/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/x64/injectors/ |
H A D | jit_uni_eltwise_injector.cpp | 287 h->uni_vpaddd(vmm_aux2, vmm_aux2, table_val(exponent_bias)); in exp_compute_vector_fwd() 296 vec_shift(vmm_aux2, vmm_aux2, true /*shift_left*/, n_mantissa_bits); in exp_compute_vector_fwd() 798 h->uni_vsubps(vmm_aux2, vmm_aux2, vmm_src); in logistic_compute_vector_fwd() 878 h->uni_vxorps(vmm_aux2, vmm_aux2, table_val(exponent_bias)); in log_compute_vector_fwd() 879 vec_shift(vmm_aux2, vmm_aux2, true, n_mantissa_bits); in log_compute_vector_fwd() 1209 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_aux0); in gelu_tanh_compute_vector_bwd() 1225 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_aux3); in gelu_tanh_compute_vector_bwd() 1227 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_src); in gelu_tanh_compute_vector_bwd() 1231 h->uni_vfnmadd231ps(vmm_aux2, vmm_aux2, vmm_src); in gelu_tanh_compute_vector_bwd() 1463 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_src); in gelu_erf_compute_vector_bwd() [all …]
|
H A D | jit_uni_eltwise_injector.hpp | 169 Vmm vmm_mask, vmm_aux0, vmm_aux1, vmm_aux2, vmm_aux3, vmm_aux4, vmm_tmp; member
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/x64/injectors/ |
H A D | jit_uni_eltwise_injector.cpp | 288 h->uni_vpaddd(vmm_aux2, vmm_aux2, table_val(exponent_bias)); in exp_compute_vector_fwd() 297 vec_shift(vmm_aux2, vmm_aux2, true /*shift_left*/, n_mantissa_bits); in exp_compute_vector_fwd() 801 h->uni_vsubps(vmm_aux2, vmm_aux2, vmm_src); in logistic_compute_vector_fwd() 881 h->uni_vxorps(vmm_aux2, vmm_aux2, table_val(exponent_bias)); in log_compute_vector_fwd() 882 vec_shift(vmm_aux2, vmm_aux2, true, n_mantissa_bits); in log_compute_vector_fwd() 1220 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_aux0); in gelu_tanh_compute_vector_bwd() 1236 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_aux3); in gelu_tanh_compute_vector_bwd() 1238 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_src); in gelu_tanh_compute_vector_bwd() 1242 h->uni_vfnmadd231ps(vmm_aux2, vmm_aux2, vmm_src); in gelu_tanh_compute_vector_bwd() 1474 h->uni_vmulps(vmm_aux2, vmm_aux2, vmm_src); in gelu_erf_compute_vector_bwd() [all …]
|
H A D | jit_uni_eltwise_injector.hpp | 173 Vmm vmm_mask, vmm_aux0, vmm_aux1, vmm_aux2, vmm_aux3, vmm_aux4, vmm_tmp; member
|
/dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/aarch64/injectors/ |
H A D | jit_uni_eltwise_injector.cpp | 738 blend_with_mask(vmm_src, vmm_aux2); in soft_relu_compute_vector_fwd() 766 h->fsub(vmm_aux2, vmm_aux2, vmm_src); in logistic_compute_vector_fwd() 771 blend_with_mask(vmm_aux2, vmm_src); in logistic_compute_vector_fwd() 914 h->mov(ZRegD(IDX(vmm_aux2)), in gelu_erf_compute_vector_fwd() 919 h->fdiv(vmm_aux4, p_all, vmm_aux2); in gelu_erf_compute_vector_fwd() 996 h->mov(ZRegD(IDX(vmm_aux2)), in gelu_tanh_compute_vector_bwd() 1006 h->fmul(vmm_aux2, vmm_aux2, vmm_aux0); in gelu_tanh_compute_vector_bwd() 1022 h->fmls(vmm_aux2, p_all / T_m, vmm_aux2, vmm_src); in gelu_tanh_compute_vector_bwd() 1183 h->fmul(vmm_aux2, vmm_aux2, in gelu_erf_compute_vector_bwd() 1185 h->fmul(vmm_aux2, vmm_aux2, vmm_src); in gelu_erf_compute_vector_bwd() [all …]
|
H A D | jit_uni_eltwise_injector.hpp | 176 TRegS vmm_mask {0}, vmm_aux0 {0}, vmm_aux1 {0}, vmm_aux2 {0}, vmm_aux3 {0}, member
|
/dports/math/onednn/oneDNN-2.5.1/src/cpu/aarch64/injectors/ |
H A D | jit_uni_eltwise_injector.cpp | 738 blend_with_mask(vmm_src, vmm_aux2); in soft_relu_compute_vector_fwd() 766 h->fsub(vmm_aux2, vmm_aux2, vmm_src); in logistic_compute_vector_fwd() 771 blend_with_mask(vmm_aux2, vmm_src); in logistic_compute_vector_fwd() 914 h->mov(ZRegD(IDX(vmm_aux2)), in gelu_erf_compute_vector_fwd() 919 h->fdiv(vmm_aux4, p_all, vmm_aux2); in gelu_erf_compute_vector_fwd() 996 h->mov(ZRegD(IDX(vmm_aux2)), in gelu_tanh_compute_vector_bwd() 1006 h->fmul(vmm_aux2, vmm_aux2, vmm_aux0); in gelu_tanh_compute_vector_bwd() 1022 h->fmls(vmm_aux2, p_all / T_m, vmm_aux2, vmm_src); in gelu_tanh_compute_vector_bwd() 1183 h->fmul(vmm_aux2, vmm_aux2, in gelu_erf_compute_vector_bwd() 1185 h->fmul(vmm_aux2, vmm_aux2, vmm_src); in gelu_erf_compute_vector_bwd() [all …]
|
H A D | jit_uni_eltwise_injector.hpp | 176 TRegS vmm_mask {0}, vmm_aux0 {0}, vmm_aux1 {0}, vmm_aux2 {0}, vmm_aux3 {0}, member
|