topic Re: Intel GPU OpenCL with g++ -O3 flag. in GPU Compute Software

Intel GPU OpenCL with g++ -O3 flag.

luisgo007 — Tue, 08 Aug 2023 12:58:40 GMT

Dear All,

I have a dynamic library (XXX.so) with a OpenCL kernel.

When I compile the library without -O3 flag the kernel works Ok.

When I compile with the -O3 flag I obtain the follow error from the run of the kernel.

-54 CL_INVALID_WORK_GROUP_SIZE

g++ -c -march=native -O3 -fPIC AoA.cpp -o AoA.o
g++ -shared -march=native  -o AoA.so AoA.o -L/usr/lib -llapack -lblas -lOpenCL

Is this normal?

Thanks,

Luís Gonçalves

Re: Intel GPU OpenCL with g++ -O3 flag.

Ben_A_Intel — Tue, 08 Aug 2023 16:11:51 GMT

Can you see what the arguments are to the call to clEnqueueNDRangeKernel that is generating the error when your program is compiled with -O3? Are they different than when your program is compiled with -O2?

You might be able to do this with the debugger, or you can use a tool like the OpenCL Intercept Layer to generate an API "call log":

https://github.com/intel/opencl-intercept-layer

https://github.com/intel/opencl-intercept-layer/blob/main/docs/controls.md#calllogging-bool

I'm guessing that aggressive compiler optimizations are causing the local_work_size to be something other than what you are expecting, which is causing the OpenCL driver to return an error.

Re: Intel GPU OpenCL with g++ -O3 flag.

luisgo007 — Wed, 09 Aug 2023 08:57:02 GMT

With -O2 happens the same.

I try to put the variables "static" but did not worked as well.

And It also does not accept "volatile" because of incompability of types.

g++ (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0

Re: Intel GPU OpenCL with g++ -O3 flag.

luisgo007 — Wed, 09 Aug 2023 09:54:13 GMT

I had a bug. Compiles ok with -O3. Sorry.