/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 266 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 269 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 298 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 301 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 335 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 338 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 405 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 408 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 383 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 386 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 404 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 407 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 468 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 471 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 524 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 527 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 643 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 646 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c8-minmax-neon-mull-padal.c | 250 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 253 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 282 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 285 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 317 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 320 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 389 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 392 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 365 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 368 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 384 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 387 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 506 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 509 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 448 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 451 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 623 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 626 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|