[WIP] Add Julia interface for cuPDLPx #2
Conversation
|
There's a lot of stuff there. You should wrap only the public C API. What functions is the user expected to call? |
| reflection_coefficient::Cdouble | ||
| end | ||
|
|
||
| struct pdhg_solver_state_t |
There was a problem hiding this comment.
For example, I assume the user shouldn't need access to this struct
| f === :csc && return Ptr{var"##Ctag#234"}(x + 0) | ||
| f === :coo && return Ptr{var"##Ctag#235"}(x + 0) | ||
| return getfield(x, f) | ||
| end |
|
|
||
| function safe_malloc(size) | ||
| ccall((:safe_malloc, libcupdlpx), Ptr{Cvoid}, (Csize_t,), size) | ||
| end |
There was a problem hiding this comment.
And this should almost certainly not be exposed
| end | ||
|
|
||
| function read_mps_file(filename) | ||
| ccall((:read_mps_file, libcupdlpx), Ptr{lp_problem_t}, (Ptr{Cchar},), filename) |
There was a problem hiding this comment.
I’m a bit confused here. Since the goal of this API is to interact with JuMP, do we still need this functionality, given that JuMP already provides the read_from_file function?
There was a problem hiding this comment.
This falls into the question of "what is your public API?"
People might want to interact with cuPDLPx directly, from Julia, and without JuMP.
Something like:
using cuPDLPx
model = cuPDLPx.read_mps_file(filename)
cuPDLPx.optimize(model)
x = cuPDLPx.get_solution(model)(Or whatever, I don't know the C API functions)
There was a problem hiding this comment.
Thanks for the input, @odow. That’s a good point — we may need to reorganize the C API functions in cuPDLPx.
|
Thank you so much for the suggestion, @odow. I wasn’t very familiar with Clang.jl before — I’ll make sure to wrap only the public C API.
|
Desciption
This PR adds the Julia interface for cuPDLPx, a GPU-accelerated first-order LP solver.
A JLL package has already been created and merged into Yggdrasil.
The next step is to build the C API wrapper using
Clang.jl, which is still a work in progress due to several CUDA header dependencies.Below is an example output from the current code generation workflow.
Any feedback or suggestions from the Julia or JuMP community would be greatly appreciated. @odow @blegat