Home
last modified time | relevance | path

Searched refs:vpsum2x1 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-neon-mull-padal.c216 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
219 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c240 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
243 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c255 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
258 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c317 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
320 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c287 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
290 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c342 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
345 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x8c8-minmax-neon-mlal-padal.c382 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
385 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c390 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
393 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c411 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
414 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c475 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
478 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c531 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
534 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c650 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
653 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-neon-mull-padal.c198 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
201 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c222 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
225 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c235 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
238 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c299 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
302 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c267 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
270 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c362 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
365 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c324 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
327 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D3x16c16-minmax-neon-mlal-padal.c372 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
375 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c391 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
394 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c513 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
516 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c455 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
458 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c630 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
633 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()