Add 'bare' interface of CUDA event

- src/event.jl: function crate_event, destroy, record, elapsed_time
- examples/performance.jl: example to check performance of kernel with CUDA event

examples/performance.jl

@@ -0,0 +1,76 @@
+# An example to benchmark with/without CUDA streaming
+
+using CUDA
+
+# array length
+len = 1024 * 1024
+
+# thread counts in thread block
+block = 256
+
+try
+	# create context
+	dev = CuDevice(0)
+	ctx = create_context(dev)
+
+	# load module from vadd.ptx
+	md = CuModule("vadd.ptx")
+
+	# get function vadd
+	f = CuFunction(md, "vadd")
+
+	# prepare arrays to compute
+	a = round(rand(Float32, len) * 100)
+	b = round(rand(Float32, len) * 100)
+
+	# create array c on GPU
+	gc = CuArray(Float32, len)
+
+	# crate event to check performance
+	e1 = create_event()
+	e2 = create_event()
+
+	# start event
+	record(e1)
+
+	# load arrays to GPU
+	ga = CuArray(a)
+	gb = CuArray(b)
+
+	# launch kernel
+	launch(f, div(len, block), block, (ga, gb, gc))
+
+	# fetch results from GPU
+	c = to_host(gc)
+
+	# end event
+	record(e2)
+
+	# elapsed time: milliseconds
+	println("Elapsed time: $(elapsed_time(e1, e2)) ms")
+
+	# destroy events
+	destroy(e1)
+	destroy(e2)
+
+	# free GPU memory
+	free(ga)
+	free(gb)
+	free(gc)
+
+	# check result
+	println("c == (a + b) ? $(c == (a + b))")
+
+	# unload module
+	unload(md)
+
+	# destroy context
+	destroy(ctx)
+
+catch err
+	if isa(err, CuDriverError)
+		println("$err: $(description(err))")
+	else
+		throw(err)
+	end
+end

src/CUDA.jl

@@ -24,6 +24,9 @@ module CUDA
     # stream
     CuStream, synchronize,
 
+    # event
+    CuEvent, create_event, destroy, record, elapsed_time, 
+
     # execution
     launch,
 
@@ -39,6 +42,7 @@ module CUDA
     include("context.jl")
     include("module.jl")
     include("stream.jl")
+    include("event.jl") 
     include("execution.jl")
 
     include("arrays.jl")

src/event.jl

@@ -0,0 +1,35 @@
+# CUDA event
+
+immutable CuEvent
+	handle::Ptr{Void}
+end
+
+const CU_EVENT_DEFAULT        = 0x0 # Default event flag
+const CU_EVENT_BLOCKING_SYNC  = 0x1 # Event uses blocking synchronization
+const CU_EVENT_DISABLE_TIMING = 0x2 # Event will not record timing data
+const CU_EVENT_INTERPROCESS   = 0x4 # Event is suitable for interprocess use. CU_EVENT_DISABLE_TIMING must be set
+
+function create_event(flags::Integer = CU_EVENT_DEFAULT)
+	ref = Array(Ptr{Void}, 1)
+	@cucall(:cuEventCreate, (Ptr{Ptr{Void}}, Cuint), ref, flags)
+	return CuEvent(ref[1])
+end
+
+function destroy(event::CuEvent)
+	@cucall(:cuEventDestroy, (Ptr{Void},), event.handle)
+end
+
+function record(event::CuEvent)
+	# TODO: support stream
+	@cucall(:cuEventRecord, 
+		(Ptr{Void}, Ptr{Void}), 
+		event.handle, 0)
+end
+
+function elapsed_time(start_event::CuEvent, end_event::CuEvent)
+	elapsed = Array(Float32, 1)
+	@cucall(:cuEventElapsedTime, 
+		(Ptr{Float32}, Ptr{Void}, Ptr{Void}), 
+		elapsed, start_event.handle, end_event.handle)
+	return elapsed[1]
+end