Profiling of OpenCL apps

[Raistmer]

@Morten
what stderr says ?

[Morten]

Stderr has not been updated during these tests - last time is days ago.
This is what .csv says:

# ProfilerVersion=2.4.1297                                                                                       
# Application=D:/ProgramData/BOINC/projects/setiathome.berkeley.edu/MB_6.10_win_x86_SSE3_OpenCL_ATi_r331.exe                                                                                       
# ApplicationArgs=                                                                                       
# Device Cayman PlatformVendor=Advanced Micro Devices    Inc.                                                                                    
# Device Cayman PlatformName=AMD Accelerated Parallel Processing                                                                                       
# Device Cayman PlatformVersion=OpenCL 1.1 AMD-APP (844.4)                                                                                       
# Device Cayman CLDriverVersion=CAL 1.4.1658 (VM)                                                                                       
# Device Cayman CLRuntimeVersion=OpenCL 1.1 AMD-APP (844.4)                                                                                       
# Device Cayman NumberAppAddressBits=32                                                                                       
# Device Intel(R) Core(TM) i7 CPU         870  @ 2.93GHz PlatformVendor=Advanced Micro Devices    Inc.                                                                                    
# Device Intel(R) Core(TM) i7 CPU         870  @ 2.93GHz PlatformName=AMD Accelerated Parallel Processing                                                                                       
# Device Intel(R) Core(TM) i7 CPU         870  @ 2.93GHz PlatformVersion=OpenCL 1.1 AMD-APP (844.4)                                                                                       
# Device Intel(R) Core(TM) i7 CPU         870  @ 2.93GHz CLDriverVersion=2.0                                                                                       
# Device Intel(R) Core(TM) i7 CPU         870  @ 2.93GHz CLRuntimeVersion=OpenCL 1.1 AMD-APP (844.4)                                                                                       
# Device Intel(R) Core(TM) i7 CPU         870  @ 2.93GHz NumberAppAddressBits=32                                                                                       
# OS=Windows 7 Build 7601 Service pack 1                                                                                       
Method     ExecutionOrder     ThreadID     CallIndex     GlobalWorkSize     WorkGroupSize     Time     LocalMemSize     VGPRs     SGPRs     ScratchRegs     FCStacks     Wavefronts     ALUInsts     FetchInsts     WriteInsts     LDSFetchInsts     LDSWriteInsts     ALUBusy     ALUFetchRatio     ALUPacking     FetchSize     CacheHit     FetchUnitBusy     FetchUnitStalled     WriteUnitStalled     FastPath     CompletePath     PathUtilization     LDSBankConflict

[Raistmer]

You should not run in BOINC directory structure.
Do the next:

run bench (offline) with VLAR task only in TestWUs.
then go to science_apps and delete state.sah
the run profiler from that directory (it should contain app executable you want to profile of course and will contain required data input for app after running benchmark).

[Morten]

I had to cancel task before its completion when it reached 2GB RAM usage. It seams in profiler mode it will generate memory leak.

Please see csv / stderr.

In this mode I did not experience sluggish system, so I hope you are able to find useful info.

[Raistmer]

hm... looks like longest kernel FindPulse partial so should respond on parameter change. Other much shorted, 18ms approx...
Perhaps it's driver/config thing still...

[Morten]

So what should be the next step here?

[Raistmer]

Await new driver/ answer from AMD...
also, if only 1 GPU working - there is lag or not ?

[Morten]

When running profiler with just one task I did not experience lagging.

Perhaps if you update the post with your findings it will be a more constructive response from AMD?

[Raistmer]

When running profiler with just one task I did not experience lagging.

Perhaps if you update the post with your findings it will be a more constructive response from AMD?

Try to run benchmark with single VLAR task but w/o profiler. Will lag there?
Profiler changes too much to judge.