Try #200:

.buildkite/pipeline.yml

@@ -3,9 +3,14 @@ steps:
     plugins:
       - JuliaCI/julia#v1:
           version: "1"
-      - JuliaCI/julia-test#v1:
       - JuliaCI/julia-coverage#v1:
           codecov: true
+    commands:
+      - julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd()))'
+      - julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))'
+      - julia -e 'using Pkg; Pkg.add(name="Run", version="0.1")'
+      - julia -e 'using Run; Run.prepare_test()'
+      - julia -e 'using Run; Run.test()'
     agents:
       queue: "juliagpu"
       cuda: "*"
@@ -15,9 +20,14 @@ steps:
     plugins:
       - JuliaCI/julia#v1:
           version: "1.6-nightly"
-      - JuliaCI/julia-test#v1:
       - JuliaCI/julia-coverage#v1:
           codecov: true
+    commands:
+      - julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd()))'
+      - julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))'
+      - julia -e 'using Pkg; Pkg.add(name="Run", version="0.1")'
+      - julia -e 'using Run; Run.prepare_test()'
+      - julia -e 'using Run; Run.test()'
     agents:
       queue: "juliagpu"
       cuda: "*"
@@ -27,9 +37,14 @@ steps:
     plugins:
       - JuliaCI/julia#v1:
           version: "nightly"
-      - JuliaCI/julia-test#v1:
       - JuliaCI/julia-coverage#v1:
           codecov: true
+    commands:
+      - julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd()))'
+      - julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))'
+      - julia -e 'using Pkg; Pkg.add(name="Run", version="0.1")'
+      - julia -e 'using Run; Run.prepare_test()'
+      - julia -e 'using Run; Run.test()'
     agents:
       queue: "juliagpu"
       cuda: "*"

.github/workflows/ci-julia-1.6-nightly.yml

@@ -37,8 +37,11 @@ jobs:
             ${{ runner.os }}-test-${{ env.cache-name }}-
             ${{ runner.os }}-test-
             ${{ runner.os }}-
-      - uses: julia-actions/julia-buildpkg@v1
-      - uses: julia-actions/julia-runtest@v1
+      - run: julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd()))'
+      - run: julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))'
+      - run: julia -e 'using Pkg; pkg"add Run@0.1"'
+      - run: julia -e 'using Run; Run.prepare_test()'
+      - run: julia -e 'using Run; Run.test()'
       - uses: julia-actions/julia-processcoverage@v1
       - uses: codecov/codecov-action@v1
         with:

.github/workflows/ci-julia-nightly.yml

@@ -37,8 +37,11 @@ jobs:
             ${{ runner.os }}-test-${{ env.cache-name }}-
             ${{ runner.os }}-test-
             ${{ runner.os }}-
-      - uses: julia-actions/julia-buildpkg@v1
-      - uses: julia-actions/julia-runtest@v1
+      - run: julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd()))'
+      - run: julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))'
+      - run: julia -e 'using Pkg; pkg"add Run@0.1"'
+      - run: julia -e 'using Run; Run.prepare_test()'
+      - run: julia -e 'using Run; Run.test()'
       - uses: julia-actions/julia-processcoverage@v1
       - uses: codecov/codecov-action@v1
         with:

.github/workflows/ci.yml

@@ -40,8 +40,11 @@ jobs:
             ${{ runner.os }}-test-${{ env.cache-name }}-
             ${{ runner.os }}-test-
             ${{ runner.os }}-
-      - uses: julia-actions/julia-buildpkg@v1
-      - uses: julia-actions/julia-runtest@v1
+      - run: julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd()))'
+      - run: julia --project=test -e 'using Pkg; Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))'
+      - run: julia -e 'using Pkg; pkg"add Run@0.1"'
+      - run: julia -e 'using Run; Run.prepare_test()'
+      - run: julia -e 'using Run; Run.test()'
       - uses: julia-actions/julia-processcoverage@v1
       - uses: codecov/codecov-action@v1
         with:
@@ -62,6 +65,7 @@ jobs:
           julia --project=docs -e '
             using Pkg
             Pkg.develop(PackageSpec(path=pwd()))
+            Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))
             Pkg.instantiate()'
       - run: julia --project=docs docs/make.jl
         env:
@@ -83,6 +87,7 @@ jobs:
           julia --project=docs -e '
             using Pkg
             Pkg.develop(PackageSpec(path=pwd()))
+            Pkg.develop(PackageSpec(path=joinpath(pwd(),"lib","CUDAKernels")))
             Pkg.instantiate()'
       - run: |
           julia --project=docs -e '

