Lines Matching refs:vget_high_s16
90 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
92 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
94 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
99 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
101 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
103 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
108 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
110 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
112 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
117 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
119 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
121 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
126 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
127 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
128 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
129 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
130 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
131 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
135 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
136 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
137 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
138 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
139 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
140 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
144 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
145 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
146 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
147 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
148 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
149 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
153 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
154 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
155 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
156 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
157 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
158 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
174 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
176 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
178 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
185 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
187 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
189 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
196 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
198 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
200 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
207 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
209 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
211 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
217 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
218 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
219 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
220 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
221 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
222 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
228 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
229 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
230 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
231 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
232 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
233 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
239 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
240 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
241 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
242 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
243 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
244 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()