Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 1 | /* |
Stephen Hines | ee7aa2e | 2012-01-12 18:56:23 -0800 | [diff] [blame] | 2 | * Copyright (C) 2011-2012 The Android Open Source Project |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 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 | |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 17 | #include "rsdCore.h" |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 18 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 19 | #include <bcc/BCCContext.h> |
Stephen Hines | 378d30b | 2012-07-27 18:11:03 -0700 | [diff] [blame] | 20 | #include <bcc/Renderscript/RSCompilerDriver.h> |
| 21 | #include <bcc/Renderscript/RSExecutable.h> |
| 22 | #include <bcc/Renderscript/RSInfo.h> |
Stephen Hines | 689821f | 2011-07-18 17:24:11 -0700 | [diff] [blame] | 23 | |
Jason Sams | dbe66d6 | 2012-09-17 13:54:41 -0700 | [diff] [blame^] | 24 | #include "rsdBcc.h" |
| 25 | #include "rsdRuntime.h" |
| 26 | #include "rsdAllocation.h" |
| 27 | #include "rsdIntrinsics.h" |
| 28 | |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 29 | #include "rsContext.h" |
Stephen Hines | 2980f07 | 2012-04-09 18:26:29 -0700 | [diff] [blame] | 30 | #include "rsElement.h" |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 31 | #include "rsScriptC.h" |
| 32 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 33 | #include "utils/Vector.h" |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 34 | #include "utils/Timers.h" |
| 35 | #include "utils/StopWatch.h" |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 36 | |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 37 | using namespace android; |
| 38 | using namespace android::renderscript; |
| 39 | |
Jason Sams | 83c451a | 2011-04-21 11:46:50 -0700 | [diff] [blame] | 40 | |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 41 | static Script * setTLS(Script *sc) { |
Jason Sams | 83c451a | 2011-04-21 11:46:50 -0700 | [diff] [blame] | 42 | ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(rsdgThreadTLSKey); |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 43 | rsAssert(tls); |
| 44 | Script *old = tls->mScript; |
| 45 | tls->mScript = sc; |
| 46 | return old; |
| 47 | } |
| 48 | |
| 49 | |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 50 | bool rsdScriptInit(const Context *rsc, |
| 51 | ScriptC *script, |
| 52 | char const *resName, |
| 53 | char const *cacheDir, |
| 54 | uint8_t const *bitcode, |
| 55 | size_t bitcodeSize, |
Jason Sams | 87fe59a | 2011-04-20 15:09:01 -0700 | [diff] [blame] | 56 | uint32_t flags) { |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 57 | //ALOGE("rsdScriptCreate %p %p %p %p %i %i %p", rsc, resName, cacheDir, bitcode, bitcodeSize, flags, lookupFunc); |
Jason Sams | 0ab9f9f | 2012-06-15 16:10:17 -0700 | [diff] [blame] | 58 | //ALOGE("rsdScriptInit %p %p", rsc, script); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 59 | |
Jason Sams | 83c451a | 2011-04-21 11:46:50 -0700 | [diff] [blame] | 60 | pthread_mutex_lock(&rsdgInitMutex); |
Logan Chien | 1415ca4 | 2011-11-27 08:41:23 +0800 | [diff] [blame] | 61 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 62 | bcc::RSExecutable *exec; |
| 63 | const bcc::RSInfo *info; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 64 | DrvScript *drv = (DrvScript *)calloc(1, sizeof(DrvScript)); |
| 65 | if (drv == NULL) { |
Jason Sams | 83c451a | 2011-04-21 11:46:50 -0700 | [diff] [blame] | 66 | goto error; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 67 | } |
| 68 | script->mHal.drv = drv; |
| 69 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 70 | drv->mCompilerContext = NULL; |
| 71 | drv->mCompilerDriver = NULL; |
