# In-place/fused broadcasting

**URL:** <https://discourse.julialang.org/t/in-place-fused-broadcasting/21067>\
**Category:** General Usage\
**Created:** [February 22, 2019, 7:37am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067 "2019-02-22T07:37:13Z")\
**Posts on this page:** 11\
**Page:** 1

<div class="post-metadata">

**Author:** ![robsmith11](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/robsmith11/32/29641_2.png) [@robsmith11](https://discourse.julialang.org/u/robsmith11)\
**Post date:** [February 22, 2019, 7:37am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/1 "2019-02-22T07:37:13Z")

</div>

I often write something like

```julia
exp.(rand(10^6))

```

which is convienent, but inefficient as it allocates 2 arrays.

What I really want is

```julia
x = rand(10^6)
map!(exp, x, x)

```

Is there a single-line/idiomatic way to write this operation without binding to a variable and still get only the 1 allocation?

---

<div class="post-metadata">

**Author:** ![improbable22](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/improbable22/32/5464_2.png) [@improbable22](https://discourse.julialang.org/u/improbable22)\
**Post date:** [February 22, 2019, 7:45am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/2 "2019-02-22T07:45:42Z")

</div>

I was going to suggest this, as `rand.()` is often a neat trick for avoiding some allocations:

```julia
exp.(rand.()) .+ 0 .* (1:10^6)

```

BUT it turns out this is an error, `ArgumentError: step cannot be zero`.

---

<div class="post-metadata">

**Author:** ![Tamas\_Papp](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/tamas_papp/32/25949_2.png) [@Tamas\_Papp](https://discourse.julialang.org/u/Tamas_Papp)\
**Post date:** [February 22, 2019, 7:53am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/3 "2019-02-22T07:53:13Z")

</div>

```julia
[exp(rand()) for _ in 1:10^6]

```

They key is that you are not really broadcasting over anything, so forcing this into the broadcasting framework is not idiomatic. You are generating a vector elementwise.

---

<div class="post-metadata">

**Author:** ![mauro3](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/mauro3/32/292_2.png) [@mauro3](https://discourse.julialang.org/u/mauro3)\
**Post date:** [February 22, 2019, 8:55am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/4 "2019-02-22T08:55:46Z")

</div>

I came up with this

```julia
exp.(rand() for i=1:10^6)

```

but it allocate also two arrays as OP’s. Any idea why? (Also note that although it has two allocations it’s as fast as Tamas’, and as fast as OP’s)

---

<div class="post-metadata">

**Author:** ![Tamas\_Papp](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/tamas_papp/32/25949_2.png) [@Tamas\_Papp](https://discourse.julialang.org/u/Tamas_Papp)\
**Post date:** [February 22, 2019, 11:07am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/5 "2019-02-22T11:07:41Z")

</div>

> [@mauro3](#):
>
> as fast as Tamas’, and as fast as OP’s

I did not profile, but I would imagine that the cost of `rand()` would dominate.

---

<div class="post-metadata">

**Author:** ![bennedich](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/bennedich/32/4894_2.png) [@bennedich](https://discourse.julialang.org/u/bennedich)\
**Post date:** [February 22, 2019, 11:30am UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/6 "2019-02-22T11:30:22Z")

</div>

> [@mauro3](#):
>
> I came up with this /…/ but it allocate also two arrays as OP’s. Any idea why?

The problem with that approach is that it’s broadcasting a generator, which ends up materializing the generator first, hence the double allocations. If you’re intent on using broadcasting, you can do better by avoiding the generator as in this example:

```julia
exp.(1:10^6 .|>_-> rand())

```

As for performance, both versions benchmark about the same if using `@btime`, BUT it’s because `@btime` reports the _minimum_ time of many invocations. Versions allocating more memory will have a higher _average_ runtime due to less cache efficiency and garbage collection. See below; first @Tamas_Papp’s version:

```julia
julia> @benchmark [exp(rand()) for _ in 1:10^6]
  memory estimate: 7.63 MiB
  --------------
  minimum time: 11.669 ms (0.00% GC)
  median time: 12.656 ms (0.00% GC)
  mean time: 13.682 ms (7.57% GC)
  maximum time: 18.180 ms (17.96% GC)
  --------------
  samples: 366

```

Next my broadcasted version:

```julia
julia> @benchmark exp.(1:10^6 .|>_-> rand())
  memory estimate: 7.63 MiB
  --------------
  minimum time: 11.263 ms (0.00% GC)
  median time: 12.247 ms (0.00% GC)
  mean time: 13.254 ms (7.66% GC)
  maximum time: 17.265 ms (18.91% GC)
  --------------
  samples: 378

```

Finally the broadcasted generator (double allocations). Note how the minimum time is about the same, but average is quite a bit higher, and there’s more time spent GCing:

```julia
julia> @benchmark exp.(rand() for i=1:10^6)
  memory estimate: 15.26 MiB
  --------------
  minimum time: 12.164 ms (0.00% GC)
  median time: 17.054 ms (15.13% GC)
  mean time: 16.380 ms (13.03% GC)
  maximum time: 20.556 ms (19.80% GC)
  --------------
  samples: 305

```

---

<div class="post-metadata">

**Author:** ![robsmith11](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/robsmith11/32/29641_2.png) [@robsmith11](https://discourse.julialang.org/u/robsmith11)\
**Post date:** [February 22, 2019, 12:27pm UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/7 "2019-02-22T12:27:56Z")

