Home
last modified time | relevance | path

Searched refs:vpsum1x4 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c154 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
158 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c169 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
173 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c191 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
195 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D2x8c8-minmax-fp32-neonv8-mlal.c230 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
234 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c229 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
233 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c229 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
233 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D3x8c16-minmax-rndnu-neon-mlal.c214 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
218 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c228 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
232 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D2x16c8-minmax-rndnu-neon-mull.c236 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
240 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D3x8c8-minmax-rndnu-neon-mlal.c292 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
296 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c170 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
174 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c185 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
189 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c245 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
249 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c245 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
249 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c246 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
250 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D3x8c8-minmax-rndnu-neon-mull.c209 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
213 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c232 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
236 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D2x16c8-minmax-rndnu-neon-mull.c252 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
256 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D4x8c8-minmax-rndnu-neon-mull.c248 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
252 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D4x8c16-minmax-rndnu-neon-mlal.c279 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
283 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D2x16c16-minmax-rndnu-neon-mlal.c283 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
287 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D2x8c8-minmax-fp32-neon-mlal.c229 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
233 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c230 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
234 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D2x8c8-minmax-fp32-neonv8-mlal.c246 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
250 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c245 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
249 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

12