Mojo function
get_rdna_fragment_layout
get_rdna_fragment_layout() -> Layout
Get the fragment layout for RDNA WMMA output fragments.
RDNA uses Wave32 (32 lanes) with 16x16 output tiles, so each lane holds 256/32 = 8 fp32 accumulator elements.
RDNA WMMA C/D register mapping: lane l, elem v -> D[row=v*2+l//16, col=l%16]. All 8 elements in a lane share the same column (l%16) but have different interleaved rows.
Returns:
Layout: Layout for 1 seq row x 8 key columns per lane.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!