</div>

Looking at the timings, I see that the two once-allocating versions suggested in this thread are actually slower than the twice-allocating version and the explicit `map!()` (which is the fastest):

```julia
julia> @benchmark exp.(rand(10^6))
BenchmarkTools.Trial:
  memory estimate: 15.26 MiB
  allocs estimate: 4
  --------------
  minimum time: 9.125 ms (0.00% GC)
  median time: 9.517 ms (2.78% GC)
  mean time: 9.590 ms (3.56% GC)
  maximum time: 50.124 ms (80.48% GC)
  --------------
  samples: 522
  evals/sample: 1

julia> @benchmark [exp(rand()) for _ in 1:10^6]
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 10.398 ms (0.00% GC)
  median time: 10.463 ms (0.00% GC)
  mean time: 10.710 ms (2.36% GC)
  maximum time: 51.215 ms (78.24% GC)
  --------------
  samples: 467
  evals/sample: 1

julia> @benchmark exp.(1:10^6 .|> _ -> rand())
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 10.348 ms (0.00% GC)
  median time: 10.412 ms (0.00% GC)
  mean time: 10.671 ms (2.41% GC)
  maximum time: 51.442 ms (78.33% GC)
  --------------
  samples: 469
  evals/sample: 1

julia> @benchmark (x = rand(10^6); map!(exp, x, x))
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 8.997 ms (0.00% GC)
  median time: 9.059 ms (0.00% GC)
  mean time: 9.350 ms (3.04% GC)
  maximum time: 49.337 ms (80.68% GC)
  --------------
  samples: 535
  evals/sample: 1

```

I guess it’s because the vectorized version of `rand()` is faster:

```julia
julia> @benchmark [rand() for _ in 1:10^6]
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 1.567 ms (0.00% GC)
  median time: 1.581 ms (0.00% GC)
  mean time: 1.673 ms (5.33% GC)
  maximum time: 42.036 ms (94.77% GC)
  --------------
  samples: 2987
  evals/sample: 1

julia> @benchmark rand(10^6)
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 716.757 μs (0.00% GC)
  median time: 731.317 μs (0.00% GC)
  mean time: 805.241 μs (8.98% GC)
  maximum time: 41.032 ms (96.89% GC)
  --------------
  samples: 6201
  evals/sample: 1

```

---

<div class="post-metadata">

**Author:** ![mauro3](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/mauro3/32/292_2.png) [@mauro3](https://discourse.julialang.org/u/mauro3)\
**Post date:** [February 22, 2019, 12:37pm UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/8 "2019-02-22T12:37:10Z")

</div>

> [@bennedich](#):
>
> which ends up materializing the generator first

Do you know why this is done/needed?

---

<div class="post-metadata">

**Author:** ![bennedich](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/bennedich/32/4894_2.png) [@bennedich](https://discourse.julialang.org/u/bennedich)\
**Post date:** [February 22, 2019, 1:09pm UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/9 "2019-02-22T13:09:19Z")

</div>

Yes. Generators are not broadcastable, so they will be materialized during broadcasting:

```julia
julia> a = Base.broadcasted(_->rand(), 1:3) # this is 1:3 .|>_-> rand()
Base.Broadcast.Broadcasted(getfield(Main, Symbol("##271#272"))(), (1:3,))

julia> Broadcast.broadcastable(a) # same object, no materialization
Base.Broadcast.Broadcasted(getfield(Main, Symbol("##271#272"))(), (1:3,))

julia> b = (rand() for i=1:3)
Base.Generator{UnitRange{Int64},getfield(Main, Symbol("##273#274"))}(getfield(Main, Symbol("##273#274"))(), 1:3)

julia> Broadcast.broadcastable(b) # materialization happens here
3-element Array{Float64,1}:
 0.9143249336534409
 0.733436533928558
 0.9617715309192958

```

As for why it’s done this way, perhaps someone more familiar with the internals can clarify, but I believe that it has to do with generators not being indexable:

```julia
julia> a[2]
0.848578901008149

julia> b[2]
ERROR: MethodError: no method matching getindex(::Base.Generator{UnitRange{Int64},getfield(Main, Symbol("##273#274"))}, ::Int64)

```

---

<div class="post-metadata">

**Author:** ![Seif\_Shebl](https://avatars.discourse-cdn.com/v4/letter/s/eada6e/32.png) [@Seif\_Shebl](https://discourse.julialang.org/u/Seif_Shebl)\
**Post date:** [February 22, 2019, 1:21pm UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/10 "2019-02-22T13:21:50Z")

</div>

Also `map(_ -> exp(rand()), 1:10^6)`, it’s elegant and has the same speed as a list comprehension.

```julia
julia> @benchmark map(_->exp(rand()),1:10^6)
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 11.519 ms (0.00% GC)
  median time: 11.833 ms (0.00% GC)
  mean time: 12.518 ms (6.99% GC)
  maximum time: 59.233 ms (80.46% GC)
  --------------
  samples: 400
  evals/sample: 1

julia> @benchmark [exp(rand()) for _ in 1:10^6]
BenchmarkTools.Trial:
  memory estimate: 7.63 MiB
  allocs estimate: 2
  --------------
  minimum time: 11.605 ms (0.00% GC)
  median time: 11.933 ms (0.00% GC)
  mean time: 12.600 ms (6.97% GC)
  maximum time: 59.423 ms (80.24% GC)
  --------------
  samples: 397
  evals/sample: 1

```

---

<div class="post-metadata">

**Author:** ![Nosferican](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/nosferican/32/9275_2.png) [@Nosferican](https://discourse.julialang.org/u/Nosferican)\
**Post date:** [February 22, 2019, 2:32pm UTC](https://discourse.julialang.org/t/in-place-fused-broadcasting/21067/11 "2019-02-22T14:32:51Z")

</div>

I would use,

```julia
rand(10^6) |> (x -> map!(exp, x, x))

```