96 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
97 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
98 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
99 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
104 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
105 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
106 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
107 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
112 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
113 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
114 …const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
115 …const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
120 …const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
121 …const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
122 …const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
123 …const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
128 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
129 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
130 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
131 …const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
136 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
137 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
138 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
139 …const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
144 …const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
145 …const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
146 …const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
147 …const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
152 …const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
153 …const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
154 …const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
155 …const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
173 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
175 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
177 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
179 …const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
181 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
183 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
185 …const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
187 …const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
196 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
198 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
200 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
202 …const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
204 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
206 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
208 …const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
210 …const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
219 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
221 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
223 …const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
225 …const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
227 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
229 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
231 …const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()
233 …const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_igemm_minmax_ukernel_2x16c2__neon_mull_padal_dup()