Merge "Migrate thread launch to driver." (37e96a15) · Commits · e / os / android_frameworks_base

libs/rs/driver/rsdBcc.cpp

+217 −4

Original line number	Diff line number	Diff line
		@@ -55,6 +55,15 @@ struct DrvScript {

		};

		static Script * setTLS(Script *sc) {
		ScriptTLSStruct * tls = (ScriptTLSStruct *)pthread_getspecific(Context::gThreadTLSKey);
		rsAssert(tls);
		Script *old = tls->mScript;
		tls->mScript = sc;
		return old;
		}


		// Input: cacheDir
		// Input: resName
		// Input: extName
		@@ -234,13 +243,215 @@ error:

		}

		typedef struct {
		Context *rsc;
		Script *script;
		const Allocation * ain;
		Allocation * aout;
		const void * usr;

		uint32_t mSliceSize;
		volatile int mSliceNum;

		const uint8_t *ptrIn;
		uint32_t eStrideIn;
		uint8_t *ptrOut;
		uint32_t eStrideOut;

		uint32_t xStart;
		uint32_t xEnd;
		uint32_t yStart;
		uint32_t yEnd;
		uint32_t zStart;
		uint32_t zEnd;
		uint32_t arrayStart;
		uint32_t arrayEnd;

		uint32_t dimX;
		uint32_t dimY;
		uint32_t dimZ;
		uint32_t dimArray;
		} MTLaunchStruct;
		typedef int (rs_t)(const void , void , const void , uint32_t, uint32_t, uint32_t, uint32_t);

		static void wc_xy(void *usr, uint32_t idx) {
		MTLaunchStruct mtls = (MTLaunchStruct )usr;

		while (1) {
		uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum);
		uint32_t yStart = mtls->yStart + slice * mtls->mSliceSize;
		uint32_t yEnd = yStart + mtls->mSliceSize;
		yEnd = rsMin(yEnd, mtls->yEnd);
		if (yEnd <= yStart) {
		return;
		}

		//LOGE("usr idx %i, x %i,%i y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
		//LOGE("usr ptr in %p, out %p", mtls->ptrIn, mtls->ptrOut);
		for (uint32_t y = yStart; y < yEnd; y++) {
		uint32_t offset = mtls->dimX * y;
		uint8_t xPtrOut = mtls->ptrOut + (mtls->eStrideOut offset);
		const uint8_t xPtrIn = mtls->ptrIn + (mtls->eStrideIn offset);

		for (uint32_t x = mtls->xStart; x < mtls->xEnd; x++) {
		((rs_t)mtls->script->mHal.info.root) (xPtrIn, xPtrOut, mtls->usr, x, y, 0, 0);
		xPtrIn += mtls->eStrideIn;
		xPtrOut += mtls->eStrideOut;
		}
		}
		}
		}

		static void wc_x(void *usr, uint32_t idx) {
		MTLaunchStruct mtls = (MTLaunchStruct )usr;

		while (1) {
		uint32_t slice = (uint32_t)android_atomic_inc(&mtls->mSliceNum);
		uint32_t xStart = mtls->xStart + slice * mtls->mSliceSize;
		uint32_t xEnd = xStart + mtls->mSliceSize;
		xEnd = rsMin(xEnd, mtls->xEnd);
		if (xEnd <= xStart) {
		return;
		}

		//LOGE("usr idx %i, x %i,%i y %i,%i", idx, mtls->xStart, mtls->xEnd, yStart, yEnd);
		//LOGE("usr ptr in %p, out %p", mtls->ptrIn, mtls->ptrOut);
		uint8_t xPtrOut = mtls->ptrOut + (mtls->eStrideOut xStart);
		const uint8_t xPtrIn = mtls->ptrIn + (mtls->eStrideIn xStart);
		for (uint32_t x = xStart; x < xEnd; x++) {
		((rs_t)mtls->script->mHal.info.root) (xPtrIn, xPtrOut, mtls->usr, x, 0, 0, 0);
		xPtrIn += mtls->eStrideIn;
		xPtrOut += mtls->eStrideOut;
		}
		}
		}

		void rsdScriptInvokeForEach(const Context *rsc,
		Script *s,
		const Allocation * ain,
		Allocation * aout,
		const void * usr,
		uint32_t usrLen,
		const RsScriptCall *sc) {

