nv_decode.h as part of CUDA - for installation on action runners

[eyalroz]

Select Topic Area

Question

Body

In my repository's .github, I have an action script which installs CUDA on the runner building my code. This all fine and good, but recently, I've noticed that a header file that's part of the CUDA distribution on Linux does not seem to be part of it on Windows, or at least - not in the packages I can notice.

The header is nv_decode.h. It is used for name demangling. As I've started rleying on its presence - all of my Windows builds are failing.

It's possible that it is actually part of the CUDA Windows distribution - but I can't figure out in which one of the packages... any insight would be greatly appreciated.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

Ping... hoping that someone can shed any light on this.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

Still waiting for someone at GitHub/Microsoft to say anything...

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

And again, still waiting for someone at GitHub/Microsoft to say anything...

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

Still waiting for a comment.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[PRIYAtechky]

nv_decode.h is not included in the standard CUDA Toolkit installation on Windows.
That header is part of the NVIDIA Video Codec SDK, not the base CUDA toolkit.

• On Linux, some distributions bundle parts of the Video Codec SDK headers with CUDA, which is why you may see it there.
• On Windows, you need to download the Video Codec SDK separately and add its include path to your build.

👉 Steps to fix your Windows builds:

1. Download the NVIDIA Video Codec SDK from NVIDIA’s developer site:
   https://developer.nvidia.com/nvidia-video-codec-sdk
2. Extract it and point your project to the Interface folder, which contains nv_decode.h and related headers.
3. Update your build workflow (in GitHub Actions) to copy or reference this SDK during setup on Windows runners.

✅ Summary:
The file is not missing—it's simply not part of the Windows CUDA Toolkit. You’ll need to explicitly include the Video Codec SDK in your build process to have nv_decode.h available.

[eyalroz]

@PRIYAtechky :

nv_decode.h is not included in the standard CUDA Toolkit installation on Windows. That header is part of the NVIDIA Video Codec SDK, not the base CUDA toolkit.

I don't think that's possible. That header has exactly one function:

char* __cu_demangle(const char *id,
                        char *output_buffer,
                        size_t *length,
                        int *status);

and this has nothing to do with video. It's possible that the video codec SDK has another header with the same basename, but from a different folder - but it's not this one.

... also, I just checked the downloadable header files, and those are: cuviddec.h, nvcuvid.h, nvEncodeAPI.h - and that's it.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

Still waiting on this.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

Still waiting on this... at the end of 2025.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[eyalroz]

Still waiting on this... in early 2026.

[github-actions[bot]]

🕒 Discussion Activity Reminder 🕒

This Discussion has been labeled as dormant by an automated system for having no activity in the last 60 days. Please consider one the following actions:

1️⃣ Close as Out of Date: If the topic is no longer relevant, close the Discussion as out of date at the bottom of the page.

2️⃣ Provide More Information: Share additional details or context — or let the community know if you've found a solution on your own.

3️⃣ Mark a Reply as Answer: If your question has been answered by a reply, mark the most helpful reply as the solution.

Note: This dormant notification will only apply to Discussions with the Question label. To learn more, see our recent announcement.

Thank you for helping bring this Discussion to a resolution! 💬

[asaddevx]

Hi @eyalroz — First, I am sorry you have been waiting on this since 2024. The previous answer was incorrect.

The Truth About nv_decode.h

You are right. The nv_decode.h header contains __cu_demangle() for name demangling. It has nothing to do with video decoding. The person who suggested Video Codec SDK confused it with nvcuvid.h / cuviddec.h, which are completely different headers.

Why It's Missing on Windows

nv_decode.h is not part of the standard CUDA Toolkit on Windows.

After extensive research, here is what I found:

• On Linux, this header sometimes gets included with CUDA Toolkit or through extra development packages
• On Windows, NVIDIA does not ship this header with the standard CUDA installer at all

The function __cu_demangle() is an internal NVIDIA function used for demangling C++ names. It was never officially documented or guaranteed to be public.

The Real Solution

Since NVIDIA does not provide nv_decode.h on Windows, you have two options:

Option 1 (Recommended): Stop using nv_decode.h

Use abi::__cxa_demangle() from <cxxabi.h> (GCC/Clang) or UnDecorateSymbolName() on Windows (DbgHelp.h). These are standard, cross-platform solutions for demangling.

Option 2: Extract the function yourself

If you really need __cu_demangle(), you can:

1. Copy the header contents from a Linux CUDA installation
2. Add it to your repository
3. Link against cudart.lib on Windows

This is a hack, but it works.

