Fix segfaults when using CUDA

[aswild]

Summary: switch from using xxd to bin2c when generating the .ptx.c files so that the PTX data can be null-terminated.

In newer drivers or cuda versions, vmaf now segfaults when trying to do anything from the GPU. The coredumps indicate that the crash happens somewhere inside the cuModuleLoadData calls in init_fex_cuda.

Documentation for cuModuleLoadData states that its image argument can be "obtained by mapping a cubin or PTX or fatbin file, [or] passing a cubin or PTX or fatbin file as a NULL-terminated text string...". It looks like VMAF is trying to do the latter, encoding PTX text files as an ASCII string using xxd, but there's no null-terminator in the data because nothing asked for one.

I'm a CUDA noob and don't know how this ever worked on older driver versions, but I tried editing the .ptx.c files by hand to add 0x00 bytes at the end and it worked!

Switch from xxd to bin2c (which is distributed with the cuda-nvcc package) that supports a --padd option to add a null byte to the PTX data, eliminating the segfaults. The arrays got renamed slightly to remove the src_ prefix, since bin2c doesn't do any automatic naming of the output array.

This should resolve #1357

[nilfm99]

Thanks for the contribution! @kylophone is this something you could easily test?

[michalkielan]

I have the same issues as described in #1357 using the latest Nvidia driver and CUDA and this fix is working for me. If testing is a blocker for this PR, I'm sharing the tests I've done to move this forward. On master both running vmaf_cuda using ffmpeg and vmaf's cuda unit tests are crashing due to CUDA_ERROR_INVALID_PTX and it with the fix there is no issue. @nilfm99 if you want to double-check something specific with ffmpeg + vmaf I can help to run it on my setup.

Tested this on: NVIDIA GeForce RTX 3060, Driver Version: 570.86.16, CUDA Version: 12.8
Without this fix (upstream master) the unit tests are failing:

$ meson libvmaf/build libvmaf -Denable_avx512=true -Denable_cuda=true
$ ninja -vC libvmaf/build test
...
1 frame  ⢀⠀ 0.00 FPScode: 218; description: CUDA_ERROR_INVALID_PTX
vmaf: ../src/feature/cuda/integer_adm_cuda.c:1026: init_fex_cuda: Assertion `0' failed.
/home/michalk/Projects/vmaf/libvmaf/tools/test/test_vmaf_cuda_gpumask.sh: line 10: 156665 Aborted                 (core dumped) ./tools/vmaf --reference /dev/zero --distorted /dev/zero --width 1920 --height 1080 --pixel_format 420 --bitdepth 8 --frame_cnt 2 --gpumask 0
...
test_cuda_no_init: pass
test_cuda_picture_preallocation_method_none: code: 218; description: CUDA_ERROR_INVALID_PTX
test_cuda_pic_preallocation: ../src/feature/cuda/integer_adm_cuda.c:1026: init_fex_cuda: Assertion `0' failed.
[1]    157225 IOT instruction (core dumped)  MSAN_OPTIONS= MALLOC_PERTURB_=221 ASAN_OPTIONS= MESON_TEST_ITERATION=1 =

With the fix (rebased to upstream master) the unit tests are passing:

...
 1/20 test_picture                        OK              0.04s
 2/20 test_feature_collector              OK              0.04s
 3/20 test_thread_pool                    OK              0.04s
 4/20 test_model                          OK              0.04s
 5/20 test_predict                        OK              0.03s
 6/20 test_dict                           OK              0.03s
 7/20 test_cpu                            OK              0.03s
 8/20 test_ref                            OK              0.02s
 9/20 test_feature                        OK              0.02s
10/20 test_ciede                          OK              0.02s
11/20 test_cambi                          OK              0.01s
12/20 test_luminance_tools                OK              0.01s
13/20 test_feature_extractor              OK              0.05s
14/20 test_cli_parse                      OK              0.01s
15/20 test_psnr                           OK              0.01s
16/20 test_propagate_metadata             OK              0.01s
17/20 test_ring_buffer                    OK              0.38s
18/20 test_vmaf_cuda_gpumask              OK              1.83s
19/20 test_cuda_pic_preallocation         OK              1.95s
20/20 test_framesync                      OK              5.01s

Ok:                 20
Expected Fail:      0
Fail:               0
Unexpected Pass:    0
Skipped:            0
Timeout:            0

ffmpeg:
Test videos generated by:

$ ffmpeg -f lavfi -i testsrc=duration=1:size=1280x720:rate=30 -pix_fmt yuv420p -c:v libx264 -preset fast -crf 23 reference.mkv
$ ffmpeg -hwaccel cuda -i reference.mkv -c:v hevc_nvenc -preset fast -qp 24 distorted.mkv

Without the fix ffmpeg is crashing at CUDA_ERROR_INVALID_PTX assert:

