lijinpei-amd wrote:

@arsenm Thanks. Before reworking, I'd like to settle where this data should 
live (cc @shiltian). I see two options:

1. **TableGen sched model.** Each MFMA SchedWrite in SISchedule.td would carry 
how long it holds plain VALU and PK_F32/DOT issue (XDL, XDL2x, SGEMM, DGEMM). 
PK_F32 and DOT would get a SchedWrite that uses that resource, and the query 
would read it from the sched model.
2. **Extend the existing `CoExecInfo` / `getMFMACoExecInfo`.** It currently 
covers gfx950 only, has a single VALU class in `CoExecMask` (no PK_F32, no 
DOT), and is a hand-written opcode switch. Extending it would mean adding 
gfx942 tables and new mask bits for PK_F32 and DOT.

I lean towards (1), since the sched model already has the pass counts and (2) 
would add more per-opcode C++ tables. Which would you prefer?


https://github.com/llvm/llvm-project/pull/233721
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to