Use a default, global context for the first device.

docs/src/man/usage.md

@@ -24,10 +24,6 @@ Using the `@cuda` macro, you can launch the kernel on a GPU of your choice:
 using CUDAdrv, CUDAnative
 using Base.Test
 
-# CUDAdrv functionality: select device, create context
-dev = CuDevice(0)
-ctx = CuContext(dev)
-
 # CUDAdrv functionality: generate and upload data
 a = round.(rand(Float32, (3, 4)) * 100)
 b = round.(rand(Float32, (3, 4)) * 100)
@@ -44,8 +40,19 @@ d_c = similar(d_a) # output array
 c = Array(d_c)
 
 @test a+b ≈ c
+```
+
+This code is executed in a default, global context for the first device in your system. The
+compiler queries the context through `CuCurrentContext`, which implies you can easily switch
+contexts (using a different device, or supplying different flags) by activating a different
+one:
 
-destroy(ctx)
+```julia
+dev = CuDevice(0)
+CuContext(dev) do ctx
+ # allocate things in this context
+ @cuda ...
+end
 ```
 
 

examples/hello_world.jl

@@ -5,11 +5,5 @@ function hello_world()
  return
 end
 
-dev = CuDevice(0)
-ctx = CuContext(dev)
-
 @cuda (2,2) hello_world()
-
 synchronize()
-
-destroy!(ctx)

examples/oob.jl

@@ -8,9 +8,6 @@
 
 using CUDAdrv, CUDAnative
 
-dev = CuDevice(0)
-ctx = CuContext(dev)
-
 a = CuArray{Float32}(10)
 
 function memset(a, val)
@@ -20,7 +17,4 @@ function memset(a, val)
 end
 
 @cuda (1,11) memset(a, 0f0)
-
 synchronize()
-
-destroy!(ctx)

examples/pairwise.jl

@@ -108,15 +108,10 @@ function pairwise_dist_gpu(lat::Vector{Float32}, lon::Vector{Float32})
 end
 
 
-const dev = CuDevice(0)
-const ctx = CuContext(dev)
-
 # generate reasonable data
 const n = 10000
 const lat = rand(Float32, n) .* 45
 const lon = rand(Float32, n) .* -120
 
 using Base.Test
 @test pairwise_dist_cpu(lat, lon) ≈ pairwise_dist_gpu(lat, lon)
-
-destroy!(ctx)

examples/reduce/benchmark.jl

@@ -4,7 +4,8 @@
 using BenchmarkTools
 include("reduce.jl")
 
-dev = CuDevice(0)
+ctx = CuCurrentContext()
+dev = device(ctx)
 @assert(capability(dev) >= v"3.0", "this example requires a newer GPU")
 
 len = 10^7
@@ -21,7 +22,6 @@ open(joinpath(@__DIR__, "reduce.jl.ptx"), "w") do f
  cap=v"6.1.0")
 end
 
-ctx = CuContext(dev)
 benchmark_gpu = @benchmarkable begin
  gpu_reduce(+, gpu_input, gpu_output)
  val = Array(gpu_output)[1]
@@ -35,7 +35,6 @@ benchmark_gpu = @benchmarkable begin
  gc()
  )
 println(run(benchmark_gpu))
-destroy!(ctx)
 
 
 ## CUDA

examples/reduce/verify.jl

@@ -1,7 +1,8 @@
 using Base.Test
 include("reduce.jl")
 
-dev = CuDevice(0)
+ctx = CuCurrentContext()
+dev = device(ctx)
 if capability(dev) < v"3.0"
  warn("this example requires a newer GPU")
  exit(0)
@@ -15,11 +16,9 @@ cpu_val = reduce(+, input)
 
 # CUDAnative
 let
- ctx = CuContext(dev)
  gpu_input = CuArray(input)
  gpu_output = similar(gpu_input)
  gpu_reduce(+, gpu_input, gpu_output)
  gpu_val = Array(gpu_output)[1]
- destroy!(ctx)
  @assert cpu_val == gpu_val
 end

examples/scan.jl

