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

bt_frag_coord

def bt_frag_coord(lane: Int, i: Int) -> IndexList[Int(2)]

(n, k) in the 16x32 right operand for element i under transpose_right=1.

Under transpose_right=1 the op reads the right operand as (N, K) instead of (K, N). The B-fragment lane->element map is the SWAP of the base bc_frag_coord: the base row-axis (frag_row_base + (sub>>2)*8, plus half*16) carries N, and the base col-axis (frag_col_base + (sub%4)) carries K. So consecutive elements walk K CONTIGUOUSLY, which is stride-1 for an (N, K) weight -> a coalesced load (the fix for the stride-K scattered NT B gather). The C-store map is UNCHANGED (bc_frag_coord); transpose_right permutes only the right operand, not C. Matched bit-exact against a host oracle.

Args:

lane (Int): Simdgroup lane index in [0, 32); determines the lane's base (n, k) in the 16x32 right operand.
i (Int): B-fragment element index in [0, 16); selects which of this lane's 16 held B elements to map to (n, k) under transpose_right=1.

Returns:

IndexList[Int(2)]