| /libCEED/include/ceed/jit-source/cuda/ ! |
| H A D | cuda-shared-basis-read-write-templates.h | 28 …(SharedData_Cuda &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in ReadElementStrided1d() argument
35 r_u[comp] = d_u[ind + comp * strides_comp]; in ReadElementStrided1d()
44 …(SharedData_Cuda &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in WriteElementStrided1d() argument
51 d_v[ind + comp * strides_comp] = r_v[comp]; in WriteElementStrided1d()
57 …(SharedData_Cuda &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in SumElementStrided1d() argument
64 d_v[ind + comp * strides_comp] += r_v[comp]; in SumElementStrided1d()
77 …(SharedData_Cuda &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in ReadElementStrided2d() argument
84 r_u[comp] = d_u[ind + comp * strides_comp]; in ReadElementStrided2d()
93 …(SharedData_Cuda &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in WriteElementStrided2d() argument
100 d_v[ind + comp * strides_comp] = r_v[comp]; in WriteElementStrided2d()
[all …]
|
| /libCEED/include/ceed/jit-source/hip/ ! |
| H A D | hip-shared-basis-read-write-templates.h | 28 …d(SharedData_Hip &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in ReadElementStrided1d() argument
35 r_u[comp] = d_u[ind + comp * strides_comp]; in ReadElementStrided1d()
44 …d(SharedData_Hip &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in WriteElementStrided1d() argument
51 d_v[ind + comp * strides_comp] = r_v[comp]; in WriteElementStrided1d()
57 …d(SharedData_Hip &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in SumElementStrided1d() argument
64 d_v[ind + comp * strides_comp] += r_v[comp]; in SumElementStrided1d()
77 …d(SharedData_Hip &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in ReadElementStrided2d() argument
84 r_u[comp] = d_u[ind + comp * strides_comp]; in ReadElementStrided2d()
93 …d(SharedData_Hip &data, const CeedInt elem, const CeedInt strides_node, const CeedInt strides_comp, in WriteElementStrided2d() argument
100 d_v[ind + comp * strides_comp] = r_v[comp]; in WriteElementStrided2d()
[all …]
|
| /libCEED/include/ceed/jit-source/sycl/ ! |
| H A D | sycl-shared-basis-read-write-templates.h | 29 … const CeedInt strides_comp, const CeedInt strides_elem, global const CeedScalar *restrict d_u, in ReadElementStrided1d() argument
38 r_u[comp] = d_u[ind + comp * strides_comp]; in ReadElementStrided1d()
47 … const CeedInt strides_comp, const CeedInt strides_elem, private const CeedScalar *restrict r_v, in WriteElementStrided1d() argument
56 d_v[ind + comp * strides_comp] = r_v[comp]; in WriteElementStrided1d()
69 … const CeedInt strides_comp, const CeedInt strides_elem, global const CeedScalar *restrict d_u, in ReadElementStrided2d() argument
79 r_u[comp] = d_u[ind + comp * strides_comp]; in ReadElementStrided2d()
88 … const CeedInt strides_comp, const CeedInt strides_elem, private const CeedScalar *restrict r_v, in WriteElementStrided2d() argument
98 d_v[ind + comp * strides_comp] = r_v[comp]; in WriteElementStrided2d()
111 … const CeedInt strides_comp, const CeedInt strides_elem, global const CeedScalar *restrict d_u, in ReadElementStrided3d() argument
122 r_u[z + comp * P_1D] = d_u[ind + comp * strides_comp]; in ReadElementStrided3d()
[all …]
|
| H A D | sycl-gen-templates.h | 33 inline void readDofsOffset1d(const CeedInt num_comp, const CeedInt strides_comp, const CeedInt P_1D… in readDofsOffset1d() argument
42 r_u[comp] = d_u[ind + strides_comp * comp]; in readDofsOffset1d()
50 …const CeedInt num_comp, const CeedInt P_1D, const CeedInt strides_node, const CeedInt strides_comp, in readDofsStrided1d() argument
60 r_u[comp] = d_u[ind + comp * strides_comp]; in readDofsStrided1d()
68 inline void writeDofsOffset1d(const CeedInt num_comp, const CeedInt strides_comp, const CeedInt P_1… in writeDofsOffset1d() argument
77 …atomic_fetch_add_explicit(&d_v[ind + strides_comp * comp], r_v[comp], memory_order_relaxed, memory… in writeDofsOffset1d()
84 …const CeedInt num_comp, const CeedInt P_1D, const CeedInt strides_node, const CeedInt strides_comp, in writeDofsStrided1d() argument
94 d_v[ind + comp * strides_comp] = r_v[comp]; in writeDofsStrided1d()
106 inline void readDofsOffset2d(const CeedInt num_comp, const CeedInt strides_comp, const CeedInt P_1D… in readDofsOffset2d() argument
115 for (CeedInt comp = 0; comp < num_comp; ++comp) r_u[comp] = d_u[ind + strides_comp * comp]; in readDofsOffset2d()
[all …]
|