Deduplicate calc_chunk_indices_kernel
#1657
Conversation
github-project-automation
bot
moved this to Todo
in Vector Search, ML, & Data Mining Release Board
jinsolp
added
non-breaking
|
@robertmaynard can you review this PR as well? I was under the impression that launching a kernel across TUs would not work in CUDA whole compilation mode but here it seems to be working. Aren't kernels supposed to have their symbols hidden too? |
robertmaynard
left a comment
robertmaynard
left a comment
There was a problem hiding this comment.
This approach is invalid and goes against the guidance in https://developer.nvidia.com/blog/cuda-c-compiler-updates-impacting-elf-visibility-and-linkage/
This currently only work in cuvs as we have failed to remove: https://github.com/rapidsai/cuvs/blob/main/cpp/cmake/modules/ConfigureCUDA.cmake#L38
|
Ohh I see okay 🥲 |
There was a problem hiding this comment.
The kernel needs to be launched in the same TU as which it is defined. We can (but should ideally avoid) pass the pointer around to other TUs but they shouldn't be attempting to launch the kernel.
cjnolet
moved this from Todo
to In Progress
in Vector Search, ML, & Data Mining Release Board
|
Hi @robertmaynard , can you check this PR? I've added the changes to follow the guidelines |
|
/merge |
github-project-automation
bot
moved this from In Progress
to Done
in Vector Search, ML, & Data Mining Release Board
3 participants
Closes #1577
Reduces binary size by deduplicating
calc_chunk_indices_kernel. This PR reduces instantiations from 62 -> 1 for each template (BlockDim=32, 64, ..., 1024)Binary Size Changes
CUDA 12.9: 1096.15MB ->
CUDA 13: 432.98 MB->