/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x8c2-minmax-neon-mlal-padal-dup.c | 108 …int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() local 113 …vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() 117 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() 271 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() local 275 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() 306 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() local 307 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
|
D | 4x8c2-minmax-neon-mull-padal-dup.c | 145 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup() local 149 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup() 180 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup() local 181 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
|
D | 4x16c2-minmax-neon-mlal-padal-dup.c | 124 …int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() local 129 …vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() 133 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() 447 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() local 451 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() 512 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() local 513 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
|
D | 4x16c2-minmax-neon-mull-padal-dup.c | 209 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() local 213 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 274 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() local 275 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_gemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x8c2-minmax-neon-mlal-padal-dup.c | 125 …int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() local 130 …vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() 134 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() 288 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() local 292 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() 323 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup() local 324 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mlal_padal_dup()
|
D | 4x8c2-minmax-neon-mull-padal-dup.c | 162 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup() local 166 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup() 197 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup() local 198 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x8c2__neon_mull_padal_dup()
|
D | 4x16c2-minmax-neon-mlal-padal-dup.c | 141 …int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() local 146 …vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() 150 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() 464 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() local 468 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() 529 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup() local 530 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
|
D | 4x16c2-minmax-neon-mull-padal-dup.c | 226 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() local 230 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() 291 …const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_… in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup() local 292 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0); in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mull_padal_dup()
|