| /libCEED/include/ceed/jit-source/hip/ |
| H A D | hip-ref-restriction-curl-oriented.h | 18 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in CurlOrientedNoTranspose() local 20 const CeedInt ind_dl = loc_node > 0 ? indices[node - 1] : 0; in CurlOrientedNoTranspose() 22 const CeedInt ind_du = loc_node < (RSTR_ELEM_SIZE - 1) ? indices[node + 1] : 0; in CurlOrientedNoTranspose() 29 value += loc_node > 0 ? u[ind_dl + comp * RSTR_COMP_STRIDE] * curl_orient_dl : 0.0; in CurlOrientedNoTranspose() 31 …value += loc_node < (RSTR_ELEM_SIZE - 1) ? u[ind_du + comp * RSTR_COMP_STRIDE] * curl_orient_du : … in CurlOrientedNoTranspose() 32 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = value; in CurlOrientedNoTranspose() 43 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in CurlOrientedUnsignedNoTranspose() local 45 const CeedInt ind_dl = loc_node > 0 ? indices[node - 1] : 0; in CurlOrientedUnsignedNoTranspose() 47 const CeedInt ind_du = loc_node < (RSTR_ELEM_SIZE - 1) ? indices[node + 1] : 0; in CurlOrientedUnsignedNoTranspose() 54 value += loc_node > 0 ? u[ind_dl + comp * RSTR_COMP_STRIDE] * curl_orient_dl : 0.0; in CurlOrientedUnsignedNoTranspose() [all …]
|
| H A D | hip-ref-restriction-strided.h | 17 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in StridedNoTranspose() local 21 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = in StridedNoTranspose() 22 u[loc_node * RSTR_STRIDE_NODES + comp * RSTR_STRIDE_COMP + elem * RSTR_STRIDE_ELEM]; in StridedNoTranspose() 32 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in StridedTranspose() local 36 v[loc_node * RSTR_STRIDE_NODES + comp * RSTR_STRIDE_COMP + elem * RSTR_STRIDE_ELEM] += in StridedTranspose() 37 u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; in StridedTranspose()
|
| H A D | hip-ref-restriction-at-points.h | 20 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in AtPointsTranspose() local 23 if (loc_node >= points_per_elem[elem]) continue; in AtPointsTranspose() 25 …atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + … in AtPointsTranspose() 44 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; in AtPointsTranspose() local 47 if (loc_node >= points_per_elem[elem]) continue; in AtPointsTranspose() 49 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; in AtPointsTranspose()
|
| H A D | hip-ref-restriction-offset.h | 18 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OffsetNoTranspose() local 22 …v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = u[ind + comp * RSTR_… in OffsetNoTranspose() 34 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OffsetTranspose() local 38 …atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + … in OffsetTranspose() 56 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; in OffsetTranspose() local 60 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; in OffsetTranspose()
|
| H A D | hip-ref-restriction-oriented.h | 20 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OrientedNoTranspose() local 24 …v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = u[ind + comp * RSTR_… in OrientedNoTranspose() 38 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OrientedTranspose() local 43 …u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * (orient ? -1.0 : 1.0… in OrientedTranspose() 63 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; in OrientedTranspose() local 67 …value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * (orie… in OrientedTranspose()
|
| /libCEED/include/ceed/jit-source/cuda/ |
