Add 32-bit in ci #57

blegat · 2023-06-27T08:07:53Z

I don't think there is added value in testing on different OS but it's failing on 32-bit and it's making SumOfSquares fail its 32-bit build

kalmarek · 2023-06-27T08:53:55Z

Why do you care about 32-bit?
I fixed one problematic line in src/Characters/GF.jl (I'm fine with that), ~~but then sparse qr fails due to some missing CHOLMOD method with UInt32. Are you willing to dive into those?~~
But fixing subsequent errors requires changing UInt32 to Int32 for Ti with sparse matrices. What's the merit of using signed integers as indices in SparseMatrixCSC except using possible range?

With this change our projections will not extend beyond size ~(46340, 46340) and that's already on the edge of what I'm sometimes doing. I'm not in favour of this just for the green CI badge on 32-bit...

blegat · 2023-06-27T09:11:50Z

No need, as long as SW compiles then it will be enough for SumOfSquares but I guess it's an issue if the tests of SW fails due to that.
Is it possible to use methods for Int32 ?

kalmarek · 2023-06-28T07:59:20Z

so there's a fix, but I'm still not really in favour; I will check how large are the problems we're dealing and if Int32 works for them (the size of PSD constraint is then bounded from 65535 to 46340 or so)

kalmarek · 2023-06-28T08:01:46Z

@piotrmizerka what are the larges psd that we solved?

codecov · 2023-06-28T08:05:22Z

Codecov Report

Patch coverage: 100.00% and no project coverage change.

Comparison is base (434ca6e) 87.45% compared to head (8436d76) 87.45%.

❗ Current head 8436d76 differs from pull request most recent head ce15824. Consider uploading reports for the commit ce15824 to get more accurate results

Additional details and impacted files

@@           Coverage Diff           @@
##           master      #57   +/-   ##
=======================================
  Coverage   87.45%   87.45%           
=======================================
  Files          20       20           
  Lines        1299     1299           
=======================================
  Hits         1136     1136           
  Misses        163      163

Flag	Coverage Δ
unittests	`87.45% <100.00%> (ø)`

Flags with carried forward coverage won't be shown. Click here to find out more.

Files Changed	Coverage Δ
src/Characters/gf.jl	`95.94% <100.00%> (ø)`

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

kalmarek · 2023-06-29T07:08:05Z

@blegat the current errors I'm unwilling to fix. Cyclotomics were designed to be lazy and delay normalization as long as possible. I'd rather keep it like this. This runs into troubles very quickly with decreased range of Rational{Int32}.

blegat · 2023-06-29T08:41:40Z

I understand, I'm good with just fixing this:

ERROR: LoadError: TypeError: in typeassert, expected UInt32, got a value of type UInt64
Stacktrace:
  [1] hashindex(key::SymbolicWedderburn.Characters.FiniteFields.GF{7}, sz::Int32)
    @ Base ./dict.jl:157
  [2] ht_keyindex2_shorthash!(h::Dict{SymbolicWedderburn.Characters.FiniteFields.GF{7}, Matrix{SymbolicWedderburn.Characters.FiniteFields.GF{7}}}, key::SymbolicWedderburn.Characters.FiniteFields.GF{7})
    @ Base ./dict.jl:293
  [3] setindex!(h::Dict{SymbolicWedderburn.Characters.FiniteFields.GF{7}, Matrix{SymbolicWedderburn.Characters.FiniteFields.GF{7}}}, v0::LinearAlgebra.Transpose{SymbolicWedderburn.Characters.FiniteFields.GF{7}, Matrix{SymbolicWedderburn.Characters.FiniteFields.GF{7}}}, key::SymbolicWedderburn.Characters.FiniteFields.GF{7})
    @ Base ./dict.jl:370
  [4] left_eigen(M::Matrix{SymbolicWedderburn.Characters.FiniteFields.GF{7}})
    @ SymbolicWedderburn.Characters ~/.julia/packages/SymbolicWedderburn/inKGG/src/Characters/eigenspacedecomposition.jl:31
  [5] eigen_decomposition!(M::Matrix{SymbolicWedderburn.Characters.FiniteFields.GF{7}})
    @ SymbolicWedderburn.Characters ~/.julia/packages/SymbolicWedderburn/inKGG/src/Characters/eigenspacedecomposition.jl:53
  [6] SymbolicWedderburn.Characters.EigenSpaceDecomposition(M::Matrix{SymbolicWedderburn.Characters.FiniteFields.GF{7}})
    @ SymbolicWedderburn.Characters ~/.julia/packages/SymbolicWedderburn/inKGG/src/Characters/eigenspacedecomposition.jl:83
  [7] common_esd(Ns::Vector{SymbolicWedderburn.Characters.CMMatrix{Int32, PermutationGroups.Orbit1{PermutationGroups.Permutation{Int32, …}, Nothing}}}, F::Type{SymbolicWedderburn.Characters.FiniteFields.GF{7}})
    @ SymbolicWedderburn.Characters ~/.julia/packages/SymbolicWedderburn/inKGG/src/Characters/dixon.jl:27
  [8] SymbolicWedderburn.Characters.CharacterTable(Fp::Type{SymbolicWedderburn.Characters.FiniteFields.GF{7}}, G::PermutationGroups.PermGroup{Int32, …}, cclasses::Vector{PermutationGroups.Orbit1{PermutationGroups.Permutation{Int32, …}, Nothing}})
    @ SymbolicWedderburn.Characters ~/.julia/packages/SymbolicWedderburn/inKGG/src/Characters/character_tables.jl:62

