In single stage pipeline, how memory loading from shared ram is overlapped with mma computation?

[wzhcz8902]

cutlass/include/cutlass/gemm/threadblock/mma_singlestage.h

Lines 198 to 201 in 5c447dd

	// Pair of fragments used to overlap shared memory loads and math instructions
	WarpFragmentA warp_frag_A;
	WarpFragmentB warp_frag_B;

As the comment say there is some overlap between memory loading from shared ram with mma, but in the loop

cutlass/include/cutlass/gemm/threadblock/mma_singlestage.h

Lines 223 to 239 in 5c447dd

	CUTLASS_PRAGMA_UNROLL
	for (int warp_mma_k = 0; warp_mma_k < Base::kWarpGemmIterations; ++warp_mma_k) {
	// Load warp-level tiles from shared memory, wrapping to k offset if this is the last group
	// as the case may be.
	this->warp_tile_iterator_A_.set_kgroup_index(warp_mma_k % Base::kWarpGemmIterations);
	this->warp_tile_iterator_B_.set_kgroup_index(warp_mma_k % Base::kWarpGemmIterations);
	this->warp_tile_iterator_A_.load(warp_frag_A);
	this->warp_tile_iterator_B_.load(warp_frag_B);
	++this->warp_tile_iterator_A_;
	++this->warp_tile_iterator_B_;
	warp_mma(accum, warp_frag_A, warp_frag_B, accum);
	}

I do not see such behavior, can you help me understand it?

[hwu36]

There is no overlap in single stage.

just

loop:
  load
  compute

It saves space at the cost of efficiency.