rsScriptC.cpp revision 900f1616bf33c7ba13cf2a737832a95bcd176388
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 23#include <GLES/gl.h> 24#include <GLES/glext.h> 25 26using namespace android; 27using namespace android::renderscript; 28 29#define GET_TLS() Context::ScriptTLSStruct * tls = \ 30 (Context::ScriptTLSStruct *)pthread_getspecific(Context::gThreadTLSKey); \ 31 Context * rsc = tls->mContext; \ 32 ScriptC * sc = (ScriptC *) tls->mScript 33 34 35ScriptC::ScriptC(Context *rsc) : Script(rsc) 36{ 37 mAllocFile = __FILE__; 38 mAllocLine = __LINE__; 39 mBccScript = NULL; 40 memset(&mProgram, 0, sizeof(mProgram)); 41} 42 43ScriptC::~ScriptC() 44{ 45 if (mBccScript) { 46 bccDeleteScript(mBccScript); 47 } 48 free(mEnviroment.mScriptText); 49 mEnviroment.mScriptText = NULL; 50} 51 52void ScriptC::setupScript(Context *rsc) 53{ 54 setupGLState(rsc); 55 mEnviroment.mStartTimeMillis 56 = nanoseconds_to_milliseconds(systemTime(SYSTEM_TIME_MONOTONIC)); 57 58 for (uint32_t ct=0; ct < mEnviroment.mFieldCount; ct++) { 59 if (mSlots[ct].get() && !mTypes[ct].get()) { 60 mTypes[ct].set(mSlots[ct]->getType()); 61 } 62 63 if (!mTypes[ct].get()) 64 continue; 65 void *ptr = NULL; 66 if (mSlots[ct].get()) { 67 ptr = mSlots[ct]->getPtr(); 68 } 69 void **dest = ((void ***)mEnviroment.mFieldAddress)[ct]; 70 //LOGE("setupScript %i %p = %p %p %i", ct, dest, ptr, mSlots[ct]->getType(), mSlots[ct]->getType()->getDimX()); 71 72 //const uint32_t *p32 = (const uint32_t *)ptr; 73 //for (uint32_t ct2=0; ct2 < mSlots[ct]->getType()->getDimX(); ct2++) { 74 //LOGE(" %i = 0x%08x ", ct2, p32[ct2]); 75 //} 76 77 if (dest) { 78 *dest = ptr; 79 } else { 80 LOGE("ScriptC::setupScript, NULL var binding address."); 81 } 82 } 83} 84 85const Allocation *ScriptC::ptrToAllocation(const void *ptr) const 86{ 87 if (!ptr) { 88 return NULL; 89 } 90 for (uint32_t ct=0; ct < mEnviroment.mFieldCount; ct++) { 91 if (!mSlots[ct].get()) 92 continue; 93 if (mSlots[ct]->getPtr() == ptr) { 94 return mSlots[ct].get(); 95 } 96 } 97 LOGE("ScriptC::ptrToAllocation, failed to find %p", ptr); 98 return NULL; 99} 100 101Script * ScriptC::setTLS(Script *sc) 102{ 103 Context::ScriptTLSStruct * tls = (Context::ScriptTLSStruct *) 104 pthread_getspecific(Context::gThreadTLSKey); 105 rsAssert(tls); 106 Script *old = tls->mScript; 107 tls->mScript = sc; 108 return old; 109} 110 111 112void ScriptC::setupGLState(Context *rsc) 113{ 114 if (mEnviroment.mFragmentStore.get()) { 115 rsc->setFragmentStore(mEnviroment.mFragmentStore.get()); 116 } 117 if (mEnviroment.mFragment.get()) { 118 rsc->setFragment(mEnviroment.mFragment.get()); 119 } 120 if (mEnviroment.mVertex.get()) { 121 rsc->setVertex(mEnviroment.mVertex.get()); 122 } 123 if (mEnviroment.mRaster.get()) { 124 rsc->setRaster(mEnviroment.mRaster.get()); 125 } 126} 127 128uint32_t