		RsHal * dc = (RsHal *)rsc->mHal.drv;

		MTLaunchStruct mtls;
		memset(&mtls, 0, sizeof(mtls));

		if (ain) {
		mtls.dimX = ain->getType()->getDimX();
		mtls.dimY = ain->getType()->getDimY();
		mtls.dimZ = ain->getType()->getDimZ();
		//mtls.dimArray = ain->getType()->getDimArray();
		} else if (aout) {
		mtls.dimX = aout->getType()->getDimX();
		mtls.dimY = aout->getType()->getDimY();
		mtls.dimZ = aout->getType()->getDimZ();
		//mtls.dimArray = aout->getType()->getDimArray();
		} else {
		rsc->setError(RS_ERROR_BAD_SCRIPT, "rsForEach called with null allocations");
		return;
		}

		if (!sc \|\| (sc->xEnd == 0)) {
		mtls.xEnd = mtls.dimX;
		} else {
		rsAssert(sc->xStart < mtls.dimX);
		rsAssert(sc->xEnd <= mtls.dimX);
		rsAssert(sc->xStart < sc->xEnd);
		mtls.xStart = rsMin(mtls.dimX, sc->xStart);
		mtls.xEnd = rsMin(mtls.dimX, sc->xEnd);
		if (mtls.xStart >= mtls.xEnd) return;
		}

		if (!sc \|\| (sc->yEnd == 0)) {
		mtls.yEnd = mtls.dimY;
		} else {
		rsAssert(sc->yStart < mtls.dimY);
		rsAssert(sc->yEnd <= mtls.dimY);
		rsAssert(sc->yStart < sc->yEnd);
		mtls.yStart = rsMin(mtls.dimY, sc->yStart);
		mtls.yEnd = rsMin(mtls.dimY, sc->yEnd);
		if (mtls.yStart >= mtls.yEnd) return;
		}

		mtls.xEnd = rsMax((uint32_t)1, mtls.xEnd);
		mtls.yEnd = rsMax((uint32_t)1, mtls.yEnd);
		mtls.zEnd = rsMax((uint32_t)1, mtls.zEnd);
		mtls.arrayEnd = rsMax((uint32_t)1, mtls.arrayEnd);

		rsAssert(ain->getType()->getDimZ() == 0);

		Context mrsc = (Context )rsc;
		Script * oldTLS = setTLS(s);

		mtls.rsc = mrsc;
		mtls.ain = ain;
		mtls.aout = aout;
		mtls.script = s;
		mtls.usr = usr;
		mtls.mSliceSize = 10;
		mtls.mSliceNum = 0;

		mtls.ptrIn = NULL;
		mtls.eStrideIn = 0;
		if (ain) {
		mtls.ptrIn = (const uint8_t *)ain->getPtr();
		mtls.eStrideIn = ain->getType()->getElementSizeBytes();
		}

		mtls.ptrOut = NULL;
		mtls.eStrideOut = 0;
		if (aout) {
		mtls.ptrOut = (uint8_t *)aout->getPtr();
		mtls.eStrideOut = aout->getType()->getElementSizeBytes();
		}

		if ((dc->mWorkers.mCount > 1) && s->mHal.info.isThreadable) {
		if (mtls.dimY > 1) {
		rsdLaunchThreads(mrsc, wc_xy, &mtls);
		} else {
		rsdLaunchThreads(mrsc, wc_x, &mtls);
		}

