AnsweredAssumed Answered

Looking for specific details of GPU_DUMP_DEVICE_KERNEL

Question asked by mfried on May 10, 2012
Latest reply on May 11, 2012 by mfried

I have some code which works fine under AMD's OpenCL implementation. It creates buffers, compiles a .cl file into a program, creates kernels, sets args on them, invokes them, measures the run time, and checks the outputs, etc, etc. I export GPU_DUMP_DEVICE_KERNEL=3 (or 1 or 2) in my bash shell or via set on Windows, and I run my program (it's a C# program with my own OpenCL wrappers around the DLL, and it has been debugged thoroughly under mono on linux and Mac OS as well as Windows with Apple, AMD, NVIDIA, and Intel OpenCL SDKs and at least 10 different GPU devices and 5 different CPU devices). I can't find any .il or .isa files.


Is there a _specific_ API call that I need to be making such as clGetBinaries?

Do I need to install the APP SDK 2.6 or later? Do I need to install the APP SDK _after_ installing Catalyst drivers?

I'm using the AMD GPU -- in this case a Radeon HD 7970 (Tahiti) from OpenSUSE 12.1 x64 Linux or a Radeon HD5870 (Cypress) on Windows 7 x64.

Do I need to be on Windows vs Linux? I tried both, no success on either box.

Is the output file stored in whatever my local directory happens to be?

If I change that current directory via an API call, can I set the folder where the file gets dumped to be a temp folder or some other place that my current user can access? On Linux, I'm running as root for the time being, so I don't expect this is a file access problem.


Thanks for any thoughts.