michael@0: #ifndef MMX_X64_H_INCLUDED
michael@0: #define MMX_X64_H_INCLUDED
michael@0: 
michael@0: /* Implementation of x64 MMX substitition functions, before
michael@0:  * pixman is reimplemented not to use __m64 type on Visual C++
michael@0:  *
michael@0:  * Copyright (C)2009 by George Yohng
michael@0:  * Released in public domain.
michael@0:  */
michael@0: 
michael@0: #include <intrin.h>
michael@0: 
michael@0: #define M64C(a) (*(const __m64 *)(&a))
michael@0: #define M64U(a) (*(const unsigned long long *)(&a))
michael@0: 
michael@0: __inline __m64
michael@0: _m_from_int (int a)
michael@0: {
michael@0:     long long i64 = a;
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _mm_setzero_si64 ()
michael@0: {
michael@0:     long long i64 = 0;
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _mm_set_pi32 (int i1,   int i0)
michael@0: {
michael@0:     unsigned long long i64 = ((unsigned)i0) + (((unsigned long long)(unsigned)i1) << 32);
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline void
michael@0: _m_empty ()
michael@0: {
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _mm_set1_pi16 (short w)
michael@0: {
michael@0:     unsigned long long i64 = ((unsigned long long)(unsigned short)(w)) * 0x0001000100010001ULL;
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline int
michael@0: _m_to_int (__m64 m)
michael@0: {
michael@0:     return m.m64_i32[0];
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _mm_movepi64_pi64 (__m128i a)
michael@0: {
michael@0:     return M64C (a.m128i_i64[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_pand (__m64 a, __m64 b)
michael@0: {
michael@0:     unsigned long long i64 = M64U (a) & M64U (b);
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_por (__m64 a, __m64 b)
michael@0: {
michael@0:     unsigned long long i64 = M64U (a) | M64U (b);
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_pxor (__m64 a, __m64 b)
michael@0: {
michael@0:     unsigned long long i64 = M64U (a) ^ M64U (b);
michael@0: 
michael@0:     return M64C (i64);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_pmulhuw (__m64 a, __m64 b)        /* unoptimized */
michael@0: {
michael@0:     unsigned short d[4] =
michael@0:     {
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0]) >> 16),
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1]) >> 16),
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2]) >> 16),
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]) >> 16)
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_pmullw2 (__m64 a, __m64 b)        /* unoptimized */
michael@0: {
michael@0:     unsigned short d[4] =
michael@0:     {
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0])),
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1])),
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2])),
michael@0: 	(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]))
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_pmullw (__m64 a, __m64 b)        /* unoptimized */
michael@0: {
michael@0:     unsigned long long x =
michael@0: 	((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0])))  +
michael@0: 	(((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1]))) << 16)  +
michael@0: 	(((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2]))) << 32)  +
michael@0: 	(((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]))) << 48);
michael@0: 
michael@0:     return M64C (x);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_paddusb (__m64 a, __m64 b)        /* unoptimized */
michael@0: {
michael@0:     unsigned long long x = (M64U (a) & 0x00FF00FF00FF00FFULL) +
michael@0:                            (M64U (b) & 0x00FF00FF00FF00FFULL);
michael@0: 
