Home
last modified time | relevance | path

Searched refs:vsum0xCD (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D1x16c8-minmax-neon-mull-padal.c126 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
131 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
158 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
160 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
D1x16c16-minmax-neon-mlal-padal.c142 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
147 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
174 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
176 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c215 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
220 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
247 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
249 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c181 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
194 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
225 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
227 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c213 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
226 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
257 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
259 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c236 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
257 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
292 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
294 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c320 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
333 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
364 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
366 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c284 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
305 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
340 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
342 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c291 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
320 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
359 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
361 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c425 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
446 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
481 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
483 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c355 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
384 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
423 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
425 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c530 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
559 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
598 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
600 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x16c8-minmax-neon-mull-padal.c140 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
145 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
172 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
174 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
D1x16c16-minmax-neon-mlal-padal.c156 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
161 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
188 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
190 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c229 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
234 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
261 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
263 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c197 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
210 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
241 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
243 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c229 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
242 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
273 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
275 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c254 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
275 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
310 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
312 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c336 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
349 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
380 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
382 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c302 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
323 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
358 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
360 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c311 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
340 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
379 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
381 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c375 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
404 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
443 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
445 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c443 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
464 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
499 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
501 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c550 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
579 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
618 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
620 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()