rsCpuCore.cpp revision 633beab5de5ffe4c5068cf1c7a1c0bd09ee9f195
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" 201ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni#include "rsCpuScriptGroup2.h" 21709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 22709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <malloc.h> 23709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "rsContext.h" 24709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 25709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sys/types.h> 26709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sys/resource.h> 27709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sched.h> 28709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <sys/syscall.h> 2911fd9ec1ab8dfa7ae45c6edeea48dddc4633efeaMatt Wala#include <stdio.h> 30709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include <string.h> 31b0934b67b95cc27e2358c2aa4db5f7c1067c8f9bStephen Hines#include <unistd.h> 320b575de8ed0b628d84d256f5846500b0385979bdTim Murray 33b0934b67b95cc27e2358c2aa4db5f7c1067c8f9bStephen Hines#if !defined(RS_SERVER) && !defined(RS_COMPATIBILITY_LIB) 340b575de8ed0b628d84d256f5846500b0385979bdTim Murray#include <cutils/properties.h> 35709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#include "utils/StopWatch.h" 360b575de8ed0b628d84d256f5846500b0385979bdTim Murray#endif 370b575de8ed0b628d84d256f5846500b0385979bdTim Murray 380b575de8ed0b628d84d256f5846500b0385979bdTim Murray#ifdef RS_SERVER 390b575de8ed0b628d84d256f5846500b0385979bdTim Murray// Android exposes gettid(), standard Linux does not 400b575de8ed0b628d84d256f5846500b0385979bdTim Murraystatic pid_t gettid() { 410b575de8ed0b628d84d256f5846500b0385979bdTim Murray return syscall(SYS_gettid); 420b575de8ed0b628d84d256f5846500b0385979bdTim Murray} 430b575de8ed0b628d84d256f5846500b0385979bdTim Murray#endif 44709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 45709a0978ae141198018ca9769f8d96292a8928e6Jason Samsusing namespace android; 46709a0978ae141198018ca9769f8d96292a8928e6Jason Samsusing namespace android::renderscript; 47709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 486760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross#define REDUCE_NEW_ALOGV(mtls, level, ...) do { if ((mtls)->logReduce >= (level)) ALOGV(__VA_ARGS__); } while(0) 4910adb0c2029f112b5738228617d5645f6ecea0c5David Gross 50709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic pthread_key_t gThreadTLSKey = 0; 51709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic uint32_t gThreadTLSKeyCount = 0; 52709a0978ae141198018ca9769f8d96292a8928e6Jason Samsstatic pthread_mutex_t gInitMutex = PTHREAD_MUTEX_INITIALIZER; 53709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 54f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsbool android::renderscript::gArchUseSIMD = false; 55f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 56709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::~RsdCpuReference() { 57709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 58709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 59709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference * RsdCpuReference::create(Context *rsc, uint32_t version_major, 60cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams uint32_t version_minor, sym_lookup_t lfn, script_lookup_t slfn 61b043df0676fef226336deb3a00ead2f31e02343fDavid Gross , RSSelectRTCallback pSelectRTCallback, 62005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines const char *pBccPluginName 63cadfac411e6690e39de36c4f9e94deb9b7d2d08eJason Sams ) { 64709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 65709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuReferenceImpl *cpu = new RsdCpuReferenceImpl(rsc); 66709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!cpu) { 6744bef6fba6244292b751387f3d6c31cca96c28adChris Wailes return nullptr; 68709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 69709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!cpu->init(version_major, version_minor, lfn, slfn)) { 70709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete cpu; 7144bef6fba6244292b751387f3d6c31cca96c28adChris Wailes return nullptr; 72709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 73f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines 741d476620399d54774e4fd386c1d23cc583d49522Stephen Hines cpu->setSelectRTCallback(pSelectRTCallback); 75005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines if (pBccPluginName) { 76005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines cpu->setBccPluginName(pBccPluginName); 77005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines } 78f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hines 79709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return cpu; 80709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 81709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 82709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 83709a0978ae141198018ca9769f8d96292a8928e6Jason SamsContext * RsdCpuReference::getTlsContext() { 84709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 85709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return tls->mContext; 86709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 87709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 88709a0978ae141198018ca9769f8d96292a8928e6Jason Samsconst Script * RsdCpuReference::getTlsScript() { 89709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 90709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return tls->mScript; 91709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 92709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 93f218bf115af4ae4fd79adbb8842608b308a4cf07Stephen Hinespthread_key_t RsdCpuReference::getThreadTLSKey(){ return gThreadTLSKey; } 94709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 95709a0978ae141198018ca9769f8d96292a8928e6Jason Sams//////////////////////////////////////////////////////////// 96709a0978ae141198018ca9769f8d96292a8928e6Jason Sams/// 97709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 98709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReferenceImpl::RsdCpuReferenceImpl(Context *rsc) { 99709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mRSC = rsc; 100709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 101709a0978ae141198018ca9769f8d96292a8928e6Jason Sams version_major = 0; 102709a0978ae141198018ca9769f8d96292a8928e6Jason Sams version_minor = 0; 10310adb0c2029f112b5738228617d5645f6ecea0c5David Gross mInKernel = false; 104709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&mWorkers, 0, sizeof(mWorkers)); 105709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&mTlsStruct, 0, sizeof(mTlsStruct)); 106709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mExit = false; 10744bef6fba6244292b751387f3d6c31cca96c28adChris Wailes mSelectRTCallback = nullptr; 1088409d6414dd4a42aa59779fcfe9fce18648cb135Stephen Hines mEmbedGlobalInfo = true; 1098409d6414dd4a42aa59779fcfe9fce18648cb135Stephen Hines mEmbedGlobalInfoSkipConstant = true; 110709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 111709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 112709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 113709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid * RsdCpuReferenceImpl::helperThreadProc(void *vrsc) { 114709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuReferenceImpl *dc = (RsdCpuReferenceImpl *)vrsc; 115709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1160b575de8ed0b628d84d256f5846500b0385979bdTim Murray uint32_t idx = __sync_fetch_and_add(&dc->mWorkers.mLaunchCount, 1); 117709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 118709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGV("RS helperThread starting %p idx=%i", dc, idx); 119709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 120709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchSignals[idx].init(); 121709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mNativeThreadId[idx] = gettid(); 122709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 123709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&dc->mTlsStruct, 0, sizeof(dc->mTlsStruct)); 124709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_setspecific(gThreadTLSKey, &dc->mTlsStruct); 125709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 126709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("pthread_setspecific %i", status); 127709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 128709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 129709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#if 0 130709a0978ae141198018ca9769f8d96292a8928e6Jason Sams typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t; 131709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpu_set_t cpuset; 132709a0978ae141198018ca9769f8d96292a8928e6Jason Sams memset(&cpuset, 0, sizeof(cpuset)); 133709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpuset.bits[idx / 64] |= 1ULL << (idx % 64); 134709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx], 135709a0978ae141198018ca9769f8d96292a8928e6Jason Sams sizeof(cpuset), &cpuset); 136709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret)); 137709a0978ae141198018ca9769f8d96292a8928e6Jason Sams#endif 138709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 139709a0978ae141198018ca9769f8d96292a8928e6Jason Sams while (!dc->mExit) { 140709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchSignals[idx].wait(); 141709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (dc->mWorkers.mLaunchCallback) { 142709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // idx +1 is used because the calling thread is always worker 0. 143709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx+1); 144709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 1450b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_fetch_and_sub(&dc->mWorkers.mRunningCount, 1); 146709a0978ae141198018ca9769f8d96292a8928e6Jason Sams dc->mWorkers.mCompleteSignal.set(); 147709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 148709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 149709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGV("RS helperThread exited %p idx=%i", dc, idx); 15044bef6fba6244292b751387f3d6c31cca96c28adChris Wailes return nullptr; 151709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 152709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 15314ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// Launch a kernel. 15414ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// The callback function is called to execute the kernel. 