How to tune sm90 kernel schedule?

[MARD1NO]

I found in cutlass sm90 there exist many kernel policy:

struct KernelTmaWarpSpecializedFP8FastAccum : KernelTmaWarpSpecialized { };
struct KernelTmaWarpSpecializedPingpongFP8FastAccum : KernelTmaWarpSpecializedPingpong { };
struct KernelTmaWarpSpecializedCooperativeFP8FastAccum: KernelTmaWarpSpecializedCooperative { };
struct KernelPtrArrayTmaWarpSpecializedCooperativeFP8FastAccum : KernelPtrArrayTmaWarpSpecializedCooperative { };

I'm not very clear that, and I found there exist performance gap. Is there any guide or document about the policy and how to select the best policy for my matmul problem?

[thakkarV]

See #1669