Made globalGridN and d_multigpu_offsets built-in parameters. Note the renaming from globalGrid.n to globalGridN.

This commit is contained in:
jpekkila
2019-08-06 16:39:15 +03:00
parent 812b5e170e
commit 3726847683
5 changed files with 16 additions and 15 deletions

View File

@@ -30,7 +30,7 @@
#define LMAGNETIC (1)
#define LENTROPY (1)
#define LTEMPERATURE (0)
#define LFORCING (0)
#define LFORCING (1)
#define LUPWD (0)
#define AC_THERMAL_CONDUCTIVITY (AcReal(0.001)) // TODO: make an actual config parameter

View File

@@ -248,9 +248,9 @@ helical_forcing(Scalar magnitude, Vector k_force, Vector xx, Vector ff_re, Vecto
Vector
forcing(int3 globalVertexIdx, Scalar dt)
{
Vector a = Scalar(.5) * (Vector){globalGrid.n.x * dsx,
globalGrid.n.y * dsy,
globalGrid.n.z * dsz}; // source (origin)
Vector a = Scalar(.5) * (Vector){globalGridN.x * dsx,
globalGridN.y * dsy,
globalGridN.z * dsz}; // source (origin)
Vector xx = (Vector){(globalVertexIdx.x - nx_min) * dsx,
(globalVertexIdx.y - ny_min) * dsy,
(globalVertexIdx.z - nz_min) * dsz}; // sink (current index)

View File

@@ -86,7 +86,9 @@ typedef struct {
FUNC(AC_nxy),\
FUNC(AC_nxyz),\
#define AC_FOR_BUILTIN_INT3_PARAM_TYPES(FUNC)
#define AC_FOR_BUILTIN_INT3_PARAM_TYPES(FUNC)\
FUNC(AC_global_grid_n),\
FUNC(AC_multigpu_offset),
#define AC_FOR_BUILTIN_REAL_PARAM_TYPES(FUNC)

View File

@@ -40,13 +40,14 @@ typedef struct {
} VertexBufferArray;
__constant__ AcMeshInfo d_mesh_info;
__constant__ int3 d_multigpu_offset;
#define DCONST_INT(X) (d_mesh_info.int_params[X])
#define DCONST_INT3(X) (d_mesh_info.int3_params[X])
#define DCONST_REAL(X) (d_mesh_info.real_params[X])
#define DCONST_REAL3(X) (d_mesh_info.real3_params[X])
#define DEVICE_VTXBUF_IDX(i, j, k) ((i) + (j)*DCONST_INT(AC_mx) + (k)*DCONST_INT(AC_mxy))
#define DEVICE_1D_COMPDOMAIN_IDX(i, j, k) ((i) + (j)*DCONST_INT(AC_nx) + (k)*DCONST_INT(AC_nxy))
#define globalGridN (d_mesh_info.int3_params[AC_global_grid_n])
#define d_multigpu_offset (d_mesh_info.int3_params[AC_multigpu_offset])
#include "kernels/kernels.cuh"
static dim3 rk3_tpb(32, 1, 4);
@@ -122,13 +123,6 @@ acDeviceCreate(const int id, const AcMeshInfo device_config, Device* device_hand
ERRCHK_CUDA_ALWAYS(cudaMemcpyToSymbol(d_mesh_info, &device_config, sizeof(device_config), 0,
cudaMemcpyHostToDevice));
// Multi-GPU offset. This is used to compute globalVertexIdx.
// Might be better to calculate this in astaroth.cu instead of here, s.t.
// everything related to the decomposition is limited to the multi-GPU layer
const int3 multigpu_offset = (int3){0, 0, device->id * device->local_config.int_params[AC_nz]};
ERRCHK_CUDA_ALWAYS(cudaMemcpyToSymbol(d_multigpu_offset, &multigpu_offset,
sizeof(multigpu_offset), 0, cudaMemcpyHostToDevice));
printf("Created device %d (%p)\n", device->id, device);
*device_handle = device;

View File

@@ -152,17 +152,22 @@ acNodeCreate(const int id, const AcMeshInfo node_config, Node* node_handle)
ERRCHK_ALWAYS(node->subgrid.n.y >= STENCIL_ORDER);
ERRCHK_ALWAYS(node->subgrid.n.z >= STENCIL_ORDER);
#if VERBOSE_PRINTING
// clang-format off
#if VERBOSE_PRINTING
printf("Grid m "); printInt3(node->grid.m); printf("\n");
printf("Grid n "); printInt3(node->grid.n); printf("\n");
printf("Subrid m "); printInt3(node->subgrid.m); printf("\n");
printf("Subrid n "); printInt3(node->subgrid.n); printf("\n");
#endif
// clang-format on
#endif
// Initialize the devices
for (int i = 0; i < node->num_devices; ++i) {
const int3 multinode_offset = (int3){0, 0, 0}; // Placeholder
const int3 multigpu_offset = (int3){0, 0, i * node->subgrid.n.z};
subgrid_config.int3_params[AC_global_grid_n] = node->grid.n;
subgrid_config.int3_params[AC_multigpu_offset] = multinode_offset + multigpu_offset;
acDeviceCreate(i, subgrid_config, &node->devices[i]);
acDevicePrintInfo(node->devices[i]);
}