106 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
107 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
108 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
109 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
114 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
115 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
116 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
117 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
122 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
123 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
124 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
125 …const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
130 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
131 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
132 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
133 …const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
138 …const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
139 …const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
140 …const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
141 …const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
146 …const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
147 …const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
148 …const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
149 …const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
154 …const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
155 …const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
156 …const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
157 …const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
162 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
163 …const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
164 …const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
165 …const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
183 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
185 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
187 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
189 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
191 …const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
193 …const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
195 …const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
197 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
204 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
206 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
208 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
210 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
212 …const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
214 …const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
216 …const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
218 …const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
225 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
227 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
229 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
231 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
233 …const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
235 …const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
237 …const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
239 …const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()