rsCpuCore.cpp revision 005113297b19ed256b6db9d6bc293ed9266899fc
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, 69005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines RSSelectRTCallback pSelectRTCallback, 70005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines const char *pBccPluginName 71cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#endif 72cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams ) { 73709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 74709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc); 75709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!cpu) { 76709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 77709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 78709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!cpu->init(version_major, version_minor, lfn, slfn)) { 79709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete cpu; 80709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 81709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 82f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines 83cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#ifndef RS_COMPATIBILITY_LIB 84f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines cpu->setLinkRuntimeCallback(pLinkRuntimeCallback); 851d476620399d54774e4fd386c1d23cc583d49522Stephen Hines cpu->setSelectRTCallback(pSelectRTCallback); 86005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines if (pBccPluginName) { 87005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines cpu->setBccPluginName(pBccPluginName); 88005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines } 89cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#endif 90f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines 91709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return cpu; 92709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 93709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 94709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 95709a0978ae141198018ca9769f8d96292a8928e6Jason SamsContext * RsdCpuReference::getTlsContext() { 96709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 97709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return tls->mContext; 98709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 99709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 100709a0978ae141198018ca9769f8d96292a8928e6Jason Samsconst Script * RsdCpuReference::getTlsScript() { 101709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 102709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return tls->mScript; 103709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 104709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 105f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hinespthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 106709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 107709a0978ae141198018ca9769f8d96292a8928e6Jason Sams//////////////////////////////////////////////////////////// 108709a0978ae141198018ca9769f8d96292a8928e6Jason Sams/// 109709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 110709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) { 111709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mRSC = rsc; 112709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 113709a0978ae141198018ca9769f8d96292a8928e6Jason Sams version_major = 0; 114709a0978ae141198018ca9769f8d96292a8928e6Jason Sams version_minor = 0; 115709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mInForEach = false; 116709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&mWorkers, 0, sizeof(mWorkers)); 117709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&mTlsStruct, 0, sizeof(mTlsStruct)); 118709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mExit = false; 119cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#ifndef RS_COMPATIBILITY_LIB 120f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines mLinkRuntimeCallback = NULL; 1211d476620399d54774e4fd386c1d23cc583d49522Stephen Hines mSelectRTCallback = NULL; 122b7d9c80c98fc96aa7c638e3124be24f13a6436b2Stephen Hines mSetupCompilerCallback = NULL; 123cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams#endif 124709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 125709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 126709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 127709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 128709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 129709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1300b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 131709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 132709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 133709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 134709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchSignals[idx].init(); 135709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mNativeThreadId[idx] = gettid(); 136709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 137709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 138709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 139709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 140709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("pthread_setspecific %i", status); 141709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 142709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 143709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#if 0 144709a0978ae141198018ca9769f8d96292a8928e6Jason Sams typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 145709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpu_set_t cpuset; 146709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&cpuset, 0, sizeof(cpuset)); 147709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 148709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 149709a0978ae141198018ca9769f8d96292a8928e6Jason Sams sizeof(cpuset), &cpuset); 150709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 151709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#endif 152709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 153709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (!dc->mExit) { 154709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchSignals[idx].wait(); 155709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (dc->mWorkers.mLaunchCallback) { 156709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // idx +1 is used because the calling thread is always worker 0. 157709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 158709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 1590b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 160709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mCompleteSignal.set(); 161709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 162709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 163709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 164709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 165709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 166709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 167709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 168709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchData = data; 169709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = cbk; 1704d252d6e807b89764dad123ac845df298c52ca97Tim Murray 1714d252d6e807b89764dad123ac845df298c52ca97Tim Murray // fast path for very small launches 1724d252d6e807b89764dad123ac845df298c52ca97Tim Murray MTLaunchStruct *mtls = (MTLaunchStruct *)data; 1734d252d6e807b89764dad123ac845df298c52ca97Tim Murray if (mtls && mtls->fep.dimY <= 1 && mtls->xEnd <= mtls->xStart + mtls->mSliceSize) { 1744d252d6e807b89764dad123ac845df298c52ca97Tim Murray if (mWorkers.mLaunchCallback) { 1754d252d6e807b89764dad123ac845df298c52ca97Tim Murray mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 1764d252d6e807b89764dad123ac845df298c52ca97Tim Murray } 1774d252d6e807b89764dad123ac845df298c52ca97Tim Murray return; 1784d252d6e807b89764dad123ac845df298c52ca97Tim Murray } 1794d252d6e807b89764dad123ac845df298c52ca97Tim Murray 1800b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 1810b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 1820b575de8ed0b628d84d256f5846500b0385979bdTim Murray 183709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 184709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals[ct].set(); 185709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 186709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 187709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // We use the calling thread as one of the workers so we can start without 188709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // the delay of the thread wakeup. 