72 …int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
74 …vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
76 …int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
78 …vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
80 …int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
82 …vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
84 …int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
86 …vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
88 …int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
90 …vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
92 …int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
94 …vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
96 …int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
98 …vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
100 …int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
102 …vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
104 …int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
106 …vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
108 …int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
110 …vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
112 …int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
114 …vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
116 …int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
118 …vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
120 …int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
122 …vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
124 …int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
126 …vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
128 …int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
130 …vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
132 …int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
134 …vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
160 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
161 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
162 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
163 …const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
168 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
169 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
170 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
171 …const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
176 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
177 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
178 …const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
179 …const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
184 …const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
185 …const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
186 …const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
187 …const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
204 …const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
206 …const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
208 …const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
210 …const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
219 …const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
221 …const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
223 …const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
225 …const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
234 …const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
236 …const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
238 …const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
240 …const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_…  in xnn_qs8_gemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()