minor changes

lancedb · Dec 8, 2023 · ce5dcb7 · ce5dcb7
1 parent bb3865d
commit ce5dcb7
Showing 1 changed file with 0 additions and 9 deletions.
diff --git a/rust/lance-linalg/src/simd/bf16.c b/rust/lance-linalg/src/simd/bf16.c
@@ -45,7 +45,6 @@ float dot_bf16(const __bf16 *a, const __bf16 *b, size_t n) {
 
     while (n > 0) {
         if (n < 32) {
-//            __mmask32 mask = d - i >= 32 ? 0xFFFFFFFF : (1u << n) - 1u;
             __mmask32 mask = (1u << n) - 1u;
             a_vec = _mm512_maskz_loadu_epi16(mask, a);
             b_vec = _mm512_maskz_loadu_epi16(mask, b);
@@ -91,7 +90,6 @@ float cosine_bf16(const __bf16 *a, const __bf16 *b, size_t n) {
 
     while (n > 0) {
         if (n < 32) {
-//            __mmask32 mask = d - i >= 32 ? 0xFFFFFFFF : (1u << n) - 1u;
             __mmask32 mask = (1u << n) - 1u;
             a_vec = _mm512_maskz_loadu_epi16(mask, a);
             b_vec = _mm512_maskz_loadu_epi16(mask, b);
@@ -110,11 +108,4 @@ float cosine_bf16(const __bf16 *a, const __bf16 *b, size_t n) {
     __bf16 a2 = _mm512_reduce_add_ph(a2_vec);
     __bf16 b2 = _mm512_reduce_add_ph(b2_vec);
     return 1 - ab / (a2 * b2);
-
-    // Compute the reciprocal square roots of a2 and b2
-    // __m128 rsqrts = __m128(_mm_set_ps(0.f, 0.f, a2 + 1.e-9f, b2 + 1.e-9f));
-    // __m128 rsqrt_a2 = _mm_cvtss_f32(rsqrts);
-    // f32_t rsqrt_b2 = _mm_cvtss_f32(_mm_shuffle_ps(rsqrts, rsqrts, _MM_SHUFFLE(0, 0, 0, 1)));
-    // return 1 - ab * rsqrt_a2 * rsqrt_b2;
-//    return 1;
 }