How to configure nvidia CUDA for VIsual Studio 2017 [duplicate] - visual-studio

Visual Studio 2017 RC includes much tighter CMake integration, allowing one to skip the intermediate step of generating project/solution files and use CMake effectively as the project file itself. There is sufficient documentation from Microsoft for using these features with regular C++ files, and there is sufficient documentation on this website (example) for making CUDA and Cmake play nicely, when it comes to linking CUDA code to C++ code.
What I can't find information on is how to make CMake, Visual Studio 2017 RC, and CUDA 8.0 all play nicely. This is a difficult problem, because 2017RC has no integration for the CUDA SDK anyways, and I was hoping to use 2017RC so that my C++ interface to the CUDA code could use C++14 and/or C++17. I'm working on the beginning of a large project that will primarily involve writing a static CUDA library that is accessed through C++: so, I'd like to get the CMake to take care of compiling my CUDA sources into a static library, and for it to help with feeding the linking information to Visual Studio. So far, I haven't had any success with using FindCUDA's various features to accomplish this, but I'm assuming that's due to a misunderstanding on my part. I've read through the documentation on separable compilation from Nvidia, but that wasn't helpful for figuring out CMake.
Further, whenever I try to use CMake in VS2017RC, I still end up with the various vcxproj files that CMake likes to spit out. Is this due to an error on my part? How do I edit the build command arguments, or CMakeLists.txt, to get the functionality demonstrated here to work?

The very short (and only at the time of writing) answer is that you can't. CUDA 8 doesn't support VS2017. Only VS2015 is presently supported.
You can always find the compiler/IDE versions which the release version of CUDA supports here
Edit to add that the CUDA 9 release will add official support for VS2017.

All you need to do is set the CUDA_HOST_COMPILER variable to a supported compiler for example the visual studio 2015 compiler.
In my case this is:
C:/Program Files (x86)/Microsoft Visual Studio 14.0/VC/bin/amd64/cl.exe
As both runtime libraries are binary compatible you can use the 2015 compiler within CUDA and compile all the rest of the application with the 2017 compiler.

Related

Cl/cl.h not found in Visual Studio Code 2019 AMD GPU Win10

I wanted to get started in GPU programming and since I have an AMD GPU I would like to start with OpenCL.
I have installed on my Windows 10 machine Visual Studio Code 2019 editor and I've also installed this https://github.com/GPUOpen-LibrariesAndSDKs/OCL-SDK. I tried to put those variables inside the settings of VC2019 but nothing happened.
Can somebody help troubleshooting this problem? How did you manage to install OpenCL?
Thanks in advance
When compiling your code, you need to tell the linker where the OpenCL headers and lib file are located. I usually put the headers and lib file inside the project directory:
g++ *.cpp -o Test.exe -I./OpenCL/include -L./OpenCL/lib -lOpenCL
But you can also make the -Ipath/to/OpenCL/include -Lpath/to/OpenCL/lib paths point to the OCL-SDK directory.
For how to setup OpenCL with VS Community, see here: https://stackoverflow.com/a/57017982/9178992
For an easy start with OpenCL, I created a wrapper that vastly simplifies the OpenCL C++ bindings and eliminates the entire code overhead that comes with it. This includes the OpenCL headers and all Visual Studio Community project settings; no additional setup required:
https://github.com/ProjectPhysX/OpenCL-Wrapper

CUDA: How to link a specific obj, ptx, cubin from a separate compilation?