		//LOGE("launch 1");
		} else {
		//LOGE("launch 3");
		for (uint32_t ar = mtls.arrayStart; ar < mtls.arrayEnd; ar++) {
		for (uint32_t z = mtls.zStart; z < mtls.zEnd; z++) {
		for (uint32_t y = mtls.yStart; y < mtls.yEnd; y++) {
		uint32_t offset = mtls.dimX * mtls.dimY * mtls.dimZ * ar +
		mtls.dimX * mtls.dimY * z +
		mtls.dimX * y;
		uint8_t xPtrOut = mtls.ptrOut + (mtls.eStrideOut offset);
		const uint8_t xPtrIn = mtls.ptrIn + (mtls.eStrideIn offset);

		for (uint32_t x = mtls.xStart; x < mtls.xEnd; x++) {
		((rs_t)s->mHal.info.root) (xPtrIn, xPtrOut, usr, x, y, z, ar);
		xPtrIn += mtls.eStrideIn;
		xPtrOut += mtls.eStrideOut;
		}
		}
		}
		}
		}

		setTLS(oldTLS);
		}


		int rsdScriptInvokeRoot(const Context dc, const Script script) {
		int rsdScriptInvokeRoot(const Context dc, Script script) {
		DrvScript drv = (DrvScript )script->mHal.drv;
		return drv->mRoot();

		Script * oldTLS = setTLS(script);
		int ret = drv->mRoot();
		setTLS(oldTLS);

		return ret;
		}

		void rsdScriptInvokeInit(const Context dc, const Script script) {
		void rsdScriptInvokeInit(const Context dc, Script script) {
		DrvScript drv = (DrvScript )script->mHal.drv;

		if (drv->mInit) {
		@@ -249,15 +460,17 @@ void rsdScriptInvokeInit(const Context dc, const Script script) {
		}


		void rsdScriptInvokeFunction(const Context dc, const Script script,
		void rsdScriptInvokeFunction(const Context dc, Script script,
		uint32_t slot,
		const void *params,
		size_t paramLength) {
		DrvScript drv = (DrvScript )script->mHal.drv;
		//LOGE("invoke %p %p %i %p %i", dc, script, slot, params, paramLength);

		Script * oldTLS = setTLS(script);
		((void ()(const void , uint32_t))
		drv->mInvokeFunctions[slot])(params, paramLength);
		setTLS(oldTLS);
		}

		void rsdScriptSetGlobalVar(const Context dc, const Script script,

libs/rs/driver/rsdBcc.h

+12 −3

Original line number	Diff line number	Diff line
		@@ -25,14 +25,23 @@ bool rsdScriptInit(const android::renderscript::Context *, android::renderscript
		uint8_t const *bitcode, size_t bitcodeSize,
		uint32_t flags, android::renderscript::RsHalSymbolLookupFunc lookupFunc);
		void rsdScriptInvokeFunction(const android::renderscript::Context *dc,
		const android::renderscript::Script *script,
		android::renderscript::Script *script,
		uint32_t slot,
		const void *params,
		size_t paramLength);

		void rsdScriptInvokeForEach(const android::renderscript::Context *rsc,
		android::renderscript::Script *s,
		const android::renderscript::Allocation * ain,
		android::renderscript::Allocation * aout,
		const void * usr,
		uint32_t usrLen,
		const RsScriptCall *sc);

		int rsdScriptInvokeRoot(const android::renderscript::Context *dc,
		const android::renderscript::Script *script);
		android::renderscript::Script *script);
		void rsdScriptInvokeInit(const android::renderscript::Context *dc,
		const android::renderscript::Script *script);
		android::renderscript::Script *script);

		void rsdScriptSetGlobalVar(const android::renderscript::Context *,
		const android::renderscript::Script *,

libs/rs/driver/rsdCore.cpp

+135 −4

Original line number	Diff line number	Diff line
		@@ -20,16 +20,29 @@
		#include <malloc.h>
		#include "rsContext.h"

		#include <sys/types.h>
		#include <sys/resource.h>
		#include <sched.h>
		#include <cutils/properties.h>
		#include <cutils/sched_policy.h>
		#include <sys/syscall.h>
		#include <string.h>

		using namespace android;
		using namespace android::renderscript;

		static void Shutdown(Context *rsc);
		static void SetPriority(const Context *rsc, int32_t priority);

		static RsdHalFunctions FunctionTable = {
		Shutdown,
		NULL,
		NULL,
		SetPriority,
		{
		rsdScriptInit,
		rsdScriptInvokeFunction,
		rsdScriptInvokeRoot,
		rsdScriptInvokeForEach,
		rsdScriptInvokeInit,
		rsdScriptSetGlobalVar,
		rsdScriptSetGlobalBind,
		@@ -39,16 +52,134 @@ static RsdHalFunctions FunctionTable = {
		};



		static void * HelperThreadProc(void *vrsc) {
		Context rsc = static_cast<Context >(vrsc);
		RsHal dc = (RsHal )rsc->mHal.drv;


		uint32_t idx = (uint32_t)android_atomic_inc(&dc->mWorkers.mLaunchCount);

