Home
last modified time | relevance | path

Searched refs:vpsum2x4 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-neon-mull-padal.c222 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
226 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c246 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
250 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c261 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
265 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c323 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
327 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c293 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
297 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c348 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
352 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x8c8-minmax-neon-mlal-padal.c388 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
392 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c396 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
400 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c417 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
421 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c481 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
485 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c537 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
541 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c656 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
660 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-neon-mull-padal.c204 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
208 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c228 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
232 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c241 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
245 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c305 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
309 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c273 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
277 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c368 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
372 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c330 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
334 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D3x16c16-minmax-neon-mlal-padal.c378 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
382 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c397 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
401 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c519 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
523 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c461 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
465 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c636 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
640 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()