Allow CPU and CUDA kernels to wait on each other

This allows the event system to be used to coordinate work done on the
host with work done on the device.

Author: lcw

src/backends/cpu.jl

@@ -41,7 +41,18 @@ end
 function __run(obj, ndrange, iterspace, args, dependencies)
     return Threads.@spawn begin
         if dependencies !== nothing
-            Base.sync_end(map(e->e.task, dependencies))
+            cpu_tasks = Core.Task[]
+            for event in dependencies
+                if event isa CPUEvent
+                    push!(cpu_tasks, event.task)
+                end
+            end
+            !isempty(cpu_tasks) && Base.sync_end(cpu_tasks)
+            for event in dependencies
+                if !(event isa CPUEvent)
+                    wait(event, ()->yield())
+                end
+            end
         end
         @sync begin
             # TODO: how do we use the information that the iteration space maps perfectly to

src/backends/cuda.jl

@@ -72,8 +72,14 @@ function (obj::Kernel{CUDA})(args...; ndrange=nothing, dependencies=nothing, wor
     stream = next_stream()
     if dependencies !== nothing
         for event in dependencies
-            @assert event isa CudaEvent
-            CUDAdrv.wait(event.event, stream)
+            if event isa CudaEvent
+                CUDAdrv.wait(event.event, stream)
+            end
+        end
+        for event in dependencies
+            if !(event isa CudaEvent)
+                wait(event, ()->yield())
+            end
         end
     end
 

test/test.jl

@@ -161,3 +161,26 @@ if has_cuda_gpu()
     wait(kernel_val!(CUDA())(A,Val(3), ndrange=size(A)))
     @test all((a)->a==3, A)
 end
+
+@kernel function kernel_empty()
+    return
+end
+if has_cuda_gpu()
+    @testset "CPU--CUDA dependencies" begin
+        event1 = kernel_empty(CPU(), 1)(ndrange=1)
+        event2 = kernel_empty(CUDA(), 1)(ndrange=1)
+        event3 = kernel_empty(CPU(), 1)(ndrange=1)
+        event4 = kernel_empty(CUDA(), 1)(ndrange=1)
+        event5 = kernel_empty(CUDA(), 1)(ndrange=1, dependencies=(event1, event2, event3, event4))
+        wait(event5)
+        @test event5 isa KernelAbstractions.Event
+
+        event1 = kernel_empty(CPU(), 1)(ndrange=1)
+        event2 = kernel_empty(CUDA(), 1)(ndrange=1)
+        event3 = kernel_empty(CPU(), 1)(ndrange=1)
+        event4 = kernel_empty(CUDA(), 1)(ndrange=1)
+        event5 = kernel_empty(CPU(), 1)(ndrange=1, dependencies=(event1, event2, event3, event4))
+        wait(event5)
+        @test event5 isa KernelAbstractions.Event
+    end
+end