189709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mWorkers.mLaunchCallback) { 1904d252d6e807b89764dad123ac845df298c52ca97Tim Murray mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 191709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 192709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1930b575de8ed0b628d84d256f5846500b0385979bdTim Murray while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 194709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCompleteSignal.wait(); 195709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 196709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 197709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 198709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 199709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::lockMutex() { 200709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_mutex_lock(&gInitMutex); 201709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 202709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 203709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::unlockMutex() { 204709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_mutex_unlock(&gInitMutex); 205709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 206709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2077b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_X86_HAVE_SSSE3) 208f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsstatic int 209f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsread_file(const char* pathname, char* buffer, size_t buffsize) 210f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams{ 211f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams int fd, len; 212f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 213f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams fd = open(pathname, O_RDONLY); 214f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams if (fd < 0) 215f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return -1; 216f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 217f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams do { 218f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams len = read(fd, buffer, buffsize); 219f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams } while (len < 0 && errno == EINTR); 220f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 221f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams close(fd); 222f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 223f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return len; 224f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams} 225f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 226f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsstatic void GetCpuInfo() { 227f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams char cpuinfo[4096]; 228f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams int cpuinfo_len; 229f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 230f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo); 231f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams if (cpuinfo_len < 0) /* should not happen */ { 232f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return; 233f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams } 234f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 2357b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#if defined(ARCH_ARM_HAVE_VFP) 236f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams gArchUseSIMD = !!strstr(cpuinfo, " neon"); 2377b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#elif defined(ARCH_X86_HAVE_SSSE3) 2387b7060c61e4182b29186849c5a857ea5f0898e56Rose, James gArchUseSIMD = !!strstr(cpuinfo, " ssse3"); 2397b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#endif 240f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams} 2417b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#endif // ARCH_ARM_HAVE_VFP || ARCH_X86_HAVE_SSSE3 242f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 243709a0978ae141198018ca9769f8d96292a8928e6Jason Samsbool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 244709a0978ae141198018ca9769f8d96292a8928e6Jason Sams sym_lookup_t lfn, script_lookup_t slfn) { 245709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 246709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mSymLookupFn = lfn; 247709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mScriptLookupFn = slfn; 248709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 249709a0978ae141198018ca9769f8d96292a8928e6Jason Sams lockMutex(); 250709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!gThreadTLSKeyCount) { 251709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_key_create(&gThreadTLSKey, NULL); 252709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 253709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Failed to init thread tls key."); 254709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 255709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return false; 256709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 257709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 258709a0978ae141198018ca9769f8d96292a8928e6Jason Sams gThreadTLSKeyCount++; 259709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 260709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 261709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mTlsStruct.mContext = mRSC; 262709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mTlsStruct.mScript = NULL; 263709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 264709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 265709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("pthread_setspecific %i", status); 266709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 267709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2687b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_X86_HAVE_SSSE3) 269f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams GetCpuInfo(); 270f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams#endif 271f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 272709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int cpu = sysconf(_SC_NPROCESSORS_ONLN); 273709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mRSC->props.mDebugMaxThreads) { 274709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpu = mRSC->props.mDebugMaxThreads; 275709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 276709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (cpu < 2) { 277709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = 0; 278709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return true; 279709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 280709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 281709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // Subtract one from the cpu count because we also use the command thread as a worker. 