Try #1204:

.gitlab-ci.yml

@@ -25,9 +25,9 @@ image: nvidia/cuda:10.1-cudnn7-devel-ubuntu18.04
 #   tags:
 #     - nvidia
 
-julia:1.3:
+julia:1.4:
   extends:
-    - .julia:1.3
+    - .julia:1.4
     - .test
   tags:
     - nvidia

.travis.yml

@@ -6,7 +6,7 @@ os:
   # - osx
 
 julia:
-  - 1.3
+  - 1.4
   - 1
   - nightly
 

Manifest.toml

@@ -34,33 +34,15 @@ uuid = "b99e7846-7c00-51b0-8f62-c81ae34c0232"
 version = "0.5.9"
 
 [[CEnum]]
-git-tree-sha1 = "1b77a77c3b28e0b3f413f7567c9bb8dd9bdccd14"
+git-tree-sha1 = "215a9aa4a1f23fbd05b92769fdd62559488d70e9"
 uuid = "fa961155-64e5-5f13-b03f-caf6b980ea82"
-version = "0.3.0"
+version = "0.4.1"
 
-[[CUDAapi]]
-deps = ["Libdl", "Logging"]
-git-tree-sha1 = "831b825d10104bd29e28f6da93312a976830717b"
-uuid = "3895d2a7-ec45-59b8-82bb-cfc6a382f9b3"
-version = "4.0.0"
-
-[[CUDAdrv]]
-deps = ["CEnum", "CUDAapi", "Printf"]
-git-tree-sha1 = "f56bbf18c86bcff7a961a32a4947a5abb2963a29"
-uuid = "c5f51814-7f29-56b8-a69c-e4d8f6be1fde"
-version = "6.3.0"
-
-[[CUDAnative]]
-deps = ["Adapt", "BinaryProvider", "CEnum", "CUDAapi", "CUDAdrv", "ExprTools", "GPUCompiler", "LLVM", "Libdl", "Pkg", "Printf"]
-git-tree-sha1 = "ac86db2b05fdfec96b011e25a504ffe7476e8a68"
-uuid = "be33ccc6-a3ff-5ff2-a52e-74243cff1e17"
-version = "3.1.0"
-
-[[CodeTracking]]
-deps = ["InteractiveUtils", "UUIDs"]
-git-tree-sha1 = "cab4da992adc0a64f63fa30d2db2fd8bec40cab4"
-uuid = "da1fd8a2-8d9e-5ec2-8556-3022fb5608a2"
-version = "0.5.11"
+[[CUDA]]
+deps = ["AbstractFFTs", "Adapt", "BinaryProvider", "CEnum", "DataStructures", "ExprTools", "GPUArrays", "GPUCompiler", "LLVM", "Libdl", "LinearAlgebra", "Logging", "MacroTools", "NNlib", "Pkg", "Printf", "Random", "Reexport", "Requires", "SparseArrays", "Statistics", "TimerOutputs"]
+git-tree-sha1 = "3d247bd2260259f86853e52584169ddeca648ebc"
+uuid = "052768ef-5323-5732-b1bb-66c8b64840ba"
+version = "0.1.0"
 
 [[CodecZlib]]
 deps = ["TranscodingStreams", "Zlib_jll"]
@@ -92,18 +74,6 @@ git-tree-sha1 = "7c4f882c41faa72118841185afc58a2eb00ef612"
 uuid = "e66e0078-7015-5450-92f7-15fbd957f2ae"
 version = "0.3.3+0"
 
-[[Cthulhu]]
-deps = ["CodeTracking", "InteractiveUtils", "REPL", "Unicode"]
-git-tree-sha1 = "a4849ec61df9659423cc63b298ed895904ee9743"
-uuid = "f68482b8-f384-11e8-15f7-abe071a5a75f"
-version = "1.0.2"
-
-[[CuArrays]]
-deps = ["AbstractFFTs", "Adapt", "CEnum", "CUDAapi", "CUDAdrv", "CUDAnative", "DataStructures", "GPUArrays", "Libdl", "LinearAlgebra", "MacroTools", "NNlib", "Pkg", "Printf", "Random", "Reexport", "Requires", "SparseArrays", "Statistics", "TimerOutputs"]
-git-tree-sha1 = "870a4ac61e99c36f42d15e496fd290c841541d90"
-uuid = "3a865a2d-5b23-5a0f-bc46-62713ec82fae"
-version = "2.2.0"
-
 [[DataAPI]]
 git-tree-sha1 = "176e23402d80e7743fc26c19c681bfb11246af32"
 uuid = "9a962f9c-6df0-11e9-0e5d-c546b8b5ee8a"
@@ -173,15 +143,15 @@ uuid = "9fa8497b-333b-5362-9e8d-4d0656e87820"
 
