/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/ |
H A D | jit_uni_eltwise_int.cpp | 271 uni_vcvtps2dq(vr_to, vr_to); in process_linear() 292 cvtps2dq(vr_to, vr_to); in process_relu() 303 vcvtps2dq(vr_to, vr_to); in process_relu() 313 vblendmps(vr_to | k_mask, vr_to, vr_from); in process_relu() 314 vcvtps2dq(vr_to, vr_to); in process_relu() 356 vpackssdw(vr_to, vr_to, vmm_zero); in store_8bit() 362 vpacksswb(vr_to, vr_to, vmm_zero); in store_8bit() 364 vpackuswb(vr_to, vr_to, vmm_zero); in store_8bit() 369 vpackssdw(vr_to, vr_to, vmm_zero); in store_8bit() 371 vpacksswb(vr_to, vr_to, vmm_zero); in store_8bit() [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/ |
H A D | jit_uni_eltwise_int.cpp | 271 uni_vcvtps2dq(vr_to, vr_to); in process_linear() 292 cvtps2dq(vr_to, vr_to); in process_relu() 303 vcvtps2dq(vr_to, vr_to); in process_relu() 313 vblendmps(vr_to | k_mask, vr_to, vr_from); in process_relu() 314 vcvtps2dq(vr_to, vr_to); in process_relu() 356 vpackssdw(vr_to, vr_to, vmm_zero); in store_8bit() 362 vpacksswb(vr_to, vr_to, vmm_zero); in store_8bit() 364 vpackuswb(vr_to, vr_to, vmm_zero); in store_8bit() 369 vpackssdw(vr_to, vr_to, vmm_zero); in store_8bit() 371 vpacksswb(vr_to, vr_to, vmm_zero); in store_8bit() [all …]
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/aarch64/ |
H A D | jit_uni_eltwise_int.cpp | 219 uni_str(vr_to, mem_to); in store_32bit() 222 st1w(vr_to.s, p_vl1, ptr(mem_to)); in store_32bit() 276 scvtf(vr_to.s, p_all_one / T_m, vr_from.s); in process_linear() 287 frinti(vr_to.s, p_all_one / T_m, vr_to.s); in process_linear() 288 fcvtzs(vr_to.s, p_all_one / T_m, vr_to.s); in process_linear() 303 fmul(vr_to.s, vr_from.s, ts_alpha); in process_relu() 307 sel(vr_to.s, p_mask / T_m, vr_from.s, vr_to.s); in process_relu() 309 frinti(vr_to.s, p_all_one / T_m, vr_to.s); in process_relu() 310 fcvtzs(vr_to.s, p_all_one / T_m, vr_to.s); in process_relu() 324 mov(t_tmp0.d, vr_to.d); in store_8bit() [all …]
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/aarch64/ |
H A D | jit_uni_eltwise_int.cpp | 219 uni_str(vr_to, mem_to); in store_32bit() 222 st1w(vr_to.s, p_vl1, ptr(mem_to)); in store_32bit() 276 scvtf(vr_to.s, p_all_one / T_m, vr_from.s); in process_linear() 287 frinti(vr_to.s, p_all_one / T_m, vr_to.s); in process_linear() 288 fcvtzs(vr_to.s, p_all_one / T_m, vr_to.s); in process_linear() 303 fmul(vr_to.s, vr_from.s, ts_alpha); in process_relu() 307 sel(vr_to.s, p_mask / T_m, vr_from.s, vr_to.s); in process_relu() 309 frinti(vr_to.s, p_all_one / T_m, vr_to.s); in process_relu() 310 fcvtzs(vr_to.s, p_all_one / T_m, vr_to.s); in process_relu() 324 mov(t_tmp0.d, vr_to.d); in store_8bit() [all …]
|
/dports/math/gravity/Gravity-da941e9/examples/Optimization/NonLinear/Power/ACOPF/ |
H A D | ACOPF_main.cpp | 145 var<> vr_from, vr_to, vi_from, vi_to; in main() local 161 vr_to = vr.to(arcs); in main() 212 Flow_P_From -= g_ft*(vr_from*vr_to + vi_from*vi_to); in main() 213 Flow_P_From -= b_ft*(vi_from*vr_to - vr_from*vi_to); in main() 225 Flow_P_To -= g_tt*(pow(vr_to, 2) + pow(vi_to, 2)); in main() 226 Flow_P_To -= g_tf*(vr_from*vr_to + vi_from*vi_to); in main() 227 Flow_P_To -= b_tf*(vi_to*vr_from - vr_to*vi_from); in main() 240 Flow_Q_From += b_ft*(vr_from*vr_to + vi_from*vi_to); in main() 253 Flow_Q_To += b_tt*(pow(vr_to, 2) + pow(vi_to, 2)); in main() 254 Flow_Q_To += b_tf*(vr_from*vr_to + vi_from*vi_to); in main() [all …]
|
/dports/math/gravity/Gravity-da941e9/examples/Optimization/NonLinear/Power/ |
H A D | PowerNet.cpp | 1282 var<> vr_from, vr_to, vi_from, vi_to; in build_ACOPF() local 1298 vr_to = vr.to(arcs); in build_ACOPF() 1349 Flow_P_From -= g_ft*(vr_from*vr_to + vi_from*vi_to); in build_ACOPF() 1350 Flow_P_From -= b_ft*(vi_from*vr_to - vr_from*vi_to); in build_ACOPF() 1362 Flow_P_To -= g_tt*(pow(vr_to, 2) + pow(vi_to, 2)); in build_ACOPF() 1363 Flow_P_To -= g_tf*(vr_from*vr_to + vi_from*vi_to); in build_ACOPF() 1364 Flow_P_To -= b_tf*(vi_to*vr_from - vr_to*vi_from); in build_ACOPF() 1377 Flow_Q_From += b_ft*(vr_from*vr_to + vi_from*vi_to); in build_ACOPF() 1390 Flow_Q_To += b_tt*(pow(vr_to, 2) + pow(vi_to, 2)); in build_ACOPF() 1391 Flow_Q_To += b_tf*(vr_from*vr_to + vi_from*vi_to); in build_ACOPF() [all …]
|