82 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
83 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
84 …int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
86 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
87 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
88 …vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
92 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
93 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
94 …int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
96 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
97 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
98 …vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
102 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
103 …int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
104 …int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
106 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
107 …vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
108 …vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
112 …int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
113 …int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
114 …int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
116 …vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
117 …vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
118 …vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
122 …int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
123 …int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
124 …int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
126 …vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
127 …vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
128 …vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
132 …int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
133 …int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
134 …int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
136 …vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
137 …vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
138 …vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
142 …int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
143 …int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
144 …int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
146 …vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
147 …vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
148 …vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
152 …int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
153 …int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
154 …int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
156 …vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
157 …vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
158 …vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
180 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
181 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
182 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
183 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
188 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
189 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
190 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
191 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
196 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
197 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
198 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
199 …const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
204 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
205 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
206 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
207 …const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
212 …const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
213 …const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
214 …const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
215 …const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
220 …const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
221 …const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
222 …const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
223 …const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
240 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
242 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
244 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
246 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
248 …const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
250 …const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
257 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
259 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
261 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
263 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
265 …const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
267 …const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
274 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
276 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
278 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
280 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
282 …const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()
284 …const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup()