1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/cairo/libpixman/src/pixman-x64-mmx-emulation.h Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,263 @@
1.4 +#ifndef MMX_X64_H_INCLUDED
1.5 +#define MMX_X64_H_INCLUDED
1.6 +
1.7 +/* Implementation of x64 MMX substitition functions, before
1.8 + * pixman is reimplemented not to use __m64 type on Visual C++
1.9 + *
1.10 + * Copyright (C)2009 by George Yohng
1.11 + * Released in public domain.
1.12 + */
1.13 +
1.14 +#include <intrin.h>
1.15 +
1.16 +#define M64C(a) (*(const __m64 *)(&a))
1.17 +#define M64U(a) (*(const unsigned long long *)(&a))
1.18 +
1.19 +__inline __m64
1.20 +_m_from_int (int a)
1.21 +{
1.22 + long long i64 = a;
1.23 +
1.24 + return M64C (i64);
1.25 +}
1.26 +
1.27 +__inline __m64
1.28 +_mm_setzero_si64 ()
1.29 +{
1.30 + long long i64 = 0;
1.31 +
1.32 + return M64C (i64);
1.33 +}
1.34 +
1.35 +__inline __m64
1.36 +_mm_set_pi32 (int i1, int i0)
1.37 +{
1.38 + unsigned long long i64 = ((unsigned)i0) + (((unsigned long long)(unsigned)i1) << 32);
1.39 +
1.40 + return M64C (i64);
1.41 +}
1.42 +
1.43 +__inline void
1.44 +_m_empty ()
1.45 +{
1.46 +}
1.47 +
1.48 +__inline __m64
1.49 +_mm_set1_pi16 (short w)
1.50 +{
1.51 + unsigned long long i64 = ((unsigned long long)(unsigned short)(w)) * 0x0001000100010001ULL;
1.52 +
1.53 + return M64C (i64);
1.54 +}
1.55 +
1.56 +__inline int
1.57 +_m_to_int (__m64 m)
1.58 +{
1.59 + return m.m64_i32[0];
1.60 +}
1.61 +
1.62 +__inline __m64
1.63 +_mm_movepi64_pi64 (__m128i a)
1.64 +{
1.65 + return M64C (a.m128i_i64[0]);
1.66 +}
1.67 +
1.68 +__inline __m64
1.69 +_m_pand (__m64 a, __m64 b)
1.70 +{
1.71 + unsigned long long i64 = M64U (a) & M64U (b);
1.72 +
1.73 + return M64C (i64);
1.74 +}
1.75 +
1.76 +__inline __m64
1.77 +_m_por (__m64 a, __m64 b)
1.78 +{
1.79 + unsigned long long i64 = M64U (a) | M64U (b);
1.80 +
1.81 + return M64C (i64);
1.82 +}
1.83 +
1.84 +__inline __m64
1.85 +_m_pxor (__m64 a, __m64 b)
1.86 +{
1.87 + unsigned long long i64 = M64U (a) ^ M64U (b);
1.88 +
1.89 + return M64C (i64);
1.90 +}
1.91 +
1.92 +__inline __m64
1.93 +_m_pmulhuw (__m64 a, __m64 b) /* unoptimized */
1.94 +{
1.95 + unsigned short d[4] =
1.96 + {
1.97 + (unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0]) >> 16),
1.98 + (unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1]) >> 16),
1.99 + (unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2]) >> 16),
1.100 + (unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]) >> 16)
1.101 + };
1.102 +
1.103 + return M64C (d[0]);
1.104 +}
1.105 +
1.106 +__inline __m64
1.107 +_m_pmullw2 (__m64 a, __m64 b) /* unoptimized */
1.108 +{
1.109 + unsigned short d[4] =
1.110 + {
1.111 + (unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0])),
1.112 + (unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1])),
1.113 + (unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2])),
1.114 + (unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]))
1.115 + };
1.116 +
1.117 + return M64C (d[0]);
1.118 +}
1.119 +
1.120 +__inline __m64
1.121 +_m_pmullw (__m64 a, __m64 b) /* unoptimized */
1.122 +{
1.123 + unsigned long long x =
1.124 + ((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0]))) +
1.125 + (((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1]))) << 16) +
1.126 + (((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2]))) << 32) +
1.127 + (((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]))) << 48);
1.128 +
1.129 + return M64C (x);
