-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Performance issues with fusing 12+ broadcasts #22255
Comments
Any chance you could make the example simpler, e.g. by removing uses of |
AFAICT you are not doing braodcasting for the slower version. Also, |
Ah, the |
Give me a little bit to get it simpler. I have to restart Atom and delete my precompilation cache every time I do something due to #21969, so this will take awhile. |
Here's the form that expands all of the broadcasts: function perform_step!(integrator,cache::Tsit5Cache,f=integrator.f)
@unpack t,dt,uprev,u,k = integrator
@unpack c1,c2,c3,c4,c5,c6,a21,a31,a32,a41,a42,a43,a51,a52,a53,a54,a61,a62,a63,a64,a65,a71,a72,a73,a74,a75,a76,b1,b2,b3,b4,b5,b6,b7 = cache.tab
@unpack k1,k2,k3,k4,k5,k6,k7,utilde,tmp,atmp = cache
a = dt*a21
tmp .= (muladd).(a, k1, uprev)
f(muladd(c1, dt, t),tmp,k2)
tmp .= (muladd).(dt, (muladd).(a31, k1, a32 .* k2), uprev)
f(muladd(c2, dt, t),tmp,k3)
tmp .= (muladd).(dt, (muladd).(a41, k1, (muladd).(a42, k2, a43 .* k3)), uprev)
f(muladd(c3, dt, t),tmp,k4)
tmp .= (muladd).(dt, (muladd).(a51, k1, (muladd).(a52, k2, (muladd).(a53, k3, a54 .* k4))), uprev)
f(muladd(c4, dt, t),tmp,k5)
tmp .= (muladd).(dt, (muladd).(a61, k1, (muladd).(a62, k2, (muladd).(a63, k3, (muladd).(a64, k4, a65 .* k5)))), uprev)
f(t+dt,tmp,k6)
u .= (muladd).(dt, (muladd).(a71, k1, (muladd).(a72, k2, (muladd).(a73, k3, (muladd).(a74, k4, (muladd).(a75, k5, a76 .* k6))))), uprev)
f(t+dt,u,k7)
if integrator.opts.adaptive
utilde .= (muladd).(dt, (muladd).(b1, k1, (muladd).(b2, k2, (muladd).(b3, k3, (muladd).(b4, k4, (muladd).(b5, k5, (muladd).(b6, k6, b7 .* k7)))))), uprev)
atmp .= ((utilde.-u)./(muladd).(max.(abs.(uprev), abs.(u)), integrator.opts.reltol, integrator.opts.abstol))
integrator.EEst = integrator.opts.internalnorm(atmp)
end
@pack integrator = t,dt,u,k
end adding integrator = init(prob,Tsit5(),dense=false,dt=1/10)
OrdinaryDiffEq.loopheader!(integrator)
@code_warntype OrdinaryDiffEq.perform_step!(integrator,integrator.cache,integrator.f) https://gist.github.com/ChrisRackauckas/11ba482f80147dd744f5c2b87e288067 |
This might be an MWE: function f(a,b,c)
a.= b.*c
end
function g(a,b,c)
@inbounds for ii in eachindex(a)
a[ii] = b[ii].*c[ii]
end
end
function f(a,b,c,d,e,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v,w,x,y,z)
a.= b.*c .+ d .* e .+ h .* i .+ j .*k .+ l .* m .+ n .* o .+ p .* q .+ r .* s .+ t .* u .+ v .* w .+ x .* y .+ z
end
function g(a,b,c,d,e,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v,w,x,y,z)
@inbounds for ii in eachindex(a)
a[ii] = b[ii].*c[ii] .+ d[ii] .* e[ii] .+ h[ii] .* i[ii] .+ j[ii] .*k[ii] .+ l[ii] .* m[ii] .+ n[ii] .* o[ii] .+ p[ii] .* q[ii] .+ r[ii] .* s[ii] .+ t[ii] .* u[ii] .+ v[ii] .* w[ii] .+ x[ii] .* y[ii] .+ z[ii]
end
end
a = rand(10)
b = rand(10)
c = rand(10)
d = rand(10)
e = rand(10)
h = rand(10)
i = rand(10)
j = rand(10)
k = rand(10)
l = rand(10)
m = rand(10)
n = rand(10)
o = rand(10)
p = rand(10)
q = rand(10)
r = rand(10)
s = rand(10)
t = rand(10)
u = rand(10)
v = rand(10)
w = rand(10)
x = rand(10)
y = rand(10)
z = rand(10)
@benchmark f($a,$b,$c)
@benchmark g($a,$b,$c)
@benchmark f($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q,$r,$s,$t,$u,$v,$w,$x,$y,$z)
@benchmark g($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q,$r,$s,$t,$u,$v,$w,$x,$y,$z) Benchmarks on the small broadcasts: @benchmark f($a,$b,$c)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0
--------------
minimum time: 16.393 ns (0.00% GC)
median time: 32.494 ns (0.00% GC)
mean time: 28.710 ns (0.00% GC)
maximum time: 225.996 ns (0.00% GC)
--------------
samples: 10000
evals/sample: 1000
@benchmark g($a,$b,$c)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0
--------------
minimum time: 7.611 ns (0.00% GC)
median time: 14.051 ns (0.00% GC)
mean time: 12.511 ns (0.00% GC)
maximum time: 192.038 ns (0.00% GC)
--------------
samples: 10000
evals/sample: 1000 and benchmarks on the long broadcasts: @benchmark f($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q,$r,$s,$t,$u,$v,$w,$x,$y,$z)
BenchmarkTools.Trial:
memory estimate: 17.77 KiB
allocs estimate: 650
--------------
minimum time: 51.229 μs (0.00% GC)
median time: 101.874 μs (0.00% GC)
mean time: 122.778 μs (3.60% GC)
maximum time: 31.670 ms (0.00% GC)
--------------
samples: 10000
evals/sample: 1
@benchmark g($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q,$r,$s,$t,$u,$v,$w,$x,$y,$z)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0
--------------
minimum time: 222.590 ns (0.00% GC)
median time: 413.482 ns (0.00% GC)
mean time: 363.430 ns (0.00% GC)
maximum time: 2.199 μs (0.00% GC)
--------------
samples: 10000
evals/sample: 434 When it gets sufficiently long, it begins to allocate quite a bit! |
Indeed get a the same kinds of differences when I use some of the broadcast expressions from the test code: function f(a,b,c,d,e,h,i,j,k,l,m,n,o,p,q)
a .= (muladd).(b, (muladd).(c, d, (muladd).(e, h, (muladd).(i, j, (muladd).(k, l, (muladd).(m, n, o .* p))))), q)
end
function g(a,b,c,d,e,h,i,j,k,l,m,n,o,p,q)
@inbounds for ii in eachindex(a)
a[ii] = (muladd)(b[ii], (muladd)(c[ii], d[ii], (muladd)(e[ii], h[ii], (muladd)(i[ii], j[ii], (muladd)(k[ii], l[ii], (muladd)(m[ii], n[ii], o[ii] .* p[ii]))))), q[ii])
end
end
@benchmark f($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q)
@benchmark g($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q) @benchmark f($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q)
BenchmarkTools.Trial:
memory estimate: 1.58 KiB
allocs estimate: 44
--------------
minimum time: 6.382 μs (0.00% GC)
median time: 12.471 μs (0.00% GC)
mean time: 11.167 μs (2.64% GC)
maximum time: 1.154 ms (94.98% GC)
--------------
samples: 10000
evals/sample: 5 @benchmark g($a,$b,$c,$d,$e,$h,$i,$j,$k,$l,$m,$n,$o,$p,$q)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0
--------------
minimum time: 48.009 ns (0.00% GC)
median time: 83.139 ns (0.00% GC)
mean time: 74.315 ns (0.00% GC)
maximum time: 1.335 μs (0.00% GC)
--------------
samples: 10000
evals/sample: 1000 Interestingly, that's only with 8 dots and 15 numbers... I would've though the magic 16 would be a limit here. |
Does it still happen when reducing the number of operands? |
No. See the "benchmarks on small broadcasts". For "sufficiently few" operands, it works just fine. |
OK. So where's the cutoff? |
With the earlier example Chris had, it happens here with 12 variables, but not with 11: function fun1!(a,b,c,d,e,f,g,h,i,j,k)
a .= b.*c .+ d.*e .+ f.*g .+ h.*i .+ j.*k
end
function fun2!(a,b,c,d,e,f,g,h,i,j,k,l)
a .= b.*c .+ d.*e .+ f.*g .+ h.*i .+ j.*k .+ l
end
a = rand(10); b = rand(10); c = rand(10); d = rand(10); e = rand(10);
f = rand(10); g = rand(10); h = rand(10); i = rand(10); j = rand(10);
k = rand(10); l = rand(10);
using BenchmarkTools
@benchmark fun1!($a,$b,$c,$d,$e,$f,$g,$h,$i,$j,$k)
@benchmark fun2!($a,$b,$c,$d,$e,$f,$g,$h,$i,$j,$k,$l)
|
Updated the title to reflect the new status of the issue. Thanks @jebej for finding that. Updating the OP with the new information. |
Does anyone know which inference option would be involved here? I could 10 |
This patch represents the combined efforts of four individuals, over 60 commits, and an iterated design over (at least) three pull requests that spanned nearly an entire year (closes #22063, #23692, #25377 by superceding them). This introduces a pure Julia data structure that represents a fused broadcast expression. For example, the expression `2 .* (x .+ 1)` lowers to: ```julia julia> Meta.@lower 2 .* (x .+ 1) :($(Expr(:thunk, CodeInfo(:(begin Core.SSAValue(0) = (Base.getproperty)(Base.Broadcast, :materialize) Core.SSAValue(1) = (Base.getproperty)(Base.Broadcast, :make) Core.SSAValue(2) = (Base.getproperty)(Base.Broadcast, :make) Core.SSAValue(3) = (Core.SSAValue(2))(+, x, 1) Core.SSAValue(4) = (Core.SSAValue(1))(*, 2, Core.SSAValue(3)) Core.SSAValue(5) = (Core.SSAValue(0))(Core.SSAValue(4)) return Core.SSAValue(5) end))))) ``` Or, slightly more readably as: ```julia using .Broadcast: materialize, make materialize(make(*, 2, make(+, x, 1))) ``` The `Broadcast.make` function serves two purposes. Its primary purpose is to construct the `Broadcast.Broadcasted` objects that hold onto the function, the tuple of arguments (potentially including nested `Broadcasted` arguments), and sometimes a set of `axes` to include knowledge of the outer shape. The secondary purpose, however, is to allow an "out" for objects that _don't_ want to participate in fusion. For example, if `x` is a range in the above `2 .* (x .+ 1)` expression, it needn't allocate an array and operate elementwise — it can just compute and return a new range. Thus custom structures are able to specialize `Broadcast.make(f, args...)` just as they'd specialize on `f` normally to return an immediate result. `Broadcast.materialize` is identity for everything _except_ `Broadcasted` objects for which it allocates an appropriate result and computes the broadcast. It does two things: it `initialize`s the outermost `Broadcasted` object to compute its axes and then `copy`s it. Similarly, an in-place fused broadcast like `y .= 2 .* (x .+ 1)` uses the exact same expression tree to compute the right-hand side of the expression as above, and then uses `materialize!(y, make(*, 2, make(+, x, 1)))` to `instantiate` the `Broadcasted` expression tree and then `copyto!` it into the given destination. All-together, this forms a complete API for custom types to extend and customize the behavior of broadcast (fixes #22060). It uses the existing `BroadcastStyle`s throughout to simplify dispatch on many arguments: * Custom types can opt-out of broadcast fusion by specializing `Broadcast.make(f, args...)` or `Broadcast.make(::BroadcastStyle, f, args...)`. * The `Broadcasted` object computes and stores the type of the combined `BroadcastStyle` of its arguments as its first type parameter, allowing for easy dispatch and specialization. * Custom Broadcast storage is still allocated via `broadcast_similar`, however instead of passing just a function as a first argument, the entire `Broadcasted` object is passed as a final argument. This potentially allows for much more runtime specialization dependent upon the exact expression given. * Custom broadcast implmentations for a `CustomStyle` are defined by specializing `copy(bc::Broadcasted{CustomStyle})` or `copyto!(dest::AbstractArray, bc::Broadcasted{CustomStyle})`. * Fallback broadcast specializations for a given output object of type `Dest` (for the `DefaultArrayStyle` or another such style that hasn't implemented assignments into such an object) are defined by specializing `copyto(dest::Dest, bc::Broadcasted{Nothing})`. As it fully supports range broadcasting, this now deprecates `(1:5) + 2` to `.