From a468b30cb75babe268c249f13606cbaa60c14bb7 Mon Sep 17 00:00:00 2001
From: Robert Knight <robertknight@gmail.com>
Date: Sun, 2 Feb 2025 21:42:40 +0000
Subject: [PATCH] Add missing `#[inline(always)]` to `Quantize` SIMD op

Fix calls to SIMD intrinsics not being inlined. Also add `target_feature` to an
AVX2 SIMD impl for consistency with other methods.
---
 rten-simd/src/arch/x86_64.rs | 1 +
 rten-vecmath/src/quantize.rs | 1 +
 2 files changed, 2 insertions(+)

diff --git a/rten-simd/src/arch/x86_64.rs b/rten-simd/src/arch/x86_64.rs
index 2cb6f2e5..4437a85e 100644
--- a/rten-simd/src/arch/x86_64.rs
+++ b/rten-simd/src/arch/x86_64.rs
@@ -171,6 +171,7 @@ impl SimdInt for __m256i {
     }
 
     #[inline]
+    #[target_feature(enable = "avx2")]
     unsafe fn saturating_cast_u8(self) -> impl Simd<Elem = u8> {
         use std::arch::x86_64::{
             __m128i, _mm256_castsi256_si128, _mm256_packus_epi16, _mm256_packus_epi32,
diff --git a/rten-vecmath/src/quantize.rs b/rten-vecmath/src/quantize.rs
index 3caffb74..6eec6a9a 100644
--- a/rten-vecmath/src/quantize.rs
+++ b/rten-vecmath/src/quantize.rs
@@ -38,6 +38,7 @@ impl<'s, 'd, To> Quantize<'s, 'd, To> {
 impl<'d> SimdOp for Quantize<'_, 'd, u8> {
     type Output = &'d mut [u8];
 
+    #[inline(always)]
     unsafe fn eval<S: SimdFloat>(self) -> Self::Output {
         let mut n = self.src.len();
         let mut src_ptr = self.src.as_ptr();