Support for `SparseMatrixCSC{Float64, Int32}`

[RoyiAvital]

See the code below:

mT = SparseMatrixCSC{Float64, int32}(sprand(1_000, 1_000, 0.01));
GBMatrix(mT); #<! Error!

The last line will just error.
It would be great to have support for SparseMatrixCSC{T, N} where T <: Union{Float16, Float32, Float64, Float128}, N <: Union{Int16, Int32, Int64}.

[rayegun]

Unfortunately that’s not possible. The underlying library only supports 64 bit indices.

…

On Sat, Sep 21, 2024 at 10:36 AM Royi ***@***.***> wrote: See the code below: mT = SparseMatrixCSC{Float64, int32}(sprand(1_000, 1_000, 0.01));GBMatrix(mT); #<! Error! The last line will just error. It would be great to have support for SparseMatrixCSC{T, N} where T <: Union{Float16, Float32, Float64, Float128}, N <: Union{Int16, Int32, Int64}. — Reply to this email directly, view it on GitHub <#144>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACG6QV5JJQIF5CO3AGJRULLZXV77TAVCNFSM6AAAAABOTUUN6SVHI2DSMVQWIX3LMV43ASLTON2WKOZSGU2DAMZSGU3DCMY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[RoyiAvital]

Support does not mean change the underlying.
Just support converting it without creating an error.

Even something like that:

GBMatrix(SparseMatrixCSC{elttype(mT), Int64}(mT));

Will do.

[rayegun]

Ah okay sure I can do that.

…

On Sat, Sep 21, 2024 at 1:44 PM Royi ***@***.***> wrote: Support does not mean change the underlying. Just support converting it without creating an error. Even something like that: GBMatrix(SparseMatrixCSC{elttype(mT), Int64}(mT)); Will do. — Reply to this email directly, view it on GitHub <#144 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACG6QVYV3VGPHZ2HUJKKZQLZXWV7LAVCNFSM6AAAAABOTUUN6SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNRVGI3DKOJQGU> . You are receiving this because you commented.Message ID: ***@***.***>

[RoyiAvital]

Is there a real reason not to support Int32 in the c code?
Does it mean a lot of work?

Just wonder if there any point in asking for it.

[rayegun]

I’ve asked and talked with Tim extensively about it. It’s maybe more possible now than it used to be, but it would be a very significant change. Fixing 32 bit indexing in SPQR alone was a ~5k line diff. This would be closer to ~40k. And there’s no way to do it in a standards compliant way. We haven’t added 32 bit indices to the standard, although I could push for it. It ends up being much more valuable for pattern matrices or Boolean than for Float64. So I can see the value add.

…

On Sat, Sep 21, 2024 at 1:51 PM Royi ***@***.***> wrote: Is there a real reason not to support Int32 in the c code? Does it mean a lot of work? Just wonder if there any point in asking for it. — Reply to this email directly, view it on GitHub <#144 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACG6QV6YQ3GH2DCOEW5STCDZXWW3HAVCNFSM6AAAAABOTUUN6SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNRVGI3DOOBVGI> . You are receiving this because you commented.Message ID: ***@***.***>

[RoyiAvital]

The added value is speed.
The throughput with 32 Bit is twice.

I don't know about GraphBLAS, but for certain in many sparse related algorithms it makes a difference since most of the data to process are indices. So the cache is twice as big effectively, the bandwidth, etc...

[rayegun]

As I mentioned it's only valuable when the data type of your values is small, and it's not twice always twice as fast. It depends on the internal representation.

I agree with you, it's very valuable especially as people often deal with smaller types these days. I'll bring it up in committee meetings but I wouldn't expect progress on that front until next year at the earliest.