From ff811f371547ed700872b79a9b5eec5e5db89ede Mon Sep 17 00:00:00 2001
From: Martin Larralde <martin.larralde@embl.de>
Date: Sat, 15 Jun 2024 09:21:05 +0200
Subject: [PATCH] Use `_mm256_permutevar8x32_ps` in AVX2 `Score` to avoid
special case for default Dna symbol
---
lightmotif/src/dense.rs | 4 +++-
lightmotif/src/pli/platform/avx2.rs | 23 ++++++-----------------
2 files changed, 9 insertions(+), 18 deletions(-)
diff --git a/lightmotif/src/dense.rs b/lightmotif/src/dense.rs
index 4fe699c..7487b2a 100644
--- a/lightmotif/src/dense.rs
+++ b/lightmotif/src/dense.rs
@@ -14,7 +14,9 @@ use crate::num::Unsigned;
// --- DefaultAlignment --------------------------------------------------------
-#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+#[cfg(target_arch = "x86_64")]
+type _DefaultAlignment = typenum::consts::U64;
+#[cfg(target_arch = "x86")]
type _DefaultAlignment = typenum::consts::U32;
#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
type _DefaultAlignment = typenum::consts::U16;
diff --git a/lightmotif/src/pli/platform/avx2.rs b/lightmotif/src/pli/platform/avx2.rs
index 367c502..a856458 100644
--- a/lightmotif/src/pli/platform/avx2.rs
+++ b/lightmotif/src/pli/platform/avx2.rs
@@ -150,23 +150,12 @@ unsafe fn score_avx2_permute<A>(
let x3 = _mm256_shuffle_epi8(x, m3);
let x4 = _mm256_shuffle_epi8(x, m4);
// load row for current weight matrix position
- let t = _mm256_broadcast_ps(&*(pssmptr as *const __m128));
- let u = _mm256_broadcast_ss(&*(pssmptr.add(<A as Alphabet>::K::USIZE - 1)));
- // check which bases from the sequence are unknown
- let unk1 = _mm256_castsi256_ps(_mm256_cmpeq_epi32(x1, n));
- let unk2 = _mm256_castsi256_ps(_mm256_cmpeq_epi32(x2, n));
- let unk3 = _mm256_castsi256_ps(_mm256_cmpeq_epi32(x3, n));
- let unk4 = _mm256_castsi256_ps(_mm256_cmpeq_epi32(x4, n));
- // index A/T/G/C lookup table with the bases
- let p1 = _mm256_permutevar_ps(t, x1);
- let p2 = _mm256_permutevar_ps(t, x2);
- let p3 = _mm256_permutevar_ps(t, x3);
- let p4 = _mm256_permutevar_ps(t, x4);
- // blend together known and unknown scores
- let b1 = _mm256_blendv_ps(p1, u, unk1);
- let b2 = _mm256_blendv_ps(p2, u, unk2);
- let b3 = _mm256_blendv_ps(p3, u, unk3);
- let b4 = _mm256_blendv_ps(p4, u, unk4);
+ let t = _mm256_load_ps(pssmptr);
+ // index A/T/G/C/N lookup table with the bases
+ let b1 = _mm256_permutevar8x32_ps(t, x1);
+ let b2 = _mm256_permutevar8x32_ps(t, x2);
+ let b3 = _mm256_permutevar8x32_ps(t, x3);
+ let b4 = _mm256_permutevar8x32_ps(t, x4);
// add log odds to the running sum
s1 = _mm256_add_ps(s1, b1);
s2 = _mm256_add_ps(s2, b2);
--
GitLab