Merge branch 'master' into forcing

src/core/astaroth.cu

@@ -77,7 +77,7 @@ acCheckDeviceAvailability(void)
 {
     int device_count; // Separate from num_devices to avoid side effects
     ERRCHK_CUDA_ALWAYS(cudaGetDeviceCount(&device_count));
-    if (device_count > 0) 
+    if (device_count > 0)
         return AC_SUCCESS;
     else
         return AC_FAILURE;
@@ -120,12 +120,14 @@ acInit(const AcMeshInfo& config)
     ERRCHK_ALWAYS(subgrid.n.y >= STENCIL_ORDER);
     ERRCHK_ALWAYS(subgrid.n.z >= STENCIL_ORDER);
 
+#if VERBOSE_PRINTING
     // clang-format off
     printf("Grid m ");   printInt3(grid.m);    printf("\n");
     printf("Grid n ");   printInt3(grid.n);    printf("\n");
     printf("Subrid m "); printInt3(subgrid.m); printf("\n");
     printf("Subrid n "); printInt3(subgrid.n); printf("\n");
     // clang-format on
+#endif
 
     // Initialize the devices
     for (int i = 0; i < num_devices; ++i) {
@@ -271,8 +273,7 @@ AcResult
 acIntegrateStep(const int& isubstep, const AcReal& dt)
 {
     const int3 start = (int3){NGHOST, NGHOST, NGHOST};
-    const int3 end   = (int3){NGHOST + subgrid.n.x, NGHOST + subgrid.n.y,
-                              NGHOST + subgrid.n.z};
+    const int3 end   = (int3){NGHOST + subgrid.n.x, NGHOST + subgrid.n.y, NGHOST + subgrid.n.z};
     for (int i = 0; i < num_devices; ++i) {
         rkStep(devices[i], STREAM_PRIMARY, isubstep, start, end, dt);
     }

src/core/device.cu

@@ -42,6 +42,10 @@ __constant__ Grid globalGrid;
 #define DEVICE_1D_COMPDOMAIN_IDX(i, j, k) ((i) + (j)*DCONST_INT(AC_nx) + (k)*DCONST_INT(AC_nxy))
 #include "kernels/kernels.cuh"
 
+#if PACKED_DATA_TRANSFERS // Defined in device.cuh
+// #include "kernels/pack_unpack.cuh"
+#endif
+
 struct device_s {
     int id;
     AcMeshInfo local_config;
@@ -53,6 +57,11 @@ struct device_s {
     VertexBufferArray vba;
     AcReal* reduce_scratchpad;
     AcReal* reduce_result;
+
+#if PACKED_DATA_TRANSFERS
+// Declare memory for buffers needed for packed data transfers here
+// AcReal* data_packing_buffer;
+#endif
 };
 
 AcResult
@@ -154,6 +163,10 @@ createDevice(const int id, const AcMeshInfo device_config, Device* device_handle
         cudaMalloc(&device->reduce_scratchpad, AC_VTXBUF_COMPDOMAIN_SIZE_BYTES(device_config)));
     ERRCHK_CUDA_ALWAYS(cudaMalloc(&device->reduce_result, sizeof(AcReal)));
 
+#if PACKED_DATA_TRANSFERS
+// Allocate data required for packed transfers here (cudaMalloc)
+#endif
+
     // Device constants
     ERRCHK_CUDA_ALWAYS(cudaMemcpyToSymbol(d_mesh_info, &device_config, sizeof(device_config), 0,
                                           cudaMemcpyHostToDevice));
@@ -184,6 +197,10 @@ destroyDevice(Device device)
     cudaFree(device->reduce_scratchpad);
     cudaFree(device->reduce_result);
 
+#if PACKED_DATA_TRANSFERS
+// Free data required for packed tranfers here (cudaFree)
+#endif
+
     // Concurrency
     for (int i = 0; i < NUM_STREAM_TYPES; ++i)
         cudaStreamDestroy(device->streams[i]);
@@ -373,3 +390,7 @@ loadGlobalGrid(const Device device, const Grid grid)
         cudaMemcpyToSymbol(globalGrid, &grid, sizeof(grid), 0, cudaMemcpyHostToDevice));
     return AC_SUCCESS;
 }
+
+#if PACKED_DATA_TRANSFERS
+// Functions for calling packed data transfers
+#endif

src/core/device.cuh

@@ -98,3 +98,8 @@ AcResult loadDeviceConstant(const Device device, const AcRealParam param, const
 
 /** */
 AcResult loadGlobalGrid(const Device device, const Grid grid);
+
+// #define PACKED_DATA_TRANSFERS (1) %JP: placeholder for optimized ghost zone packing and transfers
+#if PACKED_DATA_TRANSFERS
+// Declarations used for packed data transfers
+#endif

src/core/kernels/kernels.cuh

@@ -426,7 +426,7 @@ cross(const AcReal3& a, const AcReal3& b)
 }
 
 static __host__ __device__ __forceinline__ bool
-is_valid(const AcReal a)
+is_valid(const AcReal& a)
 {
     return !isnan(a) && !isinf(a);
 }
@@ -546,7 +546,7 @@ normalized(const AcReal3& vec)
 // Sinusoidal forcing
 // https://arxiv.org/pdf/1704.04676.pdf
 // NOTE: This method of forcing is depracated. However, it will remain in here
-// until a corresponding scheme exists in the new code. 
+// until a corresponding scheme exists in the new code.
 __constant__ AcReal3 forcing_vec;
 __constant__ AcReal forcing_phi;
 static __device__ __forceinline__ AcReal3

src/standalone/config_loader.cc

@@ -152,7 +152,7 @@ update_config(AcMeshInfo* config)
                                       config->real_params[AC_G_CONST];
     config->real_params[AC_sq2GM_star] = AcReal(sqrt(AcReal(2) * config->real_params[AC_GM_star]));
 
-#if VERBOSE_PRINTING
+#if VERBOSE_PRINTING // Defined in astaroth.h
     printf("###############################################################\n");
     printf("Config dimensions recalculated:\n");
     print(*config);

src/standalone/model/model_rk3.cc

@@ -745,7 +745,7 @@ solve_alpha_step(const int step_number, const ModelScalar dt, const int i, const
     ModelScalar rate_of_change[NUM_VTXBUF_HANDLES] = {0};
     rate_of_change[VTXBUF_LNRHO]                   = continuity(uu, lnrho);
 
-#if LINDUCTION
+#if LMAGNETIC
     const ModelVectorData aa  = read_data(i, j, k, in.vertex_buffer,
                                          (int3){VTXBUF_AX, VTXBUF_AY, VTXBUF_AZ});
     const ModelVector aa_res  = induction(uu, aa);