From f355192d5a44de2137609a25f0e48268877279e5 Mon Sep 17 00:00:00 2001
From: Martin Larralde <martin.larralde@embl.de>
Date: Fri, 21 Jun 2024 01:00:07 +0200
Subject: [PATCH] Fix AVX2 implementation of `Maximum::argmax`
---
lightmotif/src/pli/dispatch.rs | 1 -
lightmotif/src/pli/platform/avx2.rs | 4 ++--
2 files changed, 2 insertions(+), 3 deletions(-)
diff --git a/lightmotif/src/pli/dispatch.rs b/lightmotif/src/pli/dispatch.rs
index f6a4010..7fa8add 100644
--- a/lightmotif/src/pli/dispatch.rs
+++ b/lightmotif/src/pli/dispatch.rs
@@ -198,7 +198,6 @@ impl<A: Alphabet> Maximum<u8, <Dispatch as Backend>::LANES> for Pipeline<A, Disp
scores: &StripedScores<u8, <Dispatch as Backend>::LANES>,
) -> Option<MatrixCoordinates> {
match self.backend {
- // FIXME !!!!
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
Dispatch::Avx2 => Avx2::argmax_u8(scores),
// #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
diff --git a/lightmotif/src/pli/platform/avx2.rs b/lightmotif/src/pli/platform/avx2.rs
index 8f90e9d..1e731f0 100644
--- a/lightmotif/src/pli/platform/avx2.rs
+++ b/lightmotif/src/pli/platform/avx2.rs
@@ -438,8 +438,8 @@ unsafe fn argmax_u8_avx2(
let r1 = _mm256_unpacklo_epi8(r, _mm256_setzero_si256());
let r2 = _mm256_unpackhi_epi8(r, _mm256_setzero_si256());
// compare scores to local maximums
- let c1 = _mm256_or_si256(_mm256_cmpgt_epi16(r1, s1), _mm256_cmpeq_epi16(r1, s1));
- let c2 = _mm256_or_si256(_mm256_cmpgt_epi16(r2, s2), _mm256_cmpeq_epi16(r2, s2));
+ let c1 = _mm256_cmpgt_epi16(s1, r1);
+ let c2 = _mm256_cmpgt_epi16(s2, r2);
// replace indices of new local maximums
p1 = _mm256_blendv_epi8(p1, index, c1);
p2 = _mm256_blendv_epi8(p2, index, c2);
--
GitLab