		//LOGV("RS helperThread starting %p idx=%i", rsc, idx);

		dc->mWorkers.mLaunchSignals[idx].init();
		dc->mWorkers.mNativeThreadId[idx] = gettid();

		#if 0
		typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
		cpu_set_t cpuset;
		memset(&cpuset, 0, sizeof(cpuset));
		cpuset.bits[idx / 64] \|= 1ULL << (idx % 64);
		int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
		sizeof(cpuset), &cpuset);
		LOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
		#endif

		int status = pthread_setspecific(rsc->gThreadTLSKey, rsc->mTlsStruct);
		if (status) {
		LOGE("pthread_setspecific %i", status);
		}

		while (!dc->mExit) {
		dc->mWorkers.mLaunchSignals[idx].wait();
		if (dc->mWorkers.mLaunchCallback) {
		dc->mWorkers.mLaunchCallback(dc->mWorkers.mLaunchData, idx);
		}
		android_atomic_dec(&dc->mWorkers.mRunningCount);
		dc->mWorkers.mCompleteSignal.set();
		}

		//LOGV("RS helperThread exited %p idx=%i", rsc, idx);
		return NULL;
		}

		void rsdLaunchThreads(Context rsc, WorkerCallback_t cbk, void data) {
		RsHal dc = (RsHal )rsc->mHal.drv;

		dc->mWorkers.mLaunchData = data;
		dc->mWorkers.mLaunchCallback = cbk;
		android_atomic_release_store(dc->mWorkers.mCount, &dc->mWorkers.mRunningCount);
		for (uint32_t ct = 0; ct < dc->mWorkers.mCount; ct++) {
		dc->mWorkers.mLaunchSignals[ct].set();
		}
		while (android_atomic_acquire_load(&dc->mWorkers.mRunningCount) != 0) {
		dc->mWorkers.mCompleteSignal.wait();
		}
		}

		bool rsdHalInit(Context *rsc, uint32_t version_major, uint32_t version_minor) {
		rsc->mHal.funcs = FunctionTable;

		/*
		rsc->mHal.drv = (RsHal *)calloc(1, sizeof(RsHal));
		RsHal dc = (RsHal )calloc(1, sizeof(RsHal));
		if (!rsc->mHal.drv) {
		return false;
		}
		*/
		rsc->mHal.drv = dc;


		int cpu = sysconf(_SC_NPROCESSORS_ONLN);
		LOGV("RS Launching thread(s), reported CPU count %i", cpu);
		if (cpu < 2) cpu = 0;

		dc->mWorkers.mCount = (uint32_t)cpu;
		dc->mWorkers.mThreadId = (pthread_t *) calloc(dc->mWorkers.mCount, sizeof(pthread_t));
		dc->mWorkers.mNativeThreadId = (pid_t *) calloc(dc->mWorkers.mCount, sizeof(pid_t));
		dc->mWorkers.mLaunchSignals = new Signal[dc->mWorkers.mCount];
		dc->mWorkers.mLaunchCallback = NULL;

		dc->mWorkers.mCompleteSignal.init();

		android_atomic_release_store(dc->mWorkers.mCount, &dc->mWorkers.mRunningCount);
		android_atomic_release_store(0, &dc->mWorkers.mLaunchCount);

		int status;
		pthread_attr_t threadAttr;
		status = pthread_attr_init(&threadAttr);
		if (status) {
		LOGE("Failed to init thread attribute.");
		return false;
		}

		for (uint32_t ct=0; ct < dc->mWorkers.mCount; ct++) {
		status = pthread_create(&dc->mWorkers.mThreadId[ct], &threadAttr, HelperThreadProc, rsc);
		if (status) {
		dc->mWorkers.mCount = ct;
		LOGE("Created fewer than expected number of RS threads.");
		break;
		}
		}
		while (android_atomic_acquire_load(&dc->mWorkers.mRunningCount) != 0) {
		usleep(100);
		}