michael@0:     unsigned long long y = ((M64U (a) >> 8) & 0x00FF00FF00FF00FFULL) +
michael@0:                            ((M64U (b) >> 8) & 0x00FF00FF00FF00FFULL);
michael@0: 
michael@0:     x |= ((x & 0xFF00FF00FF00FF00ULL) >> 8) * 0xFF;
michael@0:     y |= ((y & 0xFF00FF00FF00FF00ULL) >> 8) * 0xFF;
michael@0: 
michael@0:     x = (x & 0x00FF00FF00FF00FFULL) | ((y & 0x00FF00FF00FF00FFULL) << 8);
michael@0: 
michael@0:     return M64C (x);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_paddusw (__m64 a, __m64 b)        /* unoptimized */
michael@0: {
michael@0:     unsigned long long x = (M64U (a) & 0x0000FFFF0000FFFFULL) +
michael@0:                            (M64U (b) & 0x0000FFFF0000FFFFULL);
michael@0: 
michael@0:     unsigned long long y = ((M64U (a) >> 16) & 0x0000FFFF0000FFFFULL) +
michael@0:                            ((M64U (b) >> 16) & 0x0000FFFF0000FFFFULL);
michael@0: 
michael@0:     x |= ((x & 0xFFFF0000FFFF0000) >> 16) * 0xFFFF;
michael@0:     y |= ((y & 0xFFFF0000FFFF0000) >> 16) * 0xFFFF;
michael@0: 
michael@0:     x = (x & 0x0000FFFF0000FFFFULL) | ((y & 0x0000FFFF0000FFFFULL) << 16);
michael@0: 
michael@0:     return M64C (x);
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_pshufw (__m64 a, int n)         /* unoptimized */
michael@0: {
michael@0:     unsigned short d[4] =
michael@0:     {
michael@0: 	a.m64_u16[n & 3],
michael@0: 	a.m64_u16[(n >> 2) & 3],
michael@0: 	a.m64_u16[(n >> 4) & 3],
michael@0: 	a.m64_u16[(n >> 6) & 3]
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline unsigned char
michael@0: sat16 (unsigned short d)
michael@0: {
michael@0:     if (d > 0xFF) return 0xFF;
michael@0:     else return d & 0xFF;
michael@0: }
michael@0: 
michael@0: __inline __m64
michael@0: _m_packuswb (__m64 m1, __m64 m2)          /* unoptimized */
michael@0: {
michael@0:     unsigned char d[8] =
michael@0:     {
michael@0: 	sat16 (m1.m64_u16[0]),
michael@0: 	sat16 (m1.m64_u16[1]),
michael@0: 	sat16 (m1.m64_u16[2]),
michael@0: 	sat16 (m1.m64_u16[3]),
michael@0: 	sat16 (m2.m64_u16[0]),
michael@0: 	sat16 (m2.m64_u16[1]),
michael@0: 	sat16 (m2.m64_u16[2]),
michael@0: 	sat16 (m2.m64_u16[3])
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64 _m_punpcklbw (__m64 m1, __m64 m2)          /* unoptimized */
michael@0: {
michael@0:     unsigned char d[8] =
michael@0:     {
michael@0: 	m1.m64_u8[0],
michael@0: 	m2.m64_u8[0],
michael@0: 	m1.m64_u8[1],
michael@0: 	m2.m64_u8[1],
michael@0: 	m1.m64_u8[2],
michael@0: 	m2.m64_u8[2],
michael@0: 	m1.m64_u8[3],
michael@0: 	m2.m64_u8[3],
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64 _m_punpckhbw (__m64 m1, __m64 m2)          /* unoptimized */
michael@0: {
michael@0:     unsigned char d[8] =
michael@0:     {
michael@0: 	m1.m64_u8[4],
michael@0: 	m2.m64_u8[4],
michael@0: 	m1.m64_u8[5],
michael@0: 	m2.m64_u8[5],
michael@0: 	m1.m64_u8[6],
michael@0: 	m2.m64_u8[6],
michael@0: 	m1.m64_u8[7],
michael@0: 	m2.m64_u8[7],
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64 _m_psrlwi (__m64 a, int n)       /* unoptimized */
michael@0: {
michael@0:     unsigned short d[4] =
michael@0:     {
michael@0: 	a.m64_u16[0] >> n,
michael@0: 	a.m64_u16[1] >> n,
michael@0: 	a.m64_u16[2] >> n,
michael@0: 	a.m64_u16[3] >> n
michael@0:     };
michael@0: 
michael@0:     return M64C (d[0]);
michael@0: }
michael@0: 
michael@0: __inline __m64 _m_psrlqi (__m64 m, int n)
michael@0: {
michael@0:     unsigned long long x = M64U (m) >> n;
michael@0: 
michael@0:     return M64C (x);
michael@0: }
michael@0: 
michael@0: __inline __m64 _m_psllqi (__m64 m, int n)
michael@0: {
michael@0:     unsigned long long x = M64U (m) << n;
michael@0: 
michael@0:     return M64C (x);
michael@0: }
michael@0: 
michael@0: #endif /* MMX_X64_H_INCLUDED */