Xilinx / llvm-aie

Fork of LLVM to support AMD AIEngine processors
Other
107 stars 12 forks source link

Martien.exhaustiveswp #242

Open martien-de-jong opened 2 days ago

martien-de-jong commented 2 days ago

NOTE: The total change is less than the sum of the absolute value of the diffs. It still needs some cleanup.

Set the baseline for GEMM and Conv2D_bf16 Add a provision to try multiple strategies for a single II. Add a critical path heuristic

For best effect on QoR, the pre-pipeliner and the prescheduler should be skipped. This is forced by

pragma clang loop pipeliner(disable)

in the source, or globally using --enable-pipeliner=0 --enable-misched=0