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

Checkpoint Z4C CUDA resident sync progress

Browse Source
This commit is contained in:
CGH0S7
2026-05-02 10:53:52 +08:00
parent ba61702fc0
commit 52beb4d153
4 changed files with 1075 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
AMSS_NCKU_source/Z4c_class.C
View File

@@ -388,41 +388,57 @@ bool z4c_cuda_interp_bh_point_resident(MyList<Patch> *PatL,
if (z4c_cuda_has_resident_state(block) &&
block->shape[0] >= ordn && block->shape[1] >= ordn && block->shape[2] >= ordn)
{
const int sx = ordn;
const int sy = ordn;
const int sz = ordn;
const int region_all = sx * sy * sz;
const int i0 = z4c_cuda_interp_tile_start(block->X[0], block->shape[0], x, DH[0], ordn);
const int j0 = z4c_cuda_interp_tile_start(block->X[1], block->shape[1], y, DH[1], ordn);
const int k0 = z4c_cuda_interp_tile_start(block->X[2], block->shape[2], z, DH[2], ordn);
double *packed_fields = new double[3 * region_all];
var *vars[3] = {forx, fory, forz};
for (int f = 0; f < 3; f++)
static int use_device_bh_interp = -1;
if (use_device_bh_interp < 0)
{
if (z4c_cuda_pack_state_region_to_host_buffer(block,
k_z4c_cuda_bh_state_indices[f],
packed_fields + f * region_all,
block->shape,
i0, j0, k0,
sx, sy, sz) != 0)
const char *env = getenv("AMSS_CUDA_Z4C_BH_INTERP_DEVICE");
use_device_bh_interp = (env && atoi(env) != 0) ? 1 : 0;
}
bool used_device_interp = false;
if (use_device_bh_interp)
{
double soa3[9];
for (int f = 0; f < 3; f++)
{
delete[] packed_fields;
cout << "CUDA Z4C BH tile download failed" << endl;
soa3[3 * f + 0] = vars[f]->SoA[0];
soa3[3 * f + 1] = vars[f]->SoA[1];
soa3[3 * f + 2] = vars[f]->SoA[2];
}
used_device_interp =
(z4c_cuda_interp_state_point3(block, block->shape,
k_z4c_cuda_bh_state_indices[0],
k_z4c_cuda_bh_state_indices[1],
k_z4c_cuda_bh_state_indices[2],
block->X[0][0], block->X[1][0], block->X[2][0],
DH[0], DH[1], DH[2],
x, y, z,
interp_ordn, interp_sym,
soa3, shellf) == 0);
}
if (!used_device_interp)
{
double *shift_views[3] = {
block->fgfs[forx->sgfn],
block->fgfs[fory->sgfn],
block->fgfs[forz->sgfn]};
if (z4c_cuda_download_state_subset(block, block->shape, 3,
k_z4c_cuda_bh_state_indices,
shift_views) != 0)
{
cout << "CUDA Z4C BH shift download failed" << endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
int tile_shape[3] = {sx, sy, sz};
f_global_interp(tile_shape,
block->X[0] + i0,
block->X[1] + j0,
block->X[2] + k0,
packed_fields + f * region_all,
shellf[f],
x, y, z,
interp_ordn,
vars[f]->SoA,
interp_sym);
f_global_interp(block->shape, block->X[0], block->X[1], block->X[2],
block->fgfs[forx->sgfn], shellf[0],
x, y, z, interp_ordn, forx->SoA, interp_sym);
f_global_interp(block->shape, block->X[0], block->X[1], block->X[2],
block->fgfs[fory->sgfn], shellf[1],
x, y, z, interp_ordn, fory->SoA, interp_sym);
f_global_interp(block->shape, block->X[0], block->X[1], block->X[2],
block->fgfs[forz->sgfn], shellf[2],
x, y, z, interp_ordn, forz->SoA, interp_sym);
}
delete[] packed_fields;
}
else
{