Searched refs:vxb01234567c2 (Results 1 – 2 of 2) sorted by relevance
/external/XNNPACK/src/q8-gemm/ |
D | 8x8-neon.c | 159 const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point)); in xnn_q8_gemm_ukernel_8x8__neon() local 161 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_q8_gemm_ukernel_8x8__neon() 162 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_q8_gemm_ukernel_8x8__neon() 163 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_q8_gemm_ukernel_8x8__neon() 164 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_q8_gemm_ukernel_8x8__neon() 165 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_q8_gemm_ukernel_8x8__neon() 166 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_q8_gemm_ukernel_8x8__neon() 167 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_q8_gemm_ukernel_8x8__neon() 168 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_q8_gemm_ukernel_8x8__neon() 169 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2); in xnn_q8_gemm_ukernel_8x8__neon() [all …]
|
D | 4x8-neon.c | 103 const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point)); in xnn_q8_gemm_ukernel_4x8__neon() local 105 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_q8_gemm_ukernel_4x8__neon() 106 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_q8_gemm_ukernel_4x8__neon() 107 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_q8_gemm_ukernel_4x8__neon() 108 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); in xnn_q8_gemm_ukernel_4x8__neon() 109 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_q8_gemm_ukernel_4x8__neon() 110 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); in xnn_q8_gemm_ukernel_4x8__neon() 111 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_q8_gemm_ukernel_4x8__neon() 112 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); in xnn_q8_gemm_ukernel_4x8__neon() 213 … const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point)); in xnn_q8_gemm_ukernel_4x8__neon() local [all …]
|