Hi all,
CUDAdrv.jl <https://github.com/JuliaGPU/CUDAdrv.jl> is Julia wrapper for
the CUDA driver API -- not to be confused with its counterpart CUDArt.jl
<https://github.com/JuliaGPU/CUDArt.jl> which wraps the slightly
higher-level CUDA runtime API.
The package doesn't feature many high-level or easy-to-use wrappers, but
focuses on providing the necessary functionality for other packages to
build upon. For example, CUDArt uses CUDAdrv for launching kernels, while
CUDAnative (the in-development native programming interface) completely
relies on CUDAdrv for all GPU interactions.
It features a ccall-like cudacall interface for launching kernels and
passing values:
using CUDAdrv
using Base.Test
dev = CuDevice(0)
ctx = CuContext(dev)
md = CuModuleFile(joinpath(dirname(@__FILE__), "vadd.ptx"))
vadd = CuFunction(md, "kernel_vadd")
dims = (3,4)
a = round(rand(Float32, dims) * 100)
b = round(rand(Float32, dims) * 100)
d_a = CuArray(a)
d_b = CuArray(b)
d_c = CuArray(Float32, dims)
len = prod(dims)
cudacall(vadd, len, 1, (DevicePtr{Cfloat},DevicePtr{Cfloat},DevicePtr{Cfloat
}), d_a, d_b, d_c)
c = Array(d_c)
@test a+b ≈ c
destroy(ctx)
For documentation, refer to the NVIDIA docs
<http://docs.nvidia.com/cuda/cuda-driver-api/>. Even though they don't
fully match what CUDAdrv.jl implements, the package is well tested, and
redocumenting the entire thing is too much work.
Current master of this package only supports 0.5, but there's a tagged
version supporting 0.4 (as CUDArt.jl does so as well). It has been tested
on CUDA 5.0 up to 8.0, but there might always be issues with certain
versions (as the wrappers aren't auto-generated, and probably will never be
due to how NVIDIA has implemented cuda.h)
Anybody thinking there's a lot of overlap between CUDArt and CUDAdrv is
completely right, but it mimics the overlap between CUDA's runtime and
driver APIs as in some cases we do specifically need one or the other (eg.,
CUDAnative wouldn't work with only the runtime API). There's also some
legacy at the Julia side: CUDAdrv.jl is based on CUDA.jl, while CUDArt.jl
has been an independent effort.
In other news, we have recently *deprecated the old CUDA.jl package*. All
users should now use either CUDArt.jl or CUDAdrv.jl, depending on what
suits them best. Neither is a drop-in replacement, but the changes should
be minor. At the very least, users will have to change the kernel launch
syntax, which should use cudacall as shown above. In the future, we might
re-use the CUDA.jl package name for the native programming interface
currently at CUDAnative.jl
Best,
Tim
