Home
last modified time | relevance | path

Searched defs:vsum1xAB (Results 1 – 18 of 18) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D2x16c16-minmax-neon-mlal-padal.c220 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
279 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
D2x16c8-minmax-neon-mull-padal.c188 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
247 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
D3x16c8-minmax-neon-mull-padal.c243 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
314 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
D2x16c8-minmax-neon-mlal-padal.c327 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
386 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
D3x16c16-minmax-neon-mlal-padal.c291 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
362 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
D4x16c16-minmax-neon-mlal-padal.c362 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
445 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
D4x16c8-minmax-neon-mull-padal.c298 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
381 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
D3x16c8-minmax-neon-mlal-padal.c432 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
503 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
D4x16c8-minmax-neon-mlal-padal.c537 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
620 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
/external/XNNPACK/src/qs8-igemm/gen/
D2x16c8-minmax-neon-mull-padal.c204 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
263 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
D2x16c16-minmax-neon-mlal-padal.c236 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
295 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
D2x16c8-minmax-neon-mlal-padal.c343 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
402 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
D3x16c8-minmax-neon-mull-padal.c261 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
332 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
D3x16c16-minmax-neon-mlal-padal.c309 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
380 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
D4x16c8-minmax-neon-mull-padal.c318 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
401 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
D4x16c16-minmax-neon-mlal-padal.c382 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
465 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
D3x16c8-minmax-neon-mlal-padal.c450 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
521 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
D4x16c8-minmax-neon-mlal-padal.c557 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
640 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local