| 72 | drv->mExecutable = NULL; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 73 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 74 | drv->mCompilerContext = new bcc::BCCContext(); |
| 75 | if (drv->mCompilerContext == NULL) { |
| 76 | ALOGE("bcc: FAILS to create compiler context (out of memory)"); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 77 | goto error; |
| 78 | } |
| 79 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 80 | drv->mCompilerDriver = new bcc::RSCompilerDriver(); |
| 81 | if (drv->mCompilerDriver == NULL) { |
| 82 | ALOGE("bcc: FAILS to create compiler driver (out of memory)"); |
Stephen Hines | 070cb23 | 2012-05-03 12:33:35 -0700 | [diff] [blame] | 83 | goto error; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 84 | } |
| 85 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 86 | script->mHal.info.isThreadable = true; |
| 87 | |
| 88 | drv->mCompilerDriver->setRSRuntimeLookupFunction(rsdLookupRuntimeStub); |
| 89 | drv->mCompilerDriver->setRSRuntimeLookupContext(script); |
| 90 | |
| 91 | exec = drv->mCompilerDriver->build(*drv->mCompilerContext, |
| 92 | cacheDir, resName, |
| 93 | (const char *)bitcode, bitcodeSize); |
| 94 | |
| 95 | if (exec == NULL) { |
| 96 | ALOGE("bcc: FAILS to prepare executable for '%s'", resName); |
Stephen Hines | 070cb23 | 2012-05-03 12:33:35 -0700 | [diff] [blame] | 97 | goto error; |
| 98 | } |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 99 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 100 | drv->mExecutable = exec; |
| 101 | |
| 102 | exec->setThreadable(script->mHal.info.isThreadable); |
| 103 | if (!exec->syncInfo()) { |
| 104 | ALOGW("bcc: FAILS to synchronize the RS info file to the disk"); |
Stephen Hines | 070cb23 | 2012-05-03 12:33:35 -0700 | [diff] [blame] | 105 | } |
| 106 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 107 | drv->mRoot = reinterpret_cast<int (*)()>(exec->getSymbolAddress("root")); |
| 108 | drv->mRootExpand = |
| 109 | reinterpret_cast<int (*)()>(exec->getSymbolAddress("root.expand")); |
| 110 | drv->mInit = reinterpret_cast<void (*)()>(exec->getSymbolAddress("init")); |
| 111 | drv->mFreeChildren = |
| 112 | reinterpret_cast<void (*)()>(exec->getSymbolAddress(".rs.dtor")); |
Stephen Hines | 070cb23 | 2012-05-03 12:33:35 -0700 | [diff] [blame] | 113 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 114 | info = &drv->mExecutable->getInfo(); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 115 | // Copy info over to runtime |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 116 | script->mHal.info.exportedFunctionCount = info->getExportFuncNames().size(); |
| 117 | script->mHal.info.exportedVariableCount = info->getExportVarNames().size(); |
| 118 | script->mHal.info.exportedPragmaCount = info->getPragmas().size(); |
| 119 | script->mHal.info.exportedPragmaKeyList = |
| 120 | const_cast<const char**>(exec->getPragmaKeys().array()); |
| 121 | script->mHal.info.exportedPragmaValueList = |
| 122 | const_cast<const char**>(exec->getPragmaValues().array()); |
Stephen Hines | ee7aa2e | 2012-01-12 18:56:23 -0800 | [diff] [blame] | 123 | |
| 124 | if (drv->mRootExpand) { |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 125 | script->mHal.info.root = drv->mRootExpand; |
Stephen Hines | ee7aa2e | 2012-01-12 18:56:23 -0800 | [diff] [blame] | 126 | } else { |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 127 | script->mHal.info.root = drv->mRoot; |
Stephen Hines | ee7aa2e | 2012-01-12 18:56:23 -0800 | [diff] [blame] | 128 | } |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 129 | |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 130 | if (script->mHal.info.exportedVariableCount) { |
| 131 | drv->mBoundAllocs = new Allocation *[script->mHal.info.exportedVariableCount]; |
| 132 | memset(drv->mBoundAllocs, 0, sizeof(void *) * script->mHal.info.exportedVariableCount); |
