Home
last modified time | relevance | path

Searched refs:vpsum1x0 (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-neon-mull-padal.c146 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
150 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c162 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
166 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c221 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
225 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c183 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
187 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c207 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
211 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c220 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
224 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c228 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
232 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c284 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
288 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c260 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
264 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c252 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
256 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c347 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
351 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c295 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
299 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c367 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
371 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-neon-mull-padal.c162 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
166 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c178 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
182 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c237 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
241 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c201 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
205 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c225 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
229 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c240 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
244 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c244 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
248 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c302 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
306 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c272 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
276 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c276 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
280 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c313 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
317 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c383 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
387 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()

12