82       int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
83       int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
88       int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
89       int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
94       int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
95       int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
100       int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
101       int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
106       int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
107       int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
112       int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
113       int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
118       int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
119       int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
124       int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
125       int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
215       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));  in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()