include/gmx_x86_sse2.h

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  * This file is part of GROMACS.
   5  * Copyright (c) 2012-
   6  *
   7  * Written by the Gromacs development team under coordination of
   8  * David van der Spoel, Berk Hess, and Erik Lindahl.
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2
  13  * of the License, or (at your option) any later version.
  14  *
  15  * To help us fund GROMACS development, we humbly ask that you cite
  16  * the research papers on the package. Check out http://www.gromacs.org
  17  *
  18  * And Hey:
  19  * Gnomes, ROck Monsters And Chili Sauce
  20  */
  21 #ifndef _gmx_x86_sse2_h_
  22 #define _gmx_x86_sse2_h_
  23
  24 #include <emmintrin.h>
  25
  26 #include <stdio.h>
  27
  28 #include "types/simple.h"
  29
  30
  31
  32 /* Create some basic definitions that are not 100% SSE2 standard and thus not
  33  * available on all compilers. These should be fairly self-evident by comparing
  34  * with an arbitrary emmintrin.h.
  35  */
  36
  37 #define gmx_mm_extract_epi32(x, imm) _mm_cvtsi128_si32(_mm_srli_si128((x), 4 * (imm)))
  38
  39 #define GMX_MM_TRANSPOSE2_PD(row0, row1) {           \
  40     __m128d __gmx_t1 = row0;                         \
  41     row0           = _mm_unpacklo_pd(row0,row1);     \
  42     row1           = _mm_unpackhi_pd(__gmx_t1,row1); \
  43 }
  44
  45
  46 #if (defined (_MSC_VER) || defined(__INTEL_COMPILER))
  47 #  define gmx_mm_castsi128_ps(a) _mm_castsi128_ps(a)
  48 #  define gmx_mm_castps_si128(a) _mm_castps_si128(a)
  49 #  define gmx_mm_castps_ps128(a) (a)
  50 #  define gmx_mm_castsi128_pd(a) _mm_castsi128_pd(a)
  51 #  define gmx_mm_castpd_si128(a) _mm_castpd_si128(a)
  52 #elif defined(__GNUC__)
  53 #  define gmx_mm_castsi128_ps(a) ((__m128)(a))
  54 #  define gmx_mm_castps_si128(a) ((__m128i)(a))
  55 #  define gmx_mm_castps_ps128(a) ((__m128)(a))
  56 #  define gmx_mm_castsi128_pd(a) ((__m128d)(a))
  57 #  define gmx_mm_castpd_si128(a) ((__m128i)(a))
  58 #else
  59 static __m128  gmx_mm_castsi128_ps(__m128i a)
  60 {
  61     return *(__m128 *) &a;
  62 }
  63 static __m128i gmx_mm_castps_si128(__m128 a)
  64 {
  65     return *(__m128i *) &a;
  66 }
  67 static __m128  gmx_mm_castps_ps128(__m128 a)
  68 {
  69     return *(__m128 *) &a;
  70 }
  71 static __m128d gmx_mm_castsi128_pd(__m128i a)
  72 {
  73     return *(__m128d *) &a;
  74 }
  75 static __m128i gmx_mm_castpd_si128(__m128d a)
  76 {
  77     return *(__m128i *) &a;
  78 }
  79 #endif
  80
  81
  82 static void
  83 gmx_mm_printxmm_ps(const char *s,__m128 xmm)
  84 {
  85     float f[4];
  86
  87     _mm_storeu_ps(f,xmm);
  88     printf("%s: %15.10e %15.10e %15.10e %15.10e\n",s,f[0],f[1],f[2],f[3]);
  89 }
  90
  91
  92 static void
  93 gmx_mm_printxmmsum_ps(const char *s,__m128 xmm)
  94 {
  95     float f[4];
  96
  97     _mm_storeu_ps(f,xmm);
  98     printf("%s (sum): %15.10g\n",s,f[0]+f[1]+f[2]+f[3]);
  99 }
 100
 101
 102 static void
 103 gmx_mm_printxmm_pd(const char *s,__m128d xmm)
 104 {
 105     double f[2];
 106
 107     _mm_storeu_pd(f,xmm);
 108     printf("%s: %30.20e %30.20e\n",s,f[0],f[1]);
 109 }
 110
 111 static void
 112 gmx_mm_printxmmsum_pd(const char *s,__m128d xmm)
 113 {
 114     double f[2];
 115
 116     _mm_storeu_pd(f,xmm);
 117     printf("%s (sum): %15.10g\n",s,f[0]+f[1]);
 118 }
 119
 120
 121 static void
 122 gmx_mm_printxmm_epi32(const char *s,__m128i xmmi)
 123 {
 124     int i[4];
 125
 126     _mm_storeu_si128((__m128i *)i,xmmi);
 127     printf("%10s: %2d %2d %2d %2d\n",s,i[0],i[1],i[2],i[3]);
 128 }
 129
 130
 131
 132 static int gmx_mm_check_and_reset_overflow(void)
 133 {
 134     int MXCSR;
 135     int sse_overflow;
 136
 137     MXCSR = _mm_getcsr();
 138     /* The overflow flag is bit 3 in the register */
 139     if (MXCSR & 0x0008)
 140     {
 141         sse_overflow = 1;
 142         /* Set the overflow flag to zero */
 143         MXCSR = MXCSR & 0xFFF7;
 144         _mm_setcsr(MXCSR);
 145     }
 146     else
 147     {
 148         sse_overflow = 0;
 149     }
 150
 151     return sse_overflow;
 152 }
 153
 154
 155 #endif /* _gmx_x86_sse2_h_ */
 156