https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116800
Bug ID: 116800
Summary: std::simd: poor code generation of AVX512 fused
multiply-add
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: pieter.p.dev at outlook dot com
Target Milestone: ---
The AVX512 implementation of the std::experimental::fma function appears to be
falling back on the AVX2 version, which results in very poor performance.
For example, compare the two functions below:
The 'fma_1' function uses standard arithmetic operators,
and the 'fma_2' function calls the std::experimental::fma function.
The 'fma_1' function compiles down to a single AVX512 vfmadd instruction (which
is the desired behavior), whereas the code for 'fma_2' moves the upper half of
the 512-bit arguments into separate 256-bit registers, then performs two
256-bit vfmadd instructions, and merges the results back into a single 512-bit
result.
#include <experimental/simd>
using simd = std::experimental::native_simd<double>;
simd fma_1(simd x, simd y, simd z) {
return x * y + z;
}
simd fma_2(simd x, simd y, simd z) {
return fma(x, y, z);
}
Compiler explorer: https://godbolt.org/z/hMM83jh4o
In GCC 11 and 12, the situation is even worse: there the 'fma_2' code actually
performs eight individual FMAs, along with over 30 extraction/insertion
instructions! GCC 13 performs four 128-bit FMAs, and GCC 14 performs two
256-bit FMAs.
