92 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
93 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
94 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
95 …int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
97 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
98 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
99 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
100 …vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
105 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
106 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
107 …int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
108 …int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
110 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
111 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
112 …vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
113 …vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
118 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
119 …int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
120 …int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
121 …int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
123 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
124 …vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
125 …vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
126 …vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
131 …int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
132 …int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
133 …int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
134 …int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
136 …vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
137 …vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
138 …vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
139 …vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
144 …int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
145 …int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
146 …int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
147 …int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
149 …vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
150 …vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
151 …vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
152 …vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
157 …int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
158 …int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
159 …int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
160 …int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
162 …vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
163 …vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
164 …vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
165 …vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
170 …int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
171 …int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
172 …int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
173 …int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
175 …vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
176 …vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
177 …vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
178 …vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
183 …int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
184 …int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
185 …int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
186 …int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
188 …vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
189 …vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
190 …vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
191 …vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
215 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
216 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
217 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
218 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
223 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
224 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
225 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
226 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
231 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
232 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
233 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
234 …const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
239 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
240 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
241 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
242 …const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
247 …const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
248 …const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
249 …const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
250 …const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
255 …const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
256 …const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
257 …const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
258 …const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
263 …const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
264 …const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
265 …const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
266 …const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
271 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
272 …const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
273 …const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
274 …const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
292 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
294 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
296 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
298 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
300 …const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
302 …const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
304 …const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
306 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
313 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
315 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
317 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
319 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
321 …const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
323 …const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
325 …const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
327 …const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
334 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
336 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
338 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
340 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
342 …const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
344 …const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
346 …const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
348 …const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()