Treewalk cuda #2
Conversation
astro-YYH
commented
Oct 15, 2024
astro-YYH
commented
- porting treewalk to cuda without any other optimizations (like query sorting);
- it compiles but not runs at present, because not all of the functions called from within the global function have been ported to cuda (mainly those the tw structure points to) yet (only device functions can be called in a global function);
astro-YYH
force-pushed
the
treewalk_cuda
branch
from
e2e2fd1 to
7ad04f5
Compare
astro-YYH
force-pushed
the
treewalk_cuda
branch
from
4b7429a to
6602817
Compare
| /* Tree freed in PM*/ | ||
| ForceTree Tree = {0}; | ||
| // ForceTree Tree = {0}; | ||
| ForceTree * Tree_ptr; |
There was a problem hiding this comment.
Let's make this stack-allocated again, but pass it by value (and probably const).
| priv.Ti_Current = Ti_Current; | ||
| priv.Accel = AccelStore; | ||
| TreeWalk *tw; | ||
| cudaMallocManaged(&tw, sizeof(TreeWalk)); // Allocate TreeWalk structure with Unified Memory |
There was a problem hiding this comment.
This one also probably stack-allocated, passed by value.
There was a problem hiding this comment.
But let me take care of this after it is merged.
libgadget/gravshort-tree.c
Outdated
|
|
||
| tw->ev_label = "GRAVTREE"; | ||
| cudaMallocManaged(&tw->ev_label, sizeof(char) * strlen("GRAVTREE") + 1); // Allocate memory for ev_label | ||
| strcpy(tw->ev_label, "GRAVTREE"); |
There was a problem hiding this comment.
Oh no! We can't have static strings on CUDA memory. Can we make it a std::string instead?
libgadget/treewalk.h
Outdated
|
|
||
| /* A pointer to the force tree structure to walk.*/ | ||
| const ForceTree * tree; | ||
| ForceTree * tree; |
There was a problem hiding this comment.
Can you check if these need to be non-const?
libgadget/treewalk.h
Outdated
|
|
||
| /* name of the evaluator (used in printing messages) */ | ||
| const char * ev_label; | ||
| char * ev_label; |
There was a problem hiding this comment.
Actually this should be "const char" because we cannot convert string literals to char * in C++. Really it should be a const std::string().
| // ForceTree Tree = {0}; | ||
| ForceTree * Tree_ptr; | ||
| cudaMallocManaged(&Tree_ptr, sizeof(ForceTree)); | ||
| cudaDeviceSynchronize(); |
There was a problem hiding this comment.
Check whether this DeviceSynchronize is needed!
2 participants