 [[GPUArrays]]
 deps = ["AbstractFFTs", "Adapt", "LinearAlgebra", "Printf", "Random", "Serialization"]
-git-tree-sha1 = "ce4579ebffef43e07318e9544ffeb6532c95d04d"
+git-tree-sha1 = "5f90fd4385ae86ab56cdd9bd4717f7bd2ff2acdf"
 uuid = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
-version = "3.3.0"
+version = "3.4.0"
 
 [[GPUCompiler]]
-deps = ["Cthulhu", "DataStructures", "InteractiveUtils", "LLVM", "Libdl", "TimerOutputs"]
-git-tree-sha1 = "5275aa268ecd09640b32560e1eae90c78816e4d1"
+deps = ["DataStructures", "InteractiveUtils", "LLVM", "Libdl", "TimerOutputs", "UUIDs"]
+git-tree-sha1 = "7ed10d9c41b7bb392f3e645233e45ecf93217130"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "0.2.0"
+version = "0.3.0"
 
 [[IRTools]]
 deps = ["InteractiveUtils", "MacroTools", "Test"]
@@ -201,9 +171,9 @@ version = "0.8.1"
 
 [[LLVM]]
 deps = ["CEnum", "Libdl", "Printf", "Unicode"]
-git-tree-sha1 = "93d2e1e960fe47db1a9015e86fad1d47cf67cf59"
+git-tree-sha1 = "dd3f584c3dbefe39b2a8fbafa1a3b77e31e21255"
 uuid = "929cbde3-209d-540e-8aea-75f648917ca0"
-version = "1.4.1"
+version = "1.5.1"
 
 [[LibGit2]]
 deps = ["Printf"]
@@ -370,9 +340,9 @@ version = "0.9.1"
 
 [[Zlib_jll]]
 deps = ["Libdl", "Pkg"]
-git-tree-sha1 = "2f6c3e15e20e036ee0a0965879b31442b7ec50fa"
+git-tree-sha1 = "a2e0d558f6031002e380a90613b199e37a8565bf"
 uuid = "83775a58-1f1d-513f-b197-d71354ab007a"
-version = "1.2.11+9"
+version = "1.2.11+10"
 
 [[Zygote]]
 deps = ["AbstractFFTs", "ArrayLayouts", "DiffRules", "FillArrays", "ForwardDiff", "Future", "IRTools", "InteractiveUtils", "LinearAlgebra", "MacroTools", "NNlib", "NaNMath", "Random", "Requires", "SpecialFunctions", "Statistics", "ZygoteRules"]

NEWS.md

@@ -1,4 +1,5 @@
 # v0.10.5
+* Moved CUDA compatibility to use [CUDA.jl instead of CuArrays.jl](https://github.com/FluxML/Flux.jl/pull/1204)
 * Add option for [same padding](https://github.com/FluxML/Flux.jl/pull/901) to conv and pooling layers by setting `pad=SamePad()`.
 * Added option to set `bias` to [Flux.Zeros](https://github.com/FluxML/Flux.jl/pull/873) to eliminating `bias` from being trained.
 * Added `GlobalMaxPool` and `GlobalMeanPool` [layers](https://github.com/FluxML/Flux.jl/pull/950) for performing global pooling operations.