		pthread_attr_destroy(&threadAttr);
		return true;
		}


		void SetPriority(const Context *rsc, int32_t priority) {
		RsHal dc = (RsHal )rsc->mHal.drv;
		for (uint32_t ct=0; ct < dc->mWorkers.mCount; ct++) {
		setpriority(PRIO_PROCESS, dc->mWorkers.mNativeThreadId[ct], priority);
		}
		}

		void Shutdown(Context *rsc) {
		RsHal dc = (RsHal )rsc->mHal.drv;

		dc->mExit = true;
		dc->mWorkers.mLaunchData = NULL;
		dc->mWorkers.mLaunchCallback = NULL;
		android_atomic_release_store(dc->mWorkers.mCount, &dc->mWorkers.mRunningCount);
		for (uint32_t ct = 0; ct < dc->mWorkers.mCount; ct++) {
		dc->mWorkers.mLaunchSignals[ct].set();
		}
		int status;
		void *res;
		for (uint32_t ct = 0; ct < dc->mWorkers.mCount; ct++) {
		status = pthread_join(dc->mWorkers.mThreadId[ct], &res);
		}
		rsAssert(android_atomic_acquire_load(&dc->mWorkers.mRunningCount) == 0);
		}

libs/rs/driver/rsdCore.h

+21 −13

Original line number	Diff line number	Diff line
		@@ -20,28 +20,36 @@
		#include <rs_hal.h>
		#include <bcc/bcc.h>

		#include "rsMutex.h"
		#include "rsSignal.h"


		typedef void (* InvokeFunc_t)(void);
		typedef void (WorkerCallback_t)(void usr, uint32_t idx);

		struct RsHalRec {
		typedef struct RsHalRec {
		uint32_t version_major;
		uint32_t version_minor;
		};

		struct RsHalProgramStoreRec {
		};

		struct RsHalProgramRasterRec {
		};

		struct RsHalProgramVertexRec {
		};

		struct RsHalProgramFragmentRec {

		struct Workers {
		volatile int mRunningCount;
		volatile int mLaunchCount;
		uint32_t mCount;
		pthread_t *mThreadId;
		pid_t *mNativeThreadId;
		android::renderscript::Signal mCompleteSignal;

		android::renderscript::Signal *mLaunchSignals;
		WorkerCallback_t mLaunchCallback;
		void *mLaunchData;
		};
		Workers mWorkers;
		bool mExit;
		} RsHal;



		void rsdLaunchThreads(android::renderscript::Context rsc, WorkerCallback_t cbk, void data);

		#endif

libs/rs/rsContext.cpp

+4 −86

Original line number	Diff line number	Diff line
		@@ -554,56 +554,6 @@ void Context::destroyWorkerThreadResources() {
		mExit = true;
		}

		void * Context::helperThreadProc(void *vrsc) {
		Context rsc = static_cast<Context >(vrsc);
		uint32_t idx = (uint32_t)android_atomic_inc(&rsc->mWorkers.mLaunchCount);

		//LOGV("RS helperThread starting %p idx=%i", rsc, idx);

		rsc->mWorkers.mLaunchSignals[idx].init();
		rsc->mWorkers.mNativeThreadId[idx] = gettid();

		#if 0
		typedef struct {uint64_t bits[1024 / 64]; } cpu_set_t;
		cpu_set_t cpuset;
		memset(&cpuset, 0, sizeof(cpuset));
		cpuset.bits[idx / 64] \|= 1ULL << (idx % 64);
		int ret = syscall(241, rsc->mWorkers.mNativeThreadId[idx],
		sizeof(cpuset), &cpuset);
		LOGE("SETAFFINITY ret = %i %s", ret, EGLUtils::strerror(ret));
		#endif

		setpriority(PRIO_PROCESS, rsc->mWorkers.mNativeThreadId[idx], rsc->mThreadPriority);
		int status = pthread_setspecific(rsc->gThreadTLSKey, rsc->mTlsStruct);
		if (status) {
		LOGE("pthread_setspecific %i", status);
		}

		while (!rsc->mExit) {
		rsc->mWorkers.mLaunchSignals[idx].wait();
		if (rsc->mWorkers.mLaunchCallback) {
		rsc->mWorkers.mLaunchCallback(rsc->mWorkers.mLaunchData, idx);
		}
		android_atomic_dec(&rsc->mWorkers.mRunningCount);
		rsc->mWorkers.mCompleteSignal.set();
		}

