| /libCEED/include/ceed/jit-source/cuda/ |
| H A D | cuda-shared-basis-tensor-at-points.h | 36 CeedScalar r_C[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; in InterpAtPoints() local 49 Interp1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_C); in InterpAtPoints() 52 InterpTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_C); in InterpAtPoints() 56 InterpTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_C); in InterpAtPoints() 67 … InterpAtPoints1d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_C, r_X, r_V); in InterpAtPoints() 69 … InterpAtPoints2d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_C, r_X, r_V); in InterpAtPoints() 71 … InterpAtPoints3d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_C, r_X, r_V); in InterpAtPoints() 92 CeedScalar r_C[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; in InterpTransposeAtPoints() local 103 for (CeedInt i = 0; i < BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1); i++) r_C[i] = 0.0; in InterpTransposeAtPoints() 125 …TransposeAtPoints1d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_U, r_X, r_C); in InterpTransposeAtPoints() [all …]
|
| H A D | cuda-shared-basis-tensor-at-points-templates.h | 44 …s1d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *__restrict__ r_C, const CeedScalar *… in InterpAtPoints1d() argument 52 if (data.t_id_x < Q_1D) data.slice[data.t_id_x] = r_C[comp]; in InterpAtPoints1d() 66 CeedScalar *__restrict__ r_C) { in InterpTransposeAtPoints1d() argument 82 if (data.t_id_x < Q_1D) r_C[comp] += data.slice[data.t_id_x]; in InterpTransposeAtPoints1d() 90 …s1d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *__restrict__ r_C, const CeedScalar *… in GradAtPoints1d() argument 99 if (data.t_id_x < Q_1D) data.slice[data.t_id_x] = r_C[comp]; in GradAtPoints1d() 113 CeedScalar *__restrict__ r_C) { in GradTransposeAtPoints1d() argument 129 if (data.t_id_x < Q_1D) r_C[comp] += data.slice[data.t_id_x]; in GradTransposeAtPoints1d() 141 …s2d(SharedData_Cuda &data, const CeedInt p, const CeedScalar *__restrict__ r_C, const CeedScalar *… in InterpAtPoints2d() argument 150 …data.t_id_x < Q_1D && data.t_id_y < Q_1D) data.slice[data.t_id_x + data.t_id_y * Q_1D] = r_C[comp]; in InterpAtPoints2d() [all …]
|
| /libCEED/include/ceed/jit-source/hip/ |
| H A D | hip-shared-basis-tensor-at-points.h | 37 CeedScalar r_C[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; in __launch_bounds__() local 50 Interp1d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_C); in __launch_bounds__() 53 InterpTensor2d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_C); in __launch_bounds__() 57 InterpTensor3d<BASIS_NUM_COMP, BASIS_P_1D, BASIS_Q_1D, BASIS_T_1D>(data, r_U, s_B, r_C); in __launch_bounds__() 68 … InterpAtPoints1d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_C, r_X, r_V); in __launch_bounds__() 70 … InterpAtPoints2d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_C, r_X, r_V); in __launch_bounds__() 72 … InterpAtPoints3d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_C, r_X, r_V); in __launch_bounds__() 93 CeedScalar r_C[BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1)]; in __launch_bounds__() local 104 for (CeedInt i = 0; i < BASIS_NUM_COMP * (BASIS_DIM > 2 ? BASIS_Q_1D : 1); i++) r_C[i] = 0.0; in __launch_bounds__() 126 …TransposeAtPoints1d<BASIS_NUM_COMP, BASIS_NUM_PTS, BASIS_P_1D, BASIS_Q_1D>(data, i, r_U, r_X, r_C); in __launch_bounds__() [all …]
|
| H A D | hip-shared-basis-tensor-at-points-templates.h | 44 …ts1d(SharedData_Hip &data, const CeedInt p, const CeedScalar *__restrict__ r_C, const CeedScalar *… in InterpAtPoints1d() argument 53 if (data.t_id_x < Q_1D) data.slice[data.t_id_x] = r_C[comp]; in InterpAtPoints1d() 67 … const CeedScalar *__restrict__ r_X, CeedScalar *__restrict__ r_C) { in InterpTransposeAtPoints1d() argument 83 if (data.t_id_x < Q_1D) r_C[comp] += data.slice[data.t_id_x]; in InterpTransposeAtPoints1d() 91 …ts1d(SharedData_Hip &data, const CeedInt p, const CeedScalar *__restrict__ r_C, const CeedScalar *… in GradAtPoints1d() argument 100 if (data.t_id_x < Q_1D) data.slice[data.t_id_x] = r_C[comp]; in GradAtPoints1d() 114 … const CeedScalar *__restrict__ r_X, CeedScalar *__restrict__ r_C) { in GradTransposeAtPoints1d() argument 130 if (data.t_id_x < Q_1D) r_C[comp] += data.slice[data.t_id_x]; in GradTransposeAtPoints1d() 142 …ts2d(SharedData_Hip &data, const CeedInt p, const CeedScalar *__restrict__ r_C, const CeedScalar *… in InterpAtPoints2d() argument 151 …data.t_id_x < Q_1D && data.t_id_y < Q_1D) data.slice[data.t_id_x + data.t_id_y * Q_1D] = r_C[comp]; in InterpAtPoints2d() [all …]
|