Home
last modified time | relevance | path

Searched refs:vsum0xCD (Results 1 – 25 of 88) sorted by relevance

1234

/external/XNNPACK/src/qs8-gemm/gen/
D1x16c8-minmax-rndnu-neon-mull.c126 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
132 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull()
159 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
161 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull()
D1x16c16-minmax-rndnu-neon-mlal.c142 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
147 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal()
174 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
176 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c215 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
221 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
248 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
250 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
D2x16c8-minmax-rndnu-neon-mull.c181 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
195 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
226 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
228 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D1x16c4s2-minmax-rndnu-neon-mull.c126 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
128 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull()
D2x16c16-minmax-rndnu-neon-mlal.c213 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
226 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
257 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
259 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
D1x16c4-minmax-rndnu-neon-mull-ld1r.c163 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
165 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r()
D1x16c4-minmax-rndnu-neon-mull-dup.c162 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
164 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup()
D1x16c4-minmax-rndnu-neon-mull-ld2r.c162 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
164 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r()
D2x16c8-minmax-rndnu-neon-mlal.c320 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
334 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal()
365 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
367 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal()
D3x16c8-minmax-rndnu-neon-mull.c236 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
258 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
293 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
295 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D2x16c4s2-minmax-rndnu-neon-mull.c178 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
180 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull()
D1x16c4s2-minmax-rndnu-neon-mlal.c215 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
217 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x16c8-minmax-rndnu-neon-mull.c140 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
146 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull()
173 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
175 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull()
D1x16c16-minmax-rndnu-neon-mlal.c156 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
161 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal()
188 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
190 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c229 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
235 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
262 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
264 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
D2x16c8-minmax-rndnu-neon-mull.c197 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
211 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
242 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
244 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D1x16c4s2-minmax-rndnu-neon-mull.c140 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull() local
142 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull()
D2x16c16-minmax-rndnu-neon-mlal.c229 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
242 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
273 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
275 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
D1x16c4-minmax-rndnu-neon-mull-ld2r.c175 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
177 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r()
D1x16c4-minmax-rndnu-neon-mull-ld1r.c176 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
178 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r()
D1x16c4-minmax-rndnu-neon-mull-dup.c175 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
177 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup()
D3x16c8-minmax-rndnu-neon-mull.c254 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
276 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
311 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
313 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D2x16c8-minmax-rndnu-neon-mlal.c336 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
350 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal()
381 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal() local
383 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mlal()
D2x16c4s2-minmax-rndnu-neon-mull.c194 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
196 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull()

1234