@@ -58,9 +58,6 @@ function gpu_accumulate!(op::Function, data::CuDeviceMatrix{T}) where {T}
  return
 end
 
-dev = CuDevice(0)
-ctx = CuContext(dev)
-
 rows = 5
 cols = 4
 
@@ -75,8 +72,6 @@ gpu_a = CuArray(a)
 using Base.Test
 @test cpu_a ≈ Array(gpu_a)
 
-destroy!(ctx)
-
 
 # FURTHER IMPROVEMENTS:
 # - work efficiency

examples/vadd.jl

@@ -8,9 +8,6 @@ function kernel_vadd(a, b, c)
  return nothing
 end
 
-dev = CuDevice(0)
-ctx = CuContext(dev)
-
 dims = (3,4)
 a = round.(rand(Float32, dims) * 100)
 b = round.(rand(Float32, dims) * 100)
@@ -23,5 +20,3 @@ len = prod(dims)
 @cuda (1,len) kernel_vadd(d_a, d_b, d_c)
 c = Array(d_c)
 @test a+b ≈ c
-
-destroy!(ctx)

src/CUDAnative.jl

@@ -25,6 +25,8 @@ include(joinpath("device", "libdevice.jl")) # so should get loaded late (JuliaL
 include("execution.jl")
 include("reflection.jl")
 
+const default_device = Ref{CuDevice}()
+const default_context = Ref{CuContext}()
 function __init__()
  if !configured
  warn("CUDAnative.jl has not been configured, and will not work properly.")
@@ -37,6 +39,16 @@ function __init__()
  VersionNumber(Base.libllvm_version) != julia_llvm_version
  error("Your set-up has changed. Please run Pkg.build(\"CUDAnative\") and restart Julia.")
  end
+
+ # instantiate a default device and context;
+ # this will be implicitly used through `CuCurrentContext`
+ # NOTE: although these conceptually match what the primary context is for,
+ # we don't use that because it is refcounted separately
+ # and might confuse / be confused by user operations
+ # (eg. calling `unsafe_reset!` on a primary context)
+ default_device[] = CuDevice(0)
+ default_context[] = CuContext(default_device[])
+
  init_jit()
 end
 

test/perf/launch_overhead/cuda.jl

@@ -10,9 +10,6 @@ const ITERATIONS = 5000
 # TODO: api-trace shows some attribute fetches, where do they come from?
 
 function main() 
- dev = CuDevice(0)
- ctx = CuContext(dev)
-
  mod = CuModuleFile("cuda.ptx")
  fun = CuFunction(mod, "kernel_dummy")
 
@@ -39,8 +36,6 @@ function main()
 
  @printf("CPU time: %.2fus\n", median(cpu_time))
  @printf("GPU time: %.2fus\n", median(gpu_time))
-
- destroy!(ctx)
 end
 
 main()

test/perf/launch_overhead/cudanative.jl

@@ -13,9 +13,6 @@ const len = 1000
 const ITERATIONS = 5000
 
 function main() 
- dev = CuDevice(0)
- ctx = CuContext(dev)
-
  cpu_time = Vector{Float64}(ITERATIONS)
  gpu_time = Vector{Float64}(ITERATIONS)
 
@@ -38,8 +35,6 @@ function main()
 
  @printf("CPU time: %.2fus\n", median(cpu_time))
  @printf("GPU time: %.2fus\n", median(gpu_time))
-
- destroy!(ctx)
 end
 
 main()

test/perf/launch_overhead/cudanative_profile.jl

@@ -14,9 +14,6 @@ const len = 1000
 const ITERATIONS = 5000
 
 function main() 
- dev = CuDevice(0)
- ctx = CuContext(dev)
-
  cpu_time = Vector{Float64}(ITERATIONS)
 
  gpu_arr = CuArray{Float32}(len)
@@ -35,8 +32,6 @@ function main()
 
  @printf("CPU time: %.2fus\n", median(cpu_time))
  CUDAnative.Profile.print()
-
- destroy!(ctx)
 end
 
 main()