| 133 | } |
| 134 | |
Jason Sams | 83c451a | 2011-04-21 11:46:50 -0700 | [diff] [blame] | 135 | pthread_mutex_unlock(&rsdgInitMutex); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 136 | return true; |
| 137 | |
| 138 | error: |
| 139 | |
Jason Sams | 83c451a | 2011-04-21 11:46:50 -0700 | [diff] [blame] | 140 | pthread_mutex_unlock(&rsdgInitMutex); |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 141 | if (drv) { |
| 142 | delete drv->mCompilerContext; |
| 143 | delete drv->mCompilerDriver; |
| 144 | delete drv->mExecutable; |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 145 | delete[] drv->mBoundAllocs; |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 146 | free(drv); |
Stephen Hines | cbb0b8a | 2011-08-01 15:02:34 -0700 | [diff] [blame] | 147 | } |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 148 | script->mHal.drv = NULL; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 149 | return false; |
| 150 | |
| 151 | } |
| 152 | |
Stephen Hines | 41d6c76 | 2012-08-21 17:07:38 -0700 | [diff] [blame] | 153 | bool rsdInitIntrinsic(const Context *rsc, Script *s, RsScriptIntrinsicID iid, Element *e) { |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 154 | pthread_mutex_lock(&rsdgInitMutex); |
| 155 | |
| 156 | DrvScript *drv = (DrvScript *)calloc(1, sizeof(DrvScript)); |
| 157 | if (drv == NULL) { |
| 158 | goto error; |
| 159 | } |
| 160 | s->mHal.drv = drv; |
| 161 | drv->mIntrinsicID = iid; |
Jason Sams | e1e08b4 | 2012-09-04 16:49:19 -0700 | [diff] [blame] | 162 | drv->mIntrinsicData = rsdIntrinsic_Init(rsc, s, iid, &drv->mIntrinsicFuncs); |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 163 | |
| 164 | pthread_mutex_unlock(&rsdgInitMutex); |
| 165 | return true; |
| 166 | |
| 167 | error: |
| 168 | pthread_mutex_unlock(&rsdgInitMutex); |
| 169 | return false; |
| 170 | } |
| 171 | |
Stephen Hines | 3815bad | 2011-08-18 19:33:01 -0700 | [diff] [blame] | 172 | typedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t); |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 173 | |
| 174 | static void wc_xy(void *usr, uint32_t idx) { |
| 175 | MTLaunchStruct *mtls = (MTLaunchStruct *)usr; |
Jason Sams | 451cf2e | 2011-08-17 13:46:46 -0700 | [diff] [blame] | 176 | RsForEachStubParamStruct p; |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 177 | memcpy(&p, &mtls->fep, sizeof(p)); |
Stephen Hines | 3815bad | 2011-08-18 19:33:01 -0700 | [diff] [blame] | 178 | RsdHal * dc = (RsdHal *)mtls->rsc->mHal.drv; |
| 179 | uint32_t sig = mtls->sig; |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 180 | |
Stephen Hines | 4419977 | 2012-02-21 20:13:12 -0800 | [diff] [blame] | 181 | outer_foreach_t fn = (outer_foreach_t) mtls->kernel; |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 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 | |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 191 | //ALOGE("usr idx %i, x %i,%i y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd); |
| 192 | //ALOGE("usr ptr in %p, out %p", mtls->ptrIn, mtls->ptrOut); |
Jason Sams | 451cf2e | 2011-08-17 13:46:46 -0700 | [diff] [blame] | 193 | for (p.y = yStart; p.y < yEnd; p.y++) { |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 194 | p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * p.y); |
| 195 | p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * p.y); |
| 196 | fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 197 | } |
| 198 | } |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 199 | } |
| 200 | |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 201 | static void wc_x(void *usr, uint32_t idx) { |
| 202 | MTLaunchStruct *mtls = (MTLaunchStruct *)usr; |
Jason Sams | 451cf2e | 2011-08-17 13:46:46 -0700 | [diff] [blame] | 203 | RsForEachStubParamStruct p; |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 204 | memcpy(&p, &mtls->fep, sizeof(p)); |
