rsCpuCore.cpp revision 39ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacd
1709a0978ae141198018ca9769f8d96292a8928e6Jason Sams/* 2709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * Copyright (C) 2012 The Android Open Source Project 3709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * 4709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * Licensed under the Apache License, Version 2.0 (the "License"); 5709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * you may not use this file except in compliance with the License. 6709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * You may obtain a copy of the License at 7709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * 8709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * http://www.apache.org/licenses/LICENSE-2.0 9709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * 10709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * Unless required by applicable law or agreed to in writing, software 11709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * distributed under the License is distributed on an "AS IS" BASIS, 12709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * See the License for the specific language governing permissions and 14709a0978ae141198018ca9769f8d96292a8928e6Jason Sams * limitations under the License. 15709a0978ae141198018ca9769f8d96292a8928e6Jason Sams */ 16709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 17709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "rsCpuCore.h" 18709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "rsCpuScript.h" 19709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "rsCpuScriptGroup.h" 20709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 21709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <malloc.h> 22709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "rsContext.h" 23709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 24709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sys/types.h> 25709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sys/resource.h> 26709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sched.h> 27709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sys/syscall.h> 28709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <string.h> 29b0934b67b95cc27e2358c2aa4db5f7c1067c8f9bStephen Hines#include <unistd.h> 300b575de8ed0b628d84d256f5846500b0385979bdTim Murray 31f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#include <stdio.h> 32f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#include <stdlib.h> 33f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#include <fcntl.h> 34f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 35b0934b67b95cc27e2358c2aa4db5f7c1067c8f9bStephen Hines#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB) 360b575de8ed0b628d84d256f5846500b0385979bdTim Murray#include <cutils/properties.h> 37709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "utils/StopWatch.h" 380b575de8ed0b628d84d256f5846500b0385979bdTim Murray#endif 390b575de8ed0b628d84d256f5846500b0385979bdTim Murray 400b575de8ed0b628d84d256f5846500b0385979bdTim Murray#ifdef RS_SERVER 410b575de8ed0b628d84d256f5846500b0385979bdTim Murray// Android exposes gettid(), standard Linux does not 420b575de8ed0b628d84d256f5846500b0385979bdTim Murraystatic pid_t gettid() { 430b575de8ed0b628d84d256f5846500b0385979bdTim Murray return syscall(SYS_gettid); 440b575de8ed0b628d84d256f5846500b0385979bdTim Murray} 450b575de8ed0b628d84d256f5846500b0385979bdTim Murray#endif 46709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 47709a0978ae141198018ca9769f8d96292a8928e6Jason Samsusing namespace android; 48709a0978ae141198018ca9769f8d96292a8928e6Jason Samsusing namespace android::renderscript; 49709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 50709a0978ae141198018ca9769f8d96292a8928e6Jason Samstypedef void (*outer_foreach_t)( 51709a0978ae141198018ca9769f8d96292a8928e6Jason Sams const android::renderscript::RsForEachStubParamStruct *, 52709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t x1, uint32_t x2, 53709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t instep, uint32_t outstep); 54709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 55709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 56709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic pthread_key_t gThreadTLSKey = 0; 57709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic uint32_t gThreadTLSKeyCount = 0; 58709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER; 59709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 60f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsbool android::renderscript::gArchUseSIMD = false; 61f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 62709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::~RsdCpuReference() { 63709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 64709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 65709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major, 66cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn 67cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#ifndef RS_COMPATIBILITY_LIB 681d476620399d54774e4fd386c1d23cc583d49522Stephen Hines , bcc::RSLinkRuntimeCallback pLinkRuntimeCallback, 691d476620399d54774e4fd386c1d23cc583d49522Stephen Hines RSSelectRTCallback pSelectRTCallback 70cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#endif 71cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams ) { 72709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 73709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc); 74709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!cpu) { 75709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 76709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 77709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!cpu->init(version_major, version_minor, lfn, slfn)) { 78709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete cpu; 79709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 80709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 81f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines 82cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#ifndef RS_COMPATIBILITY_LIB 83f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines cpu->setLinkRuntimeCallback(pLinkRuntimeCallback); 841d476620399d54774e4fd386c1d23cc583d49522Stephen Hines cpu->setSelectRTCallback(pSelectRTCallback); 85cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#endif 86f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines 87709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return cpu; 88709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 89709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 90709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 91709a0978ae141198018ca9769f8d96292a8928e6Jason SamsContext * RsdCpuReference::getTlsContext() { 92709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 93709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return tls->mContext; 94709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 95709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 96709a0978ae141198018ca9769f8d96292a8928e6Jason Samsconst Script * RsdCpuReference::getTlsScript() { 97709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 98709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return tls->mScript; 