+`, just as had been done for all `AbstractArray`s in general. As a first-mover proof of concept, LinearAlgebra uses this new system to improve broadcasting over structured arrays. Before, broadcasting over a structured matrix would result in a sparse array. Now, broadcasting over a structured matrix will _either_ return an appropriately structured matrix _or_ a dense array. This does incur a type instability (in the form of a discriminated union) in some situations, but thanks to type-based introspection of the `Broadcasted` wrapper commonly used functions can be special cased to be type stable. For example: ```julia julia> f(d) = round.(Int, d) f (generic function with 1 method) julia> @inferred f(Diagonal(rand(3))) 3×3 Diagonal{Int64,Array{Int64,1}}: 0 ⋅ ⋅ ⋅ 0 ⋅ ⋅ ⋅ 1 julia> @inferred Diagonal(rand(3)) .* 3 ERROR: return type Diagonal{Float64,Array{Float64,1}} does not match inferred return type Union{Array{Float64,2}, Diagonal{Float64,Array{Float64,1}}} Stacktrace: [1] error(::String) at ./error.jl:33 [2] top-level scope julia> @inferred Diagonal(1:4) .+ Bidiagonal(rand(4), rand(3), 'U') .* Tridiagonal(1:3, 1:4, 1:3) 4×4 Tridiagonal{Float64,Array{Float64,1}}: 1.30771 0.838589 ⋅ ⋅ 0.0 3.89109 0.0459757 ⋅ ⋅ 0.0 4.48033 2.51508 ⋅ ⋅ 0.0 6.23739 ``` In addition to the issues referenced above, it fixes: * Fixes #19313, #22053, #23445, and #24586: Literals are no longer treated specially in a fused broadcast; they're just arguments in a `Broadcasted` object like everything else. * Fixes #21094: Since broadcasting is now represented by a pure Julia datastructure it can be created within `@generated` functions and serialized. * Fixes #26097: The fallback destination-array specialization method of `copyto!` is specifically implemented as `Broadcasted{Nothing}` and will not be confused by `nothing` arguments. * Fixes the broadcast-specific element of #25499: The default base broadcast implementation no longer depends upon `Base._return_type` to allocate its array (except in the empty or concretely-type cases). Note that the sparse implementation (#19595) is still dependent upon inference and is _not_ fixed. * Fixes #25340: Functions are treated like normal values just like arguments and only evaluated once. * Fixes #22255, and is performant with 12+ fused broadcasts. Okay, that one was fixed on master already, but this fixes it now, too. * Fixes #25521. * The performance of this patch has been thoroughly tested through its iterative development process in #25377. There remain [two classes of performance regressions](#25377) that Nanosoldier flagged. * #25691: Propagation of constant literals sill lose their constant-ness upon going through the broadcast machinery. I believe quite a large number of functions would need to be marked as `@pure` to support this -- including functions that are intended to be specialized. (For bookkeeping, this is the squashed version of the [teh-jn/lazydotfuse](#25377) branch as of a1d4e7e. Squashed and separated out to make it easier to review and commit) Co-authored-by: Tim Holy <tim.holy@gmail.com> Co-authored-by: Jameson Nash <vtjnash@gmail.com> Co-authored-by: Andrew Keller <ajkeller34@users.noreply.github.com>
