RoundingVariants performance comparison

[whirlicote]

Here is the full code setup for compiling a performance comparison code example from CGAL.

Here is a collection of pointers:

• two benchmark code calls
• CGAL patches (C++ macro overrides for CGAL)
• V8 (currently forked from 326f5f8c, changes till 6d0f69f2 (instruction clones + AVX512 calls))
• chromium patches (there are no patches on chromium other than using the patched V8 version)
• build instructions: (Everything is supposed to be self-contained. So having nix installed (e.g. Single-user installation) and calling the scripts in the repository (steps_3_build.sh, steps_4_host_bench_test.sh, ...) is supposed enable you to build everything.
• for running the tests there is this fork with the RoundingVariants extension implemented

Regarding toolchain changes:

• I used the keyword __attribute__((noinline, used, visibility("default"))) for instruction stubs
• Then I used my fork of wabt to do an .wat roundtrip with patch_wasm_rounding.py
  • wasm2wat
  • patching callsites of the noinline variants with native instructions
  • wat2wasm

[whirlicote]

These are the performance numbers:

implementation	s
with RoundingVariants extension enabled (avx512)	69
software emulation	228

This implies a performance ratio on the order of 3.3.