155709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::launchThreads(WorkerCallback_t cbk, void *data) { 156709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchData = data; 157709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchCallback = cbk; 1584d252d6e807b89764dad123ac845df298c52ca97Tim Murray 1594d252d6e807b89764dad123ac845df298c52ca97Tim Murray // fast path for very small launches 16014ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala MTLaunchStructCommon *mtls = (MTLaunchStructCommon *)data; 16114ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala if (mtls && mtls->dimPtr->y <= 1 && mtls->end.x <= mtls->start.x + mtls->mSliceSize) { 1624d252d6e807b89764dad123ac845df298c52ca97Tim Murray if (mWorkers.mLaunchCallback) { 1634d252d6e807b89764dad123ac845df298c52ca97Tim Murray mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 1644d252d6e807b89764dad123ac845df298c52ca97Tim Murray } 1654d252d6e807b89764dad123ac845df298c52ca97Tim Murray return; 1664d252d6e807b89764dad123ac845df298c52ca97Tim Murray } 1674d252d6e807b89764dad123ac845df298c52ca97Tim Murray 1680b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 1690b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 1700b575de8ed0b628d84d256f5846500b0385979bdTim Murray 171709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 172709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals[ct].set(); 173709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 174709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 175709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // We use the calling thread as one of the workers so we can start without 176709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // the delay of the thread wakeup. 177709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (mWorkers.mLaunchCallback) { 1784d252d6e807b89764dad123ac845df298c52ca97Tim Murray mWorkers.mLaunchCallback(mWorkers.mLaunchData, 0); 179709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 180709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1810b575de8ed0b628d84d256f5846500b0385979bdTim Murray while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 182709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCompleteSignal.wait(); 183709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 184709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 185709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 186709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 187709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::lockMutex() { 188709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_mutex_lock(&gInitMutex); 189709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 190709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 191709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::unlockMutex() { 192709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_mutex_unlock(&gInitMutex); 193709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 194709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 19511fd9ec1ab8dfa7ae45c6edeea48dddc4633efeaMatt Wala// Determine if the CPU we're running on supports SIMD instructions. 196f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Samsstatic void GetCpuInfo() { 19711fd9ec1ab8dfa7ae45c6edeea48dddc4633efeaMatt Wala // Read the CPU flags from /proc/cpuinfo. 19811fd9ec1ab8dfa7ae45c6edeea48dddc4633efeaMatt Wala FILE *cpuinfo = fopen("/proc/cpuinfo", "r"); 199f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 20011fd9ec1ab8dfa7ae45c6edeea48dddc4633efeaMatt Wala if (!cpuinfo) { 201f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams return; 202f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams } 203f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 20411fd9ec1ab8dfa7ae45c6edeea48dddc4633efeaMatt Wala char cpuinfostr[4096]; 2055d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang // fgets() ends with newline or EOF, need to check the whole 2065d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang // "cpuinfo" file to make sure we can use SIMD or not. 2075d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang while (fgets(cpuinfostr, sizeof(cpuinfostr), cpuinfo)) { 208074424a4ac5b093331df2c92e7a5bcbfff136b71Jason Sams#if defined(ARCH_ARM_HAVE_VFP) || defined(ARCH_ARM_USE_INTRINSICS) 2095d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang gArchUseSIMD = strstr(cpuinfostr, " neon") || strstr(cpuinfostr, " asimd"); 2107b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#elif defined(ARCH_X86_HAVE_SSSE3) 2115d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang gArchUseSIMD = strstr(cpuinfostr, " ssse3"); 2127b7060c61e4182b29186849c5a857ea5f0898e56Rose, James#endif 2135d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang if (gArchUseSIMD) { 2145d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang break; 2155d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang } 2165d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang } 2175d70cb591d78d62d10839a52302ec9087c6f3350Miao Wang fclose(cpuinfo); 218f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams} 219f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 220709a0978ae141198018ca9769f8d96292a8928e6Jason Samsbool RsdCpuReferenceImpl::init(uint32_t version_major, uint32_t version_minor, 221709a0978ae141198018ca9769f8d96292a8928e6Jason Sams sym_lookup_t lfn, script_lookup_t slfn) { 222709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mSymLookupFn = lfn; 223709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mScriptLookupFn = slfn; 224709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 225709a0978ae141198018ca9769f8d96292a8928e6Jason Sams lockMutex(); 226709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!gThreadTLSKeyCount) { 22744bef6fba6244292b751387f3d6c31cca96c28adChris Wailes int status = pthread_key_create(&gThreadTLSKey, nullptr); 228709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 229709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Failed to init thread tls key."); 230709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 231709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return false; 232709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 233709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 234709a0978ae141198018ca9769f8d96292a8928e6Jason Sams gThreadTLSKeyCount++; 235709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 236709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 237709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mTlsStruct.mContext = mRSC; 23844bef6fba6244292b751387f3d6c31cca96c28adChris Wailes mTlsStruct.mScript = nullptr; 239709a0978ae141198018ca9769f8d96292a8928e6Jason Sams int status = pthread_setspecific(gThreadTLSKey, &mTlsStruct); 240709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 241709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("pthread_setspecific %i", status); 242709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 243709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 24410adb0c2029f112b5738228617d5645f6ecea0c5David Gross mPageSize = sysconf(_SC_PAGE_SIZE); 2456760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross // ALOGV("page size = %ld", mPageSize); 24610adb0c2029f112b5738228617d5645f6ecea0c5David Gross 247f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams GetCpuInfo(); 248f5ef8df639ba6363aa5d546e57ce872d04144cb6Jason Sams 24977d57a305f4134e78ebc91869011c4009988104eJason Sams int cpu = sysconf(_SC_NPROCESSORS_CONF); 250709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if(mRSC->props.mDebugMaxThreads) { 251709a0978ae141198018ca9769f8d96292a8928e6Jason Sams cpu = mRSC->props.mDebugMaxThreads; 252709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 253709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (cpu < 2) { 254709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = 0; 255709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return true; 256709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 257709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 258709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // Subtract one from the cpu count because we also use the command thread as a worker. 259709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = (uint32_t)(cpu - 1); 260709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 261633beab5de5ffe4c5068cf1c7a1c0bd09ee9f195Yang Ni if (mRSC->props.mLogScripts) { 262633beab5de5ffe4c5068cf1c7a1c0bd09ee9f195Yang Ni ALOGV("%p Launching thread(s), CPUs %i", mRSC, mWorkers.mCount + 1); 263633beab5de5ffe4c5068cf1c7a1c0bd09ee9f195Yang Ni } 264709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 265709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t)); 266709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t)); 267709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals = new Signal[mWorkers.mCount]; 26844bef6fba6244292b751387f3d6c31cca96c28adChris Wailes mWorkers.mLaunchCallback = nullptr; 269709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 270709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCompleteSignal.init(); 271709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 2720b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 2730b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mLaunchCount = 0; 2740b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 275709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 276709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_attr_t threadAttr; 277709a0978ae141198018ca9769f8d96292a8928e6Jason Sams status = pthread_attr_init(&threadAttr); 278709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 279709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Failed to init thread attribute."); 280709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return false; 281709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 282709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 283709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 284709a0978ae141198018ca9769f8d96292a8928e6Jason Sams status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this); 285709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (status) { 286709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mCount = ct; 287709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ALOGE("Created fewer than expected number of RS threads."); 288709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 289709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 290709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 2910b575de8ed0b628d84d256f5846500b0385979bdTim Murray while (__sync_fetch_and_or(&mWorkers.mRunningCount, 0) != 0) { 292709a0978ae141198018ca9769f8d96292a8928e6Jason Sams usleep(100); 293709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 294709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 295709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_attr_destroy(&threadAttr); 296709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return