Stephen Hines | 3815bad | 2011-08-18 19:33:01 -0700 | [diff] [blame] | 205 | RsdHal * dc = (RsdHal *)mtls->rsc->mHal.drv; |
| 206 | uint32_t sig = mtls->sig; |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 207 | |
Stephen Hines | 4419977 | 2012-02-21 20:13:12 -0800 | [diff] [blame] | 208 | outer_foreach_t fn = (outer_foreach_t) mtls->kernel; |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 209 | while (1) { |
| 210 | uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum); |
| 211 | uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize; |
| 212 | uint32_t xEnd = xStart + mtls->mSliceSize; |
| 213 | xEnd = rsMin(xEnd, mtls->xEnd); |
| 214 | if (xEnd <= xStart) { |
| 215 | return; |
| 216 | } |
| 217 | |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 218 | //ALOGE("usr slice %i idx %i, x %i,%i", slice, idx, xStart, xEnd); |
| 219 | //ALOGE("usr ptr in %p, out %p", mtls->ptrIn, mtls->ptrOut); |
Jason Sams | 96cd136 | 2011-10-12 18:33:01 -0700 | [diff] [blame] | 220 | |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 221 | p.out = mtls->fep.ptrOut + (mtls->fep.eStrideOut * xStart); |
| 222 | p.in = mtls->fep.ptrIn + (mtls->fep.eStrideIn * xStart); |
| 223 | fn(&p, xStart, xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 224 | } |
| 225 | } |
| 226 | |
Jason Sams | dbe66d6 | 2012-09-17 13:54:41 -0700 | [diff] [blame^] | 227 | void rsdScriptInvokeForEachMtlsSetup(const Context *rsc, |
| 228 | const Allocation * ain, |
| 229 | Allocation * aout, |
| 230 | const void * usr, |
| 231 | uint32_t usrLen, |
| 232 | const RsScriptCall *sc, |
| 233 | MTLaunchStruct *mtls) { |
| 234 | |
| 235 | memset(mtls, 0, sizeof(MTLaunchStruct)); |
| 236 | |
| 237 | if (ain) { |
| 238 | mtls->fep.dimX = ain->getType()->getDimX(); |
| 239 | mtls->fep.dimY = ain->getType()->getDimY(); |
| 240 | mtls->fep.dimZ = ain->getType()->getDimZ(); |
| 241 | //mtls->dimArray = ain->getType()->getDimArray(); |
| 242 | } else if (aout) { |
| 243 | mtls->fep.dimX = aout->getType()->getDimX(); |
| 244 | mtls->fep.dimY = aout->getType()->getDimY(); |
| 245 | mtls->fep.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->fep.dimX; |
| 254 | } else { |
| 255 | rsAssert(sc->xStart < mtls->fep.dimX); |
| 256 | rsAssert(sc->xEnd <= mtls->fep.dimX); |
| 257 | rsAssert(sc->xStart < sc->xEnd); |
| 258 | mtls->xStart = rsMin(mtls->fep.dimX, sc->xStart); |
| 259 | mtls->xEnd = rsMin(mtls->fep.dimX, sc->xEnd); |
| 260 | if (mtls->xStart >= mtls->xEnd) return; |
| 261 | } |
| 262 | |
| 263 | if (!sc || (sc->yEnd == 0)) { |
| 264 | mtls->yEnd = mtls->fep.dimY; |
| 265 | } else { |
| 266 | rsAssert(sc->yStart < mtls->fep.dimY); |
| 267 | rsAssert(sc->yEnd <= mtls->fep.dimY); |
| 268 | rsAssert(sc->yStart < sc->yEnd); |
| 269 | mtls->yStart = rsMin(mtls->fep.dimY, sc->yStart); |
| 270 | mtls->yEnd = rsMin(mtls->fep.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 || (ain->getType()->getDimZ() == 0)); |
| 280 | |
| 281 | Context *mrsc = (Context *)rsc; |
| 282 | mtls->rsc = mrsc; |
| 283 | mtls->ain = ain; |
| 284 | mtls->aout = aout; |
| 285 | mtls->fep.usr = usr; |
| 286 | mtls->fep.usrLen = usrLen; |
| 287 | mtls->mSliceSize = 10; |
| 288 | mtls->mSliceNum = 0; |
| 289 | |
| 290 | mtls->fep.ptrIn = NULL; |
| 291 | mtls->fep.eStrideIn = 0; |
| 292 | if (ain) { |
| 293 | DrvAllocation *aindrv = (DrvAllocation *)ain->mHal.drv; |
| 294 | mtls->fep.ptrIn = (const uint8_t *)aindrv->lod[0].mallocPtr; |
| 295 | mtls->fep.eStrideIn = ain->getType()->getElementSizeBytes(); |
| 296 | mtls->fep.yStrideIn = aindrv->lod[0].stride; |
| 297 | } |
| 298 | |
| 299 | mtls->fep.ptrOut = NULL; |
| 300 | mtls->fep.eStrideOut = 0; |
| 301 | if (aout) { |
