Sparse MFMA builtins

Sparse MFMA builtins#

Sparse Matrix Fused Multiply-Accumulate (SMFMAC) builtins multiply a compressed \(\pmb{A}\) tile by a dense \(\pmb{B}\) tile and accumulate the result into a \(\pmb{C}\) tile, exploiting 4:2 structured sparsity to halve the storage and bandwidth required for \(\pmb{A}\). Select a CDNA generation below for the corresponding builtin reference. For trailing parameter documentation, see Common MFMA parameters.