Write _mm_blend_epi16 emulation without multiplication

Use _mm_and_si128 and _mm_cmpeq_epi16 rather than expensive multiplication _mm_mullo_epi16 with _mm_srai_epi16 that compiler may not be able to optimize.

2 files changed, 8 insertions, 6 deletions

diff --git a/c/blake3_sse2.c b/c/blake3_sse2.c
index de33d62..1592966 100644
--- a/c/blake3_sse2.c
+++ b/c/blake3_sse2.c
@@ -79,10 +79,11 @@ INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
 }
 
 INLINE __m128i blend_epi16(__m128i a, __m128i b, const int imm8) {
-    __m128i mask = _mm_set1_epi16(imm8);
-    mask = _mm_mullo_epi16(mask, set4(0x40008000, 0x10002000, 0x04000800, 0x01000200));
-    mask = _mm_srai_epi16(mask, 15);
-    return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
+  const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
+  __m128i mask = _mm_set1_epi16(imm8);
+  mask = _mm_and_si128(mask, bits);
+  mask = _mm_cmpeq_epi16(mask, bits);
+  return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
 }
 
 INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
diff --git a/src/rust_sse2.rs b/src/rust_sse2.rs
index 3084ed1..366ab21 100644
--- a/src/rust_sse2.rs
+++ b/src/rust_sse2.rs
@@ -139,9 +139,10 @@ unsafe fn undiagonalize(row0: &mut __m128i, row2: &mut __m128i, row3: &mut __m12
 
 #[inline(always)]
 unsafe fn blend_epi16(a: __m128i, b: __m128i, imm8: i32) -> __m128i {
+    let bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
     let mut mask = _mm_set1_epi16(imm8 as i16);
-    mask = _mm_mullo_epi16(mask, set4(0x40008000, 0x10002000, 0x04000800, 0x01000200));
-    mask = _mm_srai_epi16(mask, 15);
+    mask = _mm_and_si128(mask, bits);
+    mask = _mm_cmpeq_epi16(mask, bits);
     _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a))
 }