[BUG] Sparse GEMM seems to produce wrong results for `F32` datatype and specific inputs about cutlass

[hwu36]

could you try to use small int like [-3, 3] as the input? we don't have true fp32 tensor cores. we use tf32 to compute. tf32 has only 10 explicit mantissa bits, 14 less than true fp32. fp32 are converted to tf32 first before calling tensor cores.

[hwu36]

moreover, try to use 64x32x32 warp tile size if you use 128x64x32 threadblock tile size.

[alexsamardzic]

Thanks, if dense GEMM force to tf32 then results indeed match.