diff --git a/doc/Astaroth_API_specification_and_user_manual/API_specification_and_user_manual.md b/doc/Astaroth_API_specification_and_user_manual/API_specification_and_user_manual.md
index cd27066..28c2ecb 100644
--- a/doc/Astaroth_API_specification_and_user_manual/API_specification_and_user_manual.md
+++ b/doc/Astaroth_API_specification_and_user_manual/API_specification_and_user_manual.md
@@ -171,7 +171,7 @@ AcResult acNodeLoadVertexBufferWithOffset(const Node node, const Stream stream,
                                           const AcMesh host_mesh,
                                           const VertexBufferHandle vtxbuf_handle, const int3 src,
                                           const int3 dst, const int num_vertices);
-AcResult acGridLoadMesh(const AcMesh host_mesh, const Stream stream);
+AcResult acGridLoadMesh(const Stream stream, const AcMesh host_mesh);
 ```
 
 Storing meshes and vertex buffer to host memory.
@@ -435,7 +435,7 @@ after initialization.
 
 ### Decomposition (`acNode` layer)
 
-> **Note:** This section describes implementation details specific to the acNode layer. The acGrid layer is not related to the `GridDims` structure described here. 
+> **Note:** This section describes implementation details specific to the acNode layer. The acGrid layer is not related to the `GridDims` structure described here.
 
 `GridDims` contains the dimensions of the the mesh decomposed to multiple devices.
 ```C
diff --git a/include/astaroth.h b/include/astaroth.h
index 09a1405..5eaa734 100644
--- a/include/astaroth.h
+++ b/include/astaroth.h
@@ -290,7 +290,7 @@ AcResult acGridQuit(void);
 AcResult acGridSynchronizeStream(const Stream stream);
 
 /** */
-AcResult acGridLoadMesh(const AcMesh host_mesh, const Stream stream);
+AcResult acGridLoadMesh(const Stream stream, const AcMesh host_mesh);
 
 /** */
 AcResult acGridStoreMesh(const Stream stream, AcMesh* host_mesh);
diff --git a/samples/benchmark/main.cc b/samples/benchmark/main.cc
index 733b064..4b22cc9 100644
--- a/samples/benchmark/main.cc
+++ b/samples/benchmark/main.cc
@@ -126,8 +126,12 @@ main(int argc, char** argv)
 
     // GPU alloc & compute
     acGridInit(info);
+    AcMesh model;
+    acMeshCreate(info, &model);
+    acMeshRandomize(&model);
+    acGridLoadMesh(STREAM_DEFAULT, model);
     /*
-    acGridLoadMesh(model, STREAM_DEFAULT);
+    acGridLoadMesh(STREAM_DEFAULT, model);
 
     acGridIntegrate(STREAM_DEFAULT, FLT_EPSILON);
     acGridPeriodicBoundconds(STREAM_DEFAULT);
@@ -149,7 +153,6 @@ main(int argc, char** argv)
         }
     }*/
 
-
     // Percentiles
     const size_t num_iters      = 1000;
     const double nth_percentile = 0.90;
@@ -189,11 +192,11 @@ main(int argc, char** argv)
         ERRCHK_ALWAYS(fp);
         // Format
         // nprocs, min, 50th perc, 90th perc, max
-        fprintf(fp, "%d, %g, %g, %g, %g\n", nprocs, results[0], results[0.5 * num_iters], results[nth_percentile * num_iters], results[num_iters-1]);
+        fprintf(fp, "%d, %g, %g, %g, %g\n", nprocs, results[0], results[0.5 * num_iters],
+                results[nth_percentile * num_iters], results[num_iters - 1]);
         fclose(fp);
     }
 
-
     acGridQuit();
     MPI_Finalize();
     return EXIT_SUCCESS;
diff --git a/samples/mpitest/main.cc b/samples/mpitest/main.cc
index 83a95ec..492c1df 100644
--- a/samples/mpitest/main.cc
+++ b/samples/mpitest/main.cc
@@ -51,7 +51,7 @@ main(void)
     acGridInit(info);
 
     // INTEGRATION TESTS START ---------------------------------------------------------------------
-    acGridLoadMesh(model, STREAM_DEFAULT);
+    acGridLoadMesh(STREAM_DEFAULT, model);
     acGridIntegrate(STREAM_DEFAULT, FLT_EPSILON);
     acGridPeriodicBoundconds(STREAM_DEFAULT);
     acGridStoreMesh(STREAM_DEFAULT, &candidate);
@@ -64,7 +64,7 @@ main(void)
     // INTEGRATION TESTS END -----------------------------------------------------------------------
 
     // REDUCTION TESTS START -----------------------------------------------------------------------
-    acGridLoadMesh(model, STREAM_DEFAULT);
+    acGridLoadMesh(STREAM_DEFAULT, model);
 
     std::vector<AcScalReductionTestCase> scalarReductionTests{
         acCreateScalReductionTestCase("Scalar MAX", VTXBUF_UUX, RTYPE_MAX),
diff --git a/src/core/device.cc b/src/core/device.cc
index 530d738..ae38bfe 100644
--- a/src/core/device.cc
+++ b/src/core/device.cc
@@ -1309,7 +1309,7 @@ acGridQuit(void)
 }
 
 AcResult
-acGridLoadMesh(const AcMesh host_mesh, const Stream stream)
+acGridLoadMesh(const Stream stream, const AcMesh host_mesh)
 {
     ERRCHK(grid.initialized);
     acGridSynchronizeStream(stream);