It seems that in reduce-by-key.hpp kernels are started with local sizes > global size if matrix sizes are smaller than:
SizeType kernel_WgSize = WAVESIZE * KERNELWAVES;
I don't know how the kernels work so I'm not sure if it would be enough to limit local sizes to always be smaller than global sizes.
It seems that in reduce-by-key.hpp kernels are started with local sizes > global size if matrix sizes are smaller than:
SizeType kernel_WgSize = WAVESIZE * KERNELWAVES;
I don't know how the kernels work so I'm not sure if it would be enough to limit local sizes to always be smaller than global sizes.