Mojo function

ds_read_tr16_b64_warp

ds_read_tr16_b64_warp[mma_shape: IndexList[3]](tile: TileTensor[tile.dtype, tile.LayoutType, tile.origin, address_space=AddressSpace.SHARED, linear_idx_type=tile.linear_idx_type, element_size=tile.element_size]) -> SIMD[tile.dtype, 4]

Warp-level transposed LDS read distributing across 16-lane rows.

For 32×32×16 MMA: 2×2 row distribution over 8×32 tile. For 16×16×32 MMA: 4×1 row distribution over 16×16 tile.

Returns:

SIMD

View source

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!