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

Trim GPU main-path transfer overhead

Browse Source
This commit is contained in:
CGH0S7
2026-04-08 20:16:25 +08:00
parent 01ac1f9250
commit a0af9b8804
2 changed files with 74 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
AMSS_NCKU_source/bssn_cuda_ops.cu
View File

@@ -493,9 +493,16 @@ int bssn_cuda_rk4_boundary_var(int *ex, double dT,
dim3 block(256);
dim3 grid(div_up(n, static_cast<int>(block.x)));
const bool need_bam_boundary = (lev == 0);
const bool need_coord_copy = need_bam_boundary;
const bool need_boundary_input = need_bam_boundary && (rk_stage != 0);
const bool need_stage_input = (rk_stage != 0);
const bool need_rhs_output = (rk_stage != 3);
bool ok = true;
if (cache.host_X != X || cache.host_Y != Y || cache.host_Z != Z ||
cache.nx != nx || cache.ny != ny || cache.nz != nz)
if (need_coord_copy &&
(cache.host_X != X || cache.host_Y != Y || cache.host_Z != Z ||
cache.nx != nx || cache.ny != ny || cache.nz != nz))
{
ok = copy_to_device(cache.X, X, bytes_x) &&
copy_to_device(cache.Y, Y, bytes_y) &&
@@ -512,10 +519,13 @@ int bssn_cuda_rk4_boundary_var(int *ex, double dT,
}
ok = ok &&
copy_to_device(cache.state0, state0, bytes) &&
copy_to_device(cache.boundary, boundary_src, bytes) &&
copy_to_device(cache.stage, stage_data, bytes) &&
copy_to_device(cache.rhs, rhs_accum, bytes);
copy_to_device(cache.state0, state0, bytes) &&
(!need_boundary_input || copy_to_device(cache.boundary, boundary_src, bytes)) &&
(!need_stage_input || copy_to_device(cache.stage, stage_data, bytes)) &&
copy_to_device(cache.rhs, rhs_accum, bytes);
if (ok && !need_stage_input)
ok = ensure_capacity(cache.stage, bytes);
if (!ok)
return 1;
@@ -534,7 +544,7 @@ int bssn_cuda_rk4_boundary_var(int *ex, double dT,
double soa1 = SoA[1];
double soa2 = SoA[2];
if (lev == 0)
if (need_bam_boundary)
{
int imin = 1;
int jmin = 1;
@@ -581,7 +591,7 @@ int bssn_cuda_rk4_boundary_var(int *ex, double dT,
cudaError_t err = cudaMemcpy(stage_data, cache.stage.ptr, bytes, cudaMemcpyDeviceToHost);
if (err != cudaSuccess) report_cuda_error("cudaMemcpy(D2H) stage_data", err);
ok = err == cudaSuccess;
if (ok)
if (ok && need_rhs_output)
{
err = cudaMemcpy(rhs_accum, cache.rhs.ptr, bytes, cudaMemcpyDeviceToHost);
if (err != cudaSuccess) report_cuda_error("cudaMemcpy(D2H) rhs_accum", err);