Scale Factors in TMEM (nvfp4)

M = 128, one MMA-K block (SF_K = 4)  ·  .warpx4 broadcast: R[4 : 32@TLane]

Logical SFA  (M × SF_K)

click a cell = SFA[m, sfk]  ·  color = Mgroup (m // 32)

TMEM  (128 lanes × 4 TCols × 4 bytes)

TCol 0: m 0–31 · TCol 1: m 32–63 · TCol 2: m 64–95 · TCol 3: m 96–127  ·  byte in TCol = sfk
Layout  SFA[m, sfk] → TMEM
TLane = m mod 32  ·  Mgroup = m div 32  ·  TCol = Mgroup  ·  byte = sfk  ·  byte offset = TCol·4 + byte  ·  .warpx4 broadcast → lanes { TLane, +32, +64, +96 }
click a cell on the left…