Project.toml

@@ -5,9 +5,9 @@ version = "0.10.5"
 [deps]
 AbstractTrees = "1520ce14-60c1-5f80-bbc7-55ef81b5835c"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 CodecZlib = "944b1d66-785c-5afd-91f1-9de20f533193"
 Colors = "5ae59095-9a9b-59fe-a467-6f913c188581"
-CuArrays = "3a865a2d-5b23-5a0f-bc46-62713ec82fae"
 DelimitedFiles = "8bb1440f-4735-579b-a4ab-409b98df4dab"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 Juno = "e5e0dc1b-0480-54bc-9374-aad01c23163d"
@@ -30,7 +30,7 @@ AbstractTrees = "0.2, 0.3"
 Adapt = "1"
 CodecZlib = "0.5, 0.6, 0.7"
 Colors = "0.8, 0.9, 0.10, 0.11, 0.12"
-CuArrays = "2"
+CUDA = "0"
 Functors = "0.1"
 Juno = "0.5, 0.6, 0.7, 0.8"
 MacroTools = "0.3, 0.4, 0.5"
@@ -39,7 +39,7 @@ Reexport = "0.2"
 StatsBase = "0"
 ZipFile = "0.7, 0.8, 0.9"
 Zygote = "0.4.13"
-julia = "1.3"
+julia = "1.4"
 
 [extras]
 Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4"

src/Flux.jl

@@ -25,7 +25,7 @@ export Descent, ADAM, Momentum, Nesterov, RMSProp,
   ClipValue, ClipNorm
 
 
-using CuArrays
+using CUDA
 const use_cuda = Ref(false)
 
 include("utils.jl")
@@ -46,10 +46,10 @@ include("deprecations.jl")
 include("cuda/cuda.jl")
 
 function __init__()
-  use_cuda[] = CuArrays.functional() # Can be overridden after load with `Flux.use_cuda[] = false`
-  if CuArrays.functional()
-    if !CuArrays.has_cudnn()
-      @warn "CuArrays.jl found cuda, but did not find libcudnn. Some functionality will not be available."
+  use_cuda[] = CUDA.functional() # Can be overridden after load with `Flux.use_cuda[] = false`
+  if CUDA.functional()
+    if !CUDA.has_cudnn()
+      @warn "CUDA.jl found cuda, but did not find libcudnn. Some functionality will not be available."
     end
   end
 end

src/cuda/cuda.jl

@@ -1,8 +1,8 @@
-module CUDA
+module CUDAint
 
-using ..CuArrays
+using ..CUDA
 
-using CuArrays: CUDNN
+using CUDA: CUDNN
 include("curnn.jl")
 include("cudnn.jl")
 

src/cuda/cudnn.jl

@@ -1,5 +1,5 @@
 import ..Flux: data
-import CuArrays.CUDNN: batchnorm, ∇batchnorm
+import CUDA.CUDNN: batchnorm, ∇batchnorm
 
 (BN::Flux.BatchNorm)(x::Union{CuArray{T,2},CuArray{T,4},CuArray{T,5}}, cache = nothing) where T<:Union{Float32, Float64} =
   BN.λ.(batchnorm(BN.γ, BN.β, x, BN.μ, BN.σ², BN.momentum; cache = cache, alpha = 1, beta = 0, eps = BN.ϵ, training = Flux.istraining()))

src/cuda/curnn.jl

@@ -1,5 +1,4 @@
 import ..Flux: Flux, relu
-using CuArrays.CUDAnative
 
 CuRNN{T} = Flux.RNNCell{<:Union{typeof(tanh),typeof(relu)},<:CuArray{T,2},<:CuArray{T,1}}
 CuGRU{T} = Flux.GRUCell{<:CuArray{T,2},<:CuArray{T,1}}
@@ -55,7 +54,7 @@ unbroadcast(x::AbstractArray, Δ) =
 coerce_cuda(x::Union{CuArray,Nothing}) = x
 coerce_cuda(x::Tuple) = coerce_cuda.(x)
 
-coerce_cuda(x::AbstractArray) = x .+ CuArrays.fill(0)
+coerce_cuda(x::AbstractArray) = x .+ CUDA.fill(0)
 
 function struct_grad!(cx::Zygote.Context, x, x̄)
   for f in fieldnames(typeof(x))

src/functor.jl

@@ -70,7 +70,7 @@ end
 
 cpu(m) = fmap(x -> adapt(Array, x), m)
 
-gpu(x) = use_cuda[] ? fmap(CuArrays.cu, x) : x
+gpu(x) = use_cuda[] ? fmap(CUDA.cu, x) : x
 
 # Precision
 

src/layers/stateless.jl

@@ -125,8 +125,8 @@ julia> Flux.binarycrossentropy.(σ.([-1.1491, 0.8619, 0.3127]), [1, 1, 0])
 """
 binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -xlogy(y, ŷ + ϵ) - xlogy(1 - y, 1 - ŷ + ϵ)
 
