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

Add keyed BSSN CUDA resident banks

Browse Source
This commit is contained in:
CGH0S7
2026-04-29 19:44:19 +08:00
parent 68eab03bac
commit 1ee229a91f
4 changed files with 910 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
AMSS_NCKU_source/bssn_rhs_cuda.h
View File

@@ -97,6 +97,14 @@ int bssn_cuda_pack_state_batch_to_host_buffer(void *block_tag,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_pack_state_batch_to_host_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *host_buffer,
int *ex,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
int state_count,
double *host_buffer,
@@ -104,6 +112,14 @@ int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_unpack_state_batch_from_host_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *host_buffer,
int *ex,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_pack_state_batch_to_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
@@ -111,6 +127,14 @@ int bssn_cuda_pack_state_batch_to_device_buffer(void *block_tag,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_pack_state_batch_to_device_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *device_buffer,
int *ex,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
@@ -118,6 +142,14 @@ int bssn_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_unpack_state_batch_from_device_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *device_buffer,
int *ex,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_pack_state_segments_to_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
@@ -125,6 +157,14 @@ int bssn_cuda_pack_state_segments_to_device_buffer(void *block_tag,
int segment_count,
const int *segment_meta);
int bssn_cuda_pack_state_segments_to_device_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *device_buffer,
int *ex,
int segment_count,
const int *segment_meta);
int bssn_cuda_unpack_state_segments_from_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
@@ -132,6 +172,14 @@ int bssn_cuda_unpack_state_segments_from_device_buffer(void *block_tag,
int segment_count,
const int *segment_meta);
int bssn_cuda_unpack_state_segments_from_device_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *device_buffer,
int *ex,
int segment_count,
const int *segment_meta);
int bssn_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
@@ -139,6 +187,14 @@ int bssn_cuda_restrict_state_batch_to_device_buffer(void *block_tag,
int sx, int sy, int sz,
int fi0, int fj0, int fk0);
int bssn_cuda_restrict_state_batch_to_device_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *device_buffer,
int *ex,
int sx, int sy, int sz,
int fi0, int fj0, int fk0);
int bssn_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
@@ -147,6 +203,15 @@ int bssn_cuda_prolong_state_batch_to_device_buffer(void *block_tag,
int ii0, int jj0, int kk0,
int lbc_i, int lbc_j, int lbc_k);
int bssn_cuda_prolong_state_batch_to_device_buffer_for_host_views(void *block_tag,
double **state_host_key,
int state_count,
double *device_buffer,
int *ex,
int sx, int sy, int sz,
int ii0, int jj0, int kk0,
int lbc_i, int lbc_j, int lbc_k);
int bssn_cuda_download_state_subset(void *block_tag,
int *ex,
int subset_count,
@@ -159,6 +224,15 @@ int bssn_cuda_upload_state_subset(void *block_tag,
const int *state_indices,
double **state_host_in);
int bssn_cuda_prepare_inter_time_level(void *block_tag,
int *ex,
double **src1_host_key,
double **src2_host_key,
double **src3_host_key,
double **dst_host_key,
int source_count,
int tindex);
int bssn_cuda_has_resident_state(void *block_tag);
void bssn_cuda_release_step_ctx(void *block_tag);