65       CeedCallCuda(ceed, cuModuleUnload(impl->diag->module));  in CeedOperatorDestroy_Cuda()
68       CeedCallCuda(ceed, cuModuleUnload(impl->diag->module_point_block));  in CeedOperatorDestroy_Cuda()
70     CeedCallCuda(ceed, cudaFree(impl->diag->d_eval_modes_in));  in CeedOperatorDestroy_Cuda()
71     CeedCallCuda(ceed, cudaFree(impl->diag->d_eval_modes_out));  in CeedOperatorDestroy_Cuda()
72     CeedCallCuda(ceed, cudaFree(impl->diag->d_identity));  in CeedOperatorDestroy_Cuda()
73     CeedCallCuda(ceed, cudaFree(impl->diag->d_interp_in));  in CeedOperatorDestroy_Cuda()
74     CeedCallCuda(ceed, cudaFree(impl->diag->d_interp_out));  in CeedOperatorDestroy_Cuda()
75     CeedCallCuda(ceed, cudaFree(impl->diag->d_grad_in));  in CeedOperatorDestroy_Cuda()
76     CeedCallCuda(ceed, cudaFree(impl->diag->d_grad_out));  in CeedOperatorDestroy_Cuda()
77     CeedCallCuda(ceed, cudaFree(impl->diag->d_div_in));  in CeedOperatorDestroy_Cuda()
78     CeedCallCuda(ceed, cudaFree(impl->diag->d_div_out));  in CeedOperatorDestroy_Cuda()
79     CeedCallCuda(ceed, cudaFree(impl->diag->d_curl_in));  in CeedOperatorDestroy_Cuda()
80     CeedCallCuda(ceed, cudaFree(impl->diag->d_curl_out));  in CeedOperatorDestroy_Cuda()
93     CeedCallCuda(ceed, cuModuleUnload(impl->asmb->module));  in CeedOperatorDestroy_Cuda()
94     CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_in));  in CeedOperatorDestroy_Cuda()
95     CeedCallCuda(ceed, cudaFree(impl->asmb->d_B_out));  in CeedOperatorDestroy_Cuda()
1221     CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_identity, interp_bytes));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1222 …  CeedCallCuda(ceed, cudaMemcpy(diag->d_identity, identity, interp_bytes, cudaMemcpyHostToDevice));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1242         CeedCallCuda(ceed, cudaMalloc((void **)&d_interp, interp_bytes * q_comp_interp));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1243 …CeedCallCuda(ceed, cudaMemcpy(d_interp, interp, interp_bytes * q_comp_interp, cudaMemcpyHostToDevi…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1245         CeedCallCuda(ceed, cudaMalloc((void **)&d_grad, interp_bytes * q_comp_grad));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1246 …  CeedCallCuda(ceed, cudaMemcpy(d_grad, grad, interp_bytes * q_comp_grad, cudaMemcpyHostToDevice));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1264         CeedCallCuda(ceed, cudaMalloc((void **)&d_interp, interp_bytes * q_comp_interp));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1265 …CeedCallCuda(ceed, cudaMemcpy(d_interp, interp, interp_bytes * q_comp_interp, cudaMemcpyHostToDevi…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1267         CeedCallCuda(ceed, cudaMalloc((void **)&d_div, interp_bytes * q_comp_div));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1268 …     CeedCallCuda(ceed, cudaMemcpy(d_div, div, interp_bytes * q_comp_div, cudaMemcpyHostToDevice));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1286         CeedCallCuda(ceed, cudaMalloc((void **)&d_interp, interp_bytes * q_comp_interp));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1287 …CeedCallCuda(ceed, cudaMemcpy(d_interp, interp, interp_bytes * q_comp_interp, cudaMemcpyHostToDevi…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1289         CeedCallCuda(ceed, cudaMalloc((void **)&d_curl, interp_bytes * q_comp_curl));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1290 …  CeedCallCuda(ceed, cudaMemcpy(d_curl, curl, interp_bytes * q_comp_curl, cudaMemcpyHostToDevice));  in CeedOperatorAssembleDiagonalSetup_Cuda()
1303 …CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_eval_modes_in, num_eval_modes_in * eval_modes_byte…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1304 …CeedCallCuda(ceed, cudaMemcpy(diag->d_eval_modes_in, eval_modes_in, num_eval_modes_in * eval_modes…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1305 …CeedCallCuda(ceed, cudaMalloc((void **)&diag->d_eval_modes_out, num_eval_modes_out * eval_modes_by…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1306 …CeedCallCuda(ceed, cudaMemcpy(diag->d_eval_modes_out, eval_modes_out, num_eval_modes_out * eval_mo…  in CeedOperatorAssembleDiagonalSetup_Cuda()
1392 …CeedCallCuda(ceed, CeedCompile_Cuda(ceed, diagonal_kernel_source, module, 8, "NUM_EVAL_MODES_IN", …  in CeedOperatorAssembleDiagonalSetupCompile_Cuda()
1395 …CeedCallCuda(ceed, CeedGetKernel_Cuda(ceed, *module, "LinearDiagonal", is_point_block ? &diag->Lin…  in CeedOperatorAssembleDiagonalSetupCompile_Cuda()
1646     CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_in, in_bytes));  in CeedOperatorAssembleSingleSetup_Cuda()
1658 …CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_in[i * elem_size_in * num_qpts_in], h_B_in, elem_size_in …  in CeedOperatorAssembleSingleSetup_Cuda()
1681     CeedCallCuda(ceed, cudaMalloc((void **)&asmb->d_B_out, out_bytes));  in CeedOperatorAssembleSingleSetup_Cuda()
1693 …CeedCallCuda(ceed, cudaMemcpy(&asmb->d_B_out[i * elem_size_out * num_qpts_out], h_B_out, elem_size…  in CeedOperatorAssembleSingleSetup_Cuda()