and not modify ci

piotrmizerka · 2023-06-29T23:33:43Z

The problems which I considered and were successfully solved - that is a certified spectral gap was obtained had about 200-400k constraints in total (that was without symmetrization using Wedderburn decomposition - with the symmetrization the numbers may be different - I don't remember now the size of psd variable matrices but they were at most several hundreds - rather closer to 100x100 than 1000x1000). There were also several problems I tried to solve but failed (in the sense that no positive gap was obtained) - they were similar in size to that problem (this is after symmetrization - I don't remember the sizes of blocks now):

problem: variables n: 448532, constraints m: 1197455
cones: z: primal zero / dual free vars: 748924
s: psd vars: 448531, ssize: 5
settings: eps_abs: 1.0e-07, eps_rel: 1.0e-07, eps_infeas: 1.0e-07
alpha: 1.50, scale: 1.00e-01, adaptive_scale: 1
max_iters: 10000, normalize: 1, rho_x: 1.00e-06
lin-sys: sparse-direct-amd-qdldl
nnz(A): 19168889, nnz(P): 0

I tried also to solve problems with 50-100 million constraints (no symmetrization here) but my impression is that the convergence seems too poor for them (I worked mostly with the SCS solver, I tried COSMO as well but I was even unable to see any iteration completed). Sample parameters here:

problem: variables n: 84766711, constraints m: 99781770
cones: z: primal zero / dual free vars: 15015060
s: psd vars: 84766710, ssize: 1
settings: eps_abs: 1.0e-06, eps_rel: 1.0e-06, eps_infeas: 1.0e-07
alpha: 1.50, scale: 1.00e-01, adaptive_scale: 1
max_iters: 1000000, normalize: 1, rho_x: 1.00e-06
acceleration_lookback: 10, acceleration_interval: 10
lin-sys: sparse-direct-amd-qdldl
nnz(A): 253733127, nnz(P): 0

kalmarek · 2023-06-30T05:36:41Z

@piotrmizerka yes, we're interested here in the size of the psd constraint before symmetrization; was it larger than 20k?

piotrmizerka · 2023-07-03T12:29:27Z

@piotrmizerka yes, we're interested here in the size of the psd constraint before symmetrization; was it larger than 20k?

If we would like to solve the problem for Delta_1 for SAutF_4 on half_radius 2, then we get 43k x 43k (S_n action) or 85k x 85k (wreath action) which clearly exceeds 20k. This I have never manage to launch.

I also tried to solve (but experiencing poor convergence so far) problems with PSD constraints forming matrices of size 14_000 x 14_000 and that was roughly 14_000^2/2 approx. 100_000_000 variables.

Another problem which was solved with prec. approx. 1e-06 (no positive spectral gap was obtained) had PSD constrain matrix of size 10_392 x 10_392.

The problems which were solved and gave positive spectral gap had PSD constraint matrices of sizes even less than 1_000 x 1000.

kalmarek · 2023-07-03T13:32:26Z

@piotrmizerka did we manage to do wedderburn on the 43k×43k problem?

