p, mtls.xStart, mtls.xEnd,
394                       mtls.eStrideIn, mtls.eStrideOut);
395                }
396            }
397        }
398    }
399
400    setTLS(oldTLS);
401}
402
403
404int rsdScriptInvokeRoot(const Context *dc, Script *script) {
405    DrvScript *drv = (DrvScript *)script->mHal.drv;
406
407    Script * oldTLS = setTLS(script);
408    int ret = drv->mRoot();
409    setTLS(oldTLS);
410
411    return ret;
412}
413
414void rsdScriptInvokeInit(const Context *dc, Script *script) {
415    DrvScript *drv = (DrvScript *)script->mHal.drv;
416
417    if (drv->mInit) {
418        drv->mInit();
419    }
420}
421
422void rsdScriptInvokeFreeChildren(const Context *dc, Script *script) {
423    DrvScript *drv = (DrvScript *)script->mHal.drv;
424
425    if (drv->mFreeChildren) {
426        drv->mFreeChildren();
427    }
428}
429
430void rsdScriptInvokeFunction(const Context *dc, Script *script,
431                            uint32_t slot,
432                            const void *params,
433                            size_t paramLength) {
434    DrvScript *drv = (DrvScript *)script->mHal.drv;
435    //LOGE("invoke %p %p %i %p %i", dc, script, slot, params, paramLength);
436
437    Script * oldTLS = setTLS(script);
438    ((void (*)(const void *, uint32_t))
439        drv->mInvokeFunctions[slot])(params, paramLength);
440    setTLS(oldTLS);
441}
442
443void rsdScriptSetGlobalVar(const Context *dc, const Script *script,
444                           uint32_t slot, void *data, size_t dataLength) {
445    DrvScript *drv = (DrvScript *)script->mHal.drv;
446    //rsAssert(!script->mFieldIsObject[slot]);
447    //LOGE("setGlobalVar %p %p %i %p %i", dc, script, slot, data, dataLength);
448
449    int32_t *destPtr = ((int32_t **)drv->mFieldAddress)[slot];
450    if (!destPtr) {
451        //ALOGV("Calling setVar on slot = %i which is null", slot);
452        return;
453    }
454
455    memcpy(destPtr, data, dataLength);
456}
457
458void rsdScriptSetGlobalBind(const Context *dc, const Script *script, uint32_t slot, void *data) {
459    DrvScript *drv = (DrvScript *)script->mHal.drv;
460    //rsAssert(!script->mFieldIsObject[slot]);
461    //LOGE("setGlobalBind %p %p %i %p", dc, script, slot, data);
462
463    int32_t *destPtr = ((int32_t **)drv->mFieldAddress)[slot];
464    if (!destPtr) {
465        //ALOGV("Calling setVar on slot = %i which is null", slot);
466        return;
467    }
468
469    memcpy(destPtr, &data, sizeof(void *));
470}
471
472void rsdScriptSetGlobalObj(const Context *dc, const Script *script, uint32_t slot, ObjectBase *data) {
473    DrvScript *drv = (DrvScript *)script->mHal.drv;
474    //rsAssert(script->mFieldIsObject[slot]);
475    //LOGE("setGlobalObj %p %p %i %p", dc, script, slot, data);
476
477    int32_t *destPtr = ((int32_t **)drv->mFieldAddress)[slot];
478    if (!destPtr) {
479        //ALOGV("Calling setVar on slot = %i which is null", slot);
480        return;
481    }
482
483    rsrSetObject(dc, script, (ObjectBase **)destPtr, data);
484}
485
486void rsdScriptDestroy(const Context *dc, Script *script) {
487    DrvScript *drv = (DrvScript *)script->mHal.drv;
488
489    if (drv->mFieldAddress) {
490        size_t exportVarCount = drv->ME->getExportVarCount();
491        for (size_t ct = 0; ct < exportVarCount; ct++) {
492            if (drv->mFieldIsObject[ct]) {
493                // The field address can be NULL if the script-side has
494                // optimized the corresponding global variable away.
495                if (drv->mFieldAddress[ct]) {
496                    rsrClearObject(dc, script, (ObjectBase **)drv->mFieldAddress[ct]);
497                }
498            }
499        }
500        free(drv->mFieldAddress);
501        drv->mFieldAddress = NULL;
502        free(drv->mFieldIsObject);
503        drv->mFieldIsObject = NULL;    }
504
505    if (drv->mInvokeFunctions) {
506        free(drv->mInvokeFunctions);
507        drv->mInvokeFunctions = NULL;
508    }
509
510    delete drv->ME;
511    drv->ME = NULL;
512
513    free(drv);
514    script->mHal.drv = NULL;
515
516}
517
518
519