99709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 100709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 101f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hinespthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 102709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 103709a0978ae141198018ca9769f8d96292a8928e6Jason Sams//////////////////////////////////////////////////////////// 104709a0978ae141198018ca9769f8d96292a8928e6Jason Sams/// 105709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 106709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) { 107709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mRSC = rsc; 108709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 109709a0978ae141198018ca9769f8d96292a8928e6Jason Sams version_major = 0; 110709a0978ae141198018ca9769f8d96292a8928e6Jason Sams version_minor = 0; 111709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mInForEach = false; 112709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&mWorkers, 0, sizeof(mWorkers)); 113709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&mTlsStruct, 0, sizeof(mTlsStruct)); 114709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mExit = false; 115cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#ifndef RS_COMPATIBILITY_LIB 116f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines mLinkRuntimeCallback = NULL; 1171d476620399d54774e4fd386c1d23cc583d49522Stephen Hines mSelectRTCallback = NULL; 118b7d9c80c98fc96aa7c638e3124be24f13a6436b2Stephen Hines mSetupCompilerCallback = NULL; 119cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#endif 120709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 121709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 122709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 123709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 124709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 125709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1260b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 127709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 128709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 129709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 130709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchSignals[idx].init(); 131709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mNativeThreadId[idx] = gettid(); 132709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 133709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 134709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 135709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 136709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("pthread_setspecific %i", status); 137709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 138709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 139709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#if 0 140709a0978ae141198018ca9769f8d96292a8928e6Jason Sams typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 141709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpu_set_t cpuset; 142709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&cpuset, 0, sizeof(cpuset)); 143709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 144709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 145709a0978ae141198018ca9769f8d96292a8928e6Jason Sams sizeof(cpuset), &cpuset); 146709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 147709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#endif 148709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 149709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (!dc->mExit) { 150709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchSignals[idx].wait(); 151709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (dc->mWorkers.mLaunchCallback) { 152709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // idx +1 is used because the calling thread is always worker 0. 153709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 154709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 1550b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 156709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mCompleteSignal.set(); 157709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 158709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 159709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 160709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 161709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 162709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 163709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 164709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchData = data; 165709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = cbk; 1664d252d6e807b89764dad123ac845df298c52ca97Tim Murray 1674d252d6e807b89764dad123ac845df298c52ca97Tim Murray // fast path for very small launches 1684d252d6e807b89764dad123ac845df298c52ca97Tim Murray MTLaunchStruct *mtls = (MTLaunchStruct *)data; 1694d252d6e807b89764dad123ac845df298c52ca97Tim Murray if (mtls && mtls->fep.dimY <= 1 && mtls->xEnd <= mtls->xStart + mtls->mSliceSize) { 1704d252d6e807b89764dad123ac845df298c52ca97Tim Murray if (mWorkers.mLaunchCallback) { 1714d252d6e807b89764dad123ac845df298c52ca97Tim Murray mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 1724d252d6e807b89764dad123ac845df298c52ca97Tim Murray } 1734d252d6e807b89764dad123ac845df298c52ca97Tim Murray return; 1744d252d6e807b89764dad123ac845df298c52ca97Tim Murray } 1754d252d6e807b89764dad123ac845df298c52ca97Tim Murray 1760b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 1770b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 1780b575de8ed0b628d84d256f5846500b0385979bdTim Murray 179709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 180709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals[ct].set(); 181709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 182709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 183709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // We use the calling thread as one of the workers so we can start without 184709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // the delay of the thread wakeup. 185709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mWorkers.mLaunchCallback) { 1864d252d6e807b89764dad123ac845df298c52ca97Tim Murray mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 187709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 188709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1890b575de8ed0b628d84d256f5846500b0385979bdTim Murray while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 190709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCompleteSignal.wait(); 191709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 192709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 193709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 194709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 195709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::lockMutex() { 196709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_mutex_lock(&gInitMutex); 197709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 198709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 199709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::unlockMutex() { 200709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_mutex_unlock(&gInitMutex); 201709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 