Question about the tiling size about cuda_hgemm HOT 2 CLOSED

Indeed, you need to modify the tiling size to approximate the ratio of M to K. If M is much smaller than K, you also need to consider the splitk method.

from cuda_hgemm.

And, I don't understand why you set the BLOCK_ROW_WARPS as BLOCK_COLS / WARP_COLS, not BLOCK_ROW_WARPS / WARP_ROWS.
I'm sorry to bother you with these simple questions as I am a newbie to CUDA.
Thank you.

#define BLOCK_ROWS 256
#define BLOCK_COLS 128
...
#define BLOCK_ROW_WARPS 2 // BLOCK_COLS / WARP_COLS
#define BLOCK_COL_WARPS 4 // BLOCK_ROWS / WARP_ROWS

from cuda_hgemm.