Pekka Jääskeläinen

A basic test case: ```C++ #include __host__ __device__ void fptr() { printf("Hello function pointer world!\n"); } typedef void (*FPointer)(void); __global__ void fptr_call(FPointer ptr) { ptr(); } int main() { hipLaunchKernelGGL(fptr_call,...

There was a misunderstanding from my part: CUDA/HIP doesn't support "univeral function pointers" (like I assumed would be the semantics with \_\_host\_\_ \_\_device\_\_ functions). Indirect calls should work only on...

Is the error itself this test exposes a valid issue?

@pvelesko it should, but does not (yet) guarantee the linear ids nor multidim. OpenCL doesn't even support multidimensional subgroups. In fact, it explicitly states that the subgroups are single dimensional....

I've misinterpreted the sentence of the OpenCL specs _"While sub-groups may be used in multi-dimensional work-groups, each sub-group is 1-dimensional and any given work-item may query which sub-group it is...

Not sure if we should close this yet. Although this is not a CHIP-SPV issue, the problem in OpenCL-side is not resolved before there's an extension mandating the linear id....

What will be the "fast bitcode library" here? One that calls OpenCL native_* always or the current default (which sometimes has worse accuracy than CUDA requires)?

This is an extension, meaning it's not mandatory to support it. I do not know if we can make it more graceful, as long as the user sees it's because...

All of these cases except the one reported in Issue #134 seem to be caused by the double precision issue. I'll change this ticket to fix it for 0.9 via...

Not sure if we really need this one as there's a workaround for the only known target where the issue appears.