Home
last modified time | relevance | path

Searched refs:SharedData_Cuda (Results 1 – 10 of 10) sorted by relevance

/libCEED/include/ceed/jit-source/cuda/
H A Dcuda-shared-basis-tensor-templates.h20 inline __device__ void ContractX1d(SharedData_Cuda &data, const CeedScalar *U, const CeedScalar *B,… in ContractX1d()
36 inline __device__ void ContractTransposeX1d(SharedData_Cuda &data, const CeedScalar *U, const CeedS… in ContractTransposeX1d()
52 inline __device__ void Interp1d(SharedData_Cuda &data, const CeedScalar *__restrict__ r_U, const Ce… in Interp1d()
62 inline __device__ void InterpTranspose1d(SharedData_Cuda &data, const CeedScalar *__restrict__ r_U,… in InterpTranspose1d()
73 inline __device__ void InterpCollocatedNodes1d(SharedData_Cuda &data, const CeedScalar *__restrict_… in InterpCollocatedNodes1d()
84 inline __device__ void InterpTransposeCollocatedNodes1d(SharedData_Cuda &data, const CeedScalar *__… in InterpTransposeCollocatedNodes1d()
95 inline __device__ void Grad1d(SharedData_Cuda &data, const CeedScalar *__restrict__ r_U, const Ceed… in Grad1d()
106 inline __device__ void GradTranspose1d(SharedData_Cuda &data, const CeedScalar *__restrict__ r_U, c… in GradTranspose1d()
117 inline __device__ void Weight1d(SharedData_Cuda &data, const CeedScalar *__restrict__ q_weight_1d, … in Weight1d()
129 inline __device__ void ContractX2d(SharedData_Cuda &data, const CeedScalar *U, const CeedScalar *B,… in ContractX2d()
[all …]
H A Dcuda-shared-basis-nontensor-templates.h16 inline __device__ void Contract1d(SharedData_Cuda &data, const CeedScalar *U, const CeedScalar *B, … in Contract1d()
32 inline __device__ void ContractTranspose1d(SharedData_Cuda &data, const CeedScalar *U, const CeedSc… in ContractTranspose1d()
47 inline __device__ void InterpNonTensor(SharedData_Cuda &data, const CeedScalar *__restrict__ r_U, c… in InterpNonTensor()
58 inline __device__ void InterpTransposeNonTensor(SharedData_Cuda &data, const CeedScalar *__restrict… in InterpTransposeNonTensor()
70 inline __device__ void GradNonTensor(SharedData_Cuda &data, const CeedScalar *__restrict__ r_U, con… in GradNonTensor()
82 inline __device__ void GradTransposeNonTensor(SharedData_Cuda &data, const CeedScalar *__restrict__… in GradTransposeNonTensor()
96 inline __device__ void WeightNonTensor(SharedData_Cuda &data, const CeedScalar *__restrict__ q_weig… in WeightNonTensor()
H A Dcuda-shared-basis-read-write-templates.h16 inline __device__ void LoadMatrix(SharedData_Cuda &data, const CeedScalar *__restrict__ d_B, CeedSc… in LoadMatrix()
28 inline __device__ void ReadElementStrided1d(SharedData_Cuda &data, const CeedInt elem, const CeedIn… in ReadElementStrided1d()
44 inline __device__ void WriteElementStrided1d(SharedData_Cuda &data, const CeedInt elem, const CeedI… in WriteElementStrided1d()
57 inline __device__ void SumElementStrided1d(SharedData_Cuda &data, const CeedInt elem, const CeedInt… in SumElementStrided1d()
77 inline __device__ void ReadElementStrided2d(SharedData_Cuda &data, const CeedInt elem, const CeedIn… in ReadElementStrided2d()
93 inline __device__ void WriteElementStrided2d(SharedData_Cuda &data, const CeedInt elem, const CeedI… in WriteElementStrided2d()
106 inline __device__ void SumElementStrided2d(SharedData_Cuda &data, const CeedInt elem, const CeedInt… in SumElementStrided2d()
126 inline __device__ void ReadElementStrided3d(SharedData_Cuda &data, const CeedInt elem, const CeedIn… in ReadElementStrided3d()
144 inline __device__ void WriteElementStrided3d(SharedData_Cuda &data, const CeedInt elem, const CeedI… in WriteElementStrided3d()
159 inline __device__ void SumElementStrided3d(SharedData_Cuda &data, const CeedInt elem, const CeedInt… in SumElementStrided3d()
[all …]
H A Dcuda-gen-templates.h16 inline __device__ void LoadMatrix(SharedData_Cuda &data, const CeedScalar *__restrict__ d_B, CeedSc… in LoadMatrix()
28 inline __device__ void ReadPoint(SharedData_Cuda &data, const CeedInt elem, const CeedInt p, const … in ReadPoint()
41 inline __device__ void WritePoint(SharedData_Cuda &data, const CeedInt elem, const CeedInt p, const… in WritePoint()
60 inline __device__ void SetEVecStandard1d_Single(SharedData_Cuda &data, const CeedInt n, const CeedS… in SetEVecStandard1d_Single()
73 inline __device__ void ReadLVecStandard1d(SharedData_Cuda &data, const CeedInt num_nodes, const Cee… in ReadLVecStandard1d()
87 inline __device__ void ReadLVecStrided1d(SharedData_Cuda &data, const CeedInt elem, const CeedScala… in ReadLVecStrided1d()