		//LOGV("RS helperThread exited %p idx=%i", rsc, idx);
		return NULL;
		}

		void Context::launchThreads(WorkerCallback_t cbk, void *data) {
		mWorkers.mLaunchData = data;
		mWorkers.mLaunchCallback = cbk;
		android_atomic_release_store(mWorkers.mCount, &mWorkers.mRunningCount);
		for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
		mWorkers.mLaunchSignals[ct].set();
		}
		while (android_atomic_acquire_load(&mWorkers.mRunningCount) != 0) {
		mWorkers.mCompleteSignal.wait();
		}
		}

		void Context::setPriority(int32_t p) {
		// Note: If we put this in the proper "background" policy
		// the wallpapers can become completly unresponsive at times.
		@@ -620,9 +570,6 @@ void Context::setPriority(int32_t p) {
		}
		#else
		setpriority(PRIO_PROCESS, mNativeThreadId, p);
		for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
		setpriority(PRIO_PROCESS, mWorkers.mNativeThreadId[ct], p);
		}
		#endif
		}

		@@ -691,16 +638,8 @@ bool Context::initContext(Device dev, const RsSurfaceConfig sc) {
		if (!rsdHalInit(this, 0, 0)) {
		return false;
		}
		mHal.funcs.setPriority(this, mThreadPriority);

		int cpu = sysconf(_SC_NPROCESSORS_ONLN);
		LOGV("RS Launching thread(s), reported CPU count %i", cpu);
		if (cpu < 2) cpu = 0;

		mWorkers.mCount = (uint32_t)cpu;
		mWorkers.mThreadId = (pthread_t *) calloc(mWorkers.mCount, sizeof(pthread_t));
		mWorkers.mNativeThreadId = (pid_t *) calloc(mWorkers.mCount, sizeof(pid_t));
		mWorkers.mLaunchSignals = new Signal[mWorkers.mCount];
		mWorkers.mLaunchCallback = NULL;
		status = pthread_create(&mThreadId, &threadAttr, threadProc, this);
		if (status) {
		LOGE("Failed to start rs context thread.");
		@@ -714,20 +653,6 @@ bool Context::initContext(Device dev, const RsSurfaceConfig sc) {
		return false;
		}

		mWorkers.mCompleteSignal.init();
		android_atomic_release_store(mWorkers.mCount, &mWorkers.mRunningCount);
		android_atomic_release_store(0, &mWorkers.mLaunchCount);
		for (uint32_t ct=0; ct < mWorkers.mCount; ct++) {
		status = pthread_create(&mWorkers.mThreadId[ct], &threadAttr, helperThreadProc, this);
		if (status) {
		mWorkers.mCount = ct;
		LOGE("Created fewer than expected number of RS threads.");
		break;
		}
		}
		while (android_atomic_acquire_load(&mWorkers.mRunningCount) != 0) {
		usleep(100);
		}
		pthread_attr_destroy(&threadAttr);
		return true;
		}
		@@ -744,17 +669,10 @@ Context::~Context() {
		mIO.shutdown();
		int status = pthread_join(mThreadId, &res);

		// Cleanup compute threads.
		mWorkers.mLaunchData = NULL;
		mWorkers.mLaunchCallback = NULL;
		android_atomic_release_store(mWorkers.mCount, &mWorkers.mRunningCount);
		for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
		mWorkers.mLaunchSignals[ct].set();
		}
		for (uint32_t ct = 0; ct < mWorkers.mCount; ct++) {
		status = pthread_join(mWorkers.mThreadId[ct], &res);

		if (mHal.funcs.shutdownDriver) {
		mHal.funcs.shutdownDriver(this);
		}
		rsAssert(android_atomic_acquire_load(&mWorkers.mRunningCount) == 0);

		// Global structure cleanup.
		pthread_mutex_lock(&gInitMutex);