Fix: 'scale' parameter of i32gather_ps has to be compile-time constant

2 files changed, 7 insertions, 5 deletions

diff --git a/intrinsics.h b/intrinsics.h
index 9337e35..c27ca97 100644
--- a/intrinsics.h
+++ b/intrinsics.h
@@ -132,8 +132,9 @@ INTGEMM_AVX2 static inline __m256i cvttps_epi32(__m256 a) {
 INTGEMM_AVX2 static inline __m256 div_ps(__m256 a, __m256 b) {
   return _mm256_div_ps(a, b);
 }
-INTGEMM_AVX2 static inline __m256 i32gather_ps(float const *base_addr, __m256i vindex, const int scale) {
-  return _mm256_i32gather_ps(base_addr, vindex, scale);
+template <unsigned Scale>
+INTGEMM_AVX2 static inline __m256 i32gather_ps(float const *base_addr, __m256i vindex) {
+  return _mm256_i32gather_ps(base_addr, vindex, Scale);
 }
 template <> INTGEMM_AVX2 inline __m256 loadu_ps(const float* mem_addr) {
   return _mm256_loadu_ps(mem_addr);
@@ -212,8 +213,9 @@ INTGEMM_AVX512BW static inline __m512i cvttps_epi32(__m512 a) {
 INTGEMM_AVX512BW static inline __m512 div_ps(__m512 a, __m512 b) {
   return _mm512_div_ps(a, b);
 }
-INTGEMM_AVX512BW static inline __m512 i32gather_ps(float const *base_addr, __m512i vindex, const int scale) {
-  return _mm512_i32gather_ps(vindex, base_addr, scale);
+template <unsigned Scale>
+INTGEMM_AVX512BW static inline __m512 i32gather_ps(float const *base_addr, __m512i vindex) {
+  return _mm512_i32gather_ps(vindex, base_addr, Scale);
 }
 template <> INTGEMM_AVX512BW inline __m512 loadu_ps(const float* mem_addr) {
   return _mm512_loadu_ps(mem_addr);
diff --git a/vec_utils.h b/vec_utils.h
index e2049fb..acb7d6e 100644
--- a/vec_utils.h
+++ b/vec_utils.h
@@ -119,7 +119,7 @@ Register exp_approx_taylor(Register x) {
 
   result = add_ps(result, const_one);
 
-  auto ea = i32gather_ps(EXP_LOOKUP + EXP_MAX, cvtps_epi32(a), 4);
+  auto ea = i32gather_ps<4>(EXP_LOOKUP + EXP_MAX, cvtps_epi32(a));
   return mul_ps(ea, result);
 }