true; 297709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 298709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 299709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 300709a0978ae141198018ca9769f8d96292a8928e6Jason Samsvoid RsdCpuReferenceImpl::setPriority(int32_t priority) { 301709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct=0; ct < mWorkers.mCount; ct++) { 302709a0978ae141198018ca9769f8d96292a8928e6Jason Sams setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], priority); 303709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 304709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 305709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 306709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReferenceImpl::~RsdCpuReferenceImpl() { 307709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mExit = true; 30844bef6fba6244292b751387f3d6c31cca96c28adChris Wailes mWorkers.mLaunchData = nullptr; 30944bef6fba6244292b751387f3d6c31cca96c28adChris Wailes mWorkers.mLaunchCallback = nullptr; 3100b575de8ed0b628d84d256f5846500b0385979bdTim Murray mWorkers.mRunningCount = mWorkers.mCount; 3110b575de8ed0b628d84d256f5846500b0385979bdTim Murray __sync_synchronize(); 312709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 313709a0978ae141198018ca9769f8d96292a8928e6Jason Sams mWorkers.mLaunchSignals[ct].set(); 314709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 315709a0978ae141198018ca9769f8d96292a8928e6Jason Sams void *res; 316709a0978ae141198018ca9769f8d96292a8928e6Jason Sams for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) { 317709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_join(mWorkers.mThreadId[ct], &res); 318709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 3190b575de8ed0b628d84d256f5846500b0385979bdTim Murray rsAssert(__sync_fetch_and_or(&mWorkers.mRunningCount, 0) == 0); 32007ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin free(mWorkers.mThreadId); 32107ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin free(mWorkers.mNativeThreadId); 32207ef704308b514272ed2f5c3e6a2f4c055550158Jens Gulin delete[] mWorkers.mLaunchSignals; 323709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 324709a0978ae141198018ca9769f8d96292a8928e6Jason Sams // Global structure cleanup. 325709a0978ae141198018ca9769f8d96292a8928e6Jason Sams lockMutex(); 326709a0978ae141198018ca9769f8d96292a8928e6Jason Sams --gThreadTLSKeyCount; 327709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (!gThreadTLSKeyCount) { 328709a0978ae141198018ca9769f8d96292a8928e6Jason Sams pthread_key_delete(gThreadTLSKey); 329709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 330709a0978ae141198018ca9769f8d96292a8928e6Jason Sams unlockMutex(); 331709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 332709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 333709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 33414ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// Set up the appropriate input and output pointers to the kernel driver info structure. 33514ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// Inputs: 33614ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// mtls - The MTLaunchStruct holding information about the kernel launch 33714ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// fep - The forEach parameters (driver info structure) 33814ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// x, y, z, lod, face, a1, a2, a3, a4 - The start offsets into each dimension 33914ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Walastatic inline void FepPtrSetup(const MTLaunchStructForEach *mtls, RsExpandKernelDriverInfo *fep, 340c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t x, uint32_t y, 341c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t z = 0, uint32_t lod = 0, 342c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams RsAllocationCubemapFace face = RS_ALLOCATION_CUBEMAP_FACE_POSITIVE_X, 343c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t a1 = 0, uint32_t a2 = 0, uint32_t a3 = 0, uint32_t a4 = 0) { 344c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams for (uint32_t i = 0; i < fep->inLen; i++) { 345c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams fep->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 346709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 347c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams if (mtls->aout[0] != nullptr) { 348c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams fep->outPtr[0] = (uint8_t *)mtls->aout[0]->getPointerUnchecked(x, y, z, lod, face, a1, a2, a3, a4); 349c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams } 350709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 351709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 3526c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// Set up the appropriate input and output pointers to the kernel driver info structure. 3536c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// Inputs: 3546c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// mtls - The MTLaunchStruct holding information about the kernel launch 3556c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// redp - The reduce parameters (driver info structure) 3566c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// x, y, z - The start offsets into each dimension 3576c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Grossstatic inline void RedpPtrSetup(const MTLaunchStructReduceNew *mtls, RsExpandKernelDriverInfo *redp, 3586c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross uint32_t x, uint32_t y, uint32_t z) { 3596c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross for (uint32_t i = 0; i < redp->inLen; i++) { 3606c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross redp->inPtr[i] = (const uint8_t *)mtls->ains[i]->getPointerUnchecked(x, y, z); 3616c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross } 3626c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross} 3636c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross 364bf2111d3b3de310932099514f06924e48fa1d7b2Jason Samsstatic uint32_t sliceInt(uint32_t *p, uint32_t val, uint32_t start, uint32_t end) { 365bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams if (start >= end) { 366bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams *p = start; 367bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams return val; 368bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams } 369bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 370bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t div = end - start; 371bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 372bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t n = val / div; 373bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams *p = (val - (n * div)) + start; 374bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams return n; 375bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams} 376bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 3776c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Grossstatic bool SelectOuterSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) { 378bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t r = sliceNum; 3796c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.z, r, mtls->start.z, mtls->end.z); 3806c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.lod, r, mtls->start.lod, mtls->end.lod); 3816c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.face, r, mtls->start.face, mtls->end.face); 3826c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.array[0], r, mtls->start.array[0], mtls->end.array[0]); 3836c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.array[1], r, mtls->start.array[1], mtls->end.array[1]); 3846c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.array[2], r, mtls->start.array[2], mtls->end.array[2]); 3856c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross r = sliceInt(&info->current.array[3], r, mtls->start.array[3], mtls->end.array[3]); 386bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams return r == 0; 387bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams} 388bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 389415a249300cb9ca3f560814bfde52c6796fed1f7David Grossstatic bool SelectZSlice(const MTLaunchStructCommon *mtls, RsExpandKernelDriverInfo* info, uint32_t sliceNum) { 390415a249300cb9ca3f560814bfde52c6796fed1f7David Gross return sliceInt(&info->current.z, sliceNum, mtls->start.z, mtls->end.z) == 0; 391415a249300cb9ca3f560814bfde52c6796fed1f7David Gross} 392bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 393415a249300cb9ca3f560814bfde52c6796fed1f7David Grossstatic void walk_general_foreach(void *usr, uint32_t idx) { 39414ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 395bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams RsExpandKernelDriverInfo fep = mtls->fep; 396bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams fep.lid = idx; 39714ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala ForEachFunc_t fn = mtls->kernel; 398bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 399bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams while(1) { 400bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 401bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 40259b35f29cc576243a322bf88bc16063a9810da55Jason Sams if (!SelectOuterSlice(mtls, &fep, slice)) { 403bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams return; 404bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams } 405bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 40659b35f29cc576243a322bf88bc16063a9810da55Jason Sams for (fep.current.y = mtls->start.y; fep.current.y < mtls->end.y; 40759b35f29cc576243a322bf88bc16063a9810da55Jason Sams fep.current.y++) { 408bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 40959b35f29cc576243a322bf88bc16063a9810da55Jason Sams FepPtrSetup(mtls, &fep, mtls->start.x, 41059b35f29cc576243a322bf88bc16063a9810da55Jason Sams fep.current.y, fep.current.z, fep.current.lod, 41159b35f29cc576243a322bf88bc16063a9810da55Jason Sams (RsAllocationCubemapFace)fep.current.face, 41259b35f29cc576243a322bf88bc16063a9810da55Jason Sams fep.current.array[0], fep.current.array[1], 41359b35f29cc576243a322bf88bc16063a9810da55Jason Sams fep.current.array[2], fep.current.array[3]); 414bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 41559b35f29cc576243a322bf88bc16063a9810da55Jason Sams fn(&fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 416bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams } 417bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams } 418bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams} 419c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams 420415a249300cb9ca3f560814bfde52c6796fed1f7David Grossstatic void walk_2d_foreach(void *usr, uint32_t idx) { 42114ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 422c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams RsExpandKernelDriverInfo fep = mtls->fep; 423c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams fep.lid = idx; 42414ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala ForEachFunc_t fn = mtls->kernel; 425709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 426c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams while (1) { 427c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 428bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 429c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t yEnd = yStart + mtls->mSliceSize; 430709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 431bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams yEnd = rsMin(yEnd, mtls->end.y); 432709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 433c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams if (yEnd <= yStart) { 434c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams return; 435c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams } 436709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 437c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams for (fep.current.y = yStart; fep.current.y < yEnd; fep.current.y++) { 438bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams FepPtrSetup(mtls, &fep, mtls->start.x, fep.current.y); 439709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 440b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross fn(&fep, mtls->start.x, mtls->end.x, fep.outStride[0]); 441709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 442c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams } 443f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes} 4444b2bea3dc20865f3a198797702e19912a6a2171cStephen Hines 44510adb0c2029f112b5738228617d5645f6ecea0c5David Grossstatic void walk_1d_foreach(void *usr, uint32_t idx) { 44614ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala MTLaunchStructForEach *mtls = (MTLaunchStructForEach *)usr; 447c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams RsExpandKernelDriverInfo fep = mtls->fep; 448c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams fep.lid = idx; 44914ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala ForEachFunc_t fn = mtls->kernel; 45080ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes 451c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams while (1) { 452c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 453bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 454c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t xEnd = xStart + mtls->mSliceSize; 45580ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes 456bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams xEnd = rsMin(xEnd, mtls->end.x); 45780ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes 458c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams if (xEnd <= xStart) { 459c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams return; 460c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams } 46180ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes 462bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams FepPtrSetup(mtls, &fep, xStart, 0); 46380ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes 464b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross fn(&fep, xStart, xEnd, fep.outStride[0]); 465c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams } 466709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 467709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 46810adb0c2029f112b5738228617d5645f6ecea0c5David Gross// The function format_bytes() is an auxiliary function to assist in logging. 46910adb0c2029f112b5738228617d5645f6ecea0c5David Gross// 47010adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Bytes are read from an input (inBuf) and written (as pairs of hex digits) 47110adb0c2029f112b5738228617d5645f6ecea0c5David Gross// to an output (outBuf). 47210adb0c2029f112b5738228617d5645f6ecea0c5David Gross// 47310adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Output format: 47410adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - starts with ": " 47510adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - each input byte is translated to a pair of hex digits 47610adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - bytes are separated by "." except that every fourth separator is "|" 47710adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - if the input is sufficiently long, the output is truncated and terminated with "..." 47810adb0c2029f112b5738228617d5645f6ecea0c5David Gross// 47910adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Arguments: 48010adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - outBuf -- Pointer to buffer of type "FormatBuf" into which output is written 48110adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - inBuf -- Pointer to bytes which are to be formatted into outBuf 48210adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - inBytes -- Number of bytes in inBuf 48310adb0c2029f112b5738228617d5645f6ecea0c5David Gross// 48410adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Constant: 48510adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - kFormatInBytesMax -- Only min(kFormatInBytesMax, inBytes) bytes will be read 48610adb0c2029f112b5738228617d5645f6ecea0c5David Gross// from inBuf 48710adb0c2029f112b5738228617d5645f6ecea0c5David Gross// 48810adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Return value: 48910adb0c2029f112b5738228617d5645f6ecea0c5David Gross// - pointer (const char *) to output (which is part of outBuf) 49010adb0c2029f112b5738228617d5645f6ecea0c5David Gross// 49110adb0c2029f112b5738228617d5645f6ecea0c5David Grossstatic const int kFormatInBytesMax = 16; 49210adb0c2029f112b5738228617d5645f6ecea0c5David Gross// ": " + 2 digits per byte + 1 separator between bytes + "..." + null 49310adb0c2029f112b5738228617d5645f6ecea0c5David Grosstypedef char FormatBuf[2 + kFormatInBytesMax*2 + (kFormatInBytesMax - 1) + 3 + 1]; 49410adb0c2029f112b5738228617d5645f6ecea0c5David Grossstatic const char *format_bytes(FormatBuf *outBuf, const uint8_t *inBuf, const int inBytes) { 49510adb0c2029f112b5738228617d5645f6ecea0c5David Gross strcpy(*outBuf, ": "); 49610adb0c2029f112b5738228617d5645f6ecea0c5David Gross int pos = 2; 49710adb0c2029f112b5738228617d5645f6ecea0c5David Gross const int lim = std::min(kFormatInBytesMax, inBytes); 49810adb0c2029f112b5738228617d5645f6ecea0c5David Gross for (int i = 0; i < lim; ++i) { 49910adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (i) { 50010adb0c2029f112b5738228617d5645f6ecea0c5David Gross sprintf(*outBuf + pos, (i % 4 ? "." : "|")); 50110adb0c2029f112b5738228617d5645f6ecea0c5David Gross ++pos; 50210adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 50310adb0c2029f112b5738228617d5645f6ecea0c5David Gross sprintf(*outBuf + pos, "%02x", inBuf[i]); 50410adb0c2029f112b5738228617d5645f6ecea0c5David Gross pos += 2; 50510adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 50610adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (kFormatInBytesMax < inBytes) 50710adb0c2029f112b5738228617d5645f6ecea0c5David Gross strcpy(*outBuf + pos, "..."); 50810adb0c2029f112b5738228617d5645f6ecea0c5David Gross return *outBuf; 50910adb0c2029f112b5738228617d5645f6ecea0c5David Gross} 51010adb0c2029f112b5738228617d5645f6ecea0c5David Gross 511415a249300cb9ca3f560814bfde52c6796fed1f7David Grossstatic void reduce_new_get_accumulator(uint8_t *&accumPtr, const MTLaunchStructReduceNew *mtls, 512415a249300cb9ca3f560814bfde52c6796fed1f7David Gross const char *walkerName, uint32_t threadIdx) { 513415a249300cb9ca3f560814bfde52c6796fed1f7David Gross rsAssert(!accumPtr); 514415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 515415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint32_t accumIdx = (uint32_t)__sync_fetch_and_add(&mtls->accumCount, 1); 516415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (mtls->outFunc) { 517415a249300cb9ca3f560814bfde52c6796fed1f7David Gross accumPtr = mtls->accumAlloc + mtls->accumStride * accumIdx; 518415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else { 519415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (accumIdx == 0) { 520415a249300cb9ca3f560814bfde52c6796fed1f7David Gross accumPtr = mtls->redp.outPtr[0]; 521415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else { 522415a249300cb9ca3f560814bfde52c6796fed1f7David Gross accumPtr = mtls->accumAlloc + mtls->accumStride * (accumIdx - 1); 523415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 524415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 525415a249300cb9ca3f560814bfde52c6796fed1f7David Gross REDUCE_NEW_ALOGV(mtls, 2, "%s(%p): idx = %u got accumCount %u and accumPtr %p", 526415a249300cb9ca3f560814bfde52c6796fed1f7David Gross walkerName, mtls->accumFunc, threadIdx, accumIdx, accumPtr); 527415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // initialize accumulator 528415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (mtls->initFunc) { 529415a249300cb9ca3f560814bfde52c6796fed1f7David Gross mtls->initFunc(accumPtr); 530415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else { 531415a249300cb9ca3f560814bfde52c6796fed1f7David Gross memset(accumPtr, 0, mtls->accumSize); 532415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 533415a249300cb9ca3f560814bfde52c6796fed1f7David Gross} 534415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 53510adb0c2029f112b5738228617d5645f6ecea0c5David Grossstatic void walk_1d_reduce_new(void *usr, uint32_t idx) { 53610adb0c2029f112b5738228617d5645f6ecea0c5David Gross const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr; 53710adb0c2029f112b5738228617d5645f6ecea0c5David Gross RsExpandKernelDriverInfo redp = mtls->redp; 53810adb0c2029f112b5738228617d5645f6ecea0c5David Gross 53910adb0c2029f112b5738228617d5645f6ecea0c5David Gross // find accumulator 54010adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint8_t *&accumPtr = mtls->accumPtr[idx]; 54110adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (!accumPtr) { 542415a249300cb9ca3f560814bfde52c6796fed1f7David Gross reduce_new_get_accumulator(accumPtr, mtls, __func__, idx); 54310adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 54410adb0c2029f112b5738228617d5645f6ecea0c5David Gross 54510adb0c2029f112b5738228617d5645f6ecea0c5David Gross // accumulate 54610adb0c2029f112b5738228617d5645f6ecea0c5David Gross const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc; 54710adb0c2029f112b5738228617d5645f6ecea0c5David Gross while (1) { 54810adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 54910adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint32_t xStart = mtls->start.x + slice * mtls->mSliceSize; 55010adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint32_t