202709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 203f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#if defined(ARCH_ARM_HAVE_VFP) 204f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsstatic int 205f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsread_file(const char* pathname, char* buffer, size_t buffsize) 206f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams{ 207f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams int fd, len; 208f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 209f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams fd = open(pathname, O_RDONLY); 210f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams if (fd < 0) 211f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return -1; 212f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 213f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams do { 214f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams len = read(fd, buffer, buffsize); 215f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams } while (len < 0 && errno == EINTR); 216f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 217f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams close(fd); 218f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 219f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return len; 220f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams} 221f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 222f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsstatic void GetCpuInfo() { 223f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams char cpuinfo[4096]; 224f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams int cpuinfo_len; 225f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 226f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo); 227f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams if (cpuinfo_len < 0) /* should not happen */ { 228f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return; 229f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams } 230f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 231f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams gArchUseSIMD = !!strstr(cpuinfo, " neon"); 232f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams} 233f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#endif // ARCH_ARM_HAVE_VFP 234f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 235709a0978ae141198018ca9769f8d96292a8928e6Jason Samsbool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 236709a0978ae141198018ca9769f8d96292a8928e6Jason Sams sym_lookup_t lfn, script_lookup_t slfn) { 237709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 238709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mSymLookupFn = lfn; 239709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mScriptLookupFn = slfn; 240709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 241709a0978ae141198018ca9769f8d96292a8928e6Jason Sams lockMutex(); 242709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!gThreadTLSKeyCount) { 243709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_key_create(&gThreadTLSKey, NULL); 244709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 245709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Failed to init thread tls key."); 246709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 247709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return false; 248709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 249709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 250709a0978ae141198018ca9769f8d96292a8928e6Jason Sams gThreadTLSKeyCount++; 251709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 252709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 253709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mTlsStruct.mContext = mRSC; 254709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mTlsStruct.mScript = NULL; 255709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 256709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 257709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("pthread_setspecific %i", status); 258709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 259709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 260f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#if defined(ARCH_ARM_HAVE_VFP) 261f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams GetCpuInfo(); 262f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#endif 263f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 264709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int cpu = sysconf(_SC_NPROCESSORS_ONLN); 265709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mRSC->props.mDebugMaxThreads) { 266709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpu = mRSC->props.mDebugMaxThreads; 267709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 268709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (cpu < 2) { 269709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = 0; 270709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return true; 271709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 272709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 273709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // Subtract one from the cpu count because we also use the command thread as a worker. 274709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = (uint32_t)(cpu - 1); 275709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2768ca358a2abe7e0dba23993e0fc8d64b8b55bd9caJason Sams ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 277709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 278709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 279709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 280709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 281709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = NULL; 282709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 283709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCompleteSignal.init(); 284709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2850b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 2860b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mLaunchCount = 0; 2870b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 288709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 289709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_attr_t threadAttr; 290709a0978ae141198018ca9769f8d96292a8928e6Jason Sams status = pthread_attr_init(&threadAttr); 291709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 292709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Failed to init thread attribute."); 293709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return false; 294709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 295709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 296709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 297709a0978ae141198018ca9769f8d96292a8928e6Jason Sams status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 298709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 299709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = ct; 300709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Created fewer than expected number of RS threads."); 301709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 302709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 