72 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
73 …int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
75 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
76 …vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
79 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
80 …int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
82 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
83 …vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
86 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
87 …int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
89 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
90 …vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
93 …int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
94 …int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
96 …vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
97 …vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
100 …int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
101 …int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
103 …vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
104 …vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
107 …int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
108 …int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
110 …vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
111 …vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
114 …int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
115 …int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
117 …vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
118 …vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
121 …int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
122 …int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
124 …vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
125 …vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
145 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
146 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
147 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
148 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
153 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
154 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
155 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
156 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
161 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
162 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
163 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
164 …const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
169 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
170 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
171 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
172 …const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
188 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
190 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
192 …const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
194 …const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
201 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
203 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
205 …const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
207 …const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
214 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
216 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
218 …const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()
220 …const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_2x8c2__neon_mlal_padal_dup()