xEnd = xStart + mtls->mSliceSize; 55110adb0c2029f112b5738228617d5645f6ecea0c5David Gross 55210adb0c2029f112b5738228617d5645f6ecea0c5David Gross xEnd = rsMin(xEnd, mtls->end.x); 55310adb0c2029f112b5738228617d5645f6ecea0c5David Gross 55410adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (xEnd <= xStart) { 55510adb0c2029f112b5738228617d5645f6ecea0c5David Gross return; 55610adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 55710adb0c2029f112b5738228617d5645f6ecea0c5David Gross 55810adb0c2029f112b5738228617d5645f6ecea0c5David Gross RedpPtrSetup(mtls, &redp, xStart, 0, 0); 55910adb0c2029f112b5738228617d5645f6ecea0c5David Gross fn(&redp, xStart, xEnd, accumPtr); 56010adb0c2029f112b5738228617d5645f6ecea0c5David Gross 561415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // Emit log line after slice has been run, so that we can include 562415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // the results of the run on that line. 56310adb0c2029f112b5738228617d5645f6ecea0c5David Gross FormatBuf fmt; 5646760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross if (mtls->logReduce >= 3) { 56510adb0c2029f112b5738228617d5645f6ecea0c5David Gross format_bytes(&fmt, accumPtr, mtls->accumSize); 56610adb0c2029f112b5738228617d5645f6ecea0c5David Gross } else { 56710adb0c2029f112b5738228617d5645f6ecea0c5David Gross fmt[0] = 0; 56810adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 569415a249300cb9ca3f560814bfde52c6796fed1f7David Gross REDUCE_NEW_ALOGV(mtls, 2, "walk_1d_reduce_new(%p): idx = %u, x in [%u, %u)%s", 57010adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumFunc, idx, xStart, xEnd, fmt); 57110adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 57210adb0c2029f112b5738228617d5645f6ecea0c5David Gross} 57310adb0c2029f112b5738228617d5645f6ecea0c5David Gross 574415a249300cb9ca3f560814bfde52c6796fed1f7David Grossstatic void walk_2d_reduce_new(void *usr, uint32_t idx) { 575415a249300cb9ca3f560814bfde52c6796fed1f7David Gross const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr; 576415a249300cb9ca3f560814bfde52c6796fed1f7David Gross RsExpandKernelDriverInfo redp = mtls->redp; 577415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 578415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // find accumulator 579415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint8_t *&accumPtr = mtls->accumPtr[idx]; 580415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (!accumPtr) { 581415a249300cb9ca3f560814bfde52c6796fed1f7David Gross reduce_new_get_accumulator(accumPtr, mtls, __func__, idx); 582415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 583415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 584415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // accumulate 585415a249300cb9ca3f560814bfde52c6796fed1f7David Gross const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc; 586415a249300cb9ca3f560814bfde52c6796fed1f7David Gross while (1) { 587415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 588415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint32_t yStart = mtls->start.y + slice * mtls->mSliceSize; 589415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint32_t yEnd = yStart + mtls->mSliceSize; 590415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 591415a249300cb9ca3f560814bfde52c6796fed1f7David Gross yEnd = rsMin(yEnd, mtls->end.y); 592415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 593415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (yEnd <= yStart) { 594415a249300cb9ca3f560814bfde52c6796fed1f7David Gross return; 595415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 596415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 597415a249300cb9ca3f560814bfde52c6796fed1f7David Gross for (redp.current.y = yStart; redp.current.y < yEnd; redp.current.y++) { 598415a249300cb9ca3f560814bfde52c6796fed1f7David Gross RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, 0); 599415a249300cb9ca3f560814bfde52c6796fed1f7David Gross fn(&redp, mtls->start.x, mtls->end.x, accumPtr); 600415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 601415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 602415a249300cb9ca3f560814bfde52c6796fed1f7David Gross FormatBuf fmt; 603415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (mtls->logReduce >= 3) { 604415a249300cb9ca3f560814bfde52c6796fed1f7David Gross format_bytes(&fmt, accumPtr, mtls->accumSize); 605415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else { 606415a249300cb9ca3f560814bfde52c6796fed1f7David Gross fmt[0] = 0; 607415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 608415a249300cb9ca3f560814bfde52c6796fed1f7David Gross REDUCE_NEW_ALOGV(mtls, 2, "walk_2d_reduce_new(%p): idx = %u, y in [%u, %u)%s", 609415a249300cb9ca3f560814bfde52c6796fed1f7David Gross mtls->accumFunc, idx, yStart, yEnd, fmt); 610415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 611415a249300cb9ca3f560814bfde52c6796fed1f7David Gross} 612415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 613415a249300cb9ca3f560814bfde52c6796fed1f7David Grossstatic void walk_3d_reduce_new(void *usr, uint32_t idx) { 614415a249300cb9ca3f560814bfde52c6796fed1f7David Gross const MTLaunchStructReduceNew *mtls = (const MTLaunchStructReduceNew *)usr; 615415a249300cb9ca3f560814bfde52c6796fed1f7David Gross RsExpandKernelDriverInfo redp = mtls->redp; 616415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 617415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // find accumulator 618415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint8_t *&accumPtr = mtls->accumPtr[idx]; 619415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (!accumPtr) { 620415a249300cb9ca3f560814bfde52c6796fed1f7David Gross reduce_new_get_accumulator(accumPtr, mtls, __func__, idx); 621415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 622415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 623415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // accumulate 624415a249300cb9ca3f560814bfde52c6796fed1f7David Gross const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc; 625415a249300cb9ca3f560814bfde52c6796fed1f7David Gross while (1) { 626415a249300cb9ca3f560814bfde52c6796fed1f7David Gross uint32_t slice = (uint32_t)__sync_fetch_and_add(&mtls->mSliceNum, 1); 627415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 628415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (!SelectZSlice(mtls, &redp, slice)) { 629415a249300cb9ca3f560814bfde52c6796fed1f7David Gross return; 630415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 631415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 632415a249300cb9ca3f560814bfde52c6796fed1f7David Gross for (redp.current.y = mtls->start.y; redp.current.y < mtls->end.y; redp.current.y++) { 633415a249300cb9ca3f560814bfde52c6796fed1f7David Gross RedpPtrSetup(mtls, &redp, mtls->start.x, redp.current.y, redp.current.z); 634415a249300cb9ca3f560814bfde52c6796fed1f7David Gross fn(&redp, mtls->start.x, mtls->end.x, accumPtr); 635415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 636415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 637415a249300cb9ca3f560814bfde52c6796fed1f7David Gross FormatBuf fmt; 638415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (mtls->logReduce >= 3) { 639415a249300cb9ca3f560814bfde52c6796fed1f7David Gross format_bytes(&fmt, accumPtr, mtls->accumSize); 640415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else { 641415a249300cb9ca3f560814bfde52c6796fed1f7David Gross fmt[0] = 0; 642415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 643415a249300cb9ca3f560814bfde52c6796fed1f7David Gross REDUCE_NEW_ALOGV(mtls, 2, "walk_3d_reduce_new(%p): idx = %u, z = %u%s", 644415a249300cb9ca3f560814bfde52c6796fed1f7David Gross mtls->accumFunc, idx, redp.current.z, fmt); 645415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 646415a249300cb9ca3f560814bfde52c6796fed1f7David Gross} 647415a249300cb9ca3f560814bfde52c6796fed1f7David Gross 6486c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// Launch a simple reduce-style kernel. 64914ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// Inputs: 65014ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// ain: The allocation that contains the input 65114ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// aout: The allocation that will hold the output 65214ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala// mtls: Holds launch parameters 65314ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Walavoid RsdCpuReferenceImpl::launchReduce(const Allocation *ain, 65414ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala Allocation *aout, 65514ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala MTLaunchStructReduce *mtls) { 65614ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala const uint32_t xStart = mtls->start.x; 65714ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala const uint32_t xEnd = mtls->end.x; 65814ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala 65914ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala if (xStart >= xEnd) { 66014ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala return; 66114ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala } 66214ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala 66314ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala const uint32_t startOffset = ain->getType()->getElementSizeBytes() * xStart; 66414ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala mtls->kernel(&mtls->inBuf[startOffset], mtls->outBuf, xEnd - xStart); 66514ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala} 66614ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala 6676c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// Launch a general reduce-style kernel. 6686c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// Inputs: 6696c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// ains[0..inLen-1]: Array of allocations that contain the inputs 6706c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// aout: The allocation that will hold the output 6716c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross// mtls: Holds launch parameters 6726c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Grossvoid RsdCpuReferenceImpl::launchReduceNew(const Allocation ** ains, 6736c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross uint32_t inLen, 6746c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross Allocation * aout, 6756c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross MTLaunchStructReduceNew *mtls) { 6766760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross mtls->logReduce = mRSC->props.mLogReduce; 67710adb0c2029f112b5738228617d5645f6ecea0c5David Gross if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 67810adb0c2029f112b5738228617d5645f6ecea0c5David Gross launchReduceNewParallel(ains, inLen, aout, mtls); 67910adb0c2029f112b5738228617d5645f6ecea0c5David Gross } else { 68010adb0c2029f112b5738228617d5645f6ecea0c5David Gross launchReduceNewSerial(ains, inLen, aout, mtls); 68110adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 68210adb0c2029f112b5738228617d5645f6ecea0c5David Gross} 68310adb0c2029f112b5738228617d5645f6ecea0c5David Gross 68410adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Launch a general reduce-style kernel, single-threaded. 68510adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Inputs: 68610adb0c2029f112b5738228617d5645f6ecea0c5David Gross// ains[0..inLen-1]: Array of allocations that contain the inputs 68710adb0c2029f112b5738228617d5645f6ecea0c5David Gross// aout: The allocation that will hold the output 68810adb0c2029f112b5738228617d5645f6ecea0c5David Gross// mtls: Holds launch parameters 68910adb0c2029f112b5738228617d5645f6ecea0c5David Grossvoid RsdCpuReferenceImpl::launchReduceNewSerial(const Allocation ** ains, 69010adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint32_t inLen, 69110adb0c2029f112b5738228617d5645f6ecea0c5David Gross Allocation * aout, 69210adb0c2029f112b5738228617d5645f6ecea0c5David Gross MTLaunchStructReduceNew *mtls) { 6936760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross REDUCE_NEW_ALOGV(mtls, 1, "launchReduceNewSerial(%p): %u x %u x %u", mtls->accumFunc, 6946760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z); 69510adb0c2029f112b5738228617d5645f6ecea0c5David Gross 6966c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // In the presence of outconverter, we allocate temporary memory for 6976c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // the accumulator. 6986c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // 6996c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // In the absence of outconverter, we use the output allocation as the 7006c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // accumulator. 7016c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross uint8_t *const accumPtr = (mtls->outFunc 7026c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross ? static_cast<uint8_t *>(malloc(mtls->accumSize)) 7036c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross : mtls->redp.outPtr[0]); 7046c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross 7056c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // initialize 7066c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross if (mtls->initFunc) { 7076c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross mtls->initFunc(accumPtr); 7086c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross } else { 7096c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross memset(accumPtr, 0, mtls->accumSize); 7106c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross } 7116c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross 7126c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // accumulate 7136c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross const ReduceNewAccumulatorFunc_t fn = mtls->accumFunc; 7146c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross uint32_t slice = 0; 7156c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross while (SelectOuterSlice(mtls, &mtls->redp, slice++)) { 7166c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross for (mtls->redp.current.y = mtls->start.y; 7176c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross mtls->redp.current.y < mtls->end.y; 7186c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross mtls->redp.current.y++) { 7196c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross RedpPtrSetup(mtls, &mtls->redp, mtls->start.x, mtls->redp.current.y, mtls->redp.current.z); 7206c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross fn(&mtls->redp, mtls->start.x, mtls->end.x, accumPtr); 7216c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross } 7226c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross } 7236c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross 7246c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross // outconvert 7256c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross if (mtls->outFunc) { 7266c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross mtls->outFunc(mtls->redp.outPtr[0], accumPtr); 7276c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross free(accumPtr); 7286c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross } 7296c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross} 7306c1876bbef1b2c89975dce91230a168bd2d2ce4cDavid Gross 73110adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Launch a general reduce-style kernel, multi-threaded. 73210adb0c2029f112b5738228617d5645f6ecea0c5David Gross// Inputs: 73310adb0c2029f112b5738228617d5645f6ecea0c5David Gross// ains[0..inLen-1]: Array of allocations that contain the inputs 73410adb0c2029f112b5738228617d5645f6ecea0c5David Gross// aout: The allocation that will hold the output 73510adb0c2029f112b5738228617d5645f6ecea0c5David Gross// mtls: Holds launch parameters 73610adb0c2029f112b5738228617d5645f6ecea0c5David Grossvoid RsdCpuReferenceImpl::launchReduceNewParallel(const Allocation ** ains, 73710adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint32_t inLen, 73810adb0c2029f112b5738228617d5645f6ecea0c5David Gross Allocation * aout, 73910adb0c2029f112b5738228617d5645f6ecea0c5David Gross MTLaunchStructReduceNew *mtls) { 740415a249300cb9ca3f560814bfde52c6796fed1f7David Gross // For now, we don't know how to go parallel in the absence of a combiner. 741415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (!mtls->combFunc) { 74210adb0c2029f112b5738228617d5645f6ecea0c5David Gross launchReduceNewSerial(ains, inLen, aout, mtls); 74310adb0c2029f112b5738228617d5645f6ecea0c5David Gross return; 74410adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 74510adb0c2029f112b5738228617d5645f6ecea0c5David Gross 74610adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Number of threads = "main thread" + number of other (worker) threads 74710adb0c2029f112b5738228617d5645f6ecea0c5David Gross const uint32_t numThreads = mWorkers.mCount + 1; 74810adb0c2029f112b5738228617d5645f6ecea0c5David Gross 74910adb0c2029f112b5738228617d5645f6ecea0c5David Gross // In the absence of outconverter, we use the output allocation as 75010adb0c2029f112b5738228617d5645f6ecea0c5David Gross // an accumulator, and therefore need to allocate one fewer accumulator. 75110adb0c2029f112b5738228617d5645f6ecea0c5David Gross const uint32_t numAllocAccum = numThreads - (mtls->outFunc == nullptr); 75210adb0c2029f112b5738228617d5645f6ecea0c5David Gross 75310adb0c2029f112b5738228617d5645f6ecea0c5David Gross // If mDebugReduceSplitAccum, then we want each accumulator to start 75410adb0c2029f112b5738228617d5645f6ecea0c5David Gross // on a page boundary. (TODO: Would some unit smaller than a page 75510adb0c2029f112b5738228617d5645f6ecea0c5David Gross // be sufficient to avoid false sharing?) 75610adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (mRSC->props.mDebugReduceSplitAccum) { 75710adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Round up accumulator size to an integral number of pages 75810adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumStride = 75910adb0c2029f112b5738228617d5645f6ecea0c5David Gross (unsigned(mtls->accumSize) + unsigned(mPageSize)-1) & 76010adb0c2029f112b5738228617d5645f6ecea0c5David Gross ~(unsigned(mPageSize)-1); 76110adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Each accumulator gets its own page. Alternatively, if we just 76210adb0c2029f112b5738228617d5645f6ecea0c5David Gross // wanted to make sure no two accumulators are on the same page, 76310adb0c2029f112b5738228617d5645f6ecea0c5David Gross // we could instead do 76410adb0c2029f112b5738228617d5645f6ecea0c5David Gross // allocSize = mtls->accumStride * (numAllocation - 1) + mtls->accumSize 76510adb0c2029f112b5738228617d5645f6ecea0c5David Gross const size_t allocSize = mtls->accumStride * numAllocAccum; 76610adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumAlloc = static_cast<uint8_t *>(memalign(mPageSize, allocSize)); 76710adb0c2029f112b5738228617d5645f6ecea0c5David Gross } else { 76810adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumStride = mtls->accumSize; 76910adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumAlloc = static_cast<uint8_t *>(malloc(mtls->accumStride * numAllocAccum)); 77010adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 77110adb0c2029f112b5738228617d5645f6ecea0c5David Gross 77210adb0c2029f112b5738228617d5645f6ecea0c5David Gross const size_t accumPtrArrayBytes = sizeof(uint8_t *) * numThreads; 77310adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumPtr = static_cast<uint8_t **>(malloc(accumPtrArrayBytes)); 77410adb0c2029f112b5738228617d5645f6ecea0c5David Gross memset(mtls->accumPtr, 0, accumPtrArrayBytes); 77510adb0c2029f112b5738228617d5645f6ecea0c5David Gross 77610adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumCount = 0; 77710adb0c2029f112b5738228617d5645f6ecea0c5David Gross 77810adb0c2029f112b5738228617d5645f6ecea0c5David Gross rsAssert(!mInKernel); 77910adb0c2029f112b5738228617d5645f6ecea0c5David Gross mInKernel = true; 7806760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross REDUCE_NEW_ALOGV(mtls, 1, "launchReduceNewParallel(%p): %u x %u x %u, %u threads, accumAlloc = %p", 7816760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross mtls->accumFunc, 7826760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross mtls->redp.dim.x, mtls->redp.dim.y, mtls->redp.dim.z, 7836760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross numThreads, mtls->accumAlloc); 784415a249300cb9ca3f560814bfde52c6796fed1f7David Gross if (mtls->redp.dim.z > 1) { 785415a249300cb9ca3f560814bfde52c6796fed1f7David Gross mtls->mSliceSize = 1; 786415a249300cb9ca3f560814bfde52c6796fed1f7David Gross launchThreads(walk_3d_reduce_new, mtls); 787415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else if (mtls->redp.dim.y > 1) { 788415a249300cb9ca3f560814bfde52c6796fed1f7David Gross mtls->mSliceSize = rsMax(1U, mtls->redp.dim.y / (numThreads * 4)); 789415a249300cb9ca3f560814bfde52c6796fed1f7David Gross launchThreads(walk_2d_reduce_new, mtls); 790415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } else { 791415a249300cb9ca3f560814bfde52c6796fed1f7David Gross mtls->mSliceSize = rsMax(1U, mtls->redp.dim.x / (numThreads * 4)); 792415a249300cb9ca3f560814bfde52c6796fed1f7David Gross launchThreads(walk_1d_reduce_new, mtls); 793415a249300cb9ca3f560814bfde52c6796fed1f7David Gross } 79410adb0c2029f112b5738228617d5645f6ecea0c5David Gross mInKernel = false; 79510adb0c2029f112b5738228617d5645f6ecea0c5David Gross 79610adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Combine accumulators and identify final accumulator 79710adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint8_t *finalAccumPtr = (mtls->outFunc ? nullptr : mtls->redp.outPtr[0]); 79810adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Loop over accumulators, combining into finalAccumPtr. If finalAccumPtr 79910adb0c2029f112b5738228617d5645f6ecea0c5David Gross // is null, then the first accumulator I find becomes finalAccumPtr. 