BdaBridge lets cusparseSolver copy the resulting x vector from the GPU directly into the BlockVector from ISTLSolverEbos, this only works if the underlying data from BlockVector is contiguous

opm/simulators/linalg/bda/BdaBridge.cpp

@@ -203,9 +203,7 @@ template <class BridgeVector>
 void BdaBridge::get_result(BridgeVector &x){
 #if HAVE_CUDA
 	if(use_gpu){
-		double *h_x = backend->post_process();
-        // convert flat array to blockvector
-        memcpy(&(x[0]), h_x, sizeof(double) * x.N() * x[0].dim());
+		backend->post_process(&(x[0][0]));
 	}
 #endif
 }

opm/simulators/linalg/bda/cusparseSolverBackend.cu

@@ -427,7 +427,8 @@ namespace Opm
 
 
 	// copy result to host memory
-	double* cusparseSolverBackend::post_process(){
+	// caller must be sure that x is a valid array
+	void cusparseSolverBackend::post_process(double *x){
 
 		double t1, t2;
 		if(verbosity > 2){
@@ -441,8 +442,6 @@ namespace Opm
 			t2 = second();
 			printf("cusparseSolver::post_process(): %f s\n", t2-t1);
 		}
-
-		return x;
 	} // end post_process()
 
 

opm/simulators/linalg/bda/cusparseSolverBackend.hpp

@@ -94,7 +94,7 @@ public:
     // return true iff converged
     bool solve_system(BdaResult &res);
 
-    double* post_process();
+    void post_process(double *x);
 
     bool isInitialized();