Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / legacyheaders / gmx_x86_avx_128_fma.h

diff --git a/src/gromacs/legacyheaders/gmx_x86_avx_128_fma.h b/src/gromacs/legacyheaders/gmx_x86_avx_128_fma.h
index 821924db56f1df9765f91c57502cb1f7b1d9f429..13a02078b02ca95efa2b8540efaa82110d34adae 100644 (file)
--- a/src/gromacs/legacyheaders/gmx_x86_avx_128_fma.h
+++ b/src/gromacs/legacyheaders/gmx_x86_avx_128_fma.h
@@ -34,6 +34,7 @@
 
 #define gmx_mm_extract_epi32(x, imm) _mm_cvtsi128_si32(_mm_srli_si128((x), 4 * (imm)))
 
+#define _GMX_MM_BLEND(b3,b2,b1,b0) (((b3) << 3) | ((b2) << 2) | ((b1) << 1) | ((b0)))
 
 #define _GMX_MM_PERMUTE128D(fp1,fp0)         (((fp1) << 1) | ((fp0)))
 
@@ -80,43 +81,44 @@ static __m128i gmx_mm_castpd_si128(__m128d a)
 }
 #endif
 
-#ifndef _MSC_VER
-/* The warning directive is not supported by MSVC, and that compiler
- * does not support overriding built-in functions anyway...
- */
-#if !defined(HAVE_X86INTRIN_H) || !defined(__FMA4__)
-#warning Emulating FMA instructions - this is probably not what you want!
-/* Wrapper routines so we can do test builds on non-FMA hardware */
+#if GMX_EMULATE_AMD_FMA
+/* Wrapper routines so we can do test builds on non-FMA or non-AMD hardware */
 static __m128
 _mm_macc_ps(__m128 a, __m128 b, __m128 c)
 {
-
     return _mm_add_ps(c,_mm_mul_ps(a,b));
 }
 
 static __m128
 _mm_nmacc_ps(__m128 a, __m128 b, __m128 c)
 {
-
     return _mm_sub_ps(c,_mm_mul_ps(a,b));
 }
 
+static __m128
+_mm_msub_ps(__m128 a, __m128 b, __m128 c)
+{
+    return _mm_sub_ps(_mm_mul_ps(a,b),c);
+}
+
 static __m128d
 _mm_macc_pd(__m128d a, __m128d b, __m128d c)
 {
-
     return _mm_add_pd(c,_mm_mul_pd(a,b));
 }
 
 static __m128d
 _mm_nmacc_pd(__m128d a, __m128d b, __m128d c)
 {
-
     return _mm_sub_pd(c,_mm_mul_pd(a,b));
 }
-#endif /* FMA4 support */
 
-#endif /* _MSC_VER */
+static __m128d
+_mm_msub_pd(__m128d a, __m128d b, __m128d c)
+{
+    return _mm_sub_pd(_mm_mul_pd(a,b),c);
+}
+#endif /* AMD FMA emulation support */
 
 static void
 gmx_mm_printxmm_ps(const char *s,__m128 xmm)