Home
last modified time | relevance | path

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

/libCEED/include/ceed/jit-source/cuda/
H A Dcuda-shared-basis-nontensor.h33 LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B); in Interp()
60 LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B); in InterpTranspose()
87 LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B); in InterpTransposeAdd()
116 LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G); in Grad()
143 LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G); in GradTranspose()
170 LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G); in GradTransposeAdd()
H A Dcuda-shared-basis-tensor.h33 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in Interp()
102 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in InterpTranspose()
171 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in InterpTransposeAdd()
243 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in Grad()
245 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in Grad()
286 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in GradCollocated()
326 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in GradTranspose()
328 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in GradTranspose()
369 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in GradCollocatedTranspose()
409 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in GradTransposeAdd()
[all …]
H A Dcuda-shared-basis-tensor-at-points.h41 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in InterpAtPoints()
97 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in InterpTransposeAtPoints()
167 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in InterpTransposeAddAtPoints()
228 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in GradAtPoints()
284 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in GradTransposeAtPoints()
355 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in GradTransposeAddAtPoints()
H A Dcuda-shared-basis-read-write-templates.h16 inline __device__ void LoadMatrix(SharedData_Cuda &data, const CeedScalar *__restrict__ d_B, CeedSc… in LoadMatrix() function
H A Dcuda-gen-templates.h16 inline __device__ void LoadMatrix(SharedData_Cuda &data, const CeedScalar *__restrict__ d_B, CeedSc… in LoadMatrix() function
/libCEED/include/ceed/jit-source/hip/
H A Dhip-shared-basis-nontensor.h34 LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B); in __launch_bounds__()
61 LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B); in __launch_bounds__()
88 LoadMatrix<BASIS_P, BASIS_Q>(data, c_B, s_B); in __launch_bounds__()
118 LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G); in __launch_bounds__()
145 LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G); in __launch_bounds__()
172 LoadMatrix<BASIS_P, BASIS_Q * BASIS_DIM>(data, c_G, s_G); in __launch_bounds__()
H A Dhip-shared-basis-tensor.h34 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
103 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
172 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
245 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
247 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in __launch_bounds__()
289 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in __launch_bounds__()
330 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
332 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in __launch_bounds__()
374 LoadMatrix<BASIS_Q_1D, BASIS_HAS_COLLOCATED_GRAD ? BASIS_Q_1D : BASIS_P_1D>(data, c_G, s_G); in __launch_bounds__()
415 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
[all …]
H A Dhip-shared-basis-tensor-at-points.h42 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
98 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
168 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
230 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
286 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
357 LoadMatrix<BASIS_P_1D, BASIS_Q_1D>(data, c_B, s_B); in __launch_bounds__()
H A Dhip-shared-basis-read-write-templates.h16 inline __device__ void LoadMatrix(SharedData_Hip &data, const CeedScalar *__restrict__ d_B, CeedSca… in LoadMatrix() function
H A Dhip-gen-templates.h16 inline __device__ void LoadMatrix(SharedData_Hip &data, const CeedScalar *__restrict__ d_B, CeedSca… in LoadMatrix() function