AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

P4 to Git Change 1293980 by jatang@jatang-opencl-hsa-stg3 on 2016/07/21 10:22:49

Browse Source 
SWDEV-95919 - OCL ThreadTrace on PAL.

	The PAL PerfExperiment interface is not compatible with the  OCL ThreadTrace extension, in terms of Thread Trace buffer handling. PAL PerfExperiment is expecting one buffer bound to it for all shader engines, while the OCL ThreadTrace extension is expecting one buffer for each shader engines. To accomodate the difference, we copy from the PAL PerfExperiment buffer to the  OCL ThreadTrace extension buffers.

	The next step is to support setting ThreadTrace parameters, and getting actually capture size from the layout.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palthreadtrace.cpp#2 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palthreadtrace.hpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#18 edit
This commit is contained in:
foreman
2016-07-21 10:32:54 -04:00
parent eb2b6a8f39
commit 892bda4554
3 changed files with 210 additions and 149 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/runtime/device/pal/palvirtual.cpp
+22 -64
View File
@@ -2399,34 +2399,29 @@ VirtualGPU::submitThreadTraceMemObjects(amd::ThreadTraceMemObjectsCommand& cmd)
amd::ThreadTrace* amdThreadTrace = &cmd.getThreadTrace();
ThreadTrace* threadTrace =
static_cast<ThreadTrace*>(amdThreadTrace->getDeviceThreadTrace());
Unimplemented();
/*
if (threadTrace == nullptr) {
gslQueryObject gslThreadTrace;
// Create a HW thread trace query object
gslThreadTrace = cs()->createQuery(GSL_SHADER_TRACE_BYTES_WRITTEN);
if (0 == gslThreadTrace) {
LogError("Failure in memory allocation for the GPU threadtrace");
cmd.setStatus(CL_INVALID_OPERATION);
return;
}
CalThreadTraceReference* palRef = new CalThreadTraceReference(*this,gslThreadTrace);
PalThreadTraceReference* palRef = PalThreadTraceReference::Create(*this);
if (palRef == nullptr) {
LogError("Failure in memory allocation for the GPU threadtrace");
cmd.setStatus(CL_INVALID_OPERATION);
return;
}
size_t seNum = amdThreadTrace->deviceSeNumThreadTrace();
size_t numSe = amdThreadTrace->deviceSeNumThreadTrace();
ThreadTrace* gpuThreadTrace = new ThreadTrace(
gpuDevice_,
*this,
seNum);
palRef,
cmd.getMemList(),
numSe);
if (nullptr == gpuThreadTrace) {
LogError("Failure in memory allocation for the GPU threadtrace");
cmd.setStatus(CL_INVALID_OPERATION);
return;
}
if (gpuThreadTrace->create(palRef)) {
if (gpuThreadTrace->create()) {
amdThreadTrace->setDeviceThreadTrace(gpuThreadTrace);
}
else {
@@ -2435,21 +2430,11 @@ VirtualGPU::submitThreadTraceMemObjects(amd::ThreadTraceMemObjectsCommand& cmd)
cmd.setStatus(CL_INVALID_OPERATION);
return;
}
threadTrace = gpuThreadTrace;
palRef->finalize();
palRef->release();
}
gslShaderTraceBufferObject* threadTraceBufferObjects = threadTrace->getThreadTraceBufferObjects();
const size_t memObjSize = cmd.getMemoryObjectSize();
const std::vector<amd::Memory*>& memObj = cmd.getMemList();
size_t se = 0;
for (std::vector<amd::Memory*>::const_iterator itMemObj = memObj.begin();itMemObj != memObj.end();++itMemObj,++se) {
// Find GSL Mem Object
Pal::IGpuMemory* gslMemObj = dev().getGpuMemory(*itMemObj)->iMem();
// Bind GSL MemObject to the appropriate SE Thread Trace Buffer Object
threadTraceBufferObjects[se]->attachMemObject(cs(), gslMemObj, 0, 0, memObjSize, se);
}
*/
break;
}
default:
@@ -2480,53 +2465,26 @@ VirtualGPU::submitThreadTrace(amd::ThreadTraceCommand& cmd)
return;
}
else {
Unimplemented();
/*
gslQueryObject gslThreadTrace;
gslThreadTrace = threadTrace->gslThreadTrace();
uint32_t seNum = amdThreadTrace->deviceSeNumThreadTrace();
// Find the state and sends the commands to GSL
Pal::IPerfExperiment* palPerf = threadTrace->iPerf();
if (cmd.getState() == amd::ThreadTraceCommand::Begin) {
amd::ThreadTrace::ThreadTraceConfig* traceCfg =
static_cast<amd::ThreadTrace::ThreadTraceConfig*>(cmd.threadTraceConfig());
const gslErrorCode ec = gslThreadTrace->BeginQuery(cs(),
GSL_SHADER_TRACE_BYTES_WRITTEN, 0);
assert(ec == GSL_NO_ERROR);
for (uint32_t idx = 0; idx < seNum; ++idx) {
rs()->enableShaderTrace(cs(), idx, true);
rs()->setShaderTraceComputeUnit (idx, traceCfg->cu_);
rs()->setShaderTraceShaderArray (idx, traceCfg->sh_);
rs()->setShaderTraceSIMDMask (idx, traceCfg->simdMask_);
rs()->setShaderTraceVmIdMask (idx, traceCfg->vmIdMask_);
rs()->setShaderTraceTokenMask (idx, traceCfg->tokenMask_);
rs()->setShaderTraceRegisterMask(idx, traceCfg->regMask_);
rs()->setShaderTraceIssueMask (idx, traceCfg->instMask_);
rs()->setShaderTraceRandomSeed (idx, traceCfg->randomSeed_);
rs()->setShaderTraceCaptureMode (idx, traceCfg->captureMode_);
rs()->setShaderTraceWrap (idx, traceCfg->isWrapped_);
rs()->setShaderTraceUserData (idx,
(traceCfg->isUserData_) ? traceCfg->userData_ : 0);
}
iCmd()->CmdBeginPerfExperiment(palPerf);
}
else if (cmd.getState() == amd::ThreadTraceCommand::End) {
for (uint32_t idx = 0; idx < seNum; ++idx) {
rs()->enableShaderTrace(cs(), idx, false);
}
gslThreadTrace->EndQuery(cs(), 0);
GpuEvent event;
eventBegin(MainEngine);
iCmd()->CmdEndPerfExperiment(palPerf);
threadTrace->populateUserMemory();
eventEnd(MainEngine, event);
setGpuEvent(event);
}
else if (cmd.getState() == amd::ThreadTraceCommand::Pause) {
for (uint32_t idx = 0; idx < seNum; ++idx) {
rs()->setShaderTraceIsPaused(cs(), idx, true);
}
// There's no Pause from the PerfExperiment interface
}
else if (cmd.getState() == amd::ThreadTraceCommand::Resume) {
for (uint32_t idx = 0; idx < seNum; ++idx) {
rs()->setShaderTraceIsPaused(cs(), idx, false);
}
// There's no Resume from the PerfExperiment interface
}
*/
}
break;
}