[PATCH] D103695: [WIP][RFC][Utils] Helper script to check sanity of cost tables vs scheduler models

[Haohai, Wen via Phabricator via llvm-commits]

HaohaiWen added a comment.

I think we can't completely trust reversed throughput reported by llvm-mca since some instructions' Rthroughput is not defined correctly in schedmodel.
e.g.

  $./llvm_utils_check_cost_tables.py --cpulevel=avx512  --stop-on-diff
  double fdiv double: cost (4.0 - 4.0) vs recipthroughput (3 - 3)
  skylake-avx512 : 4.0 vs 3

defines in X86SchedSkylakeServer.td:

  def SKXWriteResGroup184 : SchedWriteRes<[SKXPort0,SKXFPDivider]> {
    let Latency = 14;
    let NumMicroOps = 1;
    let ResourceCycles = [1,3];
  }
  def : SchedAlias<WriteFDiv64,  SKXWriteResGroup184>; // TODO - convert to ZnWriteResFpuPair

However, it's measured tpt is 4 from uops.info <https://uops.info/table.html?search=vdivsd&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_SKX=on&cb_measurements=on&cb_doc=on&cb_base=on&cb_avx=on&cb_avx2=on&cb_avx512=on>. llvm-exegesis tpt result is also 4.
I think uops.info/agner.org should be more accurate.

Have you verified cost diff based on uops.info/anger.org?
We have seen some regression on our internal benchmarks due to TTI cost-model patches based on this tool...


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103695/new/

https://reviews.llvm.org/D103695