update to CUDA 3.0 and use task-local stream

Author: vchuravy

lib/CUDAKernels/Project.toml

@@ -1,7 +1,7 @@
 name = "CUDAKernels"
 uuid = "72cfdca4-0801-4ab0-bf6a-d52aa10adc57"
 authors = ["Valentin Churavy <[email protected]>"]
-version = "0.1.0"
+version = "0.2.0"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
@@ -12,10 +12,10 @@ 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"
+Adapt = "3.0"
+CUDA = "3.0"
 Cassette = "0.3.3"
 KernelAbstractions = "0.6"
 SpecialFunctions = "0.10, 1.0"
 StaticArrays = "0.12, 1.0"
-julia = "1.3"
+julia = "1.6"

lib/CUDAKernels/src/CUDAKernels.jl

@@ -65,7 +65,7 @@ failed(::CudaEvent) = false
 isdone(ev::CudaEvent) = CUDA.query(ev.event)
 
 function Event(::CUDADevice)
-    stream = CUDA.CuDefaultStream()
+    stream = CUDA.stream()
     event = CUDA.CuEvent(CUDA.EVENT_DISABLE_TIMING)
     CUDA.record(event, stream)
     CudaEvent(event)
@@ -85,11 +85,11 @@ function wait(::CPU, ev::CudaEvent, progress=yield)
     end
 end
 
-# Use this to synchronize between computation using the CuDefaultStream
-wait(::CUDADevice, ev::CudaEvent, progress=nothing, stream=CUDA.CuDefaultStream()) = CUDA.wait(ev.event, stream)
+# Use this to synchronize between computation using the task local stream
+wait(::CUDADevice, ev::CudaEvent, progress=nothing, stream=CUDA.stream()) = CUDA.wait(ev.event, stream)
 wait(::CUDADevice, ev::NoneEvent, progress=nothing, stream=nothing) = nothing
 
-function wait(::CUDADevice, ev::MultiEvent, progress=nothing, stream=CUDA.CuDefaultStream())
+function wait(::CUDADevice, ev::MultiEvent, progress=nothing, stream=CUDA.stream())
     dependencies = collect(ev.events)
     cudadeps  = filter(d->d isa CudaEvent,    dependencies)
     otherdeps = filter(d->!(d isa CudaEvent), dependencies)