79 	__m128 det, tmp1;  in M_MatrixInvert44_SSE()  local
81 	tmp1	= _mm_loadh_pi(_mm_loadl_pi(tmp1, (__m64 *)(src)),  in M_MatrixInvert44_SSE()
85 	row0	= _mm_shuffle_ps(tmp1, row1, 0x88);  in M_MatrixInvert44_SSE()
86 	row1	= _mm_shuffle_ps(row1, tmp1, 0xDD);  in M_MatrixInvert44_SSE()
87 	tmp1	= _mm_loadh_pi(_mm_loadl_pi(tmp1, (__m64 *)(src+2)),  in M_MatrixInvert44_SSE()
91 	row2	= _mm_shuffle_ps(tmp1, row3, 0x88);  in M_MatrixInvert44_SSE()
92 	row3	= _mm_shuffle_ps(row3, tmp1, 0xDD);  in M_MatrixInvert44_SSE()
94 	tmp1	= _mm_mul_ps(row2, row3);  in M_MatrixInvert44_SSE()
95 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0xB1);  in M_MatrixInvert44_SSE()
96 	minor0	= _mm_mul_ps(row1, tmp1);  in M_MatrixInvert44_SSE()
97 	minor1	= _mm_mul_ps(row0, tmp1);  in M_MatrixInvert44_SSE()
98 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0x4E);  in M_MatrixInvert44_SSE()
99 	minor0	= _mm_sub_ps(_mm_mul_ps(row1, tmp1), minor0);  in M_MatrixInvert44_SSE()
100 	minor1	= _mm_sub_ps(_mm_mul_ps(row0, tmp1), minor1);  in M_MatrixInvert44_SSE()
103 	tmp1	= _mm_mul_ps(row1, row2);  in M_MatrixInvert44_SSE()
104 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0xB1);  in M_MatrixInvert44_SSE()
105 	minor0	= _mm_add_ps(_mm_mul_ps(row3, tmp1), minor0);  in M_MatrixInvert44_SSE()
106 	minor3	= _mm_mul_ps(row0, tmp1);  in M_MatrixInvert44_SSE()
107 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0x4E);  in M_MatrixInvert44_SSE()
108 	minor0	= _mm_sub_ps(minor0, _mm_mul_ps(row3, tmp1));  in M_MatrixInvert44_SSE()
109 	minor3	= _mm_sub_ps(_mm_mul_ps(row0, tmp1), minor3);  in M_MatrixInvert44_SSE()
112 	tmp1	= _mm_mul_ps(_mm_shuffle_ps(row1, row1, 0x4E), row3);  in M_MatrixInvert44_SSE()
113 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0xB1);  in M_MatrixInvert44_SSE()
115 	minor0	= _mm_add_ps(_mm_mul_ps(row2, tmp1), minor0);  in M_MatrixInvert44_SSE()
116 	minor2	= _mm_mul_ps(row0, tmp1);  in M_MatrixInvert44_SSE()
117 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0x4E);  in M_MatrixInvert44_SSE()
118 	minor0	= _mm_sub_ps(minor0, _mm_mul_ps(row2, tmp1));  in M_MatrixInvert44_SSE()
119 	minor2	= _mm_sub_ps(_mm_mul_ps(row0, tmp1), minor2);  in M_MatrixInvert44_SSE()
122 	tmp1	= _mm_mul_ps(row0, row1);  in M_MatrixInvert44_SSE()
123 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0xB1);  in M_MatrixInvert44_SSE()
124 	minor2	= _mm_add_ps(_mm_mul_ps(row3, tmp1), minor2);  in M_MatrixInvert44_SSE()
125 	minor3	= _mm_sub_ps(_mm_mul_ps(row2, tmp1), minor3);  in M_MatrixInvert44_SSE()
126 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0x4E);  in M_MatrixInvert44_SSE()
127 	minor2	= _mm_sub_ps(_mm_mul_ps(row3, tmp1), minor2);  in M_MatrixInvert44_SSE()
128 	minor3	= _mm_sub_ps(minor3, _mm_mul_ps(row2, tmp1));  in M_MatrixInvert44_SSE()
130 	tmp1	= _mm_mul_ps(row0, row3);  in M_MatrixInvert44_SSE()
131 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0xB1);  in M_MatrixInvert44_SSE()
132 	minor1	= _mm_sub_ps(minor1, _mm_mul_ps(row2, tmp1));  in M_MatrixInvert44_SSE()
133 	minor2	= _mm_add_ps(_mm_mul_ps(row1, tmp1), minor2);  in M_MatrixInvert44_SSE()
134 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0x4E);  in M_MatrixInvert44_SSE()
135 	minor1	= _mm_add_ps(_mm_mul_ps(row2, tmp1), minor1);  in M_MatrixInvert44_SSE()
136 	minor2	= _mm_sub_ps(minor2, _mm_mul_ps(row1, tmp1));  in M_MatrixInvert44_SSE()
138 	tmp1	= _mm_mul_ps(row0, row2);  in M_MatrixInvert44_SSE()
139 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0xB1);  in M_MatrixInvert44_SSE()
140 	minor1	= _mm_add_ps(_mm_mul_ps(row3, tmp1), minor1);  in M_MatrixInvert44_SSE()
141 	minor3	= _mm_sub_ps(minor3, _mm_mul_ps(row1, tmp1));  in M_MatrixInvert44_SSE()
142 	tmp1	= _mm_shuffle_ps(tmp1, tmp1, 0x4E);  in M_MatrixInvert44_SSE()
143 	minor1	= _mm_sub_ps(minor1, _mm_mul_ps(row3, tmp1));  in M_MatrixInvert44_SSE()
144 	minor3	= _mm_add_ps(_mm_mul_ps(row1, tmp1), minor3);  in M_MatrixInvert44_SSE()
149 	tmp1	= _mm_rcp_ss(det);  in M_MatrixInvert44_SSE()
150 	det	= _mm_sub_ss(_mm_add_ss(tmp1, tmp1),  in M_MatrixInvert44_SSE()
151 	                     _mm_mul_ss(det, _mm_mul_ss(tmp1,tmp1)));  in M_MatrixInvert44_SSE()