Why is CUDALink failing?

Question

I have a new windows 10 laptop with a discrete nvidia graphics card (all drivers up to date). The first example from the documention of CUDAFunctionLoad[] in v11.2 fails as shown below, even though CUDAQ[] evaluates to True and OpenCLLink appears to be working fine....

Needs["CUDALink`"]
code = "
    __global__ void addTwo(mint * in, mint * out, mint length) {
        int index = threadIdx.x + blockIdx.x*blockDim.x;
        if (index < length)
            out[index] = in[index] + 2;
    }";
cudaFun = CUDAFunctionLoad[code, 
  "addTwo", {{_Integer, _, "Input"}, {_Integer, _, 
    "Output"}, _Integer}, 256, "ShellOutputFunction" -> Print]

Here's the output: (note the message nvcc fatal: Host compiler targets unsupported OS.)

In[3]:= cudaFun = 
     CUDAFunctionLoad[code, 
      "addTwo", {{_Integer, _, "Input"}, {_Integer, _, 
        "Output"}, _Integer}, 256, "ShellOutputFunction" -> Print, 
      "ShellCommandFunction" :> Print]

    During evaluation of In[3]:= CUDAFunctionLoad::cmpf: The kernel compilation failed. Consider setting the option "ShellOutputFunction"->Print to display the compiler error message.

    During evaluation of In[3]:= call "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvarsall.bat" amd64
    "C:\Users\micha\AppData\Roaming\Mathematica\Paclets\Repository\CUDAResources-Win64-11.2.22\CUDAToolkit\bin\nvcc.exe" -cubin -m64 -arch=sm_61 -DUSING_CUDA_FUNCTION=1 -Dmint="long long" -DReal_t=double -DUSING_DOUBLE_PRECISIONQ=1 -o "C:\Users\micha\AppData\Roaming\Mathematica\ApplicationData\CUDALink\BuildFolder\sbp-14988\Working-sbp-14988-15020-1\CUDAFunction-1595.cubin" "C:\Users\micha\AppData\Roaming\Mathematica\ApplicationData\CUDALink\BuildFolder\sbp-14988\Working-sbp-14988-15020-1\CUDAFunction-1595.cu"

During evaluation of In[3]:= 
 C:\Users\micha\AppData\Roaming\Mathematica\ApplicationData\CUDALink\BuildFolder\sbp-14988\Working-sbp-14988-15020-1>call "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvarsall.bat" amd64 
    **********************************************************************
    ** Visual Studio 2017 Developer Command Prompt v15.5.2
    ** Copyright (c) 2017 Microsoft Corporation
    **********************************************************************
    [vcvarsall.bat] Environment initialized for: 'x64'
    Microsoft (R) C/C++ Optimizing Compiler Version 19.12.25831 for x64
    Copyright (C) Microsoft Corporation.  All rights reserved.

    tmpxft_00000f74_00000000-1.cpp
    **nvcc fatal   : Host compiler targets unsupported OS.**

    Out[3]= CUDAFunctionLoad["
     __global__ void addTwo(mint * in, mint * out, mint length) {
        int index = threadIdx.x + blockIdx.x*blockDim.x;
        if (index < length)
            out[index] = in[index] + 2;
     }", "addTwo", {{_Integer, _, "Input"}, {_Integer, _, 
       "Output"}, _Integer}, 256, "ShellOutputFunction" -> Print, 
     "ShellCommandFunction" :> Print]

Here's the result of my CUDAResourcesInformation[]:

{{"Name" -> "CUDAResources", "Version" -> "11.2.22", 
  "BuildNumber" -> "", "Qualifier" -> "Win64", 
  "WolframVersion" -> "11.2", "SystemID" -> {"Windows-x86-64"}, 
  "Description" -> "{ToolkitVersion -> v8.0, MinimumDriver -> 290}", 
  "Category" -> "", "Creator" -> "", "Publisher" -> "", 
  "Support" -> "", "Internal" -> False, 
  "Location" -> 
   "C:\\Users\\micha\\AppData\\Roaming\\Mathematica\\Paclets\\\
Repository\\CUDAResources-Win64-11.2.22", "Context" -> {}, 
  "Enabled" -> True, "Loading" -> Manual, 
  "Hash" -> "9c9b3bf9dfc07e0cc2376f0ef13e01d5"}}

Machine Details: I'm on Windows 10 Pro v1703 with an Nvidia GeForce 2016 card, Cuda 9.1 Toolkit, and Visual Studio installed.

References:

• https://stackoverflow.com/questions/43745099/using-cuda-with-visual-studio-2017
• https://stackoverflow.com/questions/42003849/nvcc-host-compiler-targets-unsupported-os

Answer 1

This should be working better with Visual Studio 2017 and recent Nvidia cards after updating to the latest CUDAResources for version 11.3:

PacletSiteUpdate /@ PacletSites[];

Needs["CUDALink`"]

CUDAResourcesInstall[Update -> True]

(* {Paclet[CUDAResources, 11.3.38, <>]} *)

Answer 2

The situation is that you (and me) have a card that is too new. You can try the following: Set the option "ShellCommandFunction" :> Print and run again. There, you should find the location of the nvcc, the cuda compiler. You can call this compiler on a comandline with --help. As usually a terminal is considered the evil gate to hell on a Windoze machine, here is the output:

--gpu-architecture (-arch)
Specify the name of the class of NVIDIA 'virtual' GPU architecture for which the CUDA input files must be compiled. [...] Allowed values for this option: compute_20', 'compute_30', 'compute_32', 'compute_35', 'compute_37', 'compute_50', 'compute_52', 'compute_53', 'sm_20', 'sm_21', 'sm_30', 'sm_32', 'sm_35', 'sm_37', 'sm_50', 'sm_52', 'sm_53'.

The version of the CUDA Toolkit is on my machine 7.5, while currently version 9.1 is available. I don't bother that much with CUDA from within Mathematica these days, but this is the reason.

Fix

I tested it on Linux with Mathematica Version 11.2. First, you need to run

CUDAResourcesInstall[Update -> True]

Then, there should be a new paclet in your paclet-directory. For me, this is in my home directory under

~/.Mathematica/Paclets/Repository/CUDAResources-Lin64-11.2.63

Restart your Mathematica and after that, you should be able to successfully build the CUDA kernel.