$ ffmpeg -hide_banner -hwaccel cuda -hwaccel_output_format cuda -i distorted.mkv -hwaccel cuda -hwaccel_output_format cuda -i reference.mkv -filter_complex "[0:v]scale_npp=format=yuv420p[dis];[1:v]scale_npp=format=yuv420p[ref];[dis][ref]libvmaf_cuda" -f null -
Input #0, matroska,webm, from 'distorted.mkv':
  Metadata:
    ENCODER         : Lavf61.7.100
  Duration: 00:00:01.00, start: 0.000000, bitrate: 265 kb/s
  Stream #0:0: Video: hevc (Main), yuv420p(tv, progressive), 1280x720 [SAR 1:1 DAR 16:9], 30 fps, 30 tbr, 1k tbn
      Metadata:
        ENCODER         : Lavc61.19.100 hevc_nvenc
        DURATION        : 00:00:01.000000000
Input #1, matroska,webm, from 'reference.mkv':
  Metadata:
    ENCODER         : Lavf61.7.100
  Duration: 00:00:01.00, start: 0.000000, bitrate: 203 kb/s
  Stream #1:0: Video: h264 (High), yuv420p(tv, progressive), 1280x720 [SAR 1:1 DAR 16:9], 30 fps, 30 tbr, 1k tbn
      Metadata:
        ENCODER         : Lavc61.19.100 libx264
        DURATION        : 00:00:01.000000000
Stream mapping:
  Stream #0:0 (hevc) -> scale_npp:default
  Stream #1:0 (h264) -> scale_npp:default
  libvmaf_cuda:default -> Stream #0:0 (wrapped_avframe)
Press [q] to stop, [?] for help
code: 218; description: CUDA_ERROR_INVALID_PTX
ffmpeg: ../src/feature/cuda/integer_adm_cuda.c:1026: init_fex_cuda: Assertion `0' failed.
[1]    209899 IOT instruction (core dumped)  ffmpeg -hide_banner -hwaccel cuda -hwaccel_output_format cuda -i distorted.mkv

With the fix.:

$ ffmpeg -hide_banner -hwaccel cuda -hwaccel_output_format cuda -i distorted.mkv -hwaccel cuda -hwaccel_output_format cuda -i reference.mkv -filter_complex "[0:v]scale_npp=format=yuv420p[dis];[1:v]scale_npp=format=yuv420p[ref];[dis][ref]libvmaf_cuda" -f null -
Input #0, matroska,webm, from 'distorted.mkv':
  Metadata:
    ENCODER         : Lavf61.7.100
  Duration: 00:00:01.00, start: 0.000000, bitrate: 265 kb/s
  Stream #0:0: Video: hevc (Main), yuv420p(tv, progressive), 1280x720 [SAR 1:1 DAR 16:9], 30 fps, 30 tbr, 1k tbn
      Metadata:
        ENCODER         : Lavc61.19.100 hevc_nvenc
        DURATION        : 00:00:01.000000000
Input #1, matroska,webm, from 'reference.mkv':
  Metadata:
    ENCODER         : Lavf61.7.100
  Duration: 00:00:01.00, start: 0.000000, bitrate: 203 kb/s
  Stream #1:0: Video: h264 (High), yuv420p(tv, progressive), 1280x720 [SAR 1:1 DAR 16:9], 30 fps, 30 tbr, 1k tbn
      Metadata:
        ENCODER         : Lavc61.19.100 libx264
        DURATION        : 00:00:01.000000000
Stream mapping:
  Stream #0:0 (hevc) -> scale_npp:default
  Stream #1:0 (h264) -> scale_npp:default
  libvmaf_cuda:default -> Stream #0:0 (wrapped_avframe)
Press [q] to stop, [?] for help
Output #0, null, to 'pipe:':
  Metadata:
    encoder         : Lavf61.7.100
  Stream #0:0: Video: wrapped_avframe, cuda(tv, progressive), 1280x720 [SAR 1:1 DAR 16:9], q=2-31, 200 kb/s, 30 fps, 30 tbn
      Metadata:
        encoder         : Lavc61.19.100 wrapped_avframe
[Parsed_libvmaf_cuda_2 @ 0x7335a8007280] VMAF score: 97.512795
[out#0/null @ 0x59af9bc4ea80] video:13KiB audio:0KiB subtitle:0KiB other streams:0KiB global headers:0KiB muxing overhead: unknown
frame=   30 fps=0.0 q=-0.0 Lsize=N/A time=00:00:01.00 bitrate=N/A speed=7.08x

[1480c1]

@aswild just to make sure, is bin2c distributed by cuda-nvcc the same bin2c as https://github.com/adobe/bin2c?

[1480c1]

Actually, nvm, it's looking like it's not since that bin2c does not support a lot of arguments.

[1480c1]

It seems there's a lot of implementation of bin2c. There's even an unrelated one in an ubuntu package already https://manpages.ubuntu.com/manpages/kinetic/man1/bin2c.1.html

[aswild]

@1480c1 yeah bin2c is an unfortunately generic name choice, but the hxtools package from Debian (and Ubuntu) seems to be the only other distro package providing /usr/bin/bin2c besides the relevant cuda packages that I can find.

This part of the meson configuration step is only enabled when the user explicitly builds with -Denable_cuda=true so I think it's reasonable to assume that the bin2c available will be Nvidia's version.