If not, can you try doing this again (there were multiple improvements on the "large" side of things here) and paste results here ?

piotrmizerka · 2023-07-04T21:45:16Z

@piotrmizerka did we manage to do wedderburn on the 43k×43k problem?

If not, can you try doing this again (there were multiple improvements on the "large" side of things here) and paste results here ?

Yes, we managed to do Wedderburn for the 43k x 43k problem. It resulted in the blocks of the following sizes: 3554, 5396, 1842, 1712, 5266. What I haven't done is running the actual solver for it - I was a bit sceptical to do that because the Wedderburn itself took about 2 weeks if I remember correctly.

kalmarek · 2023-07-06T15:21:10Z

@piotrmizerka I believe that now it should go much quicker; @piotrmizerka Could you open a new issue here where you list all of those problems with their numbers & times?

Indeed, it seems that we are actually hitting the typemax(Int32) boundary in some cases and UInt32 is necessary here This reverts commit ba9af48.

This reverts commit 6dfc1b0.

kalmarek · 2023-07-25T15:25:31Z

@blegat I reverted the 32bit CI; I don't intend to test the package there;
If SymbolicWedderburn doesn't load on x86 please let me know.

blegat · 2023-12-05T14:23:15Z

The compiler crashes when trying to compile SW on 32-bits but it's quite unclear why:
https://github.com/jump-dev/SumOfSquares.jl/actions/runs/7101937796/job/19331181743#step:6:2570
Maybe we can wait for Julia v1.10 to see if it's fixed

Add 32-bit in ci

6dfc1b0

kalmarek added 2 commits June 27, 2023 10:57

fix hash(::GF, ...) on 32-bit

a8d0717

use Int32 instead UInt32 as index type in (sparse) projections

ba9af48

test for Characters: interpolate Int in sprint

2acb405

kalmarek added 2 commits July 20, 2023 17:52

Revert "use Int32 instead UInt32 as index type in (sparse) projections"

8436d76

Indeed, it seems that we are actually hitting the typemax(Int32) boundary in some cases and UInt32 is necessary here This reverts commit ba9af48.

Revert "Add 32-bit in ci"

ce15824

This reverts commit 6dfc1b0.

kalmarek merged commit be9c657 into master Jul 25, 2023

kalmarek deleted the bl/32 branch July 25, 2023 15:53

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add 32-bit in ci #57

Add 32-bit in ci #57

blegat commented Jun 27, 2023

kalmarek commented Jun 27, 2023 •

edited

Loading

blegat commented Jun 27, 2023

kalmarek commented Jun 28, 2023

kalmarek commented Jun 28, 2023

codecov bot commented Jun 28, 2023 •

edited

Loading

kalmarek commented Jun 29, 2023

blegat commented Jun 29, 2023

piotrmizerka commented Jun 29, 2023 •

edited

Loading

kalmarek commented Jun 30, 2023

piotrmizerka commented Jul 3, 2023

kalmarek commented Jul 3, 2023

piotrmizerka commented Jul 4, 2023

kalmarek commented Jul 6, 2023

kalmarek commented Jul 25, 2023

blegat commented Dec 5, 2023

Add 32-bit in ci #57

Add 32-bit in ci #57

Conversation

blegat commented Jun 27, 2023

kalmarek commented Jun 27, 2023 • edited Loading

blegat commented Jun 27, 2023

kalmarek commented Jun 28, 2023

kalmarek commented Jun 28, 2023

codecov bot commented Jun 28, 2023 • edited Loading

Codecov Report

kalmarek commented Jun 29, 2023

blegat commented Jun 29, 2023

piotrmizerka commented Jun 29, 2023 • edited Loading

kalmarek commented Jun 30, 2023

piotrmizerka commented Jul 3, 2023

kalmarek commented Jul 3, 2023

piotrmizerka commented Jul 4, 2023

kalmarek commented Jul 6, 2023

kalmarek commented Jul 25, 2023

blegat commented Dec 5, 2023

kalmarek commented Jun 27, 2023 •

edited

Loading

codecov bot commented Jun 28, 2023 •

edited

Loading

piotrmizerka commented Jun 29, 2023 •

edited

Loading