This patch represents the combined efforts of four individuals, over 60 commits, and an iterated design over (at least) three pull requests that spanned nearly an entire year (closes #22063, #23692, #25377 by superceding them). This introduces a pure Julia data structure that represents a fused broadcast expression. For example, the expression `2 .* (x .+ 1)` lowers to: ```julia julia> Meta.@lower 2 .* (x .+ 1) :($(Expr(:thunk, CodeInfo(:(begin Core.SSAValue(0) = (Base.getproperty)(Base.Broadcast, :materialize) Core.SSAValue(1) = (Base.getproperty)(Base.Broadcast, :make) Core.SSAValue(2) = (Base.getproperty)(Base.Broadcast, :make) Core.SSAValue(3) = (Core.SSAValue(2))(+, x, 1) Core.SSAValue(4) = (Core.SSAValue(1))(*, 2, Core.SSAValue(3)) Core.SSAValue(5) = (Core.SSAValue(0))(Core.SSAValue(4)) return Core.SSAValue(5) end))))) ``` Or, slightly more readably as: ```julia using .Broadcast: materialize, make materialize(make(*, 2, make(+, x, 1))) ``` The `Broadcast.make` function serves two purposes. Its primary purpose is to construct the `Broadcast.Broadcasted` objects that hold onto the function, the tuple of arguments (potentially including nested `Broadcasted` arguments), and sometimes a set of `axes` to include knowledge of the outer shape. The secondary purpose, however, is to allow an "out" for objects that _don't_ want to participate in fusion. For example, if `x` is a range in the above `2 .* (x .+ 1)` expression, it needn't allocate an array and operate elementwise — it can just compute and return a new range. Thus custom structures are able to specialize `Broadcast.make(f, args...)` just as they'd specialize on `f` normally to return an immediate result. `Broadcast.materialize` is identity for everything _except_ `Broadcasted` objects for which it allocates an appropriate result and computes the broadcast. It does two things: it `initialize`s the outermost `Broadcasted` object to compute its axes and then `copy`s it. Similarly, an in-place fused broadcast like `y .= 2 .* (x .+ 1)` uses the exact same expression tree to compute the right-hand side of the expression as above, and then uses `materialize!(y, make(*, 2, make(+, x, 1)))` to `instantiate` the `Broadcasted` expression tree and then `copyto!` it into the given destination. All-together, this forms a complete API for custom types to extend and customize the behavior of broadcast (fixes #22060). It uses the existing `BroadcastStyle`s throughout to simplify dispatch on many arguments: * Custom types can opt-out of broadcast fusion by specializing `Broadcast.make(f, args...)` or `Broadcast.make(::BroadcastStyle, f, args...)`. * The `Broadcasted` object computes and stores the type of the combined `BroadcastStyle` of its arguments as its first type parameter, allowing for easy dispatch and specialization. * Custom Broadcast storage is still allocated via `broadcast_similar`, however instead of passing just a function as a first argument, the entire `Broadcasted` object is passed as a final argument. This potentially allows for much more runtime specialization dependent upon the exact expression given. * Custom broadcast implmentations for a `CustomStyle` are defined by specializing `copy(bc::Broadcasted{CustomStyle})` or `copyto!(dest::AbstractArray, bc::Broadcasted{CustomStyle})`. * Fallback broadcast specializations for a given output object of type `Dest` (for the `DefaultArrayStyle` or another such style that hasn't implemented assignments into such an object) are defined by specializing `copyto(dest::Dest, bc::Broadcasted{Nothing})`. As it fully supports range broadcasting, this now deprecates `(1:5) + 2` to `.