Home
last modified time | relevance | path

Searched refs:vpsum1x1 (Results 1 – 25 of 44) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c148 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
151 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c163 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
166 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c185 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
188 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D2x8c8-minmax-fp32-neonv8-mlal.c224 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
227 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c223 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
226 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c223 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
226 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D3x8c16-minmax-rndnu-neon-mlal.c208 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
211 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c222 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
225 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D2x16c8-minmax-rndnu-neon-mull.c230 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
233 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D3x8c8-minmax-rndnu-neon-mlal.c286 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
289 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-rndnu-neon-mull.c164 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
167 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c179 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
182 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c239 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
242 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c239 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
242 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c240 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
243 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D3x8c8-minmax-rndnu-neon-mull.c203 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
206 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D3x8c16-minmax-rndnu-neon-mlal.c226 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
229 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D2x16c8-minmax-rndnu-neon-mull.c246 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull() local
249 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull()
D4x8c8-minmax-rndnu-neon-mull.c242 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
245 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D4x8c16-minmax-rndnu-neon-mlal.c273 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
276 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D2x16c16-minmax-rndnu-neon-mlal.c277 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal() local
280 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D2x8c8-minmax-fp32-neon-mlal.c223 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
226 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c224 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
227 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D2x8c8-minmax-fp32-neonv8-mlal.c240 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
243 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c239 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
242 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

12