1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This file is part of GROMACS.
7 * Written by the Gromacs development team under coordination of
8 * David van der Spoel, Berk Hess, and Erik Lindahl.
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.
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
19 * Gnomes, ROck Monsters And Chili Sauce
21 #ifndef _gmx_x86_avx_128_fma_h_
22 #define _gmx_x86_avx_128_fma_h_
25 #include <immintrin.h>
26 #ifdef HAVE_X86INTRIN_H
27 #include <x86intrin.h> /* FMA */
30 #include <intrin.h> /* FMA MSVC */
36 #include "types/simple.h"
39 #define gmx_mm_extract_epi32(x, imm) _mm_cvtsi128_si32(_mm_srli_si128((x), 4 * (imm)))
41 #define _GMX_MM_BLEND(b3, b2, b1, b0) (((b3) << 3) | ((b2) << 2) | ((b1) << 1) | ((b0)))
43 #define _GMX_MM_PERMUTE128D(fp1, fp0) (((fp1) << 1) | ((fp0)))
46 #define GMX_MM_TRANSPOSE2_PD(row0, row1) { \
47 __m128d __gmx_t1 = row0; \
48 row0 = _mm_unpacklo_pd(row0, row1); \
49 row1 = _mm_unpackhi_pd(__gmx_t1, row1); \
53 #if (defined (_MSC_VER) || defined(__INTEL_COMPILER))
54 # define gmx_mm_castsi128_ps(a) _mm_castsi128_ps(a)
55 # define gmx_mm_castps_si128(a) _mm_castps_si128(a)
56 # define gmx_mm_castps_ps128(a) (a)
57 # define gmx_mm_castsi128_pd(a) _mm_castsi128_pd(a)
58 # define gmx_mm_castpd_si128(a) _mm_castpd_si128(a)
59 #elif defined(__GNUC__)
60 # define gmx_mm_castsi128_ps(a) ((__m128)(a))
61 # define gmx_mm_castps_si128(a) ((__m128i)(a))
62 # define gmx_mm_castps_ps128(a) ((__m128)(a))
63 # define gmx_mm_castsi128_pd(a) ((__m128d)(a))
64 # define gmx_mm_castpd_si128(a) ((__m128i)(a))
66 static __m128 gmx_mm_castsi128_ps(__m128i a)
68 return *(__m128 *) &a;
70 static __m128i gmx_mm_castps_si128(__m128 a)
72 return *(__m128i *) &a;
74 static __m128 gmx_mm_castps_ps128(__m128 a)
76 return *(__m128 *) &a;
78 static __m128d gmx_mm_castsi128_pd(__m128i a)
80 return *(__m128d *) &a;
82 static __m128i gmx_mm_castpd_si128(__m128d a)
84 return *(__m128i *) &a;
88 #if GMX_EMULATE_AMD_FMA
89 /* Wrapper routines so we can do test builds on non-FMA or non-AMD hardware */
91 _mm_macc_ps(__m128 a, __m128 b, __m128 c)
93 return _mm_add_ps(c, _mm_mul_ps(a, b));
97 _mm_nmacc_ps(__m128 a, __m128 b, __m128 c)
99 return _mm_sub_ps(c, _mm_mul_ps(a, b));
103 _mm_msub_ps(__m128 a, __m128 b, __m128 c)
105 return _mm_sub_ps(_mm_mul_ps(a, b), c);
109 _mm_macc_pd(__m128d a, __m128d b, __m128d c)
111 return _mm_add_pd(c, _mm_mul_pd(a, b));
115 _mm_nmacc_pd(__m128d a, __m128d b, __m128d c)
117 return _mm_sub_pd(c, _mm_mul_pd(a, b));
121 _mm_msub_pd(__m128d a, __m128d b, __m128d c)
123 return _mm_sub_pd(_mm_mul_pd(a, b), c);
125 #endif /* AMD FMA emulation support */
128 gmx_mm_printxmm_ps(const char *s, __m128 xmm)
132 _mm_storeu_ps(f, xmm);
133 printf("%s: %15.10e %15.10e %15.10e %15.10e\n", s, f[0], f[1], f[2], f[3]);
138 gmx_mm_printxmmsum_ps(const char *s, __m128 xmm)
142 _mm_storeu_ps(f, xmm);
143 printf("%s (sum): %15.10g\n", s, f[0]+f[1]+f[2]+f[3]);
148 gmx_mm_printxmm_pd(const char *s, __m128d xmm)
152 _mm_storeu_pd(f, xmm);
153 printf("%s: %30.20e %30.20e\n", s, f[0], f[1]);
157 gmx_mm_printxmmsum_pd(const char *s, __m128d xmm)
161 _mm_storeu_pd(f, xmm);
162 printf("%s (sum): %15.10g\n", s, f[0]+f[1]);
167 gmx_mm_printxmm_epi32(const char *s, __m128i xmmi)
171 _mm_storeu_si128((__m128i *)i, xmmi);
172 printf("%10s: %2d %2d %2d %2d\n", s, i[0], i[1], i[2], i[3]);
177 static int gmx_mm_check_and_reset_overflow(void)
182 MXCSR = _mm_getcsr();
183 /* The overflow flag is bit 3 in the register */
187 /* Set the overflow flag to zero */
188 MXCSR = MXCSR & 0xFFF7;
199 /* Work around gcc bug with wrong type for mask formal parameter to maskload/maskstore */
200 #ifdef GMX_X86_AVX_GCC_MASKLOAD_BUG
201 # define gmx_mm_maskload_ps(mem, mask) _mm_maskload_ps((mem), _mm_castsi128_ps(mask))
202 # define gmx_mm_maskstore_ps(mem, mask, x) _mm_maskstore_ps((mem), _mm_castsi128_ps(mask), (x))
203 # define gmx_mm256_maskload_ps(mem, mask) _mm256_maskload_ps((mem), _mm256_castsi256_ps(mask))
204 # define gmx_mm256_maskstore_ps(mem, mask, x) _mm256_maskstore_ps((mem), _mm256_castsi256_ps(mask), (x))
206 # define gmx_mm_maskload_ps(mem, mask) _mm_maskload_ps((mem), (mask))
207 # define gmx_mm_maskstore_ps(mem, mask, x) _mm_maskstore_ps((mem), (mask), (x))
208 # define gmx_mm256_maskload_ps(mem, mask) _mm256_maskload_ps((mem), (mask))
209 # define gmx_mm256_maskstore_ps(mem, mask, x) _mm256_maskstore_ps((mem), (mask), (x))
214 #endif /* _gmx_x86_avx_128_fma_h_ */