Windows support about cudaapi.jl HOT 9 CLOSED

This can be related:

I am working on windows support for Knet.jl. I tried to build CUDAdrv.jl on windows and it uses discovery.jl of CUDAapi.jl . I realized that names in this line should include nvcuda but it is not there. I fixed it by manually adding it. If you can guide me, I can submit a PR for this.

from cudaapi.jl.

No, nvcuda was already being searched for:

The problem was more complicated, see #10

The only thing I haven't been able to fix, is cl.exe discovery. Even after (instructing the user to) source vcvarsall.bat, I can't manage to deduce the location of cl.exe from the environment variables. The heuristics @musm has added don't seem to work on Community 2017 anymore.
I'm thinking of just having the user put cl.exe in his PATH, and be done with it. If there's no easy way to discover cl.exe, we shouldn't be trying to:

from cudaapi.jl.

https://github.com/Microsoft/vswhere

is needed for newer vc studio installations, which is in included in the 2017 visual studio installation, and can be used to find CL.exe

CUDA 9 is required for VC2017 support

from cudaapi.jl.

Also see JuliaAttic/CUDArt.jl#70

from cudaapi.jl.

Right, but how do I get the actual path to cl.exe from that. Sure, it'll help to find or activate a prompt with the VS vars sourced, but even in that prompt cl.exe doesn't seem discoverable.

And yes, CUDAapi needs a compatibility database for matching VS versions to CUDA versions.

from cudaapi.jl.

vsdevcmd.bat has cl sourced (I just tried it)

https://github.com/Microsoft/vswhere/wiki/Find-VC

from cudaapi.jl.

Mine doesn't, I must be missing a part of VS then.

Can you confirm this picks up the path to cl.exe?

vswhere = download("https://github.com/Microsoft/vswhere/releases/download/2.2.11/vswhere.exe")

vs_install_path = chomp(readstring(`$vswhere  -latest -property installationPath`))
if !isdir(vs_install_path)
    error("Cannot find a proper Visual Studio installation. Make sure Visual Studio is installed.")
end

vs_version_major = parse(split(chomp(readstring(`$vswhere  -latest -property installationVersion`)), '.')[1])
if vs_version_major >= 15
    vs_cmd_prompt = joinpath(vs_install_path, "VC", "Auxiliary", "Build", "vcvarsall.bat")
else
    vs_cmd_prompt = joinpath(vs_install_path, "VC", "vcvarsall.bat")
end

# check whether 32 or 64 bit archtecture
# NOTE: Actually, nvcc in x86 visual studio command prompt doesn't make 32-bit binary
#       It depends on whether CUDA toolkit is 32bit or 64bit
if Int == Int64
    arch = "amd64"
else
    arch = "x86"
end

run(`cmd /C "$vs_cmd_prompt" $arch \& where cl.exe`)

EDIT: yeah seems to work here after installing a boatload of VC++ tools. Figured I only needed the compiler, apparently not.

from cudaapi.jl.

@quaertym: the changes to CUDAapi in #10 should fix Windows support. What remains for Knet, from a quick first glance, is that the Knet package build scripts calls out to make instead of building directly. But you can pretty easily recreate that in Julia, see e.g.: https://github.com/JuliaGPU/CUDArt.jl/blob/51ada407672d779ee11e69ad3cbd2642364b1330/deps/build.jl#L58-L77
(but don't use the logging_run because I'll be removing that from CUDAapi)

from cudaapi.jl.

@maleadt Thanks. My problem was getting nvcuda.dll without versions at the end. #10 seems like a fix.

from cudaapi.jl.