Searched refs:vxb01234567c4 (Results 1 – 2 of 2) sorted by relevance
/external/XNNPACK/src/q8-gemm/ |
D | 8x8-neon.c | 199 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point)); in xnn_q8_gemm_ukernel_8x8__neon() local 201 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_q8_gemm_ukernel_8x8__neon() 202 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_q8_gemm_ukernel_8x8__neon() 203 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_q8_gemm_ukernel_8x8__neon() 204 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_q8_gemm_ukernel_8x8__neon() 205 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_q8_gemm_ukernel_8x8__neon() 206 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_q8_gemm_ukernel_8x8__neon() 207 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_q8_gemm_ukernel_8x8__neon() 208 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_q8_gemm_ukernel_8x8__neon() 209 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0); in xnn_q8_gemm_ukernel_8x8__neon() [all …]
|
D | 4x8-neon.c | 127 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point)); in xnn_q8_gemm_ukernel_4x8__neon() local 129 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_q8_gemm_ukernel_4x8__neon() 130 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_q8_gemm_ukernel_4x8__neon() 131 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_q8_gemm_ukernel_4x8__neon() 132 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); in xnn_q8_gemm_ukernel_4x8__neon() 133 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_q8_gemm_ukernel_4x8__neon() 134 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); in xnn_q8_gemm_ukernel_4x8__neon() 135 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_q8_gemm_ukernel_4x8__neon() 136 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); in xnn_q8_gemm_ukernel_4x8__neon() 239 … const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point)); in xnn_q8_gemm_ukernel_4x8__neon() local [all …]
|