303709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 3040b575de8ed0b628d84d256f5846500b0385979bdTim Murray while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 305709a0978ae141198018ca9769f8d96292a8928e6Jason Sams usleep(100); 306709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 307709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 308709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_attr_destroy(&threadAttr); 309709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return true; 310709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 311709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 312709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 313709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::setPriority(int32_t priority) { 314709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 315709a0978ae141198018ca9769f8d96292a8928e6Jason Sams setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 316709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 317709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 318709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 319709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 320709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mExit = true; 321709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchData = NULL; 322709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = NULL; 3230b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 3240b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 325709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 326709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals[ct].set(); 327709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 328709a0978ae141198018ca9769f8d96292a8928e6Jason Sams void *res; 329709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 330709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_join(mWorkers.mThreadId[ct], &res); 331709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 3320b575de8ed0b628d84d256f5846500b0385979bdTim Murray rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 33307ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin free(mWorkers.mThreadId); 33407ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin free(mWorkers.mNativeThreadId); 33507ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin delete[] mWorkers.mLaunchSignals; 336709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 337709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // Global structure cleanup. 338709a0978ae141198018ca9769f8d96292a8928e6Jason Sams lockMutex(); 339709a0978ae141198018ca9769f8d96292a8928e6Jason Sams --gThreadTLSKeyCount; 340709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!gThreadTLSKeyCount) { 341709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_key_delete(gThreadTLSKey); 342709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 343709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 344709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 345709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 346709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 347709a0978ae141198018ca9769f8d96292a8928e6Jason Samstypedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t); 348709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 349709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic void wc_xy(void *usr, uint32_t idx) { 350709a0978ae141198018ca9769f8d96292a8928e6Jason Sams MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 351709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsForEachStubParamStruct p; 352709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memcpy(&p, &mtls->fep, sizeof(p)); 353709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.lid = idx; 354709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t sig = mtls->sig; 355709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 356709a0978ae141198018ca9769f8d96292a8928e6Jason Sams outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 357709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (1) { 3580b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 359709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize; 360709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t yEnd = yStart + mtls->mSliceSize; 361709a0978ae141198018ca9769f8d96292a8928e6Jason Sams yEnd = rsMin(yEnd, mtls->yEnd); 362709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (yEnd <= yStart) { 363709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return; 364709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 365709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 366709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr idx %i, x %i,%i y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd); 367709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr ptr in %p, out %p", mtls->fep.ptrIn, mtls->fep.ptrOut); 368709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 369709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.y = yStart; p.y < yEnd; p.y++) { 370709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * p.y) + 371709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideOut * mtls->xStart); 372709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * p.y) + 373709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideIn * mtls->xStart); 374709a0978ae141198018ca9769f8d96292a8928e6Jason Sams fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); 375709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 376709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 377709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 378709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 379709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic void wc_x(void *usr, uint32_t idx) { 380709a0978ae141198018ca9769f8d96292a8928e6Jason Sams MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 381709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsForEachStubParamStruct p; 382709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memcpy(&p, &mtls->fep, sizeof(p)); 383709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.lid = idx; 384709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t sig = mtls->sig; 385709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 386709a0978ae141198018ca9769f8d96292a8928e6Jason Sams outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 387709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (1) { 3880b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 389709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize; 390709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t xEnd = xStart + mtls->mSliceSize; 391709a0978ae141198018ca9769f8d96292a8928e6Jason Sams xEnd = rsMin(xEnd, mtls->xEnd); 392709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (xEnd <= xStart) { 393709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return; 394709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 395709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 396709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr slice %i idx %i, x %i,%i", slice, idx, xStart, xEnd); 397709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr ptr in %p, out %p", mtls->fep.ptrIn, mtls->fep.ptrOut); 398709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 399709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.out = mtls->fep.ptrOut + (mtls->fep.eStrideOut * xStart); 400709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.in = mtls->fep.ptrIn + (mtls->fep.eStrideIn * xStart); 401709a0978ae141198018ca9769f8d96292a8928e6Jason Sams fn(&p, xStart, xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); 402709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 403709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 404709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 405709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::launchThreads(const Allocation * ain, Allocation * aout, 406709a0978ae141198018ca9769f8d96292a8928e6Jason Sams const RsScriptCall *sc, MTLaunchStruct *mtls) { 407709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 408709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //android::StopWatch kernel_time("kernel time"); 409709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 410709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) { 411709a0978ae141198018ca9769f8d96292a8928e6Jason Sams const size_t targetByteChunk = 16 * 1024; 412709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mInForEach = true; 413709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mtls->fep.dimY > 1) { 414709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s1 = mtls->fep.dimY / ((mWorkers.mCount + 1) * 4); 415709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s2 = 0; 416709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 417709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // This chooses our slice size to rate limit atomic ops to 418709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // one per 16k bytes of reads/writes. 419709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mtls->fep.yStrideOut) { 420709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.yStrideOut; 421709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 422709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.yStrideIn; 423709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 424709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = rsMin(s1, s2); 425709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 426709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mtls->mSliceSize < 1) { 427709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = 1; 428709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 429709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 430709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // mtls->mSliceSize = 2; 431709a0978ae141198018ca9769f8d96292a8928e6Jason Sams launchThreads(wc_xy, mtls); 432709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 433709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s1 = mtls->fep.dimX / ((mWorkers.mCount + 1) * 4); 434709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s2 = 0; 435709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 436709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // This chooses our slice size to rate limit atomic ops to 437709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // one per 16k bytes of reads/writes. 438709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mtls->fep.eStrideOut) { 439709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.eStrideOut; 440709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 441709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.eStrideIn; 442709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 443709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = rsMin(s1, s2); 444709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 445709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mtls->mSliceSize < 1) { 446709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = 1; 447709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 448709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 449709a0978ae141198018ca9769f8d96292a8928e6Jason Sams launchThreads(wc_x, mtls); 450709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 451709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mInForEach = false; 452709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 453709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("launch 1"); 454709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 455709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsForEachStubParamStruct p; 456709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memcpy(&p, &mtls->fep, sizeof(p)); 457709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t sig = mtls->sig; 458709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 459709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("launch 3"); 460709a0978ae141198018ca9769f8d96292a8928e6Jason Sams outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 461709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.ar[0] = mtls->arrayStart; p.ar[0] < mtls->arrayEnd; p.ar[0]++) { 462709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.z = mtls->zStart; p.z < mtls->zEnd; p.z++) { 463709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.y = mtls->yStart; p.y < mtls->yEnd; p.y++) { 464709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t offset = mtls->fep.dimY * mtls->fep.dimZ * p.ar[0] + 465709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->fep.dimY * p.z + p.y; 466709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * offset) + 467709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideOut * mtls->xStart); 468709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * offset) + 469709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideIn * mtls->xStart); 470709a0978ae141198018ca9769f8d96292a8928e6Jason Sams fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); 471709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 472709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 473709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 474709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 475709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 476709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 477709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 478709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("setTls %p", sc); 479709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 480709a0978ae141198018ca9769f8d96292a8928e6Jason Sams rsAssert(tls); 481709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *old = tls->mImpl; 482709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mImpl = sc; 483709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mContext = mRSC; 484709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (sc) { 485709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mScript = sc->getScript(); 486709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 487709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mScript = NULL; 488709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 489709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return old; 490709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 491709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 492709a0978ae141198018ca9769f8d96292a8928e6Jason Samsconst RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 493709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return mSymLookupFn(mRSC, name); 494709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 495709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 