Home
last modified time | relevance | path

Searched refs:vsum0x45 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-neon-mull-padal.c90 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
93 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
106 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
108 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
D1x8c16-minmax-neon-mlal-padal.c98 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
101 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
114 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
116 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D1x8c8-minmax-neon-mlal-padal.c139 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
142 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
155 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
157 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D2x8c8-minmax-neon-mull-padal.c121 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
128 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
143 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
145 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
D1x16c8-minmax-neon-mull-padal.c122 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
129 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
144 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
146 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c137 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
144 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
161 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c138 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
145 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
162 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c196 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
203 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
218 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
220 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c152 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
163 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
180 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
182 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c176 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
187 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
204 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
206 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c211 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
218 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
233 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
235 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c183 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
198 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
217 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
219 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-neon-mull-padal.c104 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
107 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
120 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
122 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
D1x8c16-minmax-neon-mlal-padal.c112 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
115 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
128 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
130 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D1x8c8-minmax-neon-mlal-padal.c153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
156 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
169 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
171 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D2x8c8-minmax-neon-mull-padal.c137 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
144 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
161 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
D1x16c8-minmax-neon-mull-padal.c136 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
143 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
158 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
160 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
160 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
175 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
177 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c152 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
159 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
174 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
176 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c212 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
219 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
234 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
236 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c170 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
181 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
198 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
200 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D1x16c8-minmax-neon-mlal-padal.c225 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
232 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
247 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
249 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c194 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
205 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
222 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
224 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c203 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
218 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
237 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
239 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c193 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
208 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
227 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
229 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()

12