Update kref! for GPU support

[amontoison]

close #215
Related PRs : JuliaGPU/CUDA.jl#419, JuliaGPU/CUDA.jl#427 and JuliaLang/julia#35124

[coveralls]

Coverage increased (+0.002%) to 97.426% when pulling aa27b87 on amontoison:gpu_kref! into 2e3a8ed on JuliaSmoothOptimizers:master.

[codecov]

Codecov Report

Merging #217 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #217   +/-   ##
=======================================
  Coverage   97.42%   97.42%           
=======================================
  Files          30       30           
  Lines        3190     3190           
=======================================
  Hits         3108     3108           
  Misses         82       82           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2e3a8ed...aa27b87. Read the comment docs.

[amontoison]

@mtanneau can you check that this PR solves the problem with MINRES-QLP? You need the gpu_kref! branch of amontoison/krylov.jl and the master branch of CUDA.jl. I interfaced some functions associated to CUBLAS and Tim Besard merged the PR this morning.

[mtanneau]

@mtanneau can you check that this PR solves the problem with MINRES-QLP? You need the gpu_kref! branch of amontoison/krylov.jl and the master branch of CUDA.jl. I interfaced some functions associated to CUBLAS and Tim Besard merged the PR this morning.

It works on my machine, yes.

using Random, LinearAlgebra, CUDA, Krylov

Random.seed!(0)
n = 64
A = Matrix(Symmetric(rand(n, n)))
cA = CuArray(A)

b = ones(n)
cb = CuArray(b)

x, stats = minres_qlp(A, b);

CUDA.allowscalar(false)
cx, cstats = minres_qlp(cA, cb);  # no error

Status `~/Git/Tulip.jl/Project.toml`
  [052768ef] CUDA v1.3.0 `https://github.com/JuliaGPU/CUDA.jl.git#master`
  [ba0b0d4f] Krylov v0.5.2 `https://github.com/amontoison/Krylov.jl.git#gpu_kref!`
  [40e66cde] LDLFactorizations v0.5.0
  [b8f27783] MathOptInterface v0.9.15
  [10f199a5] QPSReader v0.2.0
  [37e2e46d] LinearAlgebra
  [56ddb016] Logging
  [de0858da] Printf
  [2f01184e] SparseArrays
  [4607b0f0] SuiteSparse
  [8dfed614] Test

I have an Nvidia GTX1080 GPU, you may want to check on another machine just to be sure.

[dpo]

@amontoison give me the green light when this is ready to go.

[amontoison]

@mtanneau I checked with a Tesla P100 available at the GERAD and it works well too.
@dpo you can merge this PR.
Do you want to merge another PR like #206 and / or #128 before I create the release 0.5.3 ?

[amontoison]

using Random, LinearAlgebra, CUDA, Krylov

Random.seed!(0)
n = 2^12
A = Matrix(Symmetric(rand(n, n)))
cA = CuArray(A)

b = ones(n)
cb = CuArray(b)

minres_qlp(A, b)  # warm up
CUDA.@time minres_qlp(A, b);

-> 104.331374 seconds

CUDA.allowscalar(false)
minres_qlp(cA, cb) #warm up
CUDA.@time  minres_qlp(cA, cb);

-> 3.651324 seconds 💪

[mtanneau]

As a side remark, the CUDA version seems to be allocated some memory on the CPU.

julia> CUDA.@time minres_qlp(A, b);
 68.133280 seconds (56 CPU allocations: 578.141 KiB)

julia> CUDA.@time  minres_qlp(cA, cb);
  5.153048 seconds (2.99 M CPU allocations: 117.499 MiB, 0.21% gc time) (6 GPU allocations: 192.000 KiB, 0.00% gc time)

(these are all post warm-up runs)

No idea if it's expected nor where it comes from.

[amontoison]

Without the modifications of this PR:

CUDA.@time minres_qlp(cA, cb);

-> 860.511457 seconds 🐌

For the allocations on CPU, I don't know where the problem comes from. It's difficult to profile it by ssh.

[dpo]

Thanks. Let's release.