Lines Matching refs:vget_high_s16
75 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
77 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
79 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
84 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
86 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
88 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
93 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
95 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
97 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
104 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
106 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
111 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
112 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
113 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
114 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
115 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
116 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
120 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
121 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
122 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
123 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
124 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
125 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
129 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
130 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
131 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
132 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
133 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
134 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
138 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
139 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
140 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
141 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
142 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
143 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
159 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
161 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
163 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
170 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
172 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
174 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
181 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
183 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
185 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
192 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
194 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
196 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
202 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
203 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
204 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
205 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
206 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
207 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
213 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
214 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
215 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
216 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
217 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
218 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
224 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
225 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
226 … vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
227 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
228 … vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
229 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()