possible huge performance regression in batched_mul

[bjarthur]

i'm still looking into the cause and quantifying the magnitude of the effect, but i think it is caused by #271. has anyone else observed this?

[DhairyaLGandhi]

Can you provide an MWE?

[bjarthur]

here's an MWE showing a 1000x performance regression for batched_mul when one of the inputs is of type NNlib.BatchedTranspose. i've bisected it to somewhere between [email protected] and [email protected].

julia> VERSION
v"1.5.0"

(test0712) pkg> st
Status `~/projects/darshan/trainBalancedNet-ver2-GPU-forchris/Project.toml`
  [052768ef] CUDA v2.4.1
  [872c559c] NNlib v0.7.12 ⚲

julia> using CUDA, NNlib, BenchmarkTools

julia> x=CUDA.rand(Float32, 64,1,5000);

julia> xT=batched_transpose(x);

julia> y=CUDA.rand(Float32, 64,1,5000);

julia> @benchmark CUDA.@sync batched_mul(xT,y)
BenchmarkTools.Trial: 
  memory estimate:  256 bytes
  allocs estimate:  11
  --------------
  minimum time:     26.261 μs (0.00% GC)
  median time:      49.789 μs (0.00% GC)
  mean time:        50.305 μs (0.00% GC)
  maximum time:     1.174 ms (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

julia> xT2=CUDA.rand(Float32, 1,64,5000);

julia> @benchmark CUDA.@sync batched_mul(xT2,y)
BenchmarkTools.Trial: 
  memory estimate:  176 bytes
  allocs estimate:  7
  --------------
  minimum time:     34.175 μs (0.00% GC)
  median time:      54.682 μs (0.00% GC)
  mean time:        58.268 μs (0.00% GC)
  maximum time:     26.206 ms (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

julia> VERSION
v"1.5.0"

(test0713) pkg> st
Status `~/projects/darshan/trainBalancedNet-ver2-gpu/Project.toml`
  [052768ef] CUDA v2.4.1
  [872c559c] NNlib v0.7.13 ⚲

julia> using CUDA, NNlib, BenchmarkTools

julia> x=CUDA.rand(Float32, 64,1,5000);

julia> xT=batched_transpose(x);

julia> y=CUDA.rand(Float32, 64,1,5000);

julia> @benchmark CUDA.@sync batched_mul(xT,y)
BenchmarkTools.Trial: 
  memory estimate:  2.52 MiB
  allocs estimate:  110008
  --------------
  minimum time:     39.842 ms (0.00% GC)
  median time:      39.991 ms (0.00% GC)
  mean time:        43.148 ms (2.04% GC)
  maximum time:     98.322 ms (17.06% GC)
  --------------
  samples:          116
  evals/sample:     1

julia> xT2=CUDA.rand(Float32, 1,64,5000);

julia> @benchmark CUDA.@sync batched_mul(xT2,y)
BenchmarkTools.Trial: 
  memory estimate:  208 bytes
  allocs estimate:  9
  --------------
  minimum time:     34.570 μs (0.00% GC)
  median time:      54.919 μs (0.00% GC)
  mean time:        57.578 μs (0.00% GC)
  maximum time:     21.103 ms (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

[CarloLucibello]

cc @mcabbott

[mcabbott]

Oh that's not good. I can't try this out now, but the obvious guess is that it's using the generic fallback (which makes slices) rather than a CUBLAS call. If you run ENV["JULIA_DEBUG"] = NNlib before starting this, does it print anything about that? (There are some @debug statements, this may need to be done in a fresh session.) What does NNlib. storage_typejoin(xT,y) return?

Apart from that... does it return a CuArray at all? Does this also happen with say x=CUDA.rand(Float32, 64,3,5000); etc. to avoid simple dot case?

[mcabbott]

More compact demonstration:

using NNlib
ENV["JULIA_DEBUG"] = NNlib

x, y = randn(64,3,5), randn(64,7,5); # fine
xT = batched_transpose(x); xT2 = copy(xT);
batched_mul(xT,y) ≈ batched_mul(xT2,y)

x, y = randn(64,1,5), randn(64,7,5); # uses fallback
xT = batched_transpose(x); xT2 = copy(xT);
batched_mul(xT,y) ≈ batched_mul(xT2,y)  # still correct

Test all cases, from #268: quite a few use fallback.

using NNlib, Test
@testset "awkward strides, $T" for T in [Float64] #, ComplexF64]
    @testset "$tA(rand$((sA...,2))) ⊠ $tB(rand$((sB...,2)))" for tA in [identity, batched_adjoint, batched_transpose], sA in [(1,3), (3,1), (3,3)], tB in [identity, batched_adjoint, batched_transpose], sB in [(1,3), (3,1), (3,3)]
        A = tA(rand(T, sA..., 2))
        B = tB(rand(T, sB..., 2))
        size(A,2) == size(B,1) || continue
        C = cat(A[:,:,1] * B[:,:,1], A[:,:,2] * B[:,:,2]; dims=3)
        @test A ⊠ B ≈ C
    end
end;

[bjarthur]

thanks for looking into this!

would it be worth adding automatic performance regression testing to NNlib's CI like exists for julia base?

[mcabbott]

It might be, although tricky. There are many cases to test, and I'm not sure this would be a huge slowdown on the CPU, where the fast path isn't so different from the fallback.

Now #299 at least tests whether @debug prints anything, in a more exhaustive version of the loop just above. This tests a few hundred different cases; this one (batched_transpose(rand(n,1,b))) was among the edge cases tested for correctness, but there were no explicit tests of which routine is called.