Home
last modified time | relevance | path

Searched refs:vsum1x89 (Results 1 – 18 of 18) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D2x16c8-minmax-neon-mull-padal.c203 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
213 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
262 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
264 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c235 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
245 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
294 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
296 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c260 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
278 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
331 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
333 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c342 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
352 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
401 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
403 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c308 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
326 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
379 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
381 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c317 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
343 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
400 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
402 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c381 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
407 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
464 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
466 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c449 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
467 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
520 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
522 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c556 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
582 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
639 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
641 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-gemm/gen/
D2x16c8-minmax-neon-mull-padal.c187 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
197 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
246 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
248 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c219 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
229 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
278 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
280 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c242 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
260 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
313 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
315 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c326 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
336 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
385 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
387 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c290 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
308 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
361 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
363 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c297 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
323 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
380 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
382 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c431 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
449 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
502 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
504 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c361 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
387 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
444 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
446 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c536 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
562 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
619 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
621 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()