Simplify AVX integer load/store

Also has the potential to improve performance on some architectures
(if there is a domain crossing penalty - not sure whether any AVX capable
generation has a penalty).

Change-Id: Icc7b136571fc9ad1dbabeabe446c93e8816ec678

diff --git a/src/gromacs/simd/impl_x86_avx_256/impl_x86_avx_256.h b/src/gromacs/simd/impl_x86_avx_256/impl_x86_avx_256.h
index f6d5402aaaf26a282fc6aac5e588689ab9423ae1..e688d14c94d7beb4678777131968fea812f595de 100644 (file)
--- a/src/gromacs/simd/impl_x86_avx_256/impl_x86_avx_256.h
+++ b/src/gromacs/simd/impl_x86_avx_256/impl_x86_avx_256.h
@@ -119,11 +119,11 @@
 #define gmx_simd_set_exponent_f    gmx_simd_set_exponent_f_avx_256
 /* integer datatype corresponding to float: gmx_simd_fint32_t */
 #define gmx_simd_fint32_t          __m256i
-#define gmx_simd_load_fi(m)        _mm256_castps_si256(_mm256_load_ps((const float *)m))
+#define gmx_simd_load_fi(m)        _mm256_load_si256((__m256i const*)m)
 #define gmx_simd_set1_fi           _mm256_set1_epi32
-#define gmx_simd_store_fi(m, x)     _mm256_store_ps((float *)m, _mm256_castsi256_ps(x))
-#define gmx_simd_loadu_fi(m)       _mm256_castps_si256(_mm256_loadu_ps((const float *)m))
-#define gmx_simd_storeu_fi(m, x)    _mm256_storeu_ps((float *)m, _mm256_castsi256_ps(x))
+#define gmx_simd_store_fi(m, x)    _mm256_store_si256((__m256i *)m, x)
+#define gmx_simd_loadu_fi(m)       _mm256_loadu_si256((__m256i const*)m)
+#define gmx_simd_storeu_fi(m, x)   _mm256_storeu_si256((__m256i *)m, x)
 #define gmx_simd_setzero_fi        _mm256_setzero_si256
 #define gmx_simd_cvt_f2i           _mm256_cvtps_epi32
 #define gmx_simd_cvtt_f2i          _mm256_cvttps_epi32