64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Download staged GPU sync regions incrementally

Browse Source
This commit is contained in:
CGH0S7
2026-04-09 18:23:05 +08:00
parent 3b16795e78
commit 5bc67ded06
3 changed files with 147 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
AMSS_NCKU_source/bssn_gpu.cu
View File

@@ -914,6 +914,73 @@ int bssn_gpu_stage_upload_region(const double *host_ptr,
return 0;
}
int bssn_gpu_stage_download_region(double *host_ptr,
const int *full_shape,
const double *full_llb,
const double *full_uub,
const int *region_shape,
const double *region_llb)
{
if (!host_ptr || !full_shape || !full_llb || !full_uub || !region_shape || !region_llb)
return 1;
const double *device_ptr = bssn_gpu_find_device_buffer(host_ptr);
if (!device_ptr)
return 1;
int start[3] = {0, 0, 0};
for (int i = 0; i < 3; ++i)
{
if (full_shape[i] <= 0 || region_shape[i] <= 0)
return 1;
#ifdef Vertex
#ifdef Cell
#error Both Cell and Vertex are defined
#endif
const double dx = (full_uub[i] - full_llb[i]) / static_cast<double>(full_shape[i] - 1);
start[i] = static_cast<int>((region_llb[i] - full_llb[i]) / dx + 0.4);
#else
#ifdef Cell
const double dx = (full_uub[i] - full_llb[i]) / static_cast<double>(full_shape[i]);
start[i] = static_cast<int>((region_llb[i] - full_llb[i]) / dx + 0.4);
#else
#error Not define Vertex nor Cell
#endif
#endif
if (start[i] < 0 || start[i] + region_shape[i] > full_shape[i])
return 1;
}
cudaMemcpy3DParms parms = {};
parms.srcPtr = make_cudaPitchedPtr(const_cast<double *>(device_ptr),
static_cast<size_t>(full_shape[0]) * sizeof(double),
static_cast<size_t>(full_shape[0]),
static_cast<size_t>(full_shape[1]));
parms.dstPtr = make_cudaPitchedPtr(host_ptr,
static_cast<size_t>(full_shape[0]) * sizeof(double),
static_cast<size_t>(full_shape[0]),
static_cast<size_t>(full_shape[1]));
parms.srcPos = make_cudaPos(static_cast<size_t>(start[0]) * sizeof(double),
static_cast<size_t>(start[1]),
static_cast<size_t>(start[2]));
parms.dstPos = parms.srcPos;
parms.extent = make_cudaExtent(static_cast<size_t>(region_shape[0]) * sizeof(double),
static_cast<size_t>(region_shape[1]),
static_cast<size_t>(region_shape[2]));
parms.kind = cudaMemcpyDeviceToHost;
cudaError_t err = cudaMemcpy3D(&parms);
if (err != cudaSuccess)
{
cerr << "cudaMemcpy3D(D2H region) failed: " << cudaGetErrorString(err) << endl;
return 1;
}
return 0;
}
__global__ void test_const_address(double * testd){
int _t = blockIdx.x*blockDim.x+threadIdx.x;
if(_t == 0)