1.130 +}
1.131 +
1.132 +__inline __m64
1.133 +_m_paddusb (__m64 a, __m64 b) /* unoptimized */
1.134 +{
1.135 + unsigned long long x = (M64U (a) & 0x00FF00FF00FF00FFULL) +
1.136 + (M64U (b) & 0x00FF00FF00FF00FFULL);
1.137 +
1.138 + unsigned long long y = ((M64U (a) >> 8) & 0x00FF00FF00FF00FFULL) +
1.139 + ((M64U (b) >> 8) & 0x00FF00FF00FF00FFULL);
1.140 +
1.141 + x |= ((x & 0xFF00FF00FF00FF00ULL) >> 8) * 0xFF;
1.142 + y |= ((y & 0xFF00FF00FF00FF00ULL) >> 8) * 0xFF;
1.143 +
1.144 + x = (x & 0x00FF00FF00FF00FFULL) | ((y & 0x00FF00FF00FF00FFULL) << 8);
1.145 +
1.146 + return M64C (x);
1.147 +}
1.148 +
1.149 +__inline __m64
1.150 +_m_paddusw (__m64 a, __m64 b) /* unoptimized */
1.151 +{
1.152 + unsigned long long x = (M64U (a) & 0x0000FFFF0000FFFFULL) +
1.153 + (M64U (b) & 0x0000FFFF0000FFFFULL);
1.154 +
1.155 + unsigned long long y = ((M64U (a) >> 16) & 0x0000FFFF0000FFFFULL) +
1.156 + ((M64U (b) >> 16) & 0x0000FFFF0000FFFFULL);
1.157 +
1.158 + x |= ((x & 0xFFFF0000FFFF0000) >> 16) * 0xFFFF;
1.159 + y |= ((y & 0xFFFF0000FFFF0000) >> 16) * 0xFFFF;
1.160 +
1.161 + x = (x & 0x0000FFFF0000FFFFULL) | ((y & 0x0000FFFF0000FFFFULL) << 16);
1.162 +
1.163 + return M64C (x);
1.164 +}
1.165 +
1.166 +__inline __m64
1.167 +_m_pshufw (__m64 a, int n) /* unoptimized */
1.168 +{
1.169 + unsigned short d[4] =
1.170 + {
1.171 + a.m64_u16[n & 3],
1.172 + a.m64_u16[(n >> 2) & 3],
1.173 + a.m64_u16[(n >> 4) & 3],
1.174 + a.m64_u16[(n >> 6) & 3]
1.175 + };
1.176 +
1.177 + return M64C (d[0]);
1.178 +}
1.179 +
1.180 +__inline unsigned char
1.181 +sat16 (unsigned short d)
1.182 +{
1.183 + if (d > 0xFF) return 0xFF;
1.184 + else return d & 0xFF;
1.185 +}
1.186 +
1.187 +__inline __m64
1.188 +_m_packuswb (__m64 m1, __m64 m2) /* unoptimized */
1.189 +{
1.190 + unsigned char d[8] =
1.191 + {
1.192 + sat16 (m1.m64_u16[0]),
1.193 + sat16 (m1.m64_u16[1]),
1.194 + sat16 (m1.m64_u16[2]),
1.195 + sat16 (m1.m64_u16[3]),
1.196 + sat16 (m2.m64_u16[0]),
1.197 + sat16 (m2.m64_u16[1]),
1.198 + sat16 (m2.m64_u16[2]),
1.199 + sat16 (m2.m64_u16[3])
1.200 + };
1.201 +
1.202 + return M64C (d[0]);
1.203 +}
1.204 +
1.205 +__inline __m64 _m_punpcklbw (__m64 m1, __m64 m2) /* unoptimized */
1.206 +{
1.207 + unsigned char d[8] =
1.208 + {
1.209 + m1.m64_u8[0],
1.210 + m2.m64_u8[0],
1.211 + m1.m64_u8[1],
1.212 + m2.m64_u8[1],
1.213 + m1.m64_u8[2],
1.214 + m2.m64_u8[2],
1.215 + m1.m64_u8[3],
1.216 + m2.m64_u8[3],
1.217 + };
1.218 +
1.219 + return M64C (d[0]);
1.220 +}
1.221 +
1.222 +__inline __m64 _m_punpckhbw (__m64 m1, __m64 m2) /* unoptimized */
1.223 +{
1.224 + unsigned char d[8] =
1.225 + {
1.226 + m1.m64_u8[4],
1.227 + m2.m64_u8[4],
1.228 + m1.m64_u8[5],
1.229 + m2.m64_u8[5],
1.230 + m1.m64_u8[6],
1.231 + m2.m64_u8[6],
1.232 + m1.m64_u8[7],
1.233 + m2.m64_u8[7],
1.234 + };
1.235 +
1.236 + return M64C (d[0]);
1.237 +}
1.238 +
1.239 +__inline __m64 _m_psrlwi (__m64 a, int n) /* unoptimized */
1.240 +{
1.241 + unsigned short d[4] =
1.242 + {
1.243 + a.m64_u16[0] >> n,
1.244 + a.m64_u16[1] >> n,
1.245 + a.m64_u16[2] >> n,
1.246 + a.m64_u16[3] >> n
1.247 + };
1.248 +
1.249 + return M64C (d[0]);
1.250 +}
1.251 +
1.252 +__inline __m64 _m_psrlqi (__m64 m, int n)
1.253 +{
1.254 + unsigned long long x = M64U (m) >> n;
1.255 +
1.256 + return M64C (x);
1.257 +}
1.258 +
1.259 +__inline __m64 _m_psllqi (__m64 m, int n)
1.260 +{
1.261 + unsigned long long x = M64U (m) << n;
1.262 +
1.263 + return M64C (x);
1.264 +}
1.265 +
1.266 +#endif /* MMX_X64_H_INCLUDED */
