hpides / autovec-db Goto Github PK

Just to have it written down somewhere.

With hash_bucket in the bitmask variants, we have observed that the alignment of the hot loop in the function body impacts measured performance even though the CPU executes the exactly same bytes.

vTune implies that we have different amounts of work done in the frontend, it seems like the Decoded-Stream-Buffer / µOp-Cache doesn't work properly with certain alignments.
Investigation so far implied that:

• Per 32B instruction window, we do not use more than the 3 available cache lines in the µOp Cache. mostly, we have just 6-8 µOps in a 32B instruction window
• Up to IceLake (excluding), it could be due to the Intel Jcc-Erratum. We had a conditional jump that was bordering a 32B-boundary. With a 16B offset, this was not the case
  • However, investigation on our IceLake servers showed that we still get significant performance variance there with different alignment. Inline assembler for "16B nops" actually moves the function by 32B, and can give a 20% performance boost.

The last point needs more investigation. Maybe, there is something completely different going on, or maybe we haven't found the root cause yet.

Just for simplicity when comparing results and plotting, let's sort all benchmarks in the following order:

1. Scalar
2. Autovec
3. GCC vec
4. CPU-specific

This is the order I want in the plots.