We are pleased to announce the release of a new version of AMD APP Profiler (previously ATI Stream Profiler) version 2.2.
AMD APP Profiler is a performance analysis tool that gathers data from the OpenCL or DirectX run-time and AMD Radeon GPUs during the execution of an OpenCL or DirectCompute application.
New updates in this version include
Great news!
I was trying the old profiler with my new HD 6950 on Linux just to discover that it was not supported by the old profiler. But just a few days after that the new SDK 2.4 and profiler 2.2 arrived, yay! There was definitely some great progress on the SDK and tools side in the last ~6 months I've been using OpenCL.
I know there is frustrating things, and I also have been waiting longer for certain OpenCL features in the past, but still I want to congratulate everyone on the progress.
Hi,
I have the problem, that the values for the execution-time for the kernel-runtime differ from the profiler and the calling of the following functions in my programm:
long long kernelsStartTime;
long long kernelsEndTime;
status = clGetEventProfilingInfo(
events[0],
CL_PROFILING_COMMAND_START,
sizeof(long long),
&kernelsStartTime,
NULL);
if(
status !=
CL_SUCCESS)
printf("clGetEventProfilingInfo failed.");
//Window::Instance()->addText("clGetEventProfilingInfo failed.");
status = clGetEventProfilingInfo(
events[0],
CL_PROFILING_COMMAND_END,
sizeof(long long),
&kernelsEndTime,
NULL);
if(
status !=
CL_SUCCESS)
printf("clGetEventProfilingInfo failed.");
//Window::Instance()->addText("clGetEventProfilingInfo failed.");
/* Compute total time (also convert from nanoseconds to seconds) */
double totalKernelTime = (double)(kernelsEndTime - kernelsStartTime)/1e9;
The values differ up to 100% (i.e. 2,1xxxx sec to 4,0xxx sec). I think the profiler info will be the more exact value? But how can I implement a reliable bechmark function for my program which I need for my master-thesis. I don't have the possibility to install the VS2008 and the APP SDK on all the machines where I have to test the performance of my application.
Thanks for help!
problem is that on first run there is several factor which affect preformace. when i run my kernel first five kernels differs in 200% and then execution time is quite stable.
GPU take some time switch to performance mode and it takes some time to transfer buffers to GPU RAM as this transfer happend on first use of buffer.
The first benchmark (with the C++ programm) shows values which must be too low, for example 0,001s, although the rendered image is correct (so he is executing the kernel correctly). But this doesn't happen in every case.
The runs number 2-4 in a loop then show a realistic result of 0,1 second.
But I recognized, that the profiler result is much higher than expected. It shows 3,9 seconds, although I can see (and feel^^), that the time from opening my program to showing the rendered picture doesn't last one second at all.
hi micheal,
Please mention your system configuration: CPU,GPU,SDK,Driver & OS.
Do you see this issue only on your code or in samples also.
Please post a test case to reproduce it.
Hi
Am having a problem running the Profiler on Vista64 with the OpenCL samples in the APP SDK (2.4)
MS Visual Studio 2008 (v9.0.21022.8)
ATI Catalyst v2011.0413.4.41348
AMD FirePro V7800 Driver Packaging Version 8.801-101125a-111741C
(Beta driver, needed in order to run the SDK examples with a FirePro V7800. The non-beta version would not allow the sample code to run, complaining the CAL version was not new enough).
APP SDK 2.4
APP Profiler v2.2.633
When I try to run the profiler from either visual studio 2008 or the command line (either x64 or x86) the OpenCL examples in the SDK fail at clCreateContextFromType() and clCreateContext(). This happens when collecting GPU performance counters or an application trace. The examples run correctly if run from the command line or within visual studio without using the profiler. Both 32bit and 64bit builds fail. Is there anything extra we need to do when creating the cl context in order for the compiler to work?
Thanks,
George
Can you please provide the driver version as year.month format(11.4 or 11.5 or similar).
Hi,
Thanks for the quick reply.
Not sure I've seen anything as concise as 11.4 or 11.5. The driver was downloaded by selecting Workstation Graphics/FirePro Beta Drivers/FirePro V7800/Vista 64bit from the driver download site. This gives a driver with release date Jan 17th 2010, 8.801 Beta. Vista indicates the DLLs were made on November 25th 2010. I've checked the Catalyst information centre and it refers to 8.801-101125a-111741C as the driver packaging version. The 2D Driver Version is 8.01.01.1105. The Direct3D version is 8.14.10.0798. The OpenGL version is 6.14.10.10362.
We did previously try installing the FirePro 3D Series driver which has a release date of March 5th 2011, revision 8.773.1.2 but this gave a CAL version number mismatch when running OpenCL examples in the APP SDK v2.4.
It seems a little odd that the beta driver dated 1/17/2010 has a revision number 8.801 whereas the driver dated 5/3/2011 has a revision number 8.773.1.2 but then I may be reading too much in to the numbers.
Please let me know if there's some other version info I can give.
Thanks,
George