496709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 497709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 498709a0978ae141198018ca9769f8d96292a8928e6Jason Sams char const *resName, char const *cacheDir, 499709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint8_t const *bitcode, size_t bitcodeSize, 500709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t flags) { 501709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 502709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 503709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags)) { 504709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete i; 505709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 506709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 507709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return i; 508709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 509709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 5107c4b888f2147edf99690b6af75470774ff31c43bJason Samsextern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 5117c4b888f2147edf99690b6af75470774ff31c43bJason Sams const Script *s, const Element *e); 51260498fe9679ea25a260a503d6dfd27cbc0a0c079Matthieu Delahayeextern RsdCpuScriptImpl * rsdIntrinsic_InterPred(RsdCpuReferenceImpl *ctx, 51360498fe9679ea25a260a503d6dfd27cbc0a0c079Matthieu Delahaye const Script *s, const Element *e); 514c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 515c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 516c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 517c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 518c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 519c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 520c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 521c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 522c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 523c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 524c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 525c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 526c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 527c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 5282282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 5292282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams const Script *s, const Element *e); 5306fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahayeextern RsdCpuScriptImpl * rsdIntrinsic_LoopFilter(RsdCpuReferenceImpl *ctx, 5316fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye const Script *s, const Element *e); 53239ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 53339ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams const Script *s, const Element *e); 534709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 535709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 536709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsScriptIntrinsicID iid, Element *e) { 537709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 538709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *i = NULL; 539709a0978ae141198018ca9769f8d96292a8928e6Jason Sams switch (iid) { 5407c4b888f2147edf99690b6af75470774ff31c43bJason Sams case RS_SCRIPT_INTRINSIC_ID_3DLUT: 5417c4b888f2147edf99690b6af75470774ff31c43bJason Sams i = rsdIntrinsic_3DLUT(this, s, e); 5427c4b888f2147edf99690b6af75470774ff31c43bJason Sams break; 54383f304cb26008d3f4da154cec19c3a12fa2e6c74Jason Sams#ifndef RS_COMPATIBILITY_LIB 54460498fe9679ea25a260a503d6dfd27cbc0a0c079Matthieu Delahaye case RS_SCRIPT_INTRINSIC_ID_INTER_PRED: 54560498fe9679ea25a260a503d6dfd27cbc0a0c079Matthieu Delahaye i = rsdIntrinsic_InterPred(this, s, e); 54660498fe9679ea25a260a503d6dfd27cbc0a0c079Matthieu Delahaye break; 54783f304cb26008d3f4da154cec19c3a12fa2e6c74Jason Sams#endif 548709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 549c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Convolve3x3(this, s, e); 550709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 551709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 552c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_ColorMatrix(this, s, e); 553709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 554709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_LUT: 555c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_LUT(this, s, e); 556709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 557709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 558c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Convolve5x5(this, s, e); 559709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 560709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_BLUR: 561c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Blur(this, s, e); 562709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 563709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 564c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_YuvToRGB(this, s, e); 565709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 566709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_BLEND: 567c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Blend(this, s, e); 568709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 5692282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 5702282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams i = rsdIntrinsic_Histogram(this, s, e); 5712282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams break; 57239ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams case RS_SCRIPT_INTRINSIC_ID_RESIZE: 57339ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams i = rsdIntrinsic_Resize(this, s, e); 57439ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams break; 5756fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye#ifndef RS_COMPATIBILITY_LIB 5766fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye case RS_SCRIPT_INTRINSIC_ID_LOOP_FILTER: 5776fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye i = rsdIntrinsic_LoopFilter(this, s, e); 5786fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye break; 5796fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye#endif 580709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 581709a0978ae141198018ca9769f8d96292a8928e6Jason Sams default: 582709a0978ae141198018ca9769f8d96292a8928e6Jason Sams rsAssert(0); 583709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 584709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 585709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return i; 586709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 587709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 588709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScriptGroup * RsdCpuReferenceImpl::createScriptGroup(const ScriptGroup *sg) { 589709a0978ae141198018ca9769f8d96292a8928e6Jason Sams CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 590709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!sgi->init()) { 591709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete sgi; 592709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 593709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 594709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return sgi; 595709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 596709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 597709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 598