Home
last modified time | relevance | path

Searched refs:vpsum2x2 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-neon-mull-padal.c217 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
220 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c241 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
244 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c256 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
259 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c318 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
321 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c288 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
291 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c343 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
346 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x8c8-minmax-neon-mlal-padal.c383 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
386 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c391 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
394 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c412 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
415 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c476 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
479 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c532 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
535 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c651 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
654 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-neon-mull-padal.c199 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
202 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c223 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
226 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c236 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
239 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c300 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
303 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c268 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
271 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c363 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
366 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c325 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
328 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D3x16c16-minmax-neon-mlal-padal.c373 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
376 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c392 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
395 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c514 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
517 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c456 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
459 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c631 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
634 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()