ScriptC::run(Context *rsc) 129{ 130 if (mProgram.mRoot == NULL) { 131 rsc->setError(RS_ERROR_BAD_SCRIPT, "Attempted to run bad script"); 132 return 0; 133 } 134 135 setupScript(rsc); 136 137 uint32_t ret = 0; 138 Script * oldTLS = setTLS(this); 139 //LOGE("ScriptC::run %p", mProgram.mRoot); 140 ret = mProgram.mRoot(); 141 setTLS(oldTLS); 142 //LOGE("ScriptC::run ret %i", ret); 143 return ret; 144} 145 146 147typedef struct { 148 Context *rsc; 149 ScriptC *script; 150 const Allocation * ain; 151 Allocation * aout; 152 const void * usr; 153 154 uint32_t mSliceSize; 155 volatile int mSliceNum; 156 157 const uint8_t *ptrIn; 158 uint32_t eStrideIn; 159 uint8_t *ptrOut; 160 uint32_t eStrideOut; 161 162 uint32_t xStart; 163 uint32_t xEnd; 164 uint32_t yStart; 165 uint32_t yEnd; 166 uint32_t zStart; 167 uint32_t zEnd; 168 uint32_t arrayStart; 169 uint32_t arrayEnd; 170 171 uint32_t dimX; 172 uint32_t dimY; 173 uint32_t dimZ; 174 uint32_t dimArray; 175} MTLaunchStruct; 176typedef int (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t); 177 178static void wc_xy(void *usr, uint32_t idx) 179{ 180 MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 181 182 while (1) { 183 uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum); 184 uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize; 185 uint32_t yEnd = yStart + mtls->mSliceSize; 186 yEnd = rsMin(yEnd, mtls->yEnd); 187 if (yEnd <= yStart) { 188 return; 189 } 190 191 //LOGE("usr idx %i, x %i,%i y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd); 192 193 for (uint32_t y = yStart; y < yEnd; y++) { 194 uint32_t offset = mtls->dimX * y; 195 uint8_t *xPtrOut = mtls->ptrOut + (mtls->eStrideOut * offset); 196 const uint8_t *xPtrIn = mtls->ptrIn + (mtls->eStrideIn * offset); 197 198 for (uint32_t x = mtls->xStart; x < mtls->xEnd; x++) { 199 ((rs_t)mtls->script->mProgram.mRoot) (xPtrIn, xPtrOut, mtls->usr, x, y, 0, 0); 200 xPtrIn += mtls->eStrideIn; 201 xPtrOut += mtls->eStrideOut; 202 } 203 } 204 } 205 206} 207 208void ScriptC::runForEach(Context *rsc, 209 const Allocation * ain, 210 Allocation * aout, 211 const void * usr, 212 const RsScriptCall *sc) 213{ 214 MTLaunchStruct mtls; 215 memset(&mtls, 0, sizeof(mtls)); 216 217 if (ain) { 218 mtls.dimX = ain->getType()->getDimX(); 219 mtls.dimY = ain->getType()->getDimY(); 220 mtls.dimZ = ain->getType()->getDimZ(); 221 //mtls.dimArray = ain->getType()->getDimArray(); 222 } else if (aout) { 223 mtls.dimX = aout->getType()->getDimX(); 224 mtls.dimY = aout->getType()->getDimY(); 225 mtls.dimZ = aout->getType()->getDimZ(); 226 //mtls.dimArray = aout->getType()->getDimArray(); 227 } else { 228 rsc->setError(RS_ERROR_BAD_SCRIPT, "rsForEach called with null allocations"); 229 return; 230 } 231 232 if (!sc || (sc->xEnd == 0)) { 233 mtls.xEnd = mtls.dimX; 234 } else { 235 