[Feature Request] Add GPU optimizations

[acxz]

Feature

It would be great if a GPU accelerated sparse matrix library can be incorporated into GTSAM. These libraries include: AMD ROCM's rocSPARSE & NVIDIA CUDA's cuSPARSE.

hipSPARSE is a wrapper around both rocSPARSE and cuSPARSE that allows either one to be used as a backend depending on if you have ROCm or CUDA installed.

Motivation

Gotta go fast! NVIDIA mentions that its cuSPARSE provides up to 5x speed over CPU alternatives, see: https://developer.nvidia.com/cusparse

Pitch

Having the option to build GTSAM with hipSPARSE allows GPU accelerated spare matrix calculations to occur with both AMD and NVIDIA devices. This can provide a substantial speedup.

Alternatives

N/A

Additional context

It is interesting to note that @dongjing3309's minisam has the option to use CUDA for its sparse linear solver.

[ProfFan]

I took a look at minisam, I think they are using cuSPARSE for the Cholesky decomposition step. I think this would be very easy to add to the current codebase, as the only piece of code that needs to be modified is the Cholesky solver (plus some CMake-fu). It should be interesting to benchmark the speedup with some really large datasets.

[dellaert]

I remember that in fact the CUDA solvers were slower.

On Wed, May 6, 2020 at 19:25 Fan Jiang ***@***.***> wrote: I took a look at minisam, I think they are using cuSPARSE for the Cholesky decomposition step. I think this would be very easy to add to the current codebase, as the only piece of code that needs to be modified is the Cholesky solver (plus some CMake-fu). It should be interesting to benchmark the speedup with some really large datasets. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#291 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACQHGSJXQNFU2GWPUE4JTZLRQHWW7ANCNFSM4M22FUZA> .

-- Best ! Frank Dellaert http://frank.dellaert.com

[varunagrawal]

Perhaps Nvidia have improved their cuSparse library significantly by now.

[acxz]

Came across this: https://developer.nvidia.com/cholmod

Seems like CHOLMOD already has GPU capability built in, we just need a flag to turn it on.

[ProfFan]

Yes that is on the roadmap :) Should be easy once #111 is merged.

[ProfFan]

@acxz I have added some other solvers like CHOLMOD and cuSparse in #111. However their performance is not great compared to GTSAM's MULTIFRONTAL_CHOLESKY. There should be a lot of room for optimization though I don't have bandwidth on this currently.

[ProfFan]

BTW, CHOLMOD's GPU support is somewhat exaggerated: it only supports GPU cholesky with integers, not doubles.

[acxz]

added some other solvers like CHOLMOD and cuSparse

Nice! I am sure the optimizations can come later.

CHOLMOD's GPU support is somewhat exaggerated: it only supports GPU cholesky with integers, not doubles.

rip, makes me sad

[ProfFan]

rip, makes me sad

me bursted into tears after wasting one day building custom SuiteSparse with GPU and figuring out how to make it use the gpu...

[varunagrawal]

I spent some time looking into ROCm this past weekend and it looks to be ideal for tackling this issue in a transparent way.