| 302 | DrvAllocation *aoutdrv = (DrvAllocation *)aout->mHal.drv; |
| 303 | mtls->fep.ptrOut = (uint8_t *)aoutdrv->lod[0].mallocPtr; |
| 304 | mtls->fep.eStrideOut = aout->getType()->getElementSizeBytes(); |
| 305 | mtls->fep.yStrideOut = aoutdrv->lod[0].stride; |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | void rsdScriptLaunchThreads(const Context *rsc, |
| 310 | Script *s, |
| 311 | uint32_t slot, |
| 312 | const Allocation * ain, |
| 313 | Allocation * aout, |
| 314 | const void * usr, |
| 315 | uint32_t usrLen, |
| 316 | const RsScriptCall *sc, |
| 317 | MTLaunchStruct *mtls) { |
| 318 | |
| 319 | Script * oldTLS = setTLS(s); |
| 320 | Context *mrsc = (Context *)rsc; |
| 321 | RsdHal * dc = (RsdHal *)mtls->rsc->mHal.drv; |
| 322 | |
| 323 | if ((dc->mWorkers.mCount > 1) && s->mHal.info.isThreadable && !dc->mInForEach) { |
| 324 | dc->mInForEach = true; |
| 325 | if (mtls->fep.dimY > 1) { |
| 326 | mtls->mSliceSize = mtls->fep.dimY / (dc->mWorkers.mCount * 4); |
| 327 | if(mtls->mSliceSize < 1) { |
| 328 | mtls->mSliceSize = 1; |
| 329 | } |
| 330 | |
| 331 | rsdLaunchThreads(mrsc, wc_xy, mtls); |
| 332 | } else { |
| 333 | mtls->mSliceSize = mtls->fep.dimX / (dc->mWorkers.mCount * 4); |
| 334 | if(mtls->mSliceSize < 1) { |
| 335 | mtls->mSliceSize = 1; |
| 336 | } |
| 337 | |
| 338 | rsdLaunchThreads(mrsc, wc_x, mtls); |
| 339 | } |
| 340 | dc->mInForEach = false; |
| 341 | |
| 342 | //ALOGE("launch 1"); |
| 343 | } else { |
| 344 | RsForEachStubParamStruct p; |
| 345 | memcpy(&p, &mtls->fep, sizeof(p)); |
| 346 | uint32_t sig = mtls->sig; |
| 347 | |
| 348 | //ALOGE("launch 3"); |
| 349 | outer_foreach_t fn = (outer_foreach_t) mtls->kernel; |
| 350 | for (p.ar[0] = mtls->arrayStart; p.ar[0] < mtls->arrayEnd; p.ar[0]++) { |
| 351 | for (p.z = mtls->zStart; p.z < mtls->zEnd; p.z++) { |
| 352 | for (p.y = mtls->yStart; p.y < mtls->yEnd; p.y++) { |
| 353 | uint32_t offset = mtls->fep.dimY * mtls->fep.dimZ * p.ar[0] + |
| 354 | mtls->fep.dimY * p.z + p.y; |
| 355 | p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * offset); |
| 356 | p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * offset); |
| 357 | fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); |
| 358 | } |
| 359 | } |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | setTLS(oldTLS); |
| 364 | } |
| 365 | |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 366 | void rsdScriptInvokeForEach(const Context *rsc, |
| 367 | Script *s, |
Jason Sams | 35e429e | 2011-07-13 11:26:26 -0700 | [diff] [blame] | 368 | uint32_t slot, |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 369 | const Allocation * ain, |
| 370 | Allocation * aout, |
| 371 | const void * usr, |
| 372 | uint32_t usrLen, |
| 373 | const RsScriptCall *sc) { |
| 374 | |
Jason Sams | 87fe59a | 2011-04-20 15:09:01 -0700 | [diff] [blame] | 375 | RsdHal * dc = (RsdHal *)rsc->mHal.drv; |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 376 | |
| 377 | MTLaunchStruct mtls; |
Jason Sams | dbe66d6 | 2012-09-17 13:54:41 -0700 | [diff] [blame^] | 378 | rsdScriptInvokeForEachMtlsSetup(rsc, ain, aout, usr, usrLen, sc, &mtls); |
| 379 | mtls.script = s; |
Jason Sams | 0ab9f9f | 2012-06-15 16:10:17 -0700 | [diff] [blame] | 380 | |
Stephen Hines | 3815bad | 2011-08-18 19:33:01 -0700 | [diff] [blame] | 381 | DrvScript *drv = (DrvScript *)s->mHal.drv; |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 382 | if (drv->mIntrinsicID) { |
Jason Sams | 17f03fc | 2012-08-21 18:30:01 -0700 | [diff] [blame] | 383 | mtls.kernel = (void (*)())drv->mIntrinsicFuncs.root; |
Jason Sams | dbe66d6 | 2012-09-17 13:54:41 -0700 | [diff] [blame^] | 384 | mtls.fep.usr = drv->mIntrinsicData; |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 385 | } else { |
| 386 | rsAssert(slot < drv->mExecutable->getExportForeachFuncAddrs().size()); |
