libvmaf with CUDA -- how to build for Windows

[claudiomarassi]

I saw that many people tried but without success to compile vmaf with cuda for Windows.

Could you help us?

[kylophone]

I have personally not tried this on Windows, but AFAIK it should be possible. If anyone wants to lend a hand on this that would be welcome.

[stderr-to-devnull]

I opened #1415 since I am struggling with this. At least I know it works with CUDA disabled (default config) so this narrows it down a bit.

[xyx98]

I build vmaf with cuda successfully on windows by follow steps:

1、Install visual studio with msvc(should match your cuda toolkit version), cuda toolkit, msys2 (I'am using msys2 provide by m-ab-s) ,and also clone vmaf repo.

2、modify all source file under libvmaf/src/cuda,replace:

#include <pthread.h>

with:

#include "pthread.h"

And copy pthread.h,pthread_compat.h,pthread_signal.h,pthread_time.h,pthread_unistd.h,sched.h from msys2 to libvmaf/src/cuda folder,because nvcc(with msvc) can not find pthread.

Download nv-codec-header,and copy ffnvcodec folder to libvmaf/src folder,because nvcc(with msvc) can not find nv-codec-header.

4、modify src/feature/integer_adm.h (line 135),replace:

static const struct dwt_model_params dwt_7_9_YCbCr_threshold[3] = {
    {.a = 0.495, .k = 0.466, .f0 = 0.401, .g = {1.501, 1.0, 0.534, 1.0}},
    {.a = 1.633, .k = 0.353, .f0 = 0.209, .g = {1.520, 1.0, 0.502, 1.0}},
    {.a = 0.944, .k = 0.521, .f0 = 0.404, .g = {1.868, 1.0, 0.516, 1.0}}};

with:

static const struct dwt_model_params dwt_7_9_YCbCr_threshold[3] = {
    {0.495, 0.466, 0.401, {1.501, 1.0, 0.534, 1.0}},
    {1.633, 0.353, 0.209, {1.520, 1.0, 0.502, 1.0}},
    {0.944, 0.521, 0.404, {1.868, 1.0, 0.516, 1.0}}};

5、Start "x64 Native Tools Command Prompt for VS",then start msys2 under "x64 Native Tools Command Prompt for VS".

I'm using msys2 provide by m-ab-s,so i just switch to m-ab-s dir,and run mintty.lnk,it will start mingw64 environment.

6、Switch to libvmaf dir, just use:

meson setup build -Denable_cuda=true
ninja -C build

[Mishini]

I have compiled vmaf.exe following the tutorial above and integrated it into FFmpeg.
According to mdhiggins/sickbeard_mp4_automator#1748, I used scale_npp instead of scale_cuda.

Here is the command: ffmpeg.exe -threads:v 1 -hwaccel cuda -hwaccel_output_format cuda -i dst.mp4 -threads:v 1 -hwaccel cuda -hwaccel_output_format cuda -i ref.mp4 -filter_complex "[0:v]scale_cuda=format=yuv420p[dis];[1:v]scale_cuda=format=yuv420p[ref];[dis][ref]libvmaf_cuda=log_fmt=json:log_path=vmaf.json" -an -f null -

Then a format processing error occurs, this is the patch.

ffmpeg-0001-libvmaf-cuda-decouple-color-negotiation.patch

[xyx98]

I never use ffmpeg's vmaf integration,can not help you more.
It may not caused by complie,did you test under linux?

[Mishini]

I never use ffmpeg's vmaf integration,can not help you more. It may not caused by complie,did you test under linux?

Adding colorspace and range to scale_cuda works too. The tutorial's scale_npp implementation didn't specify that these were required, and now switching to scale_cuda, it might be the issue here

[StormBytePP]

While reviewing the current CUDA support and fixing a few related problems, I came across this issue and thought a short status update might be useful.

• In CUDA: Fix compile, SIGSEGV and tests #1573 I fixed several CUDA build/test issues (missing include paths for nvcc, a SIGSEGV in pinned picture release, and the gpumask test dependency). These are not Windows-only and also help on Linux.
• Add native MSVC build support for Windows #1477 adds native MSVC support, including the remaining pieces needed to build CUDA under MSVC (winpthreads includes/deps, _USE_MATH_DEFINES for nvcc, etc.).

With both changes, building libvmaf with CUDA on native Windows/MSVC is in much better shape than before. Sharing this here in case it helps others following this thread.

[stderr-to-devnull]

@StormBytePP thank you for your answer and for trying to help.

For myself I reached the conclusion some time ago that this is a ridiculously complicated process for non-programmers so for me VMAF CUDA is DOA. A waste of technology and resources with no proper workflow to generate a usable binary and library to use. It's not worth the time and effort for me.