include/gmx_x86_avx_128_fma.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_avx_128_fma_h_
  22 #define _gmx_x86_avx_128_fma_h_
  23
  24
  25 #include <immintrin.h>
  26 #ifdef HAVE_X86INTRIN_H
  27 #include <x86intrin.h> /* FMA */
  28 #endif
  29
  30 #include <stdio.h>
  31
  32 #include "types/simple.h"
  33
  34
  35 #define gmx_mm_extract_epi32(x, imm) _mm_cvtsi128_si32(_mm_srli_si128((x), 4 * (imm)))
  36
  37
  38 #define _GMX_MM_PERMUTE128D(fp1,fp0)         (((fp1) << 1) | ((fp0)))
  39
  40
  41 #define GMX_MM_TRANSPOSE2_PD(row0, row1) {           \
  42     __m128d __gmx_t1 = row0;                         \
  43     row0           = _mm_unpacklo_pd(row0,row1);     \
  44     row1           = _mm_unpackhi_pd(__gmx_t1,row1); \
  45 }
  46
  47
  48 #if (defined (_MSC_VER) || defined(__INTEL_COMPILER))
  49 #  define gmx_mm_castsi128_ps(a) _mm_castsi128_ps(a)
  50 #  define gmx_mm_castps_si128(a) _mm_castps_si128(a)
  51 #  define gmx_mm_castps_ps128(a) (a)
  52 #  define gmx_mm_castsi128_pd(a) _mm_castsi128_pd(a)
  53 #  define gmx_mm_castpd_si128(a) _mm_castpd_si128(a)
  54 #elif defined(__GNUC__)
  55 #  define gmx_mm_castsi128_ps(a) ((__m128)(a))
  56 #  define gmx_mm_castps_si128(a) ((__m128i)(a))
  57 #  define gmx_mm_castps_ps128(a) ((__m128)(a))
  58 #  define gmx_mm_castsi128_pd(a) ((__m128d)(a))
  59 #  define gmx_mm_castpd_si128(a) ((__m128i)(a))
  60 #else
  61 static __m128  gmx_mm_castsi128_ps(__m128i a)
  62 {
  63     return *(__m128 *) &a;
  64 }
  65 static __m128i gmx_mm_castps_si128(__m128 a)
  66 {
  67     return *(__m128i *) &a;
  68 }
  69 static __m128  gmx_mm_castps_ps128(__m128 a)
  70 {
  71     return *(__m128 *) &a;
  72 }
  73 static __m128d gmx_mm_castsi128_pd(__m128i a)
  74 {
  75     return *(__m128d *) &a;
  76 }
  77 static __m128i gmx_mm_castpd_si128(__m128d a)
  78 {
  79     return *(__m128i *) &a;
  80 }
  81 #endif
  82
  83 #ifndef _MSC_VER
  84 /* The warning directive is not supported by MSVC, and that compiler
  85  * does not support overriding built-in functions anyway...
  86  */
  87 #if !defined(HAVE_X86INTRIN_H) || !defined(__FMA4__)
  88 #warning Emulating FMA instructions - this is probably not what you want!
  89 /* Wrapper routines so we can do test builds on non-FMA hardware */
  90 static __m128
  91 _mm_macc_ps(__m128 a, __m128 b, __m128 c)
  92 {
  93
  94     return _mm_add_ps(c,_mm_mul_ps(a,b));
  95 }
  96
  97 static __m128
  98 _mm_nmacc_ps(__m128 a, __m128 b, __m128 c)
  99 {
 100
 101     return _mm_sub_ps(c,_mm_mul_ps(a,b));
 102 }
 103
 104 static __m128d
 105 _mm_macc_pd(__m128d a, __m128d b, __m128d c)
 106 {
 107
 108     return _mm_add_pd(c,_mm_mul_pd(a,b));
 109 }
 110
 111 static __m128d
 112 _mm_nmacc_pd(__m128d a, __m128d b, __m128d c)
 113 {
 114
 115     return _mm_sub_pd(c,_mm_mul_pd(a,b));
 116 }
 117 #endif /* FMA4 support */
 118
 119 #endif /* _MSC_VER */
 120
 121 static void
 122 gmx_mm_printxmm_ps(const char *s,__m128 xmm)
 123 {
 124     float f[4];
 125
 126     _mm_storeu_ps(f,xmm);
 127     printf("%s: %15.10e %15.10e %15.10e %15.10e\n",s,f[0],f[1],f[2],f[3]);
 128 }
 129
 130
 131 static void
 132 gmx_mm_printxmmsum_ps(const char *s,__m128 xmm)
 133 {
 134     float f[4];
 135
 136     _mm_storeu_ps(f,xmm);
 137     printf("%s (sum): %15.10g\n",s,f[0]+f[1]+f[2]+f[3]);
 138 }
 139
 140
 141 static void
 142 gmx_mm_printxmm_pd(const char *s,__m128d xmm)
 143 {
 144     double f[2];
 145
 146     _mm_storeu_pd(f,xmm);
 147     printf("%s: %30.20e %30.20e\n",s,f[0],f[1]);
 148 }
 149
 150 static void
 151 gmx_mm_printxmmsum_pd(const char *s,__m128d xmm)
 152 {
 153     double f[2];
 154
 155     _mm_storeu_pd(f,xmm);
 156     printf("%s (sum): %15.10g\n",s,f[0]+f[1]);
 157 }
 158
 159
 160 static void
 161 gmx_mm_printxmm_epi32(const char *s,__m128i xmmi)
 162 {
 163     int i[4];
 164
 165     _mm_storeu_si128((__m128i *)i,xmmi);
 166     printf("%10s: %2d %2d %2d %2d\n",s,i[0],i[1],i[2],i[3]);
 167 }
 168
 169
 170
 171 static int gmx_mm_check_and_reset_overflow(void)
 172 {
 173     int MXCSR;
 174     int sse_overflow;
 175
 176     MXCSR = _mm_getcsr();
 177     /* The overflow flag is bit 3 in the register */
 178     if (MXCSR & 0x0008)
 179     {
 180         sse_overflow = 1;
 181         /* Set the overflow flag to zero */
 182         MXCSR = MXCSR & 0xFFF7;
 183         _mm_setcsr(MXCSR);
 184     }
 185     else
 186     {
 187         sse_overflow = 0;
 188     }
 189
 190     return sse_overflow;
 191 }
 192
 193
 194
 195 #endif /* _gmx_x86_avx_128_fma_h_ */