282709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = (uint32_t)(cpu - 1); 283709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2848ca358a2abe7e0dba23993e0fc8d64b8b55bd9caJason Sams ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 285709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 286709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 287709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 288709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 289709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = NULL; 290709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 291709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCompleteSignal.init(); 292709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2930b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 2940b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mLaunchCount = 0; 2950b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 296709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 297709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_attr_t threadAttr; 298709a0978ae141198018ca9769f8d96292a8928e6Jason Sams status = pthread_attr_init(&threadAttr); 299709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 300709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Failed to init thread attribute."); 301709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return false; 302709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 303709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 304709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 305709a0978ae141198018ca9769f8d96292a8928e6Jason Sams status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 306709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 307709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = ct; 308709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Created fewer than expected number of RS threads."); 309709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 310709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 311709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 3120b575de8ed0b628d84d256f5846500b0385979bdTim Murray while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 313709a0978ae141198018ca9769f8d96292a8928e6Jason Sams usleep(100); 314709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 315709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 316709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_attr_destroy(&threadAttr); 317709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return true; 318709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 319709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 320709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 321709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::setPriority(int32_t priority) { 322709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 323709a0978ae141198018ca9769f8d96292a8928e6Jason Sams setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 324709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 325709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 326709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 327709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 328709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mExit = true; 329709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchData = NULL; 330709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = NULL; 3310b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 3320b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 333709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 334709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals[ct].set(); 335709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 336709a0978ae141198018ca9769f8d96292a8928e6Jason Sams void *res; 337709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 338709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_join(mWorkers.mThreadId[ct], &res); 339709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 3400b575de8ed0b628d84d256f5846500b0385979bdTim Murray rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 34107ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin free(mWorkers.mThreadId); 34207ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin free(mWorkers.mNativeThreadId); 34307ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin delete[] mWorkers.mLaunchSignals; 344709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 345709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // Global structure cleanup. 346709a0978ae141198018ca9769f8d96292a8928e6Jason Sams lockMutex(); 347709a0978ae141198018ca9769f8d96292a8928e6Jason Sams --gThreadTLSKeyCount; 348709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!gThreadTLSKeyCount) { 349709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_key_delete(gThreadTLSKey); 350709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 351709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 352709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 353709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 354709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 355709a0978ae141198018ca9769f8d96292a8928e6Jason Samstypedef void (*rs_t)(const void *, void *, const void *, uint32_t, uint32_t, uint32_t, uint32_t); 356709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 357709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic void wc_xy(void *usr, uint32_t idx) { 358709a0978ae141198018ca9769f8d96292a8928e6Jason Sams MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 359709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsForEachStubParamStruct p; 360709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memcpy(&p, &mtls->fep, sizeof(p)); 361709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.lid = idx; 362709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t sig = mtls->sig; 363709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 364709a0978ae141198018ca9769f8d96292a8928e6Jason Sams outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 365709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (1) { 3660b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 367709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize; 368709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t yEnd = yStart + mtls->mSliceSize; 369709a0978ae141198018ca9769f8d96292a8928e6Jason Sams yEnd = rsMin(yEnd, mtls->yEnd); 370709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (yEnd <= yStart) { 371709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return; 372709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 373709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 374709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr idx %i, x %i,%i y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd); 375709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr ptr in %p, out %p", mtls->fep.ptrIn, mtls->fep.ptrOut); 376709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 377709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.y = yStart; p.y < yEnd; p.y++) { 378709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * p.y) + 379709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideOut * mtls->xStart); 380709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * p.y) + 381709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideIn * mtls->xStart); 382709a0978ae141198018ca9769f8d96292a8928e6Jason Sams fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); 383709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 384709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 385709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 386709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 387709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic void wc_x(void *usr, uint32_t idx) { 388709a0978ae141198018ca9769f8d96292a8928e6Jason Sams MTLaunchStruct *mtls = (MTLaunchStruct *)usr; 389709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsForEachStubParamStruct p; 390709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memcpy(&p, &mtls->fep, sizeof(p)); 