For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

swiglu_extra_fixed_smem

def swiglu_extra_fixed_smem[c_type: DType](mma_shape: IndexList[Int(3)], cta_group: Int, AB_swapped: Bool = False) -> Int

Compute extra fixed SMEM bytes for the SwiGLU SMEM-path half tiles.

Returns the bytes needed for double-buffered half-output SMEM tiles. The tile shape is (output_tile_shape[0], output_tile_shape[1] / 2) in both AB_swapped orientations (dims are transposed but byte count is symmetric), so the same formula applies for both.

Parameters:

c_type (DType): DType of the C output elements.

Args:

mma_shape (IndexList[Int(3)]): MMA instruction shape as (MMA_M, MMA_N, MMA_K), used to derive the output tile shape.
cta_group (Int): Number of cooperating CTAs per cluster group (1 or 2).
AB_swapped (Bool): Whether the A and B operands are swapped (defaults to False).

Returns:

Int