| 387 | mtls.kernel = reinterpret_cast<ForEachFunc_t>( |
| 388 | drv->mExecutable->getExportForeachFuncAddrs()[slot]); |
| 389 | rsAssert(mtls.kernel != NULL); |
| 390 | mtls.sig = drv->mExecutable->getInfo().getExportForeachFuncs()[slot].second; |
| 391 | } |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 392 | |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 393 | |
Jason Sams | dbe66d6 | 2012-09-17 13:54:41 -0700 | [diff] [blame^] | 394 | rsdScriptLaunchThreads(rsc, s, slot, ain, aout, usr, usrLen, sc, &mtls); |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 395 | } |
| 396 | |
| 397 | |
| 398 | int rsdScriptInvokeRoot(const Context *dc, Script *script) { |
| 399 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 400 | |
| 401 | Script * oldTLS = setTLS(script); |
| 402 | int ret = drv->mRoot(); |
| 403 | setTLS(oldTLS); |
| 404 | |
| 405 | return ret; |
| 406 | } |
| 407 | |
| 408 | void rsdScriptInvokeInit(const Context *dc, Script *script) { |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 409 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 410 | |
| 411 | if (drv->mInit) { |
| 412 | drv->mInit(); |
| 413 | } |
| 414 | } |
| 415 | |
Stephen Hines | 4ee16ff | 2011-08-31 17:41:39 -0700 | [diff] [blame] | 416 | void rsdScriptInvokeFreeChildren(const Context *dc, Script *script) { |
| 417 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 418 | |
| 419 | if (drv->mFreeChildren) { |
| 420 | drv->mFreeChildren(); |
| 421 | } |
| 422 | } |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 423 | |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 424 | void rsdScriptInvokeFunction(const Context *dc, Script *script, |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 425 | uint32_t slot, |
| 426 | const void *params, |
| 427 | size_t paramLength) { |
| 428 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 429 | //ALOGE("invoke %p %p %i %p %i", dc, script, slot, params, paramLength); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 430 | |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 431 | Script * oldTLS = setTLS(script); |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 432 | reinterpret_cast<void (*)(const void *, uint32_t)>( |
| 433 | drv->mExecutable->getExportFuncAddrs()[slot])(params, paramLength); |
Jason Sams | cdfdb8f | 2011-03-17 16:12:47 -0700 | [diff] [blame] | 434 | setTLS(oldTLS); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 435 | } |
| 436 | |
| 437 | void rsdScriptSetGlobalVar(const Context *dc, const Script *script, |
| 438 | uint32_t slot, void *data, size_t dataLength) { |
| 439 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 440 | //rsAssert(!script->mFieldIsObject[slot]); |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 441 | //ALOGE("setGlobalVar %p %p %i %p %i", dc, script, slot, data, dataLength); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 442 | |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 443 | if (drv->mIntrinsicID) { |
Jason Sams | e1e08b4 | 2012-09-04 16:49:19 -0700 | [diff] [blame] | 444 | drv->mIntrinsicFuncs.setVar(dc, script, drv->mIntrinsicData, slot, data, dataLength); |
Jason Sams | 8eaba4f | 2012-08-14 14:38:05 -0700 | [diff] [blame] | 445 | return; |
| 446 | } |
| 447 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 448 | int32_t *destPtr = reinterpret_cast<int32_t *>( |
| 449 | drv->mExecutable->getExportVarAddrs()[slot]); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 450 | if (!destPtr) { |
Steve Block | 6598201 | 2011-10-20 11:56:00 +0100 | [diff] [blame] | 451 | //ALOGV("Calling setVar on slot = %i which is null", slot); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 452 | return; |
| 453 | } |
| 454 | |
| 455 | memcpy(destPtr, data, dataLength); |
