Skip to content

Latest commit

 

History

History
43 lines (31 loc) · 2.45 KB

CONTRIBUTING.md

File metadata and controls

43 lines (31 loc) · 2.45 KB
Raw

Advice:

  • Correctness with original semantics. Historically contributions to this repositery give tons of work to the maintainer because of insufficient reading of this document and general lack of care for correctness.
    • If your work is not high-quality it will be auto-closed, because fixing it may take more time than creating it.
    • Follow the rules here that stem from experience implementing SIMD intrinsics, a library that shall be very reliable.

Here are the source for semantics:

  • Intel Intrinsics Guide

  • When the guide and the instruction disagree, we look at what C++ compilers do for this intrinsic.

  • GODBOLT EVERYTHING YOU COMMIT

    • Use godbolt-template.d and modify to your wished
    • GDC (version 12 or later) with -mavx -mavx2 (the template doesn't build without -mavx)
    • LDC (version 1.24+ or later) with -mtriple arm64, -O2, -O0, -mattr=+avx2, etc.

  • Do not add AVX-512 or AVX-10 intrinsics, they are out of scope.

  • I implore you to do intrinsics one by one, not all at once. This is very detailed work, it's not possible nor desirable to go fast while writing intrinsics.

    • Please don't go fast.
    • Please make small PR because there is a lot of context to communicate.
    • Get pre-approval before working on something big.
    • If you're a first time contributor, PR with only one intrinsic.

  • Add PERF comment anywhere you feel that something could be done faster in a supported combination: DMD D_SIMD, LDC x86_64, LDC arm64, LDC x86, GDC x86_64, with or without optimizations, with or without instruction support...

    • If this is supposed returns a SIMD literal, does it inline?
    • Can this be faster in -O0?
    • If instruction support is not there, is the alternative path fast?

  • Later instruction set are allowed to use intrinsics from <= instruction sets.

  • Keep in mind all intrinsics should work, whatever the compiler and flags, with same semantics. This is the main appeal of the library.

To be merged a PR:

Your PR doesn't have to implement every intrinsic in a given instruction set. It's best to do one intrinsics right, than several half-done, giving work to the maintainer.