rsAssert(sc->xStart < mtls.dimX); 236 rsAssert(sc->xEnd <= mtls.dimX); 237 rsAssert(sc->xStart < sc->xEnd); 238 mtls.xStart = rsMin(mtls.dimX, sc->xStart); 239 mtls.xEnd = rsMin(mtls.dimX, sc->xEnd); 240 if (mtls.xStart >= mtls.xEnd) return; 241 } 242 243 if (!sc || (sc->yEnd == 0)) { 244 mtls.yEnd = mtls.dimY; 245 } else { 246 rsAssert(sc->yStart < mtls.dimY); 247 rsAssert(sc->yEnd <= mtls.dimY); 248 rsAssert(sc->yStart < sc->yEnd); 249 mtls.yStart = rsMin(mtls.dimY, sc->yStart); 250 mtls.yEnd = rsMin(mtls.dimY, sc->yEnd); 251 if (mtls.yStart >= mtls.yEnd) return; 252 } 253 254 mtls.xEnd = rsMax((uint32_t)1, mtls.xEnd); 255 mtls.yEnd = rsMax((uint32_t)1, mtls.yEnd); 256 mtls.zEnd = rsMax((uint32_t)1, mtls.zEnd); 257 mtls.arrayEnd = rsMax((uint32_t)1, mtls.arrayEnd); 258 259 rsAssert(ain->getType()->getDimZ() == 0); 260 261 setupScript(rsc); 262 Script * oldTLS = setTLS(this); 263 264 265 mtls.rsc = rsc; 266 mtls.ain = ain; 267 mtls.aout = aout; 268 mtls.script = this; 269 mtls.usr = usr; 270 mtls.mSliceSize = 10; 271 mtls.mSliceNum = 0; 272 273 mtls.ptrIn = NULL; 274 mtls.eStrideIn = 0; 275 if (ain) { 276 mtls.ptrIn = (const uint8_t *)ain->getPtr(); 277 mtls.eStrideIn = ain->getType()->getElementSizeBytes(); 278 } 279 280 mtls.ptrOut = NULL; 281 mtls.eStrideOut = 0; 282 if (aout) { 283 mtls.ptrOut = (uint8_t *)aout->getPtr(); 284 mtls.eStrideOut = aout->getType()->getElementSizeBytes(); 285 } 286 287 288 if ((rsc->getWorkerPoolSize() > 1) && mEnviroment.mIsThreadable && 289 ((mtls.dimY * mtls.dimZ * mtls.dimArray) > 1)) { 290 291 //LOGE("launch 1"); 292 rsc->launchThreads(wc_xy, &mtls); 293 //LOGE("launch 2"); 294 } else { 295 for (uint32_t ar = mtls.arrayStart; ar < mtls.arrayEnd; ar++) { 296 for (uint32_t z = mtls.zStart; z < mtls.zEnd; z++) { 297 for (uint32_t y = mtls.yStart; y < mtls.yEnd; y++) { 298 uint32_t offset = mtls.dimX * mtls.dimY * mtls.dimZ * ar + 299 mtls.dimX * mtls.dimY * z + 300 mtls.dimX * y; 301 uint8_t *xPtrOut = mtls.ptrOut + (mtls.eStrideOut * offset); 302 const uint8_t *xPtrIn = mtls.ptrIn + (mtls.eStrideIn * offset); 303 304 for (uint32_t x = mtls.xStart; x < mtls.xEnd; x++) { 305 ((rs_t)mProgram.mRoot) (xPtrIn, xPtrOut, usr, x, y, z, ar); 306 xPtrIn += mtls.eStrideIn; 307 xPtrOut += mtls.eStrideOut; 308 } 309 } 310 } 311 } 312 } 313 314 setTLS(oldTLS); 315} 316 317void ScriptC::Invoke(Context *rsc, uint32_t slot, const void *data, uint32_t len) 318{ 319 //LOGE("rsi_ScriptInvoke %i", slot); 320 if ((slot >= mEnviroment.mInvokeFunctionCount) || 321 (mEnviroment.mInvokeFunctions[slot] == NULL)) { 322 rsc->setError(RS_ERROR_BAD_SCRIPT, "Calling invoke on bad script"); 323 return; 324 } 325 setupScript(rsc); 326 Script * oldTLS = setTLS(this); 327 328 ((void (*)(const void *, uint32_t)) 329 mEnviroment.mInvokeFunctions[slot])(data, len); 330 