1 /**************************************************************************
2  *
3  * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas.
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 
28 /**
29  * @file
30  * SSE intrinsics portability header.
31  *
32  * Although the SSE intrinsics are support by all modern x86 and x86-64
33  * compilers, there are some intrisincs missing in some implementations
34  * (especially older MSVC versions). This header abstracts that away.
35  */
36 
37 #ifndef U_SSE_H_
38 #define U_SSE_H_
39 
40 #include "pipe/p_config.h"
41 
42 #if defined(PIPE_ARCH_SSE)
43 
44 #include <emmintrin.h>
45 
46 
47 /* MSVC before VC8 does not support the _mm_castxxx_yyy */
48 #if defined(_MSC_VER) && _MSC_VER < 1500
49 
50 union __declspec(align(16)) m128_types {
51    __m128 m128;
52    __m128i m128i;
53    __m128d m128d;
54 };
55 
56 static __inline __m128
_mm_castsi128_ps(__m128i a)57 _mm_castsi128_ps(__m128i a)
58 {
59    union m128_types u;
60    u.m128i = a;
61    return u.m128;
62 }
63 
64 static __inline __m128i
_mm_castps_si128(__m128 a)65 _mm_castps_si128(__m128 a)
66 {
67    union m128_types u;
68    u.m128 = a;
69    return u.m128i;
70 }
71 
72 #endif /* defined(_MSC_VER) && _MSC_VER < 1500 */
73 
74 union m128i {
75    __m128i m;
76    ubyte ub[16];
77    ushort us[8];
78    uint ui[4];
79 };
80 
u_print_epi8(const char * name,__m128i r)81 static INLINE void u_print_epi8(const char *name, __m128i r)
82 {
83    union { __m128i m; ubyte ub[16]; } u;
84    u.m = r;
85 
86    debug_printf("%s: "
87                 "%02x/"
88                 "%02x/"
89                 "%02x/"
90                 "%02x/"
91                 "%02x/"
92                 "%02x/"
93                 "%02x/"
94                 "%02x/"
95                 "%02x/"
96                 "%02x/"
97                 "%02x/"
98                 "%02x/"
99                 "%02x/"
100                 "%02x/"
101                 "%02x/"
102                 "%02x\n",
103                 name,
104                 u.ub[0],  u.ub[1],  u.ub[2],  u.ub[3],
105                 u.ub[4],  u.ub[5],  u.ub[6],  u.ub[7],
106                 u.ub[8],  u.ub[9],  u.ub[10], u.ub[11],
107                 u.ub[12], u.ub[13], u.ub[14], u.ub[15]);
108 }
109 
u_print_epi16(const char * name,__m128i r)110 static INLINE void u_print_epi16(const char *name, __m128i r)
111 {
112    union { __m128i m; ushort us[8]; } u;
113    u.m = r;
114 
115    debug_printf("%s: "
116                 "%04x/"
117                 "%04x/"
118                 "%04x/"
119                 "%04x/"
120                 "%04x/"
121                 "%04x/"
122                 "%04x/"
123                 "%04x\n",
124                 name,
125                 u.us[0],  u.us[1],  u.us[2],  u.us[3],
126                 u.us[4],  u.us[5],  u.us[6],  u.us[7]);
127 }
128 
u_print_epi32(const char * name,__m128i r)129 static INLINE void u_print_epi32(const char *name, __m128i r)
130 {
131    union { __m128i m; uint ui[4]; } u;
132    u.m = r;
133 
134    debug_printf("%s: "
135                 "%08x/"
136                 "%08x/"
137                 "%08x/"
138                 "%08x\n",
139                 name,
140                 u.ui[0],  u.ui[1],  u.ui[2],  u.ui[3]);
141 }
142 
u_print_ps(const char * name,__m128 r)143 static INLINE void u_print_ps(const char *name, __m128 r)
144 {
145    union { __m128 m; float f[4]; } u;
146    u.m = r;
147 
148    debug_printf("%s: "
149                 "%f/"
150                 "%f/"
151                 "%f/"
152                 "%f\n",
153                 name,
154                 u.f[0],  u.f[1],  u.f[2],  u.f[3]);
155 }
156 
157 
158 #define U_DUMP_EPI32(a) u_print_epi32(#a, a)
159 #define U_DUMP_EPI16(a) u_print_epi16(#a, a)
160 #define U_DUMP_EPI8(a)  u_print_epi8(#a, a)
161 #define U_DUMP_PS(a)    u_print_ps(#a, a)
162 
163 
164 
165 #if defined(PIPE_ARCH_SSSE3)
166 
167 #include <tmmintrin.h>
168 
169 #else /* !PIPE_ARCH_SSSE3 */
170 
171 /**
172  * Describe _mm_shuffle_epi8() with gcc extended inline assembly, for cases
173  * where -mssse3 is not supported/enabled.
174  *
175  * MSVC will never get in here as its intrinsics support do not rely on
176  * compiler command line options.
177  */
178 static __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_shuffle_epi8(__m128i a,__m128i mask)179 _mm_shuffle_epi8(__m128i a, __m128i mask)
180 {
181     __m128i result;
182     __asm__("pshufb %1, %0"
183             : "=x" (result)
184             : "xm" (mask), "0" (a));
185     return result;
186 }
187 
188 #endif /* !PIPE_ARCH_SSSE3 */
189 
190 
191 
192 
193 /* Provide an SSE2 implementation of _mm_mullo_epi32() in terms of
194  * _mm_mul_epu32().
195  *
196  * I suspect this works fine for us because one of our operands is
197  * always positive, but not sure that this can be used for general
198  * signed integer multiplication.
199  *
200  * This seems close enough to the speed of SSE4 and the real
201  * _mm_mullo_epi32() intrinsic as to not justify adding an sse4
202  * dependency at this point.
203  */
mm_mullo_epi32(const __m128i a,const __m128i b)204 static INLINE __m128i mm_mullo_epi32(const __m128i a, const __m128i b)
205 {
206    __m128i a4   = _mm_srli_epi64(a, 32);  /* shift by one dword */
207    __m128i b4   = _mm_srli_epi64(b, 32);  /* shift by one dword */
208    __m128i ba   = _mm_mul_epu32(b, a);   /* multply dwords 0, 2 */
209    __m128i b4a4 = _mm_mul_epu32(b4, a4); /* multiply dwords 1, 3 */
210 
211    /* Interleave the results, either with shuffles or (slightly
212     * faster) direct bit operations:
213     */
214 #if 0
215    __m128i ba8             = _mm_shuffle_epi32(ba, 8);
216    __m128i b4a48           = _mm_shuffle_epi32(b4a4, 8);
217    __m128i result          = _mm_unpacklo_epi32(ba8, b4a48);
218 #else
219    __m128i mask            = _mm_setr_epi32(~0,0,~0,0);
220    __m128i ba_mask         = _mm_and_si128(ba, mask);
221    __m128i b4a4_mask_shift = _mm_slli_epi64(b4a4, 32);
222    __m128i result          = _mm_or_si128(ba_mask, b4a4_mask_shift);
223 #endif
224 
225    return result;
226 }
227 
228 
229 static INLINE void
transpose4_epi32(const __m128i * restrict a,const __m128i * restrict b,const __m128i * restrict c,const __m128i * restrict d,__m128i * restrict o,__m128i * restrict p,__m128i * restrict q,__m128i * restrict r)230 transpose4_epi32(const __m128i * restrict a,
231                  const __m128i * restrict b,
232                  const __m128i * restrict c,
233                  const __m128i * restrict d,
234                  __m128i * restrict o,
235                  __m128i * restrict p,
236                  __m128i * restrict q,
237                  __m128i * restrict r)
238 {
239   __m128i t0 = _mm_unpacklo_epi32(*a, *b);
240   __m128i t1 = _mm_unpacklo_epi32(*c, *d);
241   __m128i t2 = _mm_unpackhi_epi32(*a, *b);
242   __m128i t3 = _mm_unpackhi_epi32(*c, *d);
243 
244   *o = _mm_unpacklo_epi64(t0, t1);
245   *p = _mm_unpackhi_epi64(t0, t1);
246   *q = _mm_unpacklo_epi64(t2, t3);
247   *r = _mm_unpackhi_epi64(t2, t3);
248 }
249 
250 #define SCALAR_EPI32(m, i) _mm_shuffle_epi32((m), _MM_SHUFFLE(i,i,i,i))
251 
252 
253 #endif /* PIPE_ARCH_SSE */
254 
255 #endif /* U_SSE_H_ */
256