391709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.lid = idx; 392709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t sig = mtls->sig; 393709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 394709a0978ae141198018ca9769f8d96292a8928e6Jason Sams outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 395709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (1) { 3960b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 397709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize; 398709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t xEnd = xStart + mtls->mSliceSize; 399709a0978ae141198018ca9769f8d96292a8928e6Jason Sams xEnd = rsMin(xEnd, mtls->xEnd); 400709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (xEnd <= xStart) { 401709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return; 402709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 403709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 404709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr slice %i idx %i, x %i,%i", slice, idx, xStart, xEnd); 405709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("usr ptr in %p, out %p", mtls->fep.ptrIn, mtls->fep.ptrOut); 406709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 407709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.out = mtls->fep.ptrOut + (mtls->fep.eStrideOut * xStart); 408709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.in = mtls->fep.ptrIn + (mtls->fep.eStrideIn * xStart); 409709a0978ae141198018ca9769f8d96292a8928e6Jason Sams fn(&p, xStart, xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); 410709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 411709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 412709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 413709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::launchThreads(const Allocation * ain, Allocation * aout, 414709a0978ae141198018ca9769f8d96292a8928e6Jason Sams const RsScriptCall *sc, MTLaunchStruct *mtls) { 415709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 416709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //android::StopWatch kernel_time("kernel time"); 417709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 418709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInForEach) { 419709a0978ae141198018ca9769f8d96292a8928e6Jason Sams const size_t targetByteChunk = 16 * 1024; 420709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mInForEach = true; 421709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mtls->fep.dimY > 1) { 422709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s1 = mtls->fep.dimY / ((mWorkers.mCount + 1) * 4); 423709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s2 = 0; 424709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 425709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // This chooses our slice size to rate limit atomic ops to 426709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // one per 16k bytes of reads/writes. 427709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mtls->fep.yStrideOut) { 428709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.yStrideOut; 429709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 430709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.yStrideIn; 431709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 432709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = rsMin(s1, s2); 433709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 434709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mtls->mSliceSize < 1) { 435709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = 1; 436709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 437709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 438709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // mtls->mSliceSize = 2; 439709a0978ae141198018ca9769f8d96292a8928e6Jason Sams launchThreads(wc_xy, mtls); 440709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 441709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s1 = mtls->fep.dimX / ((mWorkers.mCount + 1) * 4); 442709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t s2 = 0; 443709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 444709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // This chooses our slice size to rate limit atomic ops to 445709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // one per 16k bytes of reads/writes. 446709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mtls->fep.eStrideOut) { 447709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.eStrideOut; 448709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 449709a0978ae141198018ca9769f8d96292a8928e6Jason Sams s2 = targetByteChunk / mtls->fep.eStrideIn; 450709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 451709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = rsMin(s1, s2); 452709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 453709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mtls->mSliceSize < 1) { 454709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->mSliceSize = 1; 455709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 456709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 457709a0978ae141198018ca9769f8d96292a8928e6Jason Sams launchThreads(wc_x, mtls); 458709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 459709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mInForEach = false; 460709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 461709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("launch 1"); 462709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 463709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsForEachStubParamStruct p; 464709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memcpy(&p, &mtls->fep, sizeof(p)); 465709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t sig = mtls->sig; 466709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 467709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("launch 3"); 468709a0978ae141198018ca9769f8d96292a8928e6Jason Sams outer_foreach_t fn = (outer_foreach_t) mtls->kernel; 469709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.ar[0] = mtls->arrayStart; p.ar[0] < mtls->arrayEnd; p.ar[0]++) { 470709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.z = mtls->zStart; p.z < mtls->zEnd; p.z++) { 471709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (p.y = mtls->yStart; p.y < mtls->yEnd; p.y++) { 472709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t offset = mtls->fep.dimY * mtls->fep.dimZ * p.ar[0] + 473709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mtls->fep.dimY * p.z + p.y; 474709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.out = mtls->fep.ptrOut + (mtls->fep.yStrideOut * offset) + 475709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideOut * mtls->xStart); 476709a0978ae141198018ca9769f8d96292a8928e6Jason Sams p.in = mtls->fep.ptrIn + (mtls->fep.yStrideIn * offset) + 477709a0978ae141198018ca9769f8d96292a8928e6Jason Sams (mtls->fep.eStrideIn * mtls->xStart); 478709a0978ae141198018ca9769f8d96292a8928e6Jason Sams fn(&p, mtls->xStart, mtls->xEnd, mtls->fep.eStrideIn, mtls->fep.eStrideOut); 479709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 480709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 481709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 482709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 483709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 484709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 485709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 486709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("setTls %p", sc); 487709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 488709a0978ae141198018ca9769f8d96292a8928e6Jason Sams rsAssert(tls); 489709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *old = tls->mImpl; 490709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mImpl = sc; 