80010adb0c2029f112b5738228617d5645f6ecea0c5David Gross for (unsigned idx = 0; idx < mtls->accumCount; ++idx) { 80110adb0c2029f112b5738228617d5645f6ecea0c5David Gross uint8_t *const thisAccumPtr = mtls->accumPtr[idx]; 80210adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (finalAccumPtr) { 80310adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (finalAccumPtr != thisAccumPtr) { 80410adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (mtls->combFunc) { 8056760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross if (mtls->logReduce >= 3) { 80610adb0c2029f112b5738228617d5645f6ecea0c5David Gross FormatBuf fmt; 8076760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): accumulating into%s", 80810adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumFunc, 80910adb0c2029f112b5738228617d5645f6ecea0c5David Gross format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 8106760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): accumulator[%d]%s", 81110adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumFunc, idx, 81210adb0c2029f112b5738228617d5645f6ecea0c5David Gross format_bytes(&fmt, thisAccumPtr, mtls->accumSize)); 81310adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 81410adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->combFunc(finalAccumPtr, thisAccumPtr); 81510adb0c2029f112b5738228617d5645f6ecea0c5David Gross } else { 81610adb0c2029f112b5738228617d5645f6ecea0c5David Gross rsAssert(!"expected combiner"); 81710adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 81810adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 81910adb0c2029f112b5738228617d5645f6ecea0c5David Gross } else { 82010adb0c2029f112b5738228617d5645f6ecea0c5David Gross finalAccumPtr = thisAccumPtr; 82110adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 82210adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 82310adb0c2029f112b5738228617d5645f6ecea0c5David Gross rsAssert(finalAccumPtr != nullptr); 8246760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross if (mtls->logReduce >= 3) { 82510adb0c2029f112b5738228617d5645f6ecea0c5David Gross FormatBuf fmt; 8266760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): final accumulator%s", 82710adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumFunc, format_bytes(&fmt, finalAccumPtr, mtls->accumSize)); 82810adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 82910adb0c2029f112b5738228617d5645f6ecea0c5David Gross 83010adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Outconvert 83110adb0c2029f112b5738228617d5645f6ecea0c5David Gross if (mtls->outFunc) { 83210adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->outFunc(mtls->redp.outPtr[0], finalAccumPtr); 8336760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross if (mtls->logReduce >= 3) { 83410adb0c2029f112b5738228617d5645f6ecea0c5David Gross FormatBuf fmt; 8356760f7ba7934ddd51938a8d0206fc41c2a7cb419David Gross REDUCE_NEW_ALOGV(mtls, 3, "launchReduceNewParallel(%p): final outconverted result%s", 83610adb0c2029f112b5738228617d5645f6ecea0c5David Gross mtls->accumFunc, 83710adb0c2029f112b5738228617d5645f6ecea0c5David Gross format_bytes(&fmt, mtls->redp.outPtr[0], mtls->redp.outStride[0])); 83810adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 83910adb0c2029f112b5738228617d5645f6ecea0c5David Gross } 84010adb0c2029f112b5738228617d5645f6ecea0c5David Gross 84110adb0c2029f112b5738228617d5645f6ecea0c5David Gross // Clean up 84210adb0c2029f112b5738228617d5645f6ecea0c5David Gross free(mtls->accumPtr); 84310adb0c2029f112b5738228617d5645f6ecea0c5David Gross free(mtls->accumAlloc); 84410adb0c2029f112b5738228617d5645f6ecea0c5David Gross} 84510adb0c2029f112b5738228617d5645f6ecea0c5David Gross 84610adb0c2029f112b5738228617d5645f6ecea0c5David Gross 84714ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Walavoid RsdCpuReferenceImpl::launchForEach(const Allocation ** ains, 848f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes uint32_t inLen, 849f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes Allocation* aout, 850f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes const RsScriptCall* sc, 85114ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala MTLaunchStructForEach* mtls) { 8524b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 8534b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes //android::StopWatch kernel_time("kernel time"); 8544b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 855bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams bool outerDims = (mtls->start.z != mtls->end.z) || 856bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (mtls->start.face != mtls->end.face) || 857bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (mtls->start.lod != mtls->end.lod) || 858bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (mtls->start.array[0] != mtls->end.array[0]) || 859bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (mtls->start.array[1] != mtls->end.array[1]) || 860bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (mtls->start.array[2] != mtls->end.array[2]) || 861bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (mtls->start.array[3] != mtls->end.array[3]); 862bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams 86310adb0c2029f112b5738228617d5645f6ecea0c5David Gross if ((mWorkers.mCount >= 1) && mtls->isThreadable && !mInKernel) { 8644b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes const size_t targetByteChunk = 16 * 1024; 86510adb0c2029f112b5738228617d5645f6ecea0c5David Gross mInKernel = true; // NOTE: The guard immediately above ensures this was !mInKernel 866f37121300217d3b39ab66dd9c8881bcbcad932dfChris Wailes 867bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams if (outerDims) { 868bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams // No fancy logic for chunk size 869bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams mtls->mSliceSize = 1; 870415a249300cb9ca3f560814bfde52c6796fed1f7David Gross launchThreads(walk_general_foreach, mtls); 871bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams } else if (mtls->fep.dim.y > 1) { 872c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t s1 = mtls->fep.dim.y / ((mWorkers.mCount + 1) * 4); 8734b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes uint32_t s2 = 0; 8744b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 8754b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes // This chooses our slice size to rate limit atomic ops to 8764b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes // one per 16k bytes of reads/writes. 877c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams if ((mtls->aout[0] != nullptr) && mtls->aout[0]->mHal.drvState.lod[0].stride) { 878c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams s2 = targetByteChunk / mtls->aout[0]->mHal.drvState.lod[0].stride; 879a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams } else if (mtls->ains[0]) { 880c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams s2 = targetByteChunk / mtls->ains[0]->mHal.drvState.lod[0].stride; 881a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams } else { 882a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams // Launch option only case 883a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams // Use s1 based only on the dimensions 884a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams s2 = s1; 8854b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 8864b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes mtls->mSliceSize = rsMin(s1, s2); 8874b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 8884b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes if(mtls->mSliceSize < 1) { 8894b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes mtls->mSliceSize = 1; 8904b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 8914b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 892415a249300cb9ca3f560814bfde52c6796fed1f7David Gross launchThreads(walk_2d_foreach, mtls); 8934b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } else { 894c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams uint32_t s1 = mtls->fep.dim.x / ((mWorkers.mCount + 1) * 4); 8954b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes uint32_t s2 = 0; 8964b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 8974b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes // This chooses our slice size to rate limit atomic ops to 8984b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes // one per 16k bytes of reads/writes. 899c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams if ((mtls->aout[0] != nullptr) && mtls->aout[0]->getType()->getElementSizeBytes()) { 900c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams s2 = targetByteChunk / mtls->aout[0]->getType()->getElementSizeBytes(); 901a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams } else if (mtls->ains[0]) { 902c0d68470b978a79ce024fde56f23ea3690603ccdJason Sams s2 = targetByteChunk / mtls->ains[0]->getType()->getElementSizeBytes(); 903a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams } else { 904a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams // Launch option only case 905a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams // Use s1 based only on the dimensions 906a9139c724f8312b3634d213599f2d6b3b2505db2Jason Sams s2 = s1; 9074b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 9084b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes mtls->mSliceSize = rsMin(s1, s2); 9094b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 9104b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes if (mtls->mSliceSize < 1) { 9114b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes mtls->mSliceSize = 1; 9124b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 9134b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 91410adb0c2029f112b5738228617d5645f6ecea0c5David Gross launchThreads(walk_1d_foreach, mtls); 9154b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 91610adb0c2029f112b5738228617d5645f6ecea0c5David Gross mInKernel = false; 9174b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 9184b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } else { 91914ce007a633b10e3b9a3fae29d8f53a7e8c9b59fMatt Wala ForEachFunc_t fn = mtls->kernel; 920bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams uint32_t slice = 0; 92180ef693674f69c0343c41564e30f80e7fb513b60Chris Wailes 9224b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 92359b35f29cc576243a322bf88bc16063a9810da55Jason Sams while(SelectOuterSlice(mtls, &mtls->fep, slice++)) { 924bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams