Adding error checks after all kernel calls to avoid silent errors.
This is especially important for cudaErrorNoKernelImageForDevice errors. Those errors happen when the kernels are launched on a GPU which does not match the architecture the code was compiled for and the runtime has cannot jit-compile the kernels due to missing or incompatible PTX code.
ahehn-nv
commented
3 years ago
Adding error checks after all kernel calls to avoid silent errors. This is especially important for
cudaErrorNoKernelImageForDeviceerrors. Those errors happen when the kernels are launched on a GPU which does not match the architecture the code was compiled for and the runtime has cannot jit-compile the kernels due to missing or incompatible PTX code.