Implement and test a CUDA version of the module map #154
Conversation
|
This PR also depends on acts-project/vecmem#172. |
stephenswat
force-pushed
the
feat/cuda_module_map
branch
from
0ad5be2 to
5af1ef0
Compare
|
This is now ready to be merged. |
The module map was originally designed to be GPU friendly, and we now reap the fruits of that decision. This implementation of a CUDA version of that code is quite short and simple to understand. This commit adds an owning CUDA version of the module map, as well as a light-weight non-owning view object which is designed to be passed to kernels.
stephenswat
force-pushed
the
feat/cuda_module_map
branch
from
5af1ef0 to
9d35141
Compare
This test is rather simple, but effective. It loads a CPU module map from an input file, constructs a CUDA module map and then - for each module - launches a kernel which retrieves the relevant transformation matrix and sends it back to the CPU. The CPU will then check whether these transformation matrices retrieved from the GPU are correct.
stephenswat
force-pushed
the
feat/cuda_module_map
branch
from
9d35141 to
1cbed80
Compare
|
Thanks! You have a keen eye for this! 😄 |
|
Closing this as there seems to be little interest in it. |
|
Actually I need it for GPU. |
|
Mind you this is a map from module identifiers to transformation matrices, although we can easily repurpose this for spacepoint identifier to module identifier mappings, which I guess is what you need for the CKF. |
Exactly. I hope we can generalize it for multiple purposes. |
|
This should no longer be necessary. |
|
Right now you could still argue that we don't have the correct thing in place yet. But as I wrote in #592 (comment), I'll have some EDM updates in the foreseeable future, which will absolutely replace this PR. So yeah, I agree with closing it. Even if I was only going to suggest it a little later, once I pushed my EDM changes through. 😉 |
The module map was originally designed to be GPU friendly, and we now reap the fruits of that decision. This implementation of a CUDA version of that code is quite short and simple to understand. This pull request adds an owning CUDA version of the module map, as well as a light-weight non-owning view object which is designed to be passed to kernels. It also tests the class.
This pull request depends on #153.