tilelang.tileop.gemm_blockscaled¶
Common block-scaled GEMM tile op.
Submodules¶
Classes¶
Block-scaled GEMM tile op: |
Package Contents¶
- class tilelang.tileop.gemm_blockscaled.GemmBlockScaled¶
Bases:
tilelang.tileop.gemm.GemmBlock-scaled GEMM tile op:
C (+)= (A * SFA) @ (B * SFB).A
GemmNodesubclass on the C++ side, so it shares the dense GEMM’s operand layouts, warp partition and scheduling, and is lowered through the sametl.gemm.infer_layout/tl.gemm.lowerentry points, which dispatch on this Python class. The extra FFI fields aresfaRegion,sfbRegionandsfKStart. The C++ selector returns a block-scaled instruction key (cuda.tcgen05.blockscaled,cuda.mma.blockscaled) that the backend registry maps to an implementation class built onGemmBlockScaledMixin, so dense implementation classes never see the scale factors.- property SFARegion¶
- property SFBRegion¶
- property sf_k_start¶