Project.toml

@@ -1,11 +1,10 @@
 name = "KernelAbstractions"
 uuid = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 authors = ["Valentin Churavy <v.churavy@gmail.com>"]
-version = "0.5.3"
+version = "0.6.0"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 Cassette = "7057c7e9-c182-5462-911a-8362d720325c"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
@@ -15,7 +14,6 @@ UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 
 [compat]
 Adapt = "0.4, 1.0, 2.0, 3.0"
-CUDA = "~1.0, ~1.1, ~1.2, 1.3, 2"
 Cassette = "0.3.3"
 MacroTools = "0.5"
 SpecialFunctions = "0.10, 1.0"

examples/matmul.jl

@@ -1,4 +1,4 @@
-using KernelAbstractions, Test, CUDA
+using KernelAbstractions, CUDAKernels, Test, CUDA
 
 if has_cuda_gpu()
     CUDA.allowscalar(false)

examples/memcopy.jl

@@ -1,4 +1,5 @@
 using KernelAbstractions
+using CUDAKernels
 using CUDA
 using Test
 

examples/memcopy_static.jl

@@ -1,4 +1,5 @@
 using KernelAbstractions
+using CUDAKernels
 using CUDA
 using Test
 

examples/mpi.jl

@@ -1,6 +1,7 @@
 # EXCLUDE FROM TESTING
 using KernelAbstractions
-using  CUDA
+using CUDAKernels
+using CUDA
 
 if has_cuda_gpu()
     CUDA.allowscalar(false)

examples/naive_transpose.jl

@@ -1,4 +1,4 @@
-using KernelAbstractions, Test, CUDA
+using KernelAbstractions, CUDAKernels, Test, CUDA
 
 if has_cuda_gpu()
     CUDA.allowscalar(false)

examples/performance.jl

@@ -1,4 +1,4 @@
-using KernelAbstractions, CUDA, Test
+using KernelAbstractions, CUDAKernels, CUDA, Test
 using KernelAbstractions.Extras: @unroll
 
 has_cuda_gpu() || exit()
@@ -199,4 +199,4 @@ for (name, kernel) in (
             end
         end
     end
-end
+end

lib/CUDAKernels/Project.toml

@@ -0,0 +1,20 @@
+name = "CUDAKernels"
+uuid = "72cfdca4-0801-4ab0-bf6a-d52aa10adc57"
+authors = ["Valentin Churavy <v.churavy@gmail.com>"]
+version = "0.1.0"
+
+[deps]
+Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+Cassette = "7057c7e9-c182-5462-911a-8362d720325c"
+KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
+SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
+StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
+
+[compat]
+Adapt = "0.4, 1.0, 2.0, 3.0"
+CUDA = "~1.0, ~1.1, ~1.2, 1.3, 2"
+Cassette = "0.3.3"
+KernelAbstractions = "0.6"
+SpecialFunctions = "0.10, 1.0"
+StaticArrays = "0.12, 1.0"

src/backends/cuda.jl → lib/CUDAKernels/src/CUDAKernels.jl

@@ -1,5 +1,14 @@
+module CUDAKernels
+
 import CUDA
 import SpecialFunctions
+import StaticArrays
+import StaticArrays: MArray
+import Cassette
+import Adapt
+import KernelAbstractions
+
+export CUDADevice
 
 const FREE_STREAMS = CUDA.CuStream[]
 const STREAMS = CUDA.CuStream[]
@@ -44,6 +53,10 @@ function next_stream()
     end
 end
 
+import KernelAbstractions: Event, CPUEvent, NoneEvent, MultiEvent, CPU, GPU, isdone, failed
+
+struct CUDADevice <: GPU end
+
 struct CudaEvent <: Event
     event::CUDA.CuEvent
 end
@@ -58,6 +71,8 @@ function Event(::CUDADevice)
     CudaEvent(event)
 end
 
+import Base: wait
+
 wait(ev::CudaEvent, progress=yield) = wait(CPU(), ev, progress)
 
 function wait(::CPU, ev::CudaEvent, progress=yield)
@@ -113,7 +128,7 @@ function __pin!(a)
     return nothing
 end
 
-function async_copy!(::CUDADevice, A, B; dependencies=nothing, progress=yield)
+function KernelAbstractions.async_copy!(::CUDADevice, A, B; dependencies=nothing, progress=yield)
     A isa Array && __pin!(A)
     B isa Array && __pin!(B)
 
@@ -131,7 +146,7 @@ function async_copy!(::CUDADevice, A, B; dependencies=nothing, progress=yield)
     return CudaEvent(event)
 end
 
-
+import KernelAbstractions: Kernel, StaticSize, DynamicSize, partition, blocks, workitems, launch_config
 
 ###
 # Kernel launch
@@ -186,7 +201,7 @@ function (obj::Kernel{CUDADevice})(args...; ndrange=nothing, dependencies=nothin
 
     # If the kernel is statically sized we can tell the compiler about that
     if KernelAbstractions.workgroupsize(obj) <: StaticSize
-        maxthreads = prod(get(KernelAbstractions.workgroupsize(obj)))
+        maxthreads = prod(KernelAbstractions.get(KernelAbstractions.workgroupsize(obj)))
     else
         maxthreads = nothing
     end
@@ -211,8 +226,12 @@ end
 