I have a fairly large CUDA/C++ project that compiles to a static library. The toolchain is CUDA Toolkit 9.0/9.2 and VS 2017. I cannot change the company toolchain. Our most expensive kernel was hit by a nvcc compiler regression introduced in the 9.0 Toolkit. I have filed this with the Nvidia developer's website, and received confirmation of the regression. That was about a year ago, and the ticket is still open. Maybe the 10.0 Toolkit will fix it.
But I cannot wait. So my plan is to compile just this one specific kernel using the 8.0 nvcc compiler and v140 (VS 2015) compiler. It is a single .hpp file with __device__ decorator for the kernel declaration, and a .cu file with the definition. The kernel does not call other kernels; it is a rather simple kernel.
From the v140 Native Tools Command Prompt, I executed:
nvcc -x cu -arch=sm_61 -dc kernel.cu
And obtained a kernel.obj file. I have read the NVCC documentation on CUDA Compiler Driver NVCC. I confess to not entirely understanding. There are several compilation phases, and I do not see which is the correct course for my case.
My question is how to link this object file into my greater static library? If someone could point me to the correct series of commands, or better yet, how to include this into the VS Project, presumably with kernel.hpp and kernel.obj, I would be most grateful.
Following Njuffa's comment above, the simplest solution is create a static library using the earlier, performant toolchain for that kernel (VS 2015 & CUDA 8.0 Tookit). Then link that library into the greater project with the later toolchain. I did so with success.
I created a CUDA 8.0 template project in VS 2015 with only the kernel source and header. The compilation target set to static library. This created a .lib file. The .lib file and header are then added to the C++ linker settings of the greater project, using VS 2017 and CUDA 9.0. All test executables using this static library pass. This is a much simpler solution than trying to recompile using an intermediate compilation format ( ptx, cubin, etc.)
Although ultimately, the real solution was to refactor the kernel to use shared memory more efficiently, negating the need for the older nvcc version.

How to use CMAKE for C++11 on windows?

My C++11 project is currently using CMAKE, XCODE, CLANG on OSX. I wish to compile this code on Windows.
Plan is to use the same cmake settings files on windows. Best case would be to use CMAKE to generate VS projects which uses Clang or gcc for C++11 .
Seems to me, that Visual Studio is just not going to fully support C++11 for a while. So we should all try to find a general solution for cross platform C++11.
How would one use CMAKE to generate projects/makefiles which would compile C++11 code on windows?
CMake's Visual Studio generator will always use the cl compiler of Visual C++.
What you request would require writing a new Generator for CMake. That is, the problem cannot be solved by writing a clever CMakeLists.txt, but has to be solved by adding a feature to the CMake core binary itself. I agree that this could be useful once Clang achieves a suitable level of Windows support, but at the point of this writing, it is probably too early for that.
You might want to take a look at the experimental compile-features mechanism for CMake. This is not yet part of CMake 3.0, but is planned to be integrated with one of the next releases. The idea is that you just specify which C++11 features you need and CMake takes care of configuring the compiler accordingly (or gives an error if the compiler does not support the feature at all).
You can create VS projects which use clang by specifying the LLVM toolset when you generate.
http://public.kitware.com/Bug/view.php?id=14863
Eg:
cmake.exe .. -T LLVM-vs2010

Compiling Visual Studio 2010 project in VS 2012 (C++): error LNK2038: mismatch detected for '_MSC_VER': value '1600' doesn't match value '1700'

I've been sent a sample app for a newly minted SDK, and I can't compile it - it won't compile at all in Visual Studio Express for reasons of ATL, and that seems to be the only version of VS 2010 that Microsoft still makes available. So I'm trying to compile it in VS 2012 Professional, but it gives me error LNK2038. From what I can gather e.g. from this thread, the problem is down to trying to use .lib files compiled for 2010 in the 2012 version. Since I'm not the one who compiled the .lib, and I don't have the code for it, and VS 2010 is inaccessible, I'm really not sure what to do about this unless the original authors recompile it. Any suggestions? Something basic I'm missing? Many thanks for any help.
Library files cannot be reused across different versions of the compiler. You will need to do one of the following:
Compile your project with the same version of the compiler used for creating the library files you have.
Obtain new library files from the owner of the code, compiled using the same compiler you are now using.
Obtain the source code for the libraries, and compile new versions of them yourself.
Rewrite the code you own to not depend on the libraries you cannot control, since your build environment does not allow for their use.

