gpu, similar, cpu, adapt

adapt.jl

using Adapt
using CuArrays
using Flux
using InteractiveUtils

function f1(x)
    fill(1f0, size(x)) |> gpu
end

function f2(x)
    fill!(similar(x), 1f0)
end

function f3(x)
    fill(1f0, size(x)) |> cpu
end

function f4(x)
    adapt(Array, fill(1f0, size(x)))
end

function f5(x)
    cu(fill(1f0, size(x)))
end

function main()
    x = randn(Float32, 32, 32, 3, 32) |> gpu

    n = 10^4
    f1(x)
    @info "Case1" typeof(f1(x))
    @time for i in 1:n f1(x) end
    @code_warntype f1(x)

    println()

    f2(x)
    @info "Case2" typeof(f2(x))
    @time for i in 1:n f2(x) end
    @code_warntype f2(x)

    println()

    f3(x)
    @info "Case3" typeof(f3(x))
    @time for i in 1:n f3(x) end
    @code_warntype f3(x)

    println()

    f4(x)
    @info "Case4" typeof(f4(x))
    @time for i in 1:n f4(x) end
    @code_warntype f4(x)
    
    println()

    f5(x)
    @info "Case5" typeof(f5(x))
    @time for i in 1:n f5(x) end
    @code_warntype f5(x)
end

main()

CPU

[ Info: CUDAdrv.jl failed to initialize, GPU functionality unavailable (set JULIA_CUDA_SILENT or JULIA_CUDA_VERBOSE to silence or expand this message)
┌ Info: Case1
└   typeof(f1(x)) = Array{Float32,4}
  3.276545 seconds (30.00 k allocations: 3.663 GiB, 22.56% gc time)
Variables
  #self#::Core.Compiler.Const(f1, false)
  x::Array{Float32,4}

Body::ANY
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = (%2 |> Main.gpu)::ANY
└──      return %3

┌ Info: Case2
└   typeof(f2(x)) = Array{Float32,4}
  2.863216 seconds (30.00 k allocations: 3.663 GiB, 21.65% gc time)
Variables
  #self#::Core.Compiler.Const(f2, false)
  x::Array{Float32,4}

Body::Array{Float32,4}
1 ─ %1 = Main.similar(x)::Array{Float32,4}
│   %2 = Main.fill!(%1, 1.0f0)::Array{Float32,4}
└──      return %2

┌ Info: Case3
└   typeof(f3(x)) = Array{Float32,4}
  2.791182 seconds (50.00 k allocations: 3.667 GiB, 21.37% gc time)
Variables
  #self#::Core.Compiler.Const(f3, false)
  x::Array{Float32,4}

Body::ANY
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = (%2 |> Main.cpu)::ANY
└──      return %3

┌ Info: Case4
└   typeof(f4(x)) = Array{Float32,4}
  2.921619 seconds (30.00 k allocations: 3.663 GiB, 21.97% gc time)
Variables
  #self#::Core.Compiler.Const(f4, false)
  x::Array{Float32,4}

Body::Array{Float32,4}
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = Main.adapt(Main.Array, %2)::Array{Float32,4}
└──      return %3

ERROR: LoadError: could not load library "libcuda"
(omit)

GPU

┌ Info: Case1
└   typeof(f1(x)) = CuArray{Float32,4,Nothing}
  1.798704 seconds (201.18 k allocations: 7.334 GiB, 20.54% gc time)
Variables
  #self#::Core.Compiler.Const(f1, false)
  x::CuArray{Float32,4,Nothing}

Body::ANY
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = (%2 |> Main.gpu)::ANY
└──      return %3

┌ Info: Case2
└   typeof(f2(x)) = CuArray{Float32,4,Nothing}
  1.250865 seconds (130.00 k allocations: 4.765 MiB)
Variables
  #self#::Core.Compiler.Const(f2, false)
  x::CuArray{Float32,4,Nothing}

Body::CuArray{Float32,4,Nothing}
1 ─ %1 = Main.similar(x)::CuArray{Float32,4,Nothing}
│   %2 = Main.fill!(%1, 1.0f0)::CuArray{Float32,4,Nothing}
└──      return %2

┌ Info: Case3
└   typeof(f3(x)) = Array{Float32,4}
  0.668655 seconds (90.02 k allocations: 3.668 GiB, 26.45% gc time)
Variables
  #self#::Core.Compiler.Const(f3, false)
  x::CuArray{Float32,4,Nothing}

Body::ANY
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = (%2 |> Main.cpu)::ANY
└──      return %3

┌ Info: Case4
└   typeof(f4(x)) = Array{Float32,4}
  0.659487 seconds (30.00 k allocations: 3.663 GiB, 25.06% gc time)
Variables
  #self#::Core.Compiler.Const(f4, false)
  x::CuArray{Float32,4,Nothing}

Body::Array{Float32,4}
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = Main.adapt(Main.Array, %2)::Array{Float32,4}
└──      return %3

┌ Info: Case5
└   typeof(f5(x)) = CuArray{Float32,4,Nothing}
  1.932783 seconds (180.67 k allocations: 7.330 GiB, 19.17% gc time)
Variables
  #self#::Core.Compiler.Const(f5, false)
  x::CuArray{Float32,4,Nothing}

Body::UNION{ARRAY{FLOAT32,4}, CUARRAY{FLOAT32,4,NOTHING}}
1 ─ %1 = Main.size(x)::NTuple{4,Int64}
│   %2 = Main.fill(1.0f0, %1)::Array{Float32,4}
│   %3 = Main.cu(%2)::UNION{ARRAY{FLOAT32,4}, CUARRAY{FLOAT32,4,NOTHING}}
└──      return %3