Home
last modified time | relevance | path

Searched refs:vpsum1x6 (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-neon-mull-padal.c155 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
158 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c171 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
174 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c230 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
233 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c192 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
195 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c216 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
219 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c229 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
232 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c237 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
240 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c293 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
296 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c269 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
272 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c261 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
264 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c356 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
359 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c304 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
307 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c376 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
379 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-neon-mull-padal.c171 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
174 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c187 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
190 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c246 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
249 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c210 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
213 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c234 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
237 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c249 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
252 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c253 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
256 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c311 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
314 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c281 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
284 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c285 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
288 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c322 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
325 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c392 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
395 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()

12