for (mtls->fep.current.y = mtls->start.y; 925bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams mtls->fep.current.y < mtls->end.y; 926bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams mtls->fep.current.y++) { 9274b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 928bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams FepPtrSetup(mtls, &mtls->fep, mtls->start.x, 929bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams mtls->fep.current.y, mtls->fep.current.z, mtls->fep.current.lod, 930bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams (RsAllocationCubemapFace) mtls->fep.current.face, 931bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams mtls->fep.current.array[0], mtls->fep.current.array[1], 932bf2111d3b3de310932099514f06924e48fa1d7b2Jason Sams mtls->fep.current.array[2], mtls->fep.current.array[3]); 9334b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 934b0abb140ac51b93d1a85aadaa63fe057f2d29850David Gross fn(&mtls->fep, mtls->start.x, mtls->end.x, mtls->fep.outStride[0]); 9354b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 9364b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 9374b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes } 9384b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes} 9394b3c34e6833e39bc89c2128002806b654b8e623dChris Wailes 940709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuScriptImpl * RsdCpuReferenceImpl::setTLS(RsdCpuScriptImpl *sc) { 941709a0978ae141198018ca9769f8d96292a8928e6Jason Sams //ALOGE("setTls %p", sc); 942709a0978ae141198018ca9769f8d96292a8928e6Jason Sams ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(gThreadTLSKey); 943709a0978ae141198018ca9769f8d96292a8928e6Jason Sams rsAssert(tls); 944709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *old = tls->mImpl; 945709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mImpl = sc; 946709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mContext = mRSC; 947709a0978ae141198018ca9769f8d96292a8928e6Jason Sams if (sc) { 948709a0978ae141198018ca9769f8d96292a8928e6Jason Sams tls->mScript = sc->getScript(); 949709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } else { 95044bef6fba6244292b751387f3d6c31cca96c28adChris Wailes tls->mScript = nullptr; 951709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 952709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return old; 953709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 954709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 955709a0978ae141198018ca9769f8d96292a8928e6Jason Samsconst RsdCpuReference::CpuSymbol * RsdCpuReferenceImpl::symLookup(const char *name) { 956709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return mSymLookupFn(mRSC, name); 957709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 958709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 959709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 960709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScript * RsdCpuReferenceImpl::createScript(const ScriptC *s, 961709a0978ae141198018ca9769f8d96292a8928e6Jason Sams char const *resName, char const *cacheDir, 962709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint8_t const *bitcode, size_t bitcodeSize, 963709a0978ae141198018ca9769f8d96292a8928e6Jason Sams uint32_t flags) { 964709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 965709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsdCpuScriptImpl *i = new RsdCpuScriptImpl(this, s); 966005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines if (!i->init(resName, cacheDir, bitcode, bitcodeSize, flags 967005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines , getBccPluginName() 968005113297b19ed256b6db9d6bc293ed9266899fcStephen Hines )) { 969709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete i; 97044bef6fba6244292b751387f3d6c31cca96c28adChris Wailes return nullptr; 971709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 972709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return i; 973709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 974709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 9757c4b888f2147edf99690b6af75470774ff31c43bJason Samsextern RsdCpuScriptImpl * rsdIntrinsic_3DLUT(RsdCpuReferenceImpl *ctx, 9767c4b888f2147edf99690b6af75470774ff31c43bJason Sams const Script *s, const Element *e); 977c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Convolve3x3(RsdCpuReferenceImpl *ctx, 978c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 979c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_ColorMatrix(RsdCpuReferenceImpl *ctx, 980c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 981c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_LUT(RsdCpuReferenceImpl *ctx, 982c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 983c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Convolve5x5(RsdCpuReferenceImpl *ctx, 984c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 985c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Blur(RsdCpuReferenceImpl *ctx, 986c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 987c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_YuvToRGB(RsdCpuReferenceImpl *ctx, 988c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 989c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Blend(RsdCpuReferenceImpl *ctx, 990c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams const Script *s, const Element *e); 9912282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Histogram(RsdCpuReferenceImpl *ctx, 9922282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams const Script *s, const Element *e); 99339ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Samsextern RsdCpuScriptImpl * rsdIntrinsic_Resize(RsdCpuReferenceImpl *ctx, 99439ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams const Script *s, const Element *e); 99564c682b65cd04ac83b51251b40dca14423df351aTim Murrayextern RsdCpuScriptImpl * rsdIntrinsic_BLAS(RsdCpuReferenceImpl *ctx, 99664c682b65cd04ac83b51251b40dca14423df351aTim Murray const Script *s, const Element *e); 997709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 998709a0978ae141198018ca9769f8d96292a8928e6Jason SamsRsdCpuReference::CpuScript * RsdCpuReferenceImpl::createIntrinsic(const Script *s, 999709a0978ae141198018ca9769f8d96292a8928e6Jason Sams RsScriptIntrinsicID iid, Element *e) { 1000709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 100144bef6fba6244292b751387f3d6c31cca96c28adChris Wailes RsdCpuScriptImpl *i = nullptr; 1002709a0978ae141198018ca9769f8d96292a8928e6Jason Sams switch (iid) { 10037c4b888f2147edf99690b6af75470774ff31c43bJason Sams case RS_SCRIPT_INTRINSIC_ID_3DLUT: 10047c4b888f2147edf99690b6af75470774ff31c43bJason Sams i = rsdIntrinsic_3DLUT(this, s, e); 10057c4b888f2147edf99690b6af75470774ff31c43bJason Sams break; 1006709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_3x3: 1007c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Convolve3x3(this, s, e); 1008709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 1009709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_COLOR_MATRIX: 1010c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_ColorMatrix(this, s, e); 1011709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 1012709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_LUT: 1013c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_LUT(this, s, e); 1014709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 1015709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_CONVOLVE_5x5: 1016c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Convolve5x5(this, s, e); 1017709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 1018709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_BLUR: 1019c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Blur(this, s, e); 1020709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 1021709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_YUV_TO_RGB: 1022c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_YuvToRGB(this, s, e); 1023709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 1024709a0978ae141198018ca9769f8d96292a8928e6Jason Sams case RS_SCRIPT_INTRINSIC_ID_BLEND: 1025c905efd76fdcc1b8846b229bf7d991d185a7b4b7Jason Sams i = rsdIntrinsic_Blend(this, s, e); 1026709a0978ae141198018ca9769f8d96292a8928e6Jason Sams break; 10272282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams case RS_SCRIPT_INTRINSIC_ID_HISTOGRAM: 10282282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams i = rsdIntrinsic_Histogram(this, s, e); 10292282e2816ac5f5de53f9bd4f3ecbdfd6d756d120Jason Sams break; 103039ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams case RS_SCRIPT_INTRINSIC_ID_RESIZE: 103139ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams i = rsdIntrinsic_Resize(this, s, e); 103239ab94aafb7f0916a7f6e345ee1fa0f5ff3bbacdJason Sams break; 103364c682b65cd04ac83b51251b40dca14423df351aTim Murray case RS_SCRIPT_INTRINSIC_ID_BLAS: 103464c682b65cd04ac83b51251b40dca14423df351aTim Murray i = rsdIntrinsic_BLAS(this, s, e); 103564c682b65cd04ac83b51251b40dca14423df351aTim Murray break; 1036709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1037709a0978ae141198018ca9769f8d96292a8928e6Jason Sams default: 1038709a0978ae141198018ca9769f8d96292a8928e6Jason Sams rsAssert(0); 1039709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 1040709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 1041709a0978ae141198018ca9769f8d96292a8928e6Jason Sams return i; 1042709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 1043709a0978ae141198018ca9769f8d96292a8928e6Jason Sams 10441ffd86b448d78366190c540f98f8b6d641cdb6cfYang Nivoid* RsdCpuReferenceImpl::createScriptGroup(const ScriptGroupBase *sg) { 10451ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni switch (sg->getApiVersion()) { 10461ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni case ScriptGroupBase::SG_V1: { 10471ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni CpuScriptGroupImpl *sgi = new CpuScriptGroupImpl(this, sg); 10481ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni if (!sgi->init()) { 1049709a0978ae141198018ca9769f8d96292a8928e6Jason Sams delete sgi; 105044bef6fba6244292b751387f3d6c31cca96c28adChris Wailes return nullptr; 10511ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni } 10521ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni return sgi; 1053709a0978ae141198018ca9769f8d96292a8928e6Jason Sams } 10541ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni case ScriptGroupBase::SG_V2: { 10551ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni return new CpuScriptGroup2Impl(this, sg); 10561ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni } 10571ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni } 10581ffd86b448d78366190c540f98f8b6d641cdb6cfYang Ni return nullptr; 1059709a0978ae141198018ca9769f8d96292a8928e6Jason Sams} 1060