101 inline __device__ void WriteLVecStandard1d(SharedData_Cuda &data, const CeedInt num_nodes, const Ce… in WriteLVecStandard1d()
112 inline __device__ void WriteLVecStandard1d_Single(SharedData_Cuda &data, const CeedInt num_nodes, c… in WriteLVecStandard1d_Single()
129 inline __device__ void WriteLVecStandard1d_Assembly(SharedData_Cuda &data, const CeedInt num_nodes,… in WriteLVecStandard1d_Assembly()
148 inline __device__ void WriteLVecStandard1d_QFAssembly(SharedData_Cuda &data, const CeedInt num_elem… in WriteLVecStandard1d_QFAssembly()
[all …]
H A Dcuda-shared-basis-tensor-flattened-templates.h20 inline __device__ void ContractX2dFlattened(SharedData_Cuda &data, const int t_id_x, const int t_id… in ContractX2dFlattened()
37 inline __device__ void ContractY2dFlattened(SharedData_Cuda &data, const int t_id_x, const int t_id… in ContractY2dFlattened()
54 inline __device__ void ContractTransposeY2dFlattened(SharedData_Cuda &data, const int t_id_x, const… in ContractTransposeY2dFlattened()
71 inline __device__ void ContractTransposeX2dFlattened(SharedData_Cuda &data, const int t_id_x, const… in ContractTransposeX2dFlattened()
88 inline __device__ void ContractTransposeAddX2dFlattened(SharedData_Cuda &data, const int t_id_x, co… in ContractTransposeAddX2dFlattened()
104 inline __device__ void QPack2d(SharedData_Cuda &data, const int t_id_x, const int t_id_y, CeedScala… in QPack2d()
116 inline __device__ void QUnpack2d(SharedData_Cuda &data, const int t_id_x, const int t_id_y, CeedSca… in QUnpack2d()
131 inline __device__ void InterpTensor2dFlattened(SharedData_Cuda &data, CeedScalar *__restrict__ r_U,… in InterpTensor2dFlattened()
150 inline __device__ void InterpTransposeTensor2dFlattened(SharedData_Cuda &data, CeedScalar *__restri… in InterpTransposeTensor2dFlattened()
168 inline __device__ void InterpTensorCollocatedNodes2dFlattened(SharedData_Cuda &data, CeedScalar *__… in InterpTensorCollocatedNodes2dFlattened()
[all …]
H A Dcuda-shared-basis-tensor-at-points-templates.h44 inline __device__ void InterpAtPoints1d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *_… in InterpAtPoints1d()
65 inline __device__ void InterpTransposeAtPoints1d(SharedData_Cuda &data, const CeedInt p, const Ceed… in InterpTransposeAtPoints1d()
90 inline __device__ void GradAtPoints1d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *__r… in GradAtPoints1d()
112 inline __device__ void GradTransposeAtPoints1d(SharedData_Cuda &data, const CeedInt p, const CeedSc… in GradTransposeAtPoints1d()
141 inline __device__ void InterpAtPoints2d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *_… in InterpAtPoints2d()
172 inline __device__ void InterpTransposeAtPoints2d(SharedData_Cuda &data, const CeedInt p, const Ceed… in InterpTransposeAtPoints2d()
210 inline __device__ void GradAtPoints2d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *__r… in GradAtPoints2d()
245 inline __device__ void GradTransposeAtPoints2d(SharedData_Cuda &data, const CeedInt p, const CeedSc… in GradTransposeAtPoints2d()
291 inline __device__ void InterpAtPoints3d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *_… in InterpAtPoints3d()
328 inline __device__ void InterpTransposeAtPoints3d(SharedData_Cuda &data, const CeedInt p, const Ceed… in InterpTransposeAtPoints3d()
[all …]
H A Dcuda-shared-basis-nontensor.h21 SharedData_Cuda data; in Interp()
48 SharedData_Cuda data; in InterpTranspose()
75 SharedData_Cuda data; in InterpTransposeAdd()
104 SharedData_Cuda data; in Grad()
131 SharedData_Cuda data; in GradTranspose()
158 SharedData_Cuda data; in GradTransposeAdd()
187 SharedData_Cuda data; in Weight()
H A Dcuda-types.h39 } SharedData_Cuda; typedef
H A Dcuda-shared-basis-tensor.h21 SharedData_Cuda data; in Interp()
60 SharedData_Cuda data; in InterpCollocated()
90 SharedData_Cuda data; in InterpTranspose()
129 SharedData_Cuda data; in InterpCollocatedTranspose()
159 SharedData_Cuda data; in InterpTransposeAdd()
198 SharedData_Cuda data; in InterpCollocatedTransposeAdd()
231 SharedData_Cuda data; in Grad()
274 SharedData_Cuda data; in GradCollocated()
314 SharedData_Cuda data; in GradTranspose()
357 SharedData_Cuda data; in GradCollocatedTranspose()
[all …]
H A Dcuda-shared-basis-tensor-at-points.h27 SharedData_Cuda data; in InterpAtPoints()
83 SharedData_Cuda data; in InterpTransposeAtPoints()
153 SharedData_Cuda data; in InterpTransposeAddAtPoints()
214 SharedData_Cuda data; in GradAtPoints()
270 SharedData_Cuda data; in GradTransposeAtPoints()
341 SharedData_Cuda data; in GradTransposeAddAtPoints()