| 456 | } |
| 457 | |
Stephen Hines | 2980f07 | 2012-04-09 18:26:29 -0700 | [diff] [blame] | 458 | void rsdScriptSetGlobalVarWithElemDims( |
| 459 | const android::renderscript::Context *dc, |
| 460 | const android::renderscript::Script *script, |
| 461 | uint32_t slot, void *data, size_t dataLength, |
| 462 | const android::renderscript::Element *elem, |
| 463 | const size_t *dims, size_t dimLength) { |
| 464 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 465 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 466 | int32_t *destPtr = reinterpret_cast<int32_t *>( |
| 467 | drv->mExecutable->getExportVarAddrs()[slot]); |
Stephen Hines | 2980f07 | 2012-04-09 18:26:29 -0700 | [diff] [blame] | 468 | if (!destPtr) { |
| 469 | //ALOGV("Calling setVar on slot = %i which is null", slot); |
| 470 | return; |
| 471 | } |
| 472 | |
| 473 | // We want to look at dimension in terms of integer components, |
| 474 | // but dimLength is given in terms of bytes. |
| 475 | dimLength /= sizeof(int); |
| 476 | |
| 477 | // Only a single dimension is currently supported. |
| 478 | rsAssert(dimLength == 1); |
| 479 | if (dimLength == 1) { |
| 480 | // First do the increment loop. |
| 481 | size_t stride = elem->getSizeBytes(); |
| 482 | char *cVal = reinterpret_cast<char *>(data); |
| 483 | for (size_t i = 0; i < dims[0]; i++) { |
| 484 | elem->incRefs(cVal); |
| 485 | cVal += stride; |
| 486 | } |
| 487 | |
| 488 | // Decrement loop comes after (to prevent race conditions). |
| 489 | char *oldVal = reinterpret_cast<char *>(destPtr); |
| 490 | for (size_t i = 0; i < dims[0]; i++) { |
| 491 | elem->decRefs(oldVal); |
| 492 | oldVal += stride; |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | memcpy(destPtr, data, dataLength); |
| 497 | } |
| 498 | |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 499 | void rsdScriptSetGlobalBind(const Context *dc, const Script *script, uint32_t slot, Allocation *data) { |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 500 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 501 | |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 502 | //rsAssert(!script->mFieldIsObject[slot]); |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 503 | //ALOGE("setGlobalBind %p %p %i %p", dc, script, slot, data); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 504 | |
Jason Sams | 17f03fc | 2012-08-21 18:30:01 -0700 | [diff] [blame] | 505 | if (drv->mIntrinsicID) { |
Jason Sams | e1e08b4 | 2012-09-04 16:49:19 -0700 | [diff] [blame] | 506 | drv->mIntrinsicFuncs.bind(dc, script, drv->mIntrinsicData, slot, data); |
Jason Sams | 17f03fc | 2012-08-21 18:30:01 -0700 | [diff] [blame] | 507 | return; |
| 508 | } |
| 509 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 510 | int32_t *destPtr = reinterpret_cast<int32_t *>( |
| 511 | drv->mExecutable->getExportVarAddrs()[slot]); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 512 | if (!destPtr) { |
Steve Block | 6598201 | 2011-10-20 11:56:00 +0100 | [diff] [blame] | 513 | //ALOGV("Calling setVar on slot = %i which is null", slot); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 514 | return; |
| 515 | } |
| 516 | |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 517 | void *ptr = NULL; |
| 518 | drv->mBoundAllocs[slot] = data; |
| 519 | if(data) { |
| 520 | DrvAllocation *allocDrv = (DrvAllocation *)data->mHal.drv; |
| 521 | ptr = allocDrv->lod[0].mallocPtr; |
| 522 | } |
| 523 | memcpy(destPtr, &ptr, sizeof(void *)); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 524 | } |
| 525 | |
| 526 | void rsdScriptSetGlobalObj(const Context *dc, const Script *script, uint32_t slot, ObjectBase *data) { |
| 527 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 528 | //rsAssert(script->mFieldIsObject[slot]); |
