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

Trim GPU restrict sync overhead

Browse Source
This commit is contained in:
CGH0S7
2026-04-12 19:45:34 +08:00
parent ce88c18265
commit d702aa06b9
3 changed files with 178 additions and 158 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
AMSS_NCKU_source/bssn_cuda_ops.cu
View File

@@ -2016,25 +2016,6 @@ int bssn_cuda_prolong3_pack(int wei,
if (!launch_kernel(grid, block, (const void *)prolong3_cell_kernel, args))
return 1;
cudaError_t sync_err = cudaDeviceSynchronize();
if (sync_err != cudaSuccess)
{
std::fprintf(stderr,
"prolong3 debug: symmetry=%d extc=(%d,%d,%d) extf=(%d,%d,%d) "
"imino=%d imaxo=%d jmino=%d jmaxo=%d kmino=%d kmaxo=%d "
"ic_min=%d ic_max=%d jc_min=%d jc_max=%d kc_min=%d kc_max=%d "
"lbc=(%d,%d,%d) lbf=(%d,%d,%d)\n",
symmetry,
extc[0], extc[1], extc[2],
extf[0], extf[1], extf[2],
imino, imaxo, jmino, jmaxo, kmino, kmaxo,
ic_min, ic_max, jc_min, jc_max, kc_min, kc_max,
lbc[0], lbc[1], lbc[2],
lbf[0], lbf[1], lbf[2]);
report_cuda_error("cudaDeviceSynchronize prolong3", sync_err);
return 1;
}
int host_error_flag = 0;
err = cudaMemcpy(&host_error_flag, cache.error_flag.ptr, sizeof(int), cudaMemcpyDeviceToHost);
if (err != cudaSuccess)
@@ -2241,28 +2222,6 @@ int bssn_cuda_restrict3_pack(int wei,
if (!launch_kernel(grid, block, (const void *)restrict3_cell_kernel, args))
return 1;
cudaError_t sync_err = cudaDeviceSynchronize();
if (sync_err != cudaSuccess)
{
std::fprintf(stderr,
"restrict3 debug: symmetry=%d extc=(%d,%d,%d) extf=(%d,%d,%d) "
"imino=%d imaxo=%d jmino=%d jmaxo=%d kmino=%d kmaxo=%d "
"imini=%d imaxi=%d jmini=%d jmaxi=%d kmini=%d kmaxi=%d "
"lbc=(%d,%d,%d) lbf=(%d,%d,%d) "
"fi=[%d,%d] fj=[%d,%d] fk=[%d,%d] window=[%d:%d,%d:%d,%d:%d]\n",
symmetry,
extc[0], extc[1], extc[2],
extf[0], extf[1], extf[2],
imino, imaxo, jmino, jmaxo, kmino, kmaxo,
imini, imaxi, jmini, jmaxi, kmini, kmaxi,
lbc[0], lbc[1], lbc[2],
lbf[0], lbf[1], lbf[2],
fi_min, fi_max, fj_min, fj_max, fk_min, fk_max,
ii_lo, ii_hi, jj_lo, jj_hi, kk_lo, kk_hi);
report_cuda_error("cudaDeviceSynchronize restrict3", sync_err);
return 1;
}
int host_error_flag = 0;
err = cudaMemcpy(&host_error_flag, cache.error_flag.ptr, sizeof(int), cudaMemcpyDeviceToHost);
if (err != cudaSuccess)