 Cassette.@context CUDACtx
 
+import KernelAbstractions: CompilerMetadata, CompilerPass, DynamicCheck, LinearIndices
+import KernelAbstractions: __index_Local_Linear, __index_Group_Linear, __index_Global_Linear, __index_Local_Cartesian, __index_Group_Cartesian, __index_Global_Cartesian, __validindex, __print
+import KernelAbstractions: mkcontext, expand, __iterspace, __ndrange, __dynamic_checkbounds
+
 function mkcontext(kernel::Kernel{CUDADevice}, _ndrange, iterspace)
-    metadata = CompilerMetadata{ndrange(kernel), DynamicCheck}(_ndrange, iterspace)
+    metadata = CompilerMetadata{KernelAbstractions.ndrange(kernel), DynamicCheck}(_ndrange, iterspace)
     Cassette.disablehooks(CUDACtx(pass = CompilerPass, metadata=metadata))
 end
 
@@ -251,7 +270,9 @@ end
     end
 end
 
-generate_overdubs(CUDACtx)
+import KernelAbstractions: groupsize, __groupsize, __workitems_iterspace, add_float_contract, sub_float_contract, mul_float_contract
+
+KernelAbstractions.generate_overdubs(@__MODULE__, CUDACtx)
 
 ###
 # CUDA specific method rewrites
@@ -311,9 +332,12 @@ else
     const emit_shmem = CUDA._shmem
 end
 
+import KernelAbstractions: ConstAdaptor, SharedMemory, Scratchpad, __synchronize, __size
+
 ###
 # GPU implementation of shared memory
 ###
+
 @inline function Cassette.overdub(ctx::CUDACtx, ::typeof(SharedMemory), ::Type{T}, ::Val{Dims}, ::Val{Id}) where {T, Dims, Id}
     ptr = emit_shmem(Val(Id), T, Val(prod(Dims)))
     CUDA.CuDeviceArray(Dims, ptr)
@@ -341,3 +365,8 @@ end
 ###
 
 Adapt.adapt_storage(to::ConstAdaptor, a::CUDA.CuDeviceArray) = Base.Experimental.Const(a)
+
+# Argument conversion
+KernelAbstractions.argconvert(k::Kernel{CUDADevice}, arg) = CUDA.cudaconvert(arg)
+
+end

src/KernelAbstractions.jl

@@ -2,7 +2,7 @@ module KernelAbstractions
 
 export @kernel
 export @Const, @localmem, @private, @uniform, @synchronize, @index, groupsize, @print
-export Device, GPU, CPU, CUDADevice, Event, MultiEvent, NoneEvent
+export Device, GPU, CPU, Event, MultiEvent, NoneEvent
 export async_copy!
 
 
@@ -330,9 +330,6 @@ abstract type Device end
 abstract type GPU <: Device end
 
 struct CPU <: Device end
-struct CUDADevice <: GPU end
-# struct AMD <: GPU end
-# struct Intel <: GPU end
 
 include("nditeration.jl")
 using .NDIteration
@@ -462,17 +459,10 @@ end
     end
 end
 
-###
-# Backends/Implementation
-###
-
 # Utils
 __size(args::Tuple) = Tuple{args...}
 __size(i::Int) = Tuple{i}
 
-include("backends/cpu.jl")
-include("backends/cuda.jl")
-
 ###
 # Extras
 # - LoopInfo
@@ -481,4 +471,9 @@ include("backends/cuda.jl")
 include("extras/extras.jl")
 
 include("reflection.jl")
+
+# CPU backend
+
+include("cpu.jl")
+
 end #module

src/compiler.jl

@@ -30,8 +30,8 @@ end
 include("compiler/contract.jl")
 include("compiler/pass.jl")
 
-function generate_overdubs(Ctx)
-   @eval begin
+function generate_overdubs(mod, Ctx)
+   @eval mod begin
         @inline Cassette.overdub(ctx::$Ctx, ::typeof(groupsize)) = __groupsize(ctx.metadata)
         @inline Cassette.overdub(ctx::$Ctx, ::typeof(__workitems_iterspace)) = workitems(__iterspace(ctx.metadata))
 

src/backends/cpu.jl → src/cpu.jl

@@ -208,7 +208,7 @@ end
     __print(items...)
 end
 
-generate_overdubs(CPUCtx)
+generate_overdubs(@__MODULE__, CPUCtx)
 
 # Don't recurse into these functions
 const cpufuns = (:cos, :cospi, :sin, :sinpi, :tan,
@@ -263,3 +263,6 @@ end
 @inline function Cassette.overdub(ctx::CPUCtx, ::typeof(Base.getindex), A::ScratchArray{N}, idx) where N
     return @inbounds aview(A.data, ntuple(_->:, Val(N))..., idx)
 end
+
+# Argument conversion
+KernelAbstractions.argconvert(k::Kernel{CPU}, arg) = arg