Why not use sync after loading from TMEM to RMEM in example 02_mma_tma_sm100.cu

[gujiewen]

Accroding to ptx, tcgen05.ld is an async instruction, so why not use tcgen05.wait or its wrapper after loading in 02_mma_tma_sm100.cu.

cutlass/examples/cute/tutorial/blackwell/02_mma_tma_sm100.cu

Lines 354 to 360 in 3b05476

	// Load TMEM -> RMEM
	copy(tiled_t2r_copy, tDtAcc, tDrAcc);
	// AXPBY RMEM -> RMEM: tDrC = alpha * tDrAcc + beta * tDrC
	axpby(alpha, tDrAcc, beta, tDrC);
	// Store RMEM -> GMEM
	copy(tDrC, tDgD);

[hwu36]

According to https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-memory-consistency-model-canonical-sync-patterns-reg-dependency-same-thread, this case does not have to have a fence, but it is "nice" to have a fence here.

[gujiewen]

According to https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-memory-consistency-model-canonical-sync-patterns-reg-dependency-same-thread, this case does not have to have a fence, but it is "nice" to have a fence here.

That's clear.