Lines Matching refs:mem
39 __m128 num[3], den[3], mem[3];
46 mem[i] = _mm_loadu_ps(_mem+4*i);
50 mem[2] = _mm_setr_ps(_mem[8], _mem[9], 0, 0);
60 yy = _mm_add_ss(xx, mem[0]);
65 mem[0] = _mm_move_ss(mem[0], mem[1]);
66 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39);
68 mem[0] = _mm_add_ps(mem[0], _mm_mul_ps(xx, num[0]));
69 mem[0] = _mm_sub_ps(mem[0], _mm_mul_ps(yy, den[0]));
71 mem[1] = _mm_move_ss(mem[1], mem[2]);
72 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39);
74 mem[1] = _mm_add_ps(mem[1], _mm_mul_ps(xx, num[1]));
75 mem[1] = _mm_sub_ps(mem[1], _mm_mul_ps(yy, den[1]));
77 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0xfd);
79 mem[2] = _mm_add_ps(mem[2], _mm_mul_ps(xx, num[2]));
80 mem[2] = _mm_sub_ps(mem[2], _mm_mul_ps(yy, den[2]));
83 _mm_storeu_ps(_mem, mem[0]);
84 _mm_storeu_ps(_mem+4, mem[1]);
85 _mm_store_ss(_mem+8, mem[2]);
86 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0x55);
87 _mm_store_ss(_mem+9, mem[2]);
92 __m128 num[2], den[2], mem[2];
99 mem[i] = _mm_loadu_ps(_mem+4*i);
110 yy = _mm_add_ss(xx, mem[0]);
115 mem[0] = _mm_move_ss(mem[0], mem[1]);
116 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39);
118 mem[0] = _mm_add_ps(mem[0], _mm_mul_ps(xx, num[0]));
119 mem[0] = _mm_sub_ps(mem[0], _mm_mul_ps(yy, den[0]));
121 mem[1] = _mm_sub_ss(mem[1], mem[1]);
122 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39);
124 mem[1] = _mm_add_ps(mem[1], _mm_mul_ps(xx, num[1]));
125 mem[1] = _mm_sub_ps(mem[1], _mm_mul_ps(yy, den[1]));
128 _mm_storeu_ps(_mem, mem[0]);
129 _mm_storeu_ps(_mem+4, mem[1]);
146 __m128 den[3], mem[3];
153 mem[i] = _mm_loadu_ps(_mem+4*i);
156 mem[2] = _mm_setr_ps(_mem[8], _mem[9], 0, 0);
165 yy = _mm_add_ss(xx, mem[0]);
170 mem[0] = _mm_move_ss(mem[0], mem[1]);
171 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39);
173 mem[0] = _mm_sub_ps(mem[0], _mm_mul_ps(yy, den[0]));
175 mem[1] = _mm_move_ss(mem[1], mem[2]);
176 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39);
178 mem[1] = _mm_sub_ps(mem[1], _mm_mul_ps(yy, den[1]));
180 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0xfd);
182 mem[2] = _mm_sub_ps(mem[2], _mm_mul_ps(yy, den[2]));
185 _mm_storeu_ps(_mem, mem[0]);
186 _mm_storeu_ps(_mem+4, mem[1]);
187 _mm_store_ss(_mem+8, mem[2]);
188 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0x55);
189 _mm_store_ss(_mem+9, mem[2]);
195 __m128 den[2], mem[2];
202 mem[i] = _mm_loadu_ps(_mem+4*i);
212 yy = _mm_add_ss(xx, mem[0]);
217 mem[0] = _mm_move_ss(mem[0], mem[1]);
218 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39);
220 mem[0] = _mm_sub_ps(mem[0], _mm_mul_ps(yy, den[0]));
222 mem[1] = _mm_sub_ss(mem[1], mem[1]);
223 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39);
225 mem[1] = _mm_sub_ps(mem[1], _mm_mul_ps(yy, den[1]));
228 _mm_storeu_ps(_mem, mem[0]);
229 _mm_storeu_ps(_mem+4, mem[1]);
244 __m128 num[3], mem[3];
251 mem[i] = _mm_loadu_ps(_mem+4*i);
254 mem[2] = _mm_setr_ps(_mem[8], _mem[9], 0, 0);
263 yy = _mm_add_ss(xx, mem[0]);
268 mem[0] = _mm_move_ss(mem[0], mem[1]);
269 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39);
271 mem[0] = _mm_add_ps(mem[0], _mm_mul_ps(xx, num[0]));
273 mem[1] = _mm_move_ss(mem[1], mem[2]);
274 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39);
276 mem[1] = _mm_add_ps(mem[1], _mm_mul_ps(xx, num[1]));
278 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0xfd);
280 mem[2] = _mm_add_ps(mem[2], _mm_mul_ps(xx, num[2]));
283 _mm_storeu_ps(_mem, mem[0]);
284 _mm_storeu_ps(_mem+4, mem[1]);
285 _mm_store_ss(_mem+8, mem[2]);
286 mem[2] = _mm_shuffle_ps(mem[2], mem[2], 0x55);
287 _mm_store_ss(_mem+9, mem[2]);
292 __m128 num[2], mem[2];
299 mem[i] = _mm_loadu_ps(_mem+4*i);
309 yy = _mm_add_ss(xx, mem[0]);
314 mem[0] = _mm_move_ss(mem[0], mem[1]);
315 mem[0] = _mm_shuffle_ps(mem[0], mem[0], 0x39);
317 mem[0] = _mm_add_ps(mem[0], _mm_mul_ps(xx, num[0]));
319 mem[1] = _mm_sub_ss(mem[1], mem[1]);
320 mem[1] = _mm_shuffle_ps(mem[1], mem[1], 0x39);
322 mem[1] = _mm_add_ps(mem[1], _mm_mul_ps(xx, num[1]));
325 _mm_storeu_ps(_mem, mem[0]);
326 _mm_storeu_ps(_mem+4, mem[1]);
