Searched refs:vxb01234567c6 (Results 1 – 2 of 2) sorted by relevance
/external/XNNPACK/src/q8-gemm/ |
D | 8x8-neon.c | 239 const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point)); in xnn_q8_gemm_ukernel_8x8__neon() local 241 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_q8_gemm_ukernel_8x8__neon() 242 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_q8_gemm_ukernel_8x8__neon() 243 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_q8_gemm_ukernel_8x8__neon() 244 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_q8_gemm_ukernel_8x8__neon() 245 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_q8_gemm_ukernel_8x8__neon() 246 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_q8_gemm_ukernel_8x8__neon() 247 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_q8_gemm_ukernel_8x8__neon() 248 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_q8_gemm_ukernel_8x8__neon() 249 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2); in xnn_q8_gemm_ukernel_8x8__neon() [all …]
|
D | 4x8-neon.c | 151 const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point)); in xnn_q8_gemm_ukernel_4x8__neon() local 153 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_q8_gemm_ukernel_4x8__neon() 154 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); in xnn_q8_gemm_ukernel_4x8__neon() 155 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_q8_gemm_ukernel_4x8__neon() 156 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); in xnn_q8_gemm_ukernel_4x8__neon() 157 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_q8_gemm_ukernel_4x8__neon() 158 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); in xnn_q8_gemm_ukernel_4x8__neon() 159 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_q8_gemm_ukernel_4x8__neon() 160 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); in xnn_q8_gemm_ukernel_4x8__neon() 265 … const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point)); in xnn_q8_gemm_ukernel_4x8__neon() local [all …]
|