Lines Matching refs:xv

72 	__m256          xv[8];  in bli_sswapv_zen_int8()  local
85 xv[0] = _mm256_loadu_ps( x0 + 0*n_elem_per_reg ); in bli_sswapv_zen_int8()
86 xv[1] = _mm256_loadu_ps( x0 + 1*n_elem_per_reg ); in bli_sswapv_zen_int8()
87 xv[2] = _mm256_loadu_ps( x0 + 2*n_elem_per_reg ); in bli_sswapv_zen_int8()
88 xv[3] = _mm256_loadu_ps( x0 + 3*n_elem_per_reg ); in bli_sswapv_zen_int8()
89 xv[4] = _mm256_loadu_ps( x0 + 4*n_elem_per_reg ); in bli_sswapv_zen_int8()
90 xv[5] = _mm256_loadu_ps( x0 + 5*n_elem_per_reg ); in bli_sswapv_zen_int8()
91 xv[6] = _mm256_loadu_ps( x0 + 6*n_elem_per_reg ); in bli_sswapv_zen_int8()
92 xv[7] = _mm256_loadu_ps( x0 + 7*n_elem_per_reg ); in bli_sswapv_zen_int8()
112 _mm256_storeu_ps( (y0 + 0*n_elem_per_reg), xv[0]); in bli_sswapv_zen_int8()
113 _mm256_storeu_ps( (y0 + 1*n_elem_per_reg), xv[1]); in bli_sswapv_zen_int8()
114 _mm256_storeu_ps( (y0 + 2*n_elem_per_reg), xv[2]); in bli_sswapv_zen_int8()
115 _mm256_storeu_ps( (y0 + 3*n_elem_per_reg), xv[3]); in bli_sswapv_zen_int8()
116 _mm256_storeu_ps( (y0 + 4*n_elem_per_reg), xv[4]); in bli_sswapv_zen_int8()
117 _mm256_storeu_ps( (y0 + 5*n_elem_per_reg), xv[5]); in bli_sswapv_zen_int8()
118 _mm256_storeu_ps( (y0 + 6*n_elem_per_reg), xv[6]); in bli_sswapv_zen_int8()
119 _mm256_storeu_ps( (y0 + 7*n_elem_per_reg), xv[7]); in bli_sswapv_zen_int8()
127 xv[0] = _mm256_loadu_ps( x0 + 0*n_elem_per_reg ); in bli_sswapv_zen_int8()
128 xv[1] = _mm256_loadu_ps( x0 + 1*n_elem_per_reg ); in bli_sswapv_zen_int8()
129 xv[2] = _mm256_loadu_ps( x0 + 2*n_elem_per_reg ); in bli_sswapv_zen_int8()
130 xv[3] = _mm256_loadu_ps( x0 + 3*n_elem_per_reg ); in bli_sswapv_zen_int8()
137 _mm256_storeu_ps( (y0 + 0*n_elem_per_reg), xv[0]); in bli_sswapv_zen_int8()
138 _mm256_storeu_ps( (y0 + 1*n_elem_per_reg), xv[1]); in bli_sswapv_zen_int8()
139 _mm256_storeu_ps( (y0 + 2*n_elem_per_reg), xv[2]); in bli_sswapv_zen_int8()
140 _mm256_storeu_ps( (y0 + 3*n_elem_per_reg), xv[3]); in bli_sswapv_zen_int8()
153 xv[0] = _mm256_loadu_ps( x0 + 0*n_elem_per_reg ); in bli_sswapv_zen_int8()
154 xv[1] = _mm256_loadu_ps( x0 + 1*n_elem_per_reg ); in bli_sswapv_zen_int8()
159 _mm256_storeu_ps( (y0 + 0*n_elem_per_reg), xv[0]); in bli_sswapv_zen_int8()
160 _mm256_storeu_ps( (y0 + 1*n_elem_per_reg), xv[1]); in bli_sswapv_zen_int8()
171 xv[0] = _mm256_loadu_ps( x0 + 0*n_elem_per_reg ); in bli_sswapv_zen_int8()
177 _mm256_storeu_ps( (y0 + 0*n_elem_per_reg), xv[0]); in bli_sswapv_zen_int8()
