MCPDLP #231
MCPDLP #231
Conversation
…r of primal step size and dual step size, update the kernels to launch multiple threads and support a very wide batch size accordingly
… if batch is called with trust region restart
…less movement member
…d dual and just make them wider
Kh4ster
added
feature request
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
| namespace cuopt::linear_programming::detail { | ||
|
|
||
| // This class is used to start a batched dot product | ||
| // With large problem size (>10K) and small batch size (<100), this is faster than using Segmented Reduce |
There was a problem hiding this comment.
Come to think of it I'm not surprised, iirc SegmentedReduce does a 1 block:1 segment mapping and in your case that's pretty terrible, I'm not surprised parallel device-wide blasdot calls beats it
There was a problem hiding this comment.
Although I just realized they added a new overload optimized for fixed sizes, I wasn't aware of it, maybe this performs better?
NVIDIA/cccl#3969
There was a problem hiding this comment.
Very good catch!! I will test that right away. It might make my life way simpler
There was a problem hiding this comment.
This is still slower than using multiple dot products :(
There was a problem hiding this comment.
Dang .-.
Looking at their benchmarks they only test for segment sizes up to 1024 so I guess they don't optimize whatsoever for few-segments scenarios. Would be nice if they said so in their docs!
|
This is possibly a candidate for 25.10 but may still make 25.08 |
|
@Kh4ster Shall we move this to 25.10 |
I'm going to move this to 25.10, we can move it back if it gets finished |
|
De-prioritized for 25.10 and slating for 25.12 release |
This is still under a lot of work.
This is to allow preliminary reviews.