How do I set up CUDA v4.0 to work nicely with Visual Studio 2010? [duplicate]

Direct Question: How do I create a simple hello world CUDA project within visual studio 2010?
Background: I've written CUDA kernels. I'm intimately familiar with the .vcproj files from Visual Studio 2005 -- tweaked several by hand. In VS 2005, if I want to build a CUDA kernel, I add a custom build rule and then explicitly define the nvcc call to build the files.
I have migrated to Win 7, and VS 2010 because I really want to try out nSight. I have nSight 1.5 installed. But this is where I'm totally lost. If I proceed as before, nvcc reports that it only supports msvc 8.0 & 9.0. But the website clearly states that it supports VS 2010.
I read somewhere else that I need to have VS 2008 (msvc 9.0) also installed -- my word. Doing so now.
But I'm guessing that at least part of my problems stem from the homegrown custom build tool specifications. Several websites talk about adding a *.rules file to the build, but I've gathered that this is only applicable to VS 2008. Under "Build Customizations" I see CUDA 3.1 and 3.2, but when I add kernels to the project they aren't built. Another website proclaims that the key is three files: Cuda.props Cuda.xml Cuda.targets, but it doesn't say how or where to add these files -- or rather I'll gamble that I just don't understand the notes referenced in the website.
So does anyone know how to create a simple project in VS 2010 which builds a CUDA kernel -- using either the nSight 1.5 setup or the NvCudaRuntimeApi.v3.2.rules file which ships with the CUDA 3.2 RC?
Thanks in advance! I'd offer a bounty, but I only have 65 points total.
CUDA TOOLKIT 4.0 and later
The build customisations file (installed into the Program Files\MSBuild\Microsoft.Cpp\v4.0\BuildCustomizations directory) "teaches" Visual Studio how to compile and link any .cu files in your project into your application. If you chose to skip installing the customisations, or if you installed VS2010 after CUDA, you can add them later by following the instructions in Program Files\NVIDIA GPU Computing Toolkit\CUDA\v4.0\extras\visual_studio_integration.
Create a new project using the standard MS wizards (e.g. an empty console project)
Implement your host (serial) code in .c or .cpp files
Add the NVIDIA build customisation (right click on the project, Build customizations, tick the relevant CUDA box)
See note 1 if using CUDA 4.0
Implement your wrappers and kernels in .cu files
If you added .cu files before the build customisations, then you'll need to set the type of the .cu files to CUDA C/C++ (right-click on the file, Properties, set Item Type)
Add the CUDA runtime library (right click on the project and choose Properties, then in Linker -> Input add cudart.lib to the Additional Dependencies)
Then just build your project and the .cu files will be compiled to .obj and added to the link automatically
Incidentally I would advocate avoiding cutil if possible, instead roll your own checking. Cutil is not supported by NVIDIA, it's just used to try to keep the examples in the SDK focussed on the actual program and algorithm design and avoid repeating the same things in every example (e.g. command line parsing). If you write your own then you will have much better control and will know what is happening. For example, the cutilSafeCall wrapper calls exit() if the function fails - a real application (as opposed to a sample) should probably handle the failure more elegantly!
NOTE
For CUDA 4.0 only you may need to apply this fix to the build customisations. This patch fixes the following message:
The result "" of evaluating the value "$(CudaBuildTasksPath)" of the "AssemblyFile" attribute in the element is not valid
This answer applies to CUDA 3.2, from 4.0 onwards CUDA supports the VC 10 compiler directly, see other answers for more information
You need either VS 2008 or the 6.1 Windows SDK installed. That's because NSight 1.5 RC or the CUDA 3.2 SDK use the VC 9 compiler under the hood. I've got this working successfully with 2008 installed and am told it should work with the SDK but haven't tried.
With NSight 1.5 and/or the CUDA 3.2 SDK you shouldn't need to muck with any custom build rules. I've been there and it's painful. With the latest builds all that goes away:
Create your VC++ project.
Add a .CU file to it.
Select the project file in the Solution Explorer.
Open Project | Build Customizations...
Check the "CUDA 3.2 (.targets,
.props)" customization.
Select a .CU file in your project and hit Alt-Enter to show it's properties.
Make sure it's Item Type is set to "CUDA C/C++"
It should just build. Let me know if this helps and if you run into problems as this is from memory.
The good news it getting CUDA working with VS 2010 just got much easier.
Ade
BTW: I'll update my blog post.
Another Good tutorial here:
http://www.stevenmarkford.com/installing-nvidia-cuda-with-visual-studio-2010/
if you get an error about '<' note this step (from a previous answer):
If you added .cu files before the build customisations, then you'll need to set the type of the .cu files to CUDA C/C++ (right-click on the file, Properties, set Item Type)
But if you follow their steps, it should work!

Resources