491709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mContext = mRSC; 492709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (sc) { 493709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mScript = sc->getScript(); 494709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 495709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mScript = NULL; 496709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 497709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return old; 498709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 499709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 500709a0978ae141198018ca9769f8d96292a8928e6Jason Samsconst RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 501709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return mSymLookupFn(mRSC, name); 502709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 503709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 504709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 505709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 506709a0978ae141198018ca9769f8d96292a8928e6Jason Sams char const *resName, char const *cacheDir, 507709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint8_t const *bitcode, size_t bitcodeSize, 508709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t flags) { 509709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 510709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 511005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 512005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines#ifndef RS_COMPATIBILITY_LIB 513005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines , getBccPluginName() 514005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines#endif 515005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines )) { 516709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete i; 517709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 518709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 519709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return i; 520709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 521709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 5227c4b888f2147edf99690b6af75470774ff31c43bJason Samsextern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 5237c4b888f2147edf99690b6af75470774ff31c43bJason Sams const Script *s, const Element *e); 524c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 525c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 526c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 527c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 528c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 529c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 530c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 531c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 532c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 533c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 534c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 535c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 536c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 537c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 5382282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 5392282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams const Script *s, const Element *e); 5406fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahayeextern RsdCpuScriptImpl * rsdIntrinsic_LoopFilter(RsdCpuReferenceImpl *ctx, 5416fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye const Script *s, const Element *e); 54239ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 54339ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams const Script *s, const Element *e); 544709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 545709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 546709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsScriptIntrinsicID iid, Element *e) { 547709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 548709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *i = NULL; 549709a0978ae141198018ca9769f8d96292a8928e6Jason Sams switch (iid) { 5507c4b888f2147edf99690b6af75470774ff31c43bJason Sams case RS_SCRIPT_INTRINSIC_ID_3DLUT: 5517c4b888f2147edf99690b6af75470774ff31c43bJason Sams i = rsdIntrinsic_3DLUT(this, s, e); 5527c4b888f2147edf99690b6af75470774ff31c43bJason Sams break; 553709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 554c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Convolve3x3(this, s, e); 555709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 556709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 557c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_ColorMatrix(this, s, e); 558709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 559709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_LUT: 560c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_LUT(this, s, e); 561709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 562709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 563c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Convolve5x5(this, s, e); 564709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 565709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_BLUR: 566c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Blur(this, s, e); 567709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 568709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 569c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_YuvToRGB(this, s, e); 570709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 571709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_BLEND: 572c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Blend(this, s, e); 573709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 5742282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 5752282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams i = rsdIntrinsic_Histogram(this, s, e); 5762282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams break; 57739ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams case RS_SCRIPT_INTRINSIC_ID_RESIZE: 57839ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams i = rsdIntrinsic_Resize(this, s, e); 57939ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams break; 5806fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye#ifndef RS_COMPATIBILITY_LIB 5816fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye case RS_SCRIPT_INTRINSIC_ID_LOOP_FILTER: 5826fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye i = rsdIntrinsic_LoopFilter(this, s, e); 5836fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye break; 5846fc3e12b8912458cb4adcfd32e2f53d76b0cc737Matthieu Delahaye#endif 585709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 586709a0978ae141198018ca9769f8d96292a8928e6Jason Sams default: 587709a0978ae141198018ca9769f8d96292a8928e6Jason Sams rsAssert(0); 588709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 589709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 590709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return i; 591709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 592709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 593709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScriptGroup * RsdCpuReferenceImpl::createScriptGroup(const ScriptGroup *sg) { 594709a0978ae141198018ca9769f8d96292a8928e6Jason Sams CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 595709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!sgi->init()) { 596709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete sgi; 597709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return NULL; 598709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 599709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return sgi; 600709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 601709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 602709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 603