| /libCEED/backends/hip-shared/ |
| H A D | ceed-hip-shared-basis.c | 125 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; in CeedBasisApplyTensorCore_Hip_shared() local
135 elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Hip_shared()
137 …hared_Hip(ceed, data->Interp, NULL, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Hip_shared()
147 … thread_1d, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Hip_shared()
149 …p(ceed, data->Interp, NULL, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Hip_shared()
158 … thread_1d, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Hip_shared()
160 …p(ceed, data->Interp, NULL, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Hip_shared()
388 …void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem… in CeedBasisApplyAtPointsCore_Hip_shared() local
398 … thread_1d, 1, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyAtPointsCore_Hip_shared()
400 …p(ceed, data->InterpAtPoints, NULL, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyAtPointsCore_Hip_shared()
[all …]
|
| /libCEED/backends/cuda-shared/ |
| H A D | ceed-cuda-shared-basis.c | 58 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; in CeedBasisApplyTensorCore_Cuda_shared() local
68 elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Cuda_shared()
70 …ared_Cuda(ceed, data->Interp, NULL, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Cuda_shared()
81 … thread_1d, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Cuda_shared()
84 interp_args)); in CeedBasisApplyTensorCore_Cuda_shared()
93 … thread_1d, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyTensorCore_Cuda_shared()
96 interp_args)); in CeedBasisApplyTensorCore_Cuda_shared()
322 …void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, … in CeedBasisApplyAtPointsCore_Cuda_shared() local
332 … thread_1d, 1, elems_per_block, shared_mem, interp_args)); in CeedBasisApplyAtPointsCore_Cuda_shared()
335 interp_args)); in CeedBasisApplyAtPointsCore_Cuda_shared()
[all …]
|
| /libCEED/backends/hip-ref/ |
| H A D | ceed-hip-ref-basis.c | 51 …void *interp_args[] = {(void *)&num_elem, (void *)&is_transpose, &data->d_interp_1d, &d_u,… in CeedBasisApplyCore_Hip() local
54 CeedCallBackend(CeedRunKernel_Hip(ceed, data->Interp, num_elem, block_size, interp_args)); in CeedBasisApplyCore_Hip()
208 …void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_pe… in CeedBasisApplyAtPointsCore_Hip() local
212 interp_args)); in CeedBasisApplyAtPointsCore_Hip()
284 void *interp_args[] = {(void *)&num_elem, &data->d_interp, &d_u, &d_v}; in CeedBasisApplyNonTensorCore_Hip() local
288 …unKernelDim_Hip(ceed, data->InterpTranspose, grid, block_size_x, 1, elems_per_block, interp_args)); in CeedBasisApplyNonTensorCore_Hip()
290 …end(CeedRunKernelDim_Hip(ceed, data->Interp, grid, block_size_x, 1, elems_per_block, interp_args)); in CeedBasisApplyNonTensorCore_Hip()
|
| /libCEED/backends/cuda-ref/ |
| H A D | ceed-cuda-ref-basis.c | 51 …void *interp_args[] = {(void *)&num_elem, (void *)&is_transpose, &data->d_interp_1d, &d_u,… in CeedBasisApplyCore_Cuda() local
54 CeedCallBackend(CeedRunKernel_Cuda(ceed, data->Interp, num_elem, block_size, interp_args)); in CeedBasisApplyCore_Cuda()
209 …void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_pe… in CeedBasisApplyAtPointsCore_Cuda() local
213 interp_args)); in CeedBasisApplyAtPointsCore_Cuda()
285 void *interp_args[] = {(void *)&num_elem, &data->d_interp, &d_u, &d_v}; in CeedBasisApplyNonTensorCore_Cuda() local
289 …nKernelDim_Cuda(ceed, data->InterpTranspose, grid, block_size_x, 1, elems_per_block, interp_args)); in CeedBasisApplyNonTensorCore_Cuda()
291 …nd(CeedRunKernelDim_Cuda(ceed, data->Interp, grid, block_size_x, 1, elems_per_block, interp_args)); in CeedBasisApplyNonTensorCore_Cuda()
|