Mojo module
block_scaled_smem
Shared memory layout for block-scaled SM100 matmul.
Extends standard SMEM with scaling factor tile storage (SFA, SFB) following MXFP8 layout conventions. Also includes all pipeline barriers and TMEM state.
Structs
-
BlockScaledSmem: SMEM struct containing A/B tiles, scaling factors, C output, and barriers.
Functions
-
get_sfa_num_cols: Get the number of TMEM columns needed for A scaling factors. -
get_sfa_smem_layout: Get the SMEM layout for A scaling factors. -
get_sfb_num_cols: Get the number of TMEM columns needed for B scaling factors. -
get_sfb_smem_layout: Get the SMEM layout for B scaling factors.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!