331 setTLS(oldTLS); 332} 333 334ScriptCState::ScriptCState() 335{ 336 mScript = NULL; 337 clear(); 338} 339 340ScriptCState::~ScriptCState() 341{ 342 delete mScript; 343 mScript = NULL; 344} 345 346void ScriptCState::clear() 347{ 348 for (uint32_t ct=0; ct < MAX_SCRIPT_BANKS; ct++) { 349 mConstantBufferTypes[ct].clear(); 350 mSlotWritable[ct] = false; 351 } 352 353 delete mScript; 354 mScript = new ScriptC(NULL); 355} 356 357static BCCvoid* symbolLookup(BCCvoid* pContext, const BCCchar* name) 358{ 359 const ScriptCState::SymbolTable_t *sym; 360 ScriptC *s = (ScriptC *)pContext; 361 sym = ScriptCState::lookupSymbol(name); 362 if (sym) { 363 return sym->mPtr; 364 } 365 s->mEnviroment.mIsThreadable = false; 366 sym = ScriptCState::lookupSymbolCL(name); 367 if (sym) { 368 return sym->mPtr; 369 } 370 sym = ScriptCState::lookupSymbolGL(name); 371 if (sym) { 372 return sym->mPtr; 373 } 374 LOGE("ScriptC sym lookup failed for %s", name); 375 return NULL; 376} 377 378void ScriptCState::runCompiler(Context *rsc, ScriptC *s) 379{ 380 LOGV("ScriptCState::runCompiler "); 381 382 s->mBccScript = bccCreateScript(); 383 s->mEnviroment.mIsThreadable = true; 384 bccScriptBitcode(s->mBccScript, s->mEnviroment.mScriptText, s->mEnviroment.mScriptTextLength); 385 bccRegisterSymbolCallback(s->mBccScript, symbolLookup, s); 386 bccCompileScript(s->mBccScript); 387 bccGetScriptLabel(s->mBccScript, "root", (BCCvoid**) &s->mProgram.mRoot); 388 bccGetScriptLabel(s->mBccScript, "init", (BCCvoid**) &s->mProgram.mInit); 389 LOGV("root %p, init %p", s->mProgram.mRoot, s->mProgram.mInit); 390 391 if (s->mProgram.mInit) { 392 s->mProgram.mInit(); 393 } 394 395 bccGetExportFuncs(s->mBccScript, (BCCsizei*) &s->mEnviroment.mInvokeFunctionCount, 0, NULL); 396 if(s->mEnviroment.mInvokeFunctionCount <= 0) 397 s->mEnviroment.mInvokeFunctions = NULL; 398 else { 399 s->mEnviroment.mInvokeFunctions = (Script::InvokeFunc_t*) calloc(s->mEnviroment.mInvokeFunctionCount, sizeof(Script::InvokeFunc_t)); 400 bccGetExportFuncs(s->mBccScript, NULL, s->mEnviroment.mInvokeFunctionCount, (BCCvoid **) s->mEnviroment.mInvokeFunctions); 401 } 402 403 bccGetExportVars(s->mBccScript, (BCCsizei*) &s->mEnviroment.mFieldCount, 0, NULL); 404 if(s->mEnviroment.mFieldCount <= 0) 405 s->mEnviroment.mFieldAddress = NULL; 406 else { 407 s->mEnviroment.mFieldAddress = (void **) calloc(s->mEnviroment.mFieldCount, sizeof(void *)); 408 bccGetExportVars(s->mBccScript, NULL, s->mEnviroment.mFieldCount, (BCCvoid **) s->mEnviroment.mFieldAddress); 409 } 410 //for (int ct2=0; ct2 < s->mEnviroment.mFieldCount; ct2++ ) { 411 //LOGE("Script field %i = %p", ct2, s->mEnviroment.mFieldAddress[ct2]); 412 //} 413 414 s->mEnviroment.mFragment.set(rsc->getDefaultProgramFragment()); 415 s->mEnviroment.mVertex.set(rsc->getDefaultProgramVertex()); 416 