We are pleased to announce the availability of AMD APP Profiler v2.5. For more information, please visit the product page at: http://developer.amd.com/TOOLS/AMDAPPPROFILER/Pages/default.aspx
New features and updates in this version include:
Please post any feedback here.
Hey,
thanks for this version. The occupancy calculator seems to work under linux now (at least it gives some numbers), but there are still two issues:
btw. I'm using Linux
Catalyst 12.4
sprofile - V2.5.1804
AMDAPP SDK v2.7
HD 7970
1. The generation of the occupancy display page does not work (or I'm missing an option I have to use...):
I start it, as given in the help, with
sprofile -P "/path/to/occupancy/params/file.txt" -o "path/to/output.html"
and I get the error:
Error generating occupancy display file.
Unable to copy required files to output directory
The permissions for writing etc. are set correctly. I can see that there are some *.js files created.
My file.txt looks like
.ThreadID=17469
CallIndex=67
KernelName=pot_kernel
DeviceName=Tahiti
ComputeUnits=32
MaxWavesPerComputeUnit=40
MaxWorkGroupPerComputeUnit=40
MaxVGPRs=256
MaxSGPRs=512
MaxLDS=32768
UsedVGPRs=25
UsedSGPRs=80
UsedLDS=1536
WavefrontSize=64
WorkGroupSize=64
WavesPerWorkGroup=1
MaxWorkGroupSize=256
MaxWavesPerWorkGroup=4
GlobalWorkSize=81920
MaxGlobalWorkSize=16777216
WavesLimitedByVGPR=40
WavesLimitedBySGPR=24
WavesLimitedByLDS=21
WavesLimitedByWorkgroup=40
Occupancy=52.5
I got the numbers from the *.occupancy file. The only thing I added is the "call Index" that is not present there. (Is there any real meaning behind this number?)
2. The profiler says I have 32kB of local memory. Also the device query say I have 32kB of local memory. But as I'm using a HD 7970 I'm expecting 64kB of local memory. Is it a driver problem or a display problem?
Ralph Welsch wrote:
2. The profiler says I have 32kB of local memory. Also the device query say I have 32kB of local memory. But as I'm using a HD 7970 I'm expecting 64kB of local memory. Is it a driver problem or a display problem?
I think I got the problem. As the new Programming guide explains, there is a limit of 32 kB of LDS for each wave-front and probably that is the number displayed. But in fact the card has 64 kB of LDS/CU available and so a kernel consisting of several wave-fronts (at least two) can take full account of the 64 kB and this is not considered in the new version of the profiler, right?
So I can multiply the "WavesLimitedByLDS" by two and get the correct result if I have multiple wave-fronts per CU.
Hi all,
I also got the same problem on this :
sprofile -P "/path/to/occupancy/params/file.txt" -o "path/to/output.html"
same error message as rwelsch, any idea ?
My environment :
Ubuntu 10.04
GPU : AMD HD 7750
GPU Driver : AMD Catalyst 12.6 Proprietary Linux x86 Display Driver (amd-driver-installer-12-6-x86.x86_64.run)
CPU : Intel(R) Core(TM) i5 CPU 760 @ 2.80GHz
AMDAPP SDK v2.7
Thanks for reporting this, this is a bug in the APP Profiler, the next version of APP Profiler will have the fix. For now, you can only use Windows version of APP Profiler to generate occupancy charts.
Thanks !
By the way, can Visual Studio generate occupancy charts according to ".atp" or ".csv" ?
Or I had to manually add something in ".atp" or ".csv" in order to do that ?
HI , I am a new learner to AMD APP Profile and I met a problem.
how can I creat file.txt while follow the command line:
sprofile -P "/path/to/occupancy/params/file.txt" -o "path/to/output.html"
thaks!
APP Profiler is superseded by CodeXL. Please download CodeXL which includes the same command tool sprofile with bug fixes that should address your problem.
Driver limitation on SDK 2.7 doesn't allow me to switch to it but I took APP Profiler from it, v2.5
I would like to share some observations of 2.4 -> 2.5 APP Profiler change.
1. Bug with ICC-compiled binaries trace is fixed for VS plugin, thanks a lot!
2. Bug with loading old binary at race start is fixed too.
3. 2.5 can't load 2.4 created trace3 session fully - it can't show occupancy data for kernels. i checked, newly generated sessions show
that data ok, but it doesn't for old sessions created with v2.4.
4. Maybe the reason in change for device parameters: v2.4 said Loveland C-60 has wavefrong of 32, v2.5 says it has wavefrong of 64.
Who is right who is wrong?
Max number of wavefronts per CU changed acordingly: v2.4 said 64 wavefronts per CU is max, v2.5 says 32 wavefronts per CU is max.
Again, who is right in this?
At least number of workgroups, max workgroup size and CU number remained the same ;).
Driver used: 11.12 mobile version (it's C-60 based netbook).
Hi Raistmer
Thanks for using APP Profiler.
3. We've added a few new fields to occupancy output file which caused the incompatibility. Sorry for the inconvenience.
4. Loveland has wave size of 64. Max wave per CU is 49.
Thanks for info.
1. Bug with ICC-compiled binaries trace is fixed for VS plugin, thanks a lot!
Unfortunately, I was too optimistic in that. Actually bug still here, but I unintentionally found workaround it appears.
What bug: if ICC project is startup one APP Profiler plugin for VS2008 can't start API trace session. In this APP Profiler version too.
And what workaround: Luckely I have solution with few projects inside and one of the projects is MS VC based one, not ICC one. So when MS VC project is startup one trace session dialog opens OK and then one can chose ICC compiled binary of different project instead of binary of startup project. This bug is clearly plugin-based and has nothing to do with profiler itself, but adds some inconvience (and w/o such workaround it makes impossible to run trace session for ICC project in VS). Hope it will be eliminated in next release.
i have AMD Sapphire HD 7950 OC when i open Amd SADK APP V 2.7 it never recognize my Kernel .
Am i install AMD APP Profiler v2.5 To Compile my kernel or any solutions
i looked in its option i found that the latest card is 6970 i never found the 7000 series or my kernel name "Tahiti"
i have amd catalyst suit v12.4
HELP me plz with the right steps
i want to render on my gpu
Sorry, but who do not recognize your kernel ?
Profiler is used to profiling your OpenCL program on AMD platform, its optional package, and it is included in APP SDK.
Try newer GPU driver again, e.g. 12.6 or 12.7
Hello, using the profiler for GPU performance counters doens't work for me with Catalyst 12.10 APP SDK 2.7 Ubuntu 12.04 x64 and a Radeon 7870 or Radeon 5870. The command line profile yust quits without any message when reaching the kernel execution and when I use CodeXL the only error message is: "Failed to generate profile result".
I tried different kernels. CL_QUEUE_PROFILING_ENABLE was also set when creating the command queues.
Any suggestions?
Update: I tried some of the samples. The SimpleGL example could be profiled. Nbody and Fluid2D not.
Hi Chevydevil
Could you please try our latest developer tool CodeXL? We've fixed a few issues in the GPU profiling back end related to Catalyst compatibility. CodeXL also provides a nice GUI for the GPU profiling on Linux.
As I wrote above, I tried CodeXL already. It is not working and gives the error message above.
With the 12.11 beta driver, I get some relsults when using CodeXL. The performance counter stops after two or three iterations though. But better then nothing!
Could you please try our latest developer tool CodeXL? We've fixed a few issues in the GPU profiling back end related to Catalyst compatibility. CodeXL also provides a nice GUI for the GPU profiling on Linux. Now it is easy to use Linux for all purposes. ------------------------- |