For Your GitHub Actions Workflow

If you choose Option 2, add this to your workflow to copy the header:

- name: Copy nv_decode.h
  run: |
    curl -o nv_decode.h https://raw.githubusercontent.com/path-to-header/...
    move nv_decode.h "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.x\include\"

But honestly, Option 1 is the correct long-term fix. Relying on an internal NVIDIA header that doesn't exist on Windows is going to break again.

I hope this finally gives you the answer you have been looking for.

[eyalroz]

Hi @eyalroz — First, I am sorry you have been waiting on this since 2024.

Thanks, but - I don't think you have to apologize, I mean, it's not your responsibility. In fact, given the styling and phrasing of your answer, I suspect your answer is the output of some LLM.

• On Linux, this header sometimes gets included with CUDA Toolkit or through extra development packages

That's not the case. This header is part of the CUDA Toolkit distribution on Linux, period. Not sometimes - always.

• On Windows, NVIDIA does not ship this header with the standard CUDA installer at all

Given what you said about Linux, it's difficult for me to just accept your statement at face value.

The function __cu_demangle() is an internal NVIDIA function used for demangling C++ names. It was never officially documented or guaranteed to be public.

Again, that's not true. It is not really internal but has an exposed specialty header for us to use; it's not for C++ only but for CUDA names; it is very much officially documented, and is "guaranteed to be public" since it's part of the CUDA toolkit, at least on Linux.

Since NVIDIA does not provide nv_decode.h on Windows, you have two options:

Option 1 (Recommended): Stop using nv_decode.h

Use abi::__cxa_demangle() from <cxxabi.h> (GCC/Clang) or UnDecorateSymbolName() on Windows (DbgHelp.h). These are standard, cross-platform solutions for demangling.

That does not demangle CUDA symbols, only C++ symbols. CUDA has extra mangling syntax.

Option 2: Extract the function yourself

This is a hack, but it works.

This is interesting... but - is it legal for me to redistribute this header with my API wrappers library? I wonder.

After extensive research

If you conducted extensive research, how could you make those mistakes? I am again left thinking I've been burned by this age of LLM halucinations.

[Gecko51]

The header is genuinely absent from the Windows CUDA Toolkit distribution. It's not in any of the standard Windows CUDA component packages (cudart, nvcc, cuda-toolkit, etc.). This is a real asymmetry between Linux and Windows that NVIDIA has never fixed.

The cleanest fix that avoids any redistribution concern: create a thin compatibility shim in your repo. Since the function signature is stable, forward declare it conditionally:

#if defined(__has_include) && __has_include(<nv_decode.h>)
  #include <nv_decode.h>
#else
  // Absent from Windows CUDA distribution; forward declare the symbol.
  extern "C" char* __cu_demangle(
      const char* id,
      char* output_buffer,
      size_t* length,
      int* status
  );
#endif

This doesn't require redistributing any NVIDIA files. Whether __cu_demangle is actually exported from cudart.lib on Windows is worth verifying on a local install first: dumpbin /exports cudart.lib | findstr demangle. If the export is there, the forward declaration is all you need and Windows builds will link fine without the header.

If the symbol isn't exported on Windows, the remaining options are less clean: vendor the Linux header (NVIDIA has historically been permissive about header redistribution in their EULA, but worth checking the current version), or platform-guard the demangling code and fall back to something else on Windows for your specific use case.

[eyalroz]

I think I will need a CMake-based mechanism rather than __has_include, since my library is C++11-based. But I think I'll go with this solution basically. Or rather, I'll check with dumpbin as you suggested and see.

But it's so weird that this is missing from the Windows distribution, and for so long.

[Gecko51]

Since you're on C++11, here's the CMake-based version of what Gecko51 described.

Add to your CMakeLists.txt:

include(CheckIncludeFileCXX)
check_include_file_cxx("nv_decode.h" HAVE_NV_DECODE_H)

if(NOT HAVE_NV_DECODE_H)
  target_compile_definitions(your_target PRIVATE CUDA_DEMANGLE_NEEDS_SHIM)
endif()

Then in a header (e.g. cuda_demangle.h):

#ifdef CUDA_DEMANGLE_NEEDS_SHIM
extern "C" char* __cu_demangle(
    const char* id,
    char* output_buffer,
    size_t* length,
    int* status
);
#else
#include <nv_decode.h>
#endif

Before shipping this, run the dumpbin check on a Windows runner to confirm the symbol actually exports from cudart.lib:

dumpbin /exports "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vX.Y\lib\x64\cudart.lib" | findstr /i demangle

If the symbol shows up, the shim is all you need and you avoid any redistribution questions entirely. If it doesn't export on Windows, you'd need to platform-guard the demangling code and skip it on Windows for now.

[eyalroz]

Actually, it's going to be slightly more complicated than that, because I need to be using the INTERFACE_INCLUDE_DIRECTORIES of the relevant target(s), but yes, that will basically be it.