s->mEnviroment.mFragmentStore.set(rsc->getDefaultProgramStore()); 417 s->mEnviroment.mRaster.set(rsc->getDefaultProgramRaster()); 418 419 if (s->mProgram.mRoot) { 420 const static int pragmaMax = 16; 421 BCCsizei pragmaCount; 422 BCCchar * str[pragmaMax]; 423 bccGetPragmas(s->mBccScript, &pragmaCount, pragmaMax, &str[0]); 424 425 for (int ct=0; ct < pragmaCount; ct+=2) { 426 //LOGE("pragme %s %s", str[ct], str[ct+1]); 427 if (!strcmp(str[ct], "version")) { 428 continue; 429 } 430 431 if (!strcmp(str[ct], "stateVertex")) { 432 if (!strcmp(str[ct+1], "default")) { 433 continue; 434 } 435 if (!strcmp(str[ct+1], "parent")) { 436 s->mEnviroment.mVertex.clear(); 437 continue; 438 } 439 LOGE("Unreconized value %s passed to stateVertex", str[ct+1]); 440 } 441 442 if (!strcmp(str[ct], "stateRaster")) { 443 if (!strcmp(str[ct+1], "default")) { 444 continue; 445 } 446 if (!strcmp(str[ct+1], "parent")) { 447 s->mEnviroment.mRaster.clear(); 448 continue; 449 } 450 LOGE("Unreconized value %s passed to stateRaster", str[ct+1]); 451 } 452 453 if (!strcmp(str[ct], "stateFragment")) { 454 if (!strcmp(str[ct+1], "default")) { 455 continue; 456 } 457 if (!strcmp(str[ct+1], "parent")) { 458 s->mEnviroment.mFragment.clear(); 459 continue; 460 } 461 LOGE("Unreconized value %s passed to stateFragment", str[ct+1]); 462 } 463 464 if (!strcmp(str[ct], "stateStore")) { 465 if (!strcmp(str[ct+1], "default")) { 466 continue; 467 } 468 if (!strcmp(str[ct+1], "parent")) { 469 s->mEnviroment.mFragmentStore.clear(); 470 continue; 471 } 472 LOGE("Unreconized value %s passed to stateStore", str[ct+1]); 473 } 474 475 } 476 477 478 } else { 479 // Deal with an error. 480 } 481} 482 483 484 485namespace android { 486namespace renderscript { 487 488void rsi_ScriptCBegin(Context * rsc) 489{ 490 ScriptCState *ss = &rsc->mScriptC; 491 ss->clear(); 492} 493 494void rsi_ScriptCSetScript(Context * rsc, void *vp) 495{ 496 rsAssert(0); 497 //ScriptCState *ss = &rsc->mScriptC; 498 //ss->mProgram.mScript = reinterpret_cast<ScriptC::RunScript_t>(vp); 499} 500 501void rsi_ScriptCSetText(Context *rsc, const char *text, uint32_t len) 502{ 503 ScriptCState *ss = &rsc->mScriptC; 504 505 char *t = (char *)malloc(len + 1); 506 memcpy(t, text, len); 507 t[len] = 0; 508 ss->mScript->mEnviroment.mScriptText = t; 509 ss->mScript->mEnviroment.mScriptTextLength = len; 510} 511 512 513RsScript rsi_ScriptCCreate(Context * rsc) 514{ 515 ScriptCState *ss = &rsc->mScriptC; 516 517 ScriptC *s = ss->mScript; 518 ss->mScript = NULL; 519 520 ss->runCompiler(rsc, s); 521 s->incUserRef(); 522 s->setContext(rsc); 523 for (int ct=0; ct < MAX_SCRIPT_BANKS; ct++) { 524 s->mTypes[ct].set(ss->mConstantBufferTypes[ct].get()); 525 s->mSlotWritable[ct] = ss->mSlotWritable[ct]; 526 } 527 528 ss->clear(); 529 return s; 530} 531 532} 533} 534 535 536