Extensible broadcast fusion

11 May 2018 | Matt Bauman (JuliaHub)

Julia version 0.7 brings with it an exciting new feature: the ability to customize broadcast fusion! This recently-merged change is the culmination of a long iterative design process that involved many members of the community. We have converged on a highly extensible interface that should satisfy many use-cases. In this blog post I'll explain why this is a big deal by reviewing some of the key features and just scratch the surface of all that is possible with this new design. I'm quite certain that our enterprising community will come up with many more clever ways to exploit this new structure in the future.

Broadcasting is a core feature of Julia: it allows you to compactly and efficiently express an elementwise operation over containers and scalars by annotating operators and function calls with a .. In cases where the sizes don't match, broadcasting will virtually extend missing dimensions or "singleton" dimensions (which contain only one value) by repeating them to fill the outer shape. For example, the expression ([1, 2, 3] .+ [10 20 30 40]) ./ 10 combines a 3-element column vector, a 1x4 matrix, and a scalar to compute a 3x4 result. I imagine this as "extruding" the vector across the columns of the one-row matrix and spreading the division by 10 across the entire result:

julia> ([1, 2, 3] .+ [10 20 30 40]) ./ 10
3×4 Array{Float64,2}:
 1.1  2.1  3.1  4.1
 1.2  2.2  3.2  4.2
 1.3  2.3  3.3  4.3

Since version 0.6, Julia has executed this expression by "fusing" the two operations into a single kernel. That is, instead of first constructing an integer matrix resulting from the addition ([11 21 31 41; 12 22 ...]) and then subsequently using a second loop to divide each element by 10, Julia does both the addition and division for each element at the same time, making just one pass through the output array and skipping intermediate array(s) entirely. This fusion optimization happens as a syntax-level transformation so it is guaranteed to occur and easy to reason about. Version 0.7 adds an extensible API on top of this feature that allows arrays to customize precisely how this operates.

Julia now uses a first-class data structure to "lazily" represent a fused broadcast expression before executing it. If you're not a package developer this may not mean all that much to you, but you'll still reap the many rewards. In Base Julia and its standard libraries alone, this means:

julia> using BenchmarkTools, Random
       srand(0)
       A = bitrand(10^6)
       B = bitrand(10^6)
       @benchmark $A .& .!$B
BenchmarkTools.Trial:
  memory estimate:  122.23 KiB
  allocs estimate:  3
  ----------
  minimum time:     7.891 μs (0.00% GC)
  median time:      13.152 μs (0.00% GC)
  mean time:        17.826 μs (9.62% GC)
  maximum time:     591.497 μs (95.73% GC)
  ----------
  samples:          10000
  evals/sample:     1

Compare this to previously on 0.6:

BenchmarkTools.Trial:
  memory estimate:  126.45 KiB
  allocs estimate:  6
  ----------
  minimum time:     3.615 ms (0.00% GC)
  median time:      3.741 ms (0.00% GC)
  mean time:        3.764 ms (0.18% GC)
  maximum time:     7.744 ms (50.18% GC)
  ----------
  samples:          1328
  evals/sample:     1

That's upwards of a 450x performance gain.

julia> using LinearAlgebra
       d = Diagonal(1:3)
3×3 Diagonal{Int64,UnitRange{Int64}}:
 1  ⋅  ⋅
 ⋅  2  ⋅
 ⋅  ⋅  3

julia> d ./ 10
3×3 Diagonal{Float64,Array{Float64,1}}:
 0.1   ⋅    ⋅
  ⋅   0.2   ⋅
  ⋅    ⋅   0.3

julia> t = d .+ LowerTriangular(rand(3,3))
3×3 LowerTriangular{Float64,Array{Float64,2}}:
 1.5446     ⋅         ⋅
 0.529211  2.26095    ⋅
 0.533674  0.373702  3.88408

julia> t .+ 100
3×3 Array{Float64,2}:
 101.327  100.0    100.0
 100.85   102.138  100.0
 100.575  100.129  103.839

Previously, Julia 0.6 would have returned a SparseMatrixCSC for d ./ 10 and an Array for d .+ LowerTriangular(rand(3,3)).

# Previously:
julia> y = rand(1000);

julia> @time y .*= 2;
  0.032775 seconds (17.43 k allocations: 947.671 KiB)

julia> @time y .*= 2;
  0.020790 seconds (4.27 k allocations: 233.457 KiB)

# Now on Julia 0.7:
julia> y = rand(1000);

julia> @time y .*= 2;
  0.060494 seconds (90.41 k allocations: 5.529 MiB)

julia> @time y .*= 2;
  0.000020 seconds (6 allocations: 208 bytes)

I'll now dive deeper into exactly how this new API works.

The representation of a fused broadcast

You can see precisely how a fused broadcast is represented with Meta.@lower, but in simpler terms the expression ([1, 2, 3] .+ [10 20 30 40]) ./ 10 is effectively a syntax transformation for:

julia> using .Broadcast: materialize, broadcasted
       bc = broadcasted(/, broadcasted(+, [1, 2, 3], [10 20 30 40]), 10)
       materialize(bc)
3×4 Array{Float64,2}:
 1.1  2.1  3.1  4.1
 1.2  2.2  3.2  4.2
 1.3  2.3  3.3  4.3

In this case, that bc object is an instance of a Broadcasted struct. It just holds onto the function and its arguments — and its arguments may include other nested Broadcasted structs. The materialize function does a bit of pre-processing and then calls copy(bc), which allocates the result and then finally loops over the result and executes the functions.

Each step along the way is extensible, leveraging the power of Julia's multiple dispatch, inlining and argument specialization for near-zero overhead. With this basic framework in mind, you can begin to see how the built-in and standard library arrays are able to implement all those new features mentioned above:

Looking to the future

Of course, this is all documented and available to packages. There's no secret sauce that we're hoarding for ourselves. It'll be exciting to see how the many creative minds in the package ecosystem manage to take advantage of all this new functionality. Some of the places where I see this yielding a significant benefit include machine learning applications and computation on GPUs.

Historically, some of the machine learning libraries have put in place hacks to allow their arrays to opt-out of fusion. The new broadcasting API instead allows them to opt-out of fusion with a first-class and fully supported mechanism that is much simpler to implement. Further, many deep learning techniques are heavily dependent upon differentiation; in some cases introspecting a broadcasted kernel will allow them to use an exact analytical result instead of computing differentials.

Finally, fusing a complicated broadcast expression into a single kernel has already been a huge boon to the performance of arrays on GPUs. Most GPU programming packages won't necessarily need to introspect or customize the broadcasted expression, but they're looking to the future to a potential extension that would allow fusing a reduction with the broadcasted expression. While not possible yet, the bulk of the machinery is in place to operate directly upon the Broadcasted lazy wrapper instead of allocating the intermediate array in an expression like sum(X.^2 .+ Y.^2).