| H A D | cuda-ref-restriction-curl-oriented.h | 18 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in CurlOrientedNoTranspose() local 20 const CeedInt ind_dl = loc_node > 0 ? indices[node - 1] : 0; in CurlOrientedNoTranspose() 22 const CeedInt ind_du = loc_node < (RSTR_ELEM_SIZE - 1) ? indices[node + 1] : 0; in CurlOrientedNoTranspose() 29 value += loc_node > 0 ? u[ind_dl + comp * RSTR_COMP_STRIDE] * curl_orient_dl : 0.0; in CurlOrientedNoTranspose() 31 …value += loc_node < (RSTR_ELEM_SIZE - 1) ? u[ind_du + comp * RSTR_COMP_STRIDE] * curl_orient_du : … in CurlOrientedNoTranspose() 32 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = value; in CurlOrientedNoTranspose() 43 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in CurlOrientedUnsignedNoTranspose() local 45 const CeedInt ind_dl = loc_node > 0 ? indices[node - 1] : 0; in CurlOrientedUnsignedNoTranspose() 47 const CeedInt ind_du = loc_node < (RSTR_ELEM_SIZE - 1) ? indices[node + 1] : 0; in CurlOrientedUnsignedNoTranspose() 54 value += loc_node > 0 ? u[ind_dl + comp * RSTR_COMP_STRIDE] * curl_orient_dl : 0.0; in CurlOrientedUnsignedNoTranspose() [all …]
|
| H A D | cuda-ref-restriction-strided.h | 17 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in StridedNoTranspose() local 21 v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = in StridedNoTranspose() 22 u[loc_node * RSTR_STRIDE_NODES + comp * RSTR_STRIDE_COMP + elem * RSTR_STRIDE_ELEM]; in StridedNoTranspose() 32 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in StridedTranspose() local 36 v[loc_node * RSTR_STRIDE_NODES + comp * RSTR_STRIDE_COMP + elem * RSTR_STRIDE_ELEM] += in StridedTranspose() 37 u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; in StridedTranspose()
|
| H A D | cuda-ref-restriction-at-points.h | 20 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in AtPointsTranspose() local 23 if (loc_node >= points_per_elem[elem]) continue; in AtPointsTranspose() 25 …atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + … in AtPointsTranspose() 44 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; in AtPointsTranspose() local 47 if (loc_node >= points_per_elem[elem]) continue; in AtPointsTranspose() 49 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; in AtPointsTranspose()
|
| H A D | cuda-ref-restriction-offset.h | 18 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OffsetNoTranspose() local 22 …v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = u[ind + comp * RSTR_… in OffsetNoTranspose() 34 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OffsetTranspose() local 38 …atomicAdd(&v[ind + comp * RSTR_COMP_STRIDE], u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + … in OffsetTranspose() 56 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; in OffsetTranspose() local 60 value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE]; in OffsetTranspose()
|
| H A D | cuda-ref-restriction-oriented.h | 20 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OrientedNoTranspose() local 24 …v[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] = u[ind + comp * RSTR_… in OrientedNoTranspose() 38 const CeedInt loc_node = node % RSTR_ELEM_SIZE; in OrientedTranspose() local 43 …u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * (orient ? -1.0 : 1.0… in OrientedTranspose() 63 const CeedInt loc_node = t_ind % RSTR_ELEM_SIZE; in OrientedTranspose() local 67 …value[comp] += u[loc_node + comp * RSTR_ELEM_SIZE * RSTR_NUM_ELEM + elem * RSTR_ELEM_SIZE] * (orie… in OrientedTranspose()
|
| /libCEED/backends/sycl-ref/ |
| H A D | ceed-sycl-restriction.sycl.cpp | 41 const CeedInt loc_node = node % elem_size; in CeedElemRestrictionStridedNoTranspose_Sycl() local 45 …v[loc_node + comp * elem_size * num_elem + elem * elem_size] = u[loc_node * stride_nodes + comp * … in CeedElemRestrictionStridedNoTranspose_Sycl() 69 const CeedInt loc_node = node % elem_size; in CeedElemRestrictionOffsetNoTranspose_Sycl() local 73 v[loc_node + comp * elem_size * num_elem + elem * elem_size] = u[ind + comp * comp_stride]; in CeedElemRestrictionOffsetNoTranspose_Sycl() 97 const CeedInt loc_node = node % elem_size; in CeedElemRestrictionStridedTranspose_Sycl() local 101 …v[loc_node * stride_nodes + comp * stride_comp + elem * stride_elem] += u[loc_node + comp * elem_s… in CeedElemRestrictionStridedTranspose_Sycl() 136 CeedInt loc_node = t_ind % elem_size; in CeedElemRestrictionOffsetTranspose_Sycl() local 139 value += u[loc_node + comp * elem_size * num_elem + elem * elem_size]; in CeedElemRestrictionOffsetTranspose_Sycl()
|