Steve Block | af12ac6 | 2012-01-06 19:20:56 +0000 | [diff] [blame] | 529 | //ALOGE("setGlobalObj %p %p %i %p", dc, script, slot, data); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 530 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 531 | int32_t *destPtr = reinterpret_cast<int32_t *>( |
| 532 | drv->mExecutable->getExportVarAddrs()[slot]); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 533 | if (!destPtr) { |
Steve Block | 6598201 | 2011-10-20 11:56:00 +0100 | [diff] [blame] | 534 | //ALOGV("Calling setVar on slot = %i which is null", slot); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 535 | return; |
| 536 | } |
| 537 | |
Jason Sams | 87fe59a | 2011-04-20 15:09:01 -0700 | [diff] [blame] | 538 | rsrSetObject(dc, script, (ObjectBase **)destPtr, data); |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 539 | } |
| 540 | |
| 541 | void rsdScriptDestroy(const Context *dc, Script *script) { |
| 542 | DrvScript *drv = (DrvScript *)script->mHal.drv; |
| 543 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 544 | if (drv == NULL) { |
| 545 | return; |
| 546 | } |
| 547 | |
| 548 | if (drv->mExecutable) { |
| 549 | Vector<void *>::const_iterator var_addr_iter = |
| 550 | drv->mExecutable->getExportVarAddrs().begin(); |
| 551 | Vector<void *>::const_iterator var_addr_end = |
| 552 | drv->mExecutable->getExportVarAddrs().end(); |
| 553 | |
| 554 | bcc::RSInfo::ObjectSlotListTy::const_iterator is_object_iter = |
| 555 | drv->mExecutable->getInfo().getObjectSlots().begin(); |
| 556 | bcc::RSInfo::ObjectSlotListTy::const_iterator is_object_end = |
| 557 | drv->mExecutable->getInfo().getObjectSlots().end(); |
| 558 | |
| 559 | while ((var_addr_iter != var_addr_end) && |
| 560 | (is_object_iter != is_object_end)) { |
| 561 | // The field address can be NULL if the script-side has optimized |
| 562 | // the corresponding global variable away. |
| 563 | ObjectBase **obj_addr = |
| 564 | reinterpret_cast<ObjectBase **>(*var_addr_iter); |
| 565 | if (*is_object_iter) { |
| 566 | if (*var_addr_iter != NULL) { |
| 567 | rsrClearObject(dc, script, obj_addr); |
Stephen Hines | 8d43eaf | 2011-03-24 20:03:49 -0700 | [diff] [blame] | 568 | } |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 569 | } |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 570 | var_addr_iter++; |
| 571 | is_object_iter++; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 572 | } |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 573 | } |
Stephen Hines | 689821f | 2011-07-18 17:24:11 -0700 | [diff] [blame] | 574 | |
Shih-wei Liao | a3af2cd | 2012-04-25 04:18:31 -0700 | [diff] [blame] | 575 | delete drv->mCompilerContext; |
| 576 | delete drv->mCompilerDriver; |
| 577 | delete drv->mExecutable; |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 578 | delete[] drv->mBoundAllocs; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 579 | free(drv); |
| 580 | script->mHal.drv = NULL; |
Jason Sams | bad8074 | 2011-03-16 16:29:28 -0700 | [diff] [blame] | 581 | } |
Jason Sams | 807fdc4 | 2012-07-25 17:55:39 -0700 | [diff] [blame] | 582 | |
| 583 | Allocation * rsdScriptGetAllocationForPointer(const android::renderscript::Context *dc, |
| 584 | const android::renderscript::Script *sc, |
| 585 | const void *ptr) { |
| 586 | DrvScript *drv = (DrvScript *)sc->mHal.drv; |
| 587 | if (!ptr) { |
| 588 | return NULL; |
| 589 | } |
| 590 | |
| 591 | for (uint32_t ct=0; ct < sc->mHal.info.exportedVariableCount; ct++) { |
| 592 | Allocation *a = drv->mBoundAllocs[ct]; |
| 593 | if (!a) continue; |
| 594 | DrvAllocation *adrv = (DrvAllocation *)a->mHal.drv; |
| 595 | if (adrv->lod[0].mallocPtr == ptr) { |
| 596 | return a; |
| 597 | } |
| 598 | } |
| 599 | ALOGE("rsGetAllocation, failed to find %p", ptr); |
| 600 | return NULL; |
| 601 | } |
| 602 | |