Lines Matching refs:vget_low_s16
99 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_q8_igemm_ukernel_4x8__neon()
100 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_q8_igemm_ukernel_4x8__neon()
101 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_q8_igemm_ukernel_4x8__neon()
102 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_q8_igemm_ukernel_4x8__neon()
103 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_q8_igemm_ukernel_4x8__neon()
104 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_q8_igemm_ukernel_4x8__neon()
105 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_q8_igemm_ukernel_4x8__neon()
106 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_q8_igemm_ukernel_4x8__neon()
113 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1); in xnn_q8_igemm_ukernel_4x8__neon()
114 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1); in xnn_q8_igemm_ukernel_4x8__neon()
115 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 1); in xnn_q8_igemm_ukernel_4x8__neon()
116 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 1); in xnn_q8_igemm_ukernel_4x8__neon()
117 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 1); in xnn_q8_igemm_ukernel_4x8__neon()
118 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 1); in xnn_q8_igemm_ukernel_4x8__neon()
119 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 1); in xnn_q8_igemm_ukernel_4x8__neon()
120 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 1); in xnn_q8_igemm_ukernel_4x8__neon()
127 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 2); in xnn_q8_igemm_ukernel_4x8__neon()
128 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 2); in xnn_q8_igemm_ukernel_4x8__neon()
129 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 2); in xnn_q8_igemm_ukernel_4x8__neon()
130 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 2); in xnn_q8_igemm_ukernel_4x8__neon()
131 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 2); in xnn_q8_igemm_ukernel_4x8__neon()
132 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 2); in xnn_q8_igemm_ukernel_4x8__neon()
133 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 2); in xnn_q8_igemm_ukernel_4x8__neon()
134 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 2); in xnn_q8_igemm_ukernel_4x8__neon()
141 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 3); in xnn_q8_igemm_ukernel_4x8__neon()
142 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 3); in xnn_q8_igemm_ukernel_4x8__neon()
143 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 3); in xnn_q8_igemm_ukernel_4x8__neon()
144 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 3); in xnn_q8_igemm_ukernel_4x8__neon()
145 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 3); in xnn_q8_igemm_ukernel_4x8__neon()
146 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 3); in xnn_q8_igemm_ukernel_4x8__neon()
147 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 3); in xnn_q8_igemm_ukernel_4x8__neon()
148 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 3); in xnn_q8_igemm_ukernel_4x8__neon()
155 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 0); in xnn_q8_igemm_ukernel_4x8__neon()
157 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 0); in xnn_q8_igemm_ukernel_4x8__neon()
159 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 0); in xnn_q8_igemm_ukernel_4x8__neon()
161 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 0); in xnn_q8_igemm_ukernel_4x8__neon()
169 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 1); in xnn_q8_igemm_ukernel_4x8__neon()
171 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 1); in xnn_q8_igemm_ukernel_4x8__neon()
173 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 1); in xnn_q8_igemm_ukernel_4x8__neon()
175 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 1); in xnn_q8_igemm_ukernel_4x8__neon()
183 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 2); in xnn_q8_igemm_ukernel_4x8__neon()
185 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 2); in xnn_q8_igemm_ukernel_4x8__neon()
187 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 2); in xnn_q8_igemm_ukernel_4x8__neon()
189 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 2); in xnn_q8_igemm_ukernel_4x8__neon()
197 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 3); in xnn_q8_igemm_ukernel_4x8__neon()
199 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 3); in xnn_q8_igemm_ukernel_4x8__neon()
201 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 3); in xnn_q8_igemm_ukernel_4x8__neon()
203 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 3); in xnn_q8_igemm_ukernel_4x8__neon()
223 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_q8_igemm_ukernel_4x8__neon()
224 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_q8_igemm_ukernel_4x8__neon()
225 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_q8_igemm_ukernel_4x8__neon()
226 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_q8_igemm_ukernel_4x8__neon()
227 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_q8_igemm_ukernel_4x8__neon()
228 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_q8_igemm_ukernel_4x8__neon()
229 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_q8_igemm_ukernel_4x8__neon()
230 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_q8_igemm_ukernel_4x8__neon()
237 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1); in xnn_q8_igemm_ukernel_4x8__neon()
238 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1); in xnn_q8_igemm_ukernel_4x8__neon()
239 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 1); in xnn_q8_igemm_ukernel_4x8__neon()
240 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 1); in xnn_q8_igemm_ukernel_4x8__neon()
241 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 1); in xnn_q8_igemm_ukernel_4x8__neon()
242 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 1); in xnn_q8_igemm_ukernel_4x8__neon()
243 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 1); in xnn_q8_igemm_ukernel_4x8__neon()
244 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 1); in xnn_q8_igemm_ukernel_4x8__neon()
250 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 2); in xnn_q8_igemm_ukernel_4x8__neon()
251 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 2); in xnn_q8_igemm_ukernel_4x8__neon()
252 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 2); in xnn_q8_igemm_ukernel_4x8__neon()
253 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 2); in xnn_q8_igemm_ukernel_4x8__neon()
254 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 2); in xnn_q8_igemm_ukernel_4x8__neon()
255 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 2); in xnn_q8_igemm_ukernel_4x8__neon()
256 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 2); in xnn_q8_igemm_ukernel_4x8__neon()
257 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 2); in xnn_q8_igemm_ukernel_4x8__neon()
263 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 3); in xnn_q8_igemm_ukernel_4x8__neon()
264 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 3); in xnn_q8_igemm_ukernel_4x8__neon()
265 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 3); in xnn_q8_igemm_ukernel_4x8__neon()
266 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 3); in xnn_q8_igemm_ukernel_4x8__neon()
267 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 3); in xnn_q8_igemm_ukernel_4x8__neon()
268 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 3); in xnn_q8_igemm_ukernel_4x8__neon()
269 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 3); in xnn_q8_igemm_ukernel_4x8__neon()
270 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 3); in xnn_q8_igemm_ukernel_4x8__neon()
276 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 0); in xnn_q8_igemm_ukernel_4x8__neon()
278 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 0); in xnn_q8_igemm_ukernel_4x8__neon()
280 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 0); in xnn_q8_igemm_ukernel_4x8__neon()
282 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 0); in xnn_q8_igemm_ukernel_4x8__neon()
289 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 1); in xnn_q8_igemm_ukernel_4x8__neon()
291 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 1); in xnn_q8_igemm_ukernel_4x8__neon()
293 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 1); in xnn_q8_igemm_ukernel_4x8__neon()
295 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 1); in xnn_q8_igemm_ukernel_4x8__neon()
302 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 2); in xnn_q8_igemm_ukernel_4x8__neon()
304 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 2); in xnn_q8_igemm_ukernel_4x8__neon()
306 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 2); in xnn_q8_igemm_ukernel_4x8__neon()
308 … vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 2); in xnn_q8_igemm_ukernel_4x8__neon()