5-arg mul gives wrong result

[lkapelevich]

MWE:

using LinearAlgebra, Random

Random.seed!(1)
n = 100
A = rand(n, n)
B = rand(n, n)
C = zeros(n, n)
mul!(C, B, A, -1, 0)
D = -B * A
@show norm(D - C) # 2534.227986442133

This also happens for sparse arrays.

[chriscoey]

@tkf this occurs on Julia 1.4 master, for a variety of matrix types. Any idea why?

[tkf]

Thanks for the heads up. It turned out to be a stupid mistake (alpha and beta swapped). #33218 should fix it.

[lkapelevich]

I only just hit another problem where alpha and beta were both floats and the output was incorrect- makes sense that it's not an issue with integers :-). In case it's useful to test,

using Random, SparseArrays, LinearAlgebra
Random.seed!(1)
n = 100;
A = Symmetric(rand(n, n))
B = sparse(rand(n, n))
C = zeros(n, n)
mul!(C, B, A, -1.0, 0.0)
D = -B * A
@show norm(D - C) # 2543.524362254463

The error is not triggered without the Symmetric or without the sparse.

[tkf]

This also happens for sparse arrays.

I forgot to address this part. @lkapelevich Can you give me an example? Is the snippet you just posted the same as what you mentioned in the OP?

[tkf]

#33214 (comment) dispatches to the same method I fixed in #33218

[lkapelevich]

What I meant in the description was that I was also getting incorrect results when A and B were sparse, e.g.

Random.seed!(1)
n = 100
A = sparse(rand(n, n))
B = sparse(rand(n, n))
C = spzeros(n, n)
mul!(C, B, A, -1, 0)
D = -B * A
@show norm(D - C) # 2534.227986442133

[lkapelevich]

(I didn't notice that alpha and beta need not be integers)

[tkf]

Thanks. Fortunately, the new snippet (#33214 (comment)) is also handled by #33218, too (which is a bit surprising; I guess dispatch plumbing can be improved a lot...).

[fredrikekre]

Why did this hit generic_matmul at all? Seems like a huge performance trap if you just forget to use 1.0 and 0.0 instead of 1 and 0?

julia> @btime mul!($C, $B, $A, -1, 0);
  835.482 μs (69 allocations: 3.56 KiB)

julia> @btime mul!($C, $B, $A, -1.0, 0.0);
  52.149 μs (0 allocations: 0 bytes)