-# Re-definition to fix interaction with CuArrays.
-CuArrays.@cufunc binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -y*log(ŷ + ϵ) - (1 - y)*log(1 - ŷ + ϵ)
+# Re-definition to fix interaction with CUDA.
+CUDA.@cufunc binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -y*log(ŷ + ϵ) - (1 - y)*log(1 - ŷ + ϵ)
 
 """
     logitbinarycrossentropy(ŷ, y)
@@ -147,8 +147,8 @@ julia> Flux.logitbinarycrossentropy.([-1.1491, 0.8619, 0.3127], [1, 1, 0])
 """
 logitbinarycrossentropy(ŷ, y) = (1 - y)*ŷ - logσ(ŷ)
 
-# Re-definition to fix interaction with CuArrays.
-CuArrays.@cufunc logitbinarycrossentropy(ŷ, y) = (1 - y)*ŷ - logσ(ŷ)
+# Re-definition to fix interaction with CUDA.
+CUDA.@cufunc logitbinarycrossentropy(ŷ, y) = (1 - y)*ŷ - logσ(ŷ)
 
 """
     normalise(x; dims=1)
@@ -271,7 +271,7 @@ function xlogx(x)
   result = x * log(x)
   ifelse(iszero(x), zero(result), result)
 end
-CuArrays.@cufunc function xlogx(x)
+CUDA.@cufunc function xlogx(x)
   result = x * log(x)
   ifelse(iszero(x), zero(result), result)
 end
@@ -284,7 +284,7 @@ function xlogy(x, y)
   result = x * log(y)
   ifelse(iszero(x), zero(result), result)
 end
-CuArrays.@cufunc function xlogy(x, y)
+CUDA.@cufunc function xlogy(x, y)
   result = x * log(y)
   ifelse(iszero(x), zero(result), result)
 end

src/onehot.jl

@@ -37,7 +37,7 @@ import Adapt: adapt, adapt_structure
 
 adapt_structure(T, xs::OneHotMatrix) = OneHotMatrix(xs.height, adapt(T, xs.data))
 
-import .CuArrays: CuArray, CuArrayStyle, cudaconvert
+import .CUDA: CuArray, CuArrayStyle, cudaconvert
 import Base.Broadcast: BroadcastStyle, ArrayStyle
 BroadcastStyle(::Type{<:OneHotMatrix{<:CuArray}}) = CuArrayStyle{2}()
 cudaconvert(x::OneHotMatrix{<:CuArray}) = OneHotMatrix(x.height, cudaconvert(x.data))

test/cuda/cuda.jl

@@ -1,12 +1,12 @@
 using Flux, Test
-using Flux.CuArrays
+using Flux.CUDA
 using Flux: gpu
 
 @info "Testing GPU Support"
 
-@testset "CuArrays" begin
+@testset "CUDA" begin
 
-CuArrays.allowscalar(false)
+CUDA.allowscalar(false)
 
 x = randn(5, 5)
 cx = gpu(x)
@@ -65,7 +65,7 @@ end
   @test gradient(foo, cu(rand(1)))[1] isa CuArray
 end
 
-if CuArrays.has_cudnn()
+if CUDA.has_cudnn()
   @info "Testing Flux/CUDNN"
   include("cudnn.jl")
   include("curnn.jl")

test/cuda/cudnn.jl

@@ -1,4 +1,4 @@
-using Flux, CuArrays, Test
+using Flux, CUDA, Test
 using Flux: pullback
 
 @testset "CUDNN BatchNorm" begin

test/cuda/curnn.jl

@@ -1,4 +1,4 @@
-using Flux, CuArrays, Test
+using Flux, CUDA, Test
 using Flux: pullback
 
 @testset for R in [RNN, GRU, LSTM]

test/cuda/layers.jl

@@ -35,7 +35,7 @@ function gradtest(name::String, layers::Vector, xs = nothing, args...)
 
           # Handle pooling layers
           if !isempty(ps)
-            @test gs[first(ps)] isa Flux.CuArrays.CuArray
+            @test gs[first(ps)] isa Flux.CUDA.CuArray
           end
         end
       end