+`, just as had been done for all `AbstractArray`s in general. As a first-mover proof of concept, LinearAlgebra uses this new system to improve broadcasting over structured arrays. Before, broadcasting over a structured matrix would result in a sparse array. Now, broadcasting over a structured matrix will _either_ return an appropriately structured matrix _or_ a dense array. This does incur a type instability (in the form of a discriminated union) in some situations, but thanks to type-based introspection of the `Broadcasted` wrapper commonly used functions can be special cased to be type stable. For example: ```julia julia> f(d) = round.(Int, d) f (generic function with 1 method) julia> @inferred f(Diagonal(rand(3))) 3×3 Diagonal{Int64,Array{Int64,1}}: 0 ⋅ ⋅ ⋅ 0 ⋅ ⋅ ⋅ 1 julia> @inferred Diagonal(rand(3)) .* 3 ERROR: return type Diagonal{Float64,Array{Float64,1}} does not match inferred return type Union{Array{Float64,2}, Diagonal{Float64,Array{Float64,1}}} Stacktrace: [1] error(::String) at ./error.jl:33 [2] top-level scope julia> @inferred Diagonal(1:4) .+ Bidiagonal(rand(4), rand(3), 'U') .* Tridiagonal(1:3, 1:4, 1:3) 4×4 Tridiagonal{Float64,Array{Float64,1}}: 1.30771 0.838589 ⋅ ⋅ 0.0 3.89109 0.0459757 ⋅ ⋅ 0.0 4.48033 2.51508 ⋅ ⋅ 0.0 6.23739 ``` In addition to the issues referenced above, it fixes: * Fixes #19313, #22053, #23445, and #24586: Literals are no longer treated specially in a fused broadcast; they're just arguments in a `Broadcasted` object like everything else. * Fixes #21094: Since broadcasting is now represented by a pure Julia datastructure it can be created within `@generated` functions and serialized. * Fixes #26097: The fallback destination-array specialization method of `copyto!` is specifically implemented as `Broadcasted{Nothing}` and will not be confused by `nothing` arguments. * Fixes the broadcast-specific element of #25499: The default base broadcast implementation no longer depends upon `Base._return_type` to allocate its array (except in the empty or concretely-type cases). Note that the sparse implementation (#19595) is still dependent upon inference and is _not_ fixed. * Fixes #25340: Functions are treated like normal values just like arguments and only evaluated once. * Fixes #22255, and is performant with 12+ fused broadcasts. Okay, that one was fixed on master already, but this fixes it now, too. * Fixes #25521. * The performance of this patch has been thoroughly tested through its iterative development process in #25377. There remain [two classes of performance regressions](#25377) that Nanosoldier flagged. * #25691: Propagation of constant literals sill lose their constant-ness upon going through the broadcast machinery. I believe quite a large number of functions would need to be marked as `@pure` to support this -- including functions that are intended to be specialized. (For bookkeeping, this is the squashed version of the [teh-jn/lazydotfuse](#25377) branch as of a1d4e7e. Squashed and separated out to make it easier to review and commit) Co-authored-by: Tim Holy <tim.holy@gmail.com> Co-authored-by: Jameson Nash <vtjnash@gmail.com> Co-authored-by: Andrew Keller <ajkeller34@users.noreply.github.com>
Updated OP
While working through the issue, @jebej identified that the problem is fusing 12+ broadcasts in this comment (#22255 (comment)) which contains an MWE.
Original OP
In OrdinaryDiffEq.jl, I see a 10x performance regression due to using broadcast. With the testing code:
I get a 10x regression by changing the inner loop from:
to:
I.e. all that's changed are loops to broadcast. The input array is
y0
which is length 2. For reference, the@muladd
macro acts like:The profile is here: https://ufile.io/2lu0f
The benchmark results are using loops:
and using broadcast:
Am I hitting some broadcasting splatting penalty or something?
The text was updated successfully, but these errors were encountered: