Strange clBuildProgram Error 'anonymous_jit_identity'

[RyFo18]

I’m getting a strange error after calling clBuildProgram for my program. Here is the output from the build log:

ptxas ptx input, line 2376; fatal : Parsing error near ‘.’: syntax error

: Retrieving binary for ‘anonymous_jit_identity’, for gpu=‘sm_13’, usage mode=’

I’m taking the following explanation from the Khronos forums, but I wanted to post it here in case it is a compiler bug:

ptxas sounds like a program which is an assembler for PTX code, which is nvidia’s low level representation code for their GPUs. On Mac OS there’s a compiler which compiles from OpenCL to PTX, and then something else which goes from PTX to GPU specific code. It wouldn’t surprise me if it was similar on other platforms too.

This means that that line number doesn’t refer to your code, but to a line in the PTX code which you don’t have access too. It’s basically useless information for you.

[REMOVED 2 LINES]

This sounds like a compiler bug to me, as the system shouldn’t be generating PTX code that can’t be built.

Does anyone else have any ideas as to what this may be? It sounds like there is some error at line 2376 of this file near “.”. Very descriptive I must say External Media . I tried completely reinstalling the SDK and that doesn’t seem to work.

http://www.khronos.org/message_boards/view…f=28&t=2211

[Keldor314]

That seems like a compiler bug to me. OpenCL is rather a minefield of those right now External Image.

If you could post the code here, it might help the people at NV fix the compiler bug. In the meanwhile, try using clGetProgramInfo to get the binary from the built code. “Binary” in this case actually means PTX, so hopefully you can look at the PTX and at least get some idea of what operation it was attempting to do which got maimed by the compiler.

[RyFo18]

Thanks for your response Keldor. I actually figured this one out. I had an initial OpenCL implementation that I was using for a CPU. When I went to the NVIDIA SDK for the GPU, it apparently doesn’t support function calls within the CL code (nor does the AMD SDK, I got another strange error from their SDK). So for now I’ve lumped all of my code into one kernel with no function calls.