217 __m256d xv[8]; in bli_dswapv_zen_int8() local
230 xv[0] = _mm256_loadu_pd( x0 + 0*n_elem_per_reg ); in bli_dswapv_zen_int8()
231 xv[1] = _mm256_loadu_pd( x0 + 1*n_elem_per_reg ); in bli_dswapv_zen_int8()
232 xv[2] = _mm256_loadu_pd( x0 + 2*n_elem_per_reg ); in bli_dswapv_zen_int8()
233 xv[3] = _mm256_loadu_pd( x0 + 3*n_elem_per_reg ); in bli_dswapv_zen_int8()
234 xv[4] = _mm256_loadu_pd( x0 + 4*n_elem_per_reg ); in bli_dswapv_zen_int8()
235 xv[5] = _mm256_loadu_pd( x0 + 5*n_elem_per_reg ); in bli_dswapv_zen_int8()
236 xv[6] = _mm256_loadu_pd( x0 + 6*n_elem_per_reg ); in bli_dswapv_zen_int8()
237 xv[7] = _mm256_loadu_pd( x0 + 7*n_elem_per_reg ); in bli_dswapv_zen_int8()
257 _mm256_storeu_pd( (y0 + 0*n_elem_per_reg), xv[0]); in bli_dswapv_zen_int8()
258 _mm256_storeu_pd( (y0 + 1*n_elem_per_reg), xv[1]); in bli_dswapv_zen_int8()
259 _mm256_storeu_pd( (y0 + 2*n_elem_per_reg), xv[2]); in bli_dswapv_zen_int8()
260 _mm256_storeu_pd( (y0 + 3*n_elem_per_reg), xv[3]); in bli_dswapv_zen_int8()
261 _mm256_storeu_pd( (y0 + 4*n_elem_per_reg), xv[4]); in bli_dswapv_zen_int8()
262 _mm256_storeu_pd( (y0 + 5*n_elem_per_reg), xv[5]); in bli_dswapv_zen_int8()
263 _mm256_storeu_pd( (y0 + 6*n_elem_per_reg), xv[6]); in bli_dswapv_zen_int8()
264 _mm256_storeu_pd( (y0 + 7*n_elem_per_reg), xv[7]); in bli_dswapv_zen_int8()
272 xv[0] = _mm256_loadu_pd( x0 + 0*n_elem_per_reg ); in bli_dswapv_zen_int8()
273 xv[1] = _mm256_loadu_pd( x0 + 1*n_elem_per_reg ); in bli_dswapv_zen_int8()
274 xv[2] = _mm256_loadu_pd( x0 + 2*n_elem_per_reg ); in bli_dswapv_zen_int8()
275 xv[3] = _mm256_loadu_pd( x0 + 3*n_elem_per_reg ); in bli_dswapv_zen_int8()
282 _mm256_storeu_pd( (y0 + 0*n_elem_per_reg), xv[0]); in bli_dswapv_zen_int8()
283 _mm256_storeu_pd( (y0 + 1*n_elem_per_reg), xv[1]); in bli_dswapv_zen_int8()
284 _mm256_storeu_pd( (y0 + 2*n_elem_per_reg), xv[2]); in bli_dswapv_zen_int8()
285 _mm256_storeu_pd( (y0 + 3*n_elem_per_reg), xv[3]); in bli_dswapv_zen_int8()
298 xv[0] = _mm256_loadu_pd( x0 + 0*n_elem_per_reg ); in bli_dswapv_zen_int8()
299 xv[1] = _mm256_loadu_pd( x0 + 1*n_elem_per_reg ); in bli_dswapv_zen_int8()
304 _mm256_storeu_pd( (y0 + 0*n_elem_per_reg), xv[0]); in bli_dswapv_zen_int8()
305 _mm256_storeu_pd( (y0 + 1*n_elem_per_reg), xv[1]); in bli_dswapv_zen_int8()
316 xv[0] = _mm256_loadu_pd( x0 + 0*n_elem_per_reg ); in bli_dswapv_zen_int8()
320 _mm256_storeu_pd( (y0 + 0*n_elem_per_reg), xv[0]); in bli_dswapv_zen_int8()