Searched refs:vxa4 (Results 1 – 2 of 2) sorted by relevance
/external/XNNPACK/src/q8-igemm/ |
D | 8x8-neon.c | 138 const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4)); in xnn_q8_igemm_ukernel_8x8__neon() local 155 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 0); in xnn_q8_igemm_ukernel_8x8__neon() 156 … vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 0); in xnn_q8_igemm_ukernel_8x8__neon() 177 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 1); in xnn_q8_igemm_ukernel_8x8__neon() 178 … vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 1); in xnn_q8_igemm_ukernel_8x8__neon() 199 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 2); in xnn_q8_igemm_ukernel_8x8__neon() 200 … vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 2); in xnn_q8_igemm_ukernel_8x8__neon() 221 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 3); in xnn_q8_igemm_ukernel_8x8__neon() 222 … vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 3); in xnn_q8_igemm_ukernel_8x8__neon() 243 … vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 0); in xnn_q8_igemm_ukernel_8x8__neon() [all …]
|
/external/XNNPACK/src/q8-gemm/ |
D | 8x8-neon.c | 110 const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4)); a4 += 8; in xnn_q8_gemm_ukernel_8x8__neon() local 129 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0); in xnn_q8_gemm_ukernel_8x8__neon() 130 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0); in xnn_q8_gemm_ukernel_8x8__neon() 149 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1); in xnn_q8_gemm_ukernel_8x8__neon() 150 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1); 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() 170 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2); in xnn_q8_gemm_ukernel_8x8__neon() 189 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3); in xnn_q8_gemm_ukernel_8x8__neon() 190 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3); 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 …]
|