Home
last modified time | relevance | path

Searched refs:vpsum1x2 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c149 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
152 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c164 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
167 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c186 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
189 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D2x8c8-minmax-fp32-neonv8-mlal.c225 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
228 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c224 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
227 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c224 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
227 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D3x8c16-minmax-rndnu-neon-mlal.c209 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
212 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c223 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
226 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D2x16c8-minmax-rndnu-neon-mull.c231 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
234 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D3x8c8-minmax-rndnu-neon-mlal.c287 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
290 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c165 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
168 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c180 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
183 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c240 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
243 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c240 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
243 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c241 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
244 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D3x8c8-minmax-rndnu-neon-mull.c204 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
207 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c227 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
230 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D2x16c8-minmax-rndnu-neon-mull.c247 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
250 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D4x8c8-minmax-rndnu-neon-mull.c243 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
246 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D4x8c16-minmax-rndnu-neon-mlal.c274 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
277 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D2x16c16-minmax-rndnu-neon-mlal.c278 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
281 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D2x8c8-minmax-fp32-neon-mlal.c224 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
227 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c225 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
228 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D2x8c8-minmax-fp32-neonv8-mlal.c241 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
244 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c240 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
243 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

12