Mojo struct
MiniIterSpec
struct MiniIterSpec
One mini-iter of the 4-wave body: (DRAM prefetch, frag-load, MMA).
Fieldsβ
- βload_tag (
Int):LOAD_AorLOAD_Bβ channel-A vs channel-B prefetch. - βload_channel (
Int): 0 (A) or 1 (B). Redundant withload_tagbut needed by the framework's edge derivation. - βload_subtile (
Int): Which sub-tile of the source SMEM half this prefetch writes (0 or 1). - βfrag_tag (
Int):MMA_LOAD_AorMMA_LOAD_Bβ register frag-load. - βfrag_channel (
Int): 0 (A) or 1 (B) for the frag-load. - βfrag_subtile (
Int): Which sub-tile of the SMEM stage to read into the fragment register. - βfrag_cross_stage (
Bool): True iff the frag-load reads from the cross K-partition's stage (cross-stage rotation pre-loads the next partition's leading quadrants while this partition's MMAs are still issuing). False = read from the same stage as the prefetches in this partition. - βmma_m_quad (
Int): Whichm_quadof the warp's 2x2 quadrant grid this MMA computes. - βmma_n_quad (
Int): Whichn_quadof the warp's 2x2 quadrant grid this MMA computes.
Implemented traitsβ
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!