Fix #120077: Video output artifacts due to threaded YUV conversion #120168

Aras Pranckevicius · 2024-04-02T11:06:10+02:00

2024-04-02 11:06:10 +02:00

ffmpeg AVFrame objects should use correct alignment between image
rows, or otherwise bad things might happen. In this particular
case, multi-threaded libswscale RGB->YUV conversion was trampling
over 4 bytes of V plane, for each thread boundary.

ffmpeg AVFrame objects should use correct alignment between image rows, or otherwise bad things might happen. In this particular case, multi-threaded libswscale RGB->YUV conversion was trampling over 4 bytes of V plane, for each thread boundary.

Aras Pranckevicius added 1 commit 2024-04-02 11:06:22 +02:00

buildbot/vexp-code-patch-lint Build done. Details

buildbot/vexp-code-patch-linux-x86_64 Build done. Details

buildbot/vexp-code-patch-darwin-x86_64 Build done. Details

buildbot/vexp-code-patch-darwin-arm64 Build done. Details

buildbot/vexp-code-patch-windows-amd64 Build done. Details

buildbot/vexp-code-patch-coordinator Build done. Details

83fe771184 Fix #120077 : Video output artifacts due to threaded YUV conversion

ffmpeg AVFrame objects should use correct alignment between image
rows, or otherwise bad things might happen. In this particular
case, multi-threaded libswscale RGB->YUV conversion was tramping
over 4 bytes of V plane, for each thread boundary.

Aras Pranckevicius commented

2024-04-02 11:06:43 +02:00

@blender-bot build

Aras Pranckevicius requested review from Sergey Sharybin 2024-04-02 11:10:02 +02:00

Sergey Sharybin approved these changes 2024-04-02 12:31:53 +02:00

Sergey Sharybin left a comment

It is a bit hard to do a proper review here, as it seems i can't reproduce the issue on my macOS arm64. The code and explanation looks clear though.

What is not clear is the relation of the issue with the previous refactor. it does not seem we ever passed proper alignment to av_image_get_buffer_size/av_image_fill_arrays, so not really clear why 422dd9404f introduced the issue?

It is a bit hard to do a proper review here, as it seems i can't reproduce the issue on my macOS arm64. The code and explanation looks clear though. What is not clear is the relation of the issue with the previous refactor. it does not seem we ever passed proper alignment to `av_image_get_buffer_size`/`av_image_fill_arrays`, so not really clear why 422dd9404f introduced the issue?

Aras Pranckevicius commented

2024-04-02 13:02:39 +02:00

What is not clear is the relation of the issue with the previous refactor. it does not seem we ever passed proper alignment to av_image_get_buffer_size/av_image_fill_arrays, so not really clear why 422dd9404f introduced the issue?

From what I can tell, what happens is that each "slice" of image processed by sws_scale overwrites 4 bytes past the "image portion", at least on my machine for this video resolution. Without alignment, and with multiple threads, this actually adds "number of threads" artifacts right exactly where each thread slice ends, i.e. first 4 pixels of the next slice. With just one thread (as previously), I don't know if it actually overwrites past the array, or whether there is no overwriting happening. It could have been the former, just memory allocation / page granularity happens to be good enough that it was not triggering an actual crash.

And btw with most typical video resolutions the issue would not be showing up, since e.g. 1920x1080 has linesize that is already multiple of 16. But the repro video resolution has width 1080, which is not multiple of 16.

> What is not clear is the relation of the issue with the previous refactor. it does not seem we ever passed proper alignment to `av_image_get_buffer_size`/`av_image_fill_arrays`, so not really clear why 422dd9404f introduced the issue? From what I can tell, what happens is that each "slice" of image processed by sws_scale overwrites 4 bytes past the "image portion", at least on my machine for this video resolution. Without alignment, and with multiple threads, this actually adds "number of threads" artifacts right exactly where each thread slice ends, i.e. first 4 pixels of the next slice. With just one thread (as previously), I don't know if it actually overwrites past the array, or whether there is no overwriting happening. It could have been the former, just memory allocation / page granularity happens to be good enough that it was not triggering an actual crash. And btw with most typical video resolutions the issue would not be showing up, since e.g. 1920x1080 has linesize that is already multiple of 16. But the repro video resolution has width 1080, which is not multiple of 16.

Aras Pranckevicius merged commit 19ce05971d into main

2024-04-02 13:03:17 +02:00

Aras Pranckevicius deleted branch swscale_align

2024-04-02 13:03:19 +02:00

Aras Pranckevicius referenced this issue from a commit

2024-04-02 13:03:21 +02:00

Fix #120077: Video output artifacts due to threaded YUV conversion

Sergey Sharybin commented

2024-04-02 13:26:46 +02:00

@aras_p Ah, I see. Thanks for the explanation.

Ruoyu-Lin commented

2024-04-02 17:47:45 +02:00

it seems that on my computer (EndeavourOS) this change will generate strange output video (but in the preview windows it seems fine)

Ruoyu-Lin commented

2024-04-02 17:48:24 +02:00

No description provided.

video0001-0015.mp4

417 KiB

Ruoyu-Lin commented

2024-04-02 19:42:10 +02:00

In the code above, the rgb_frame's linesize is after aligning (which is 4352 in this case(my computer's align size is 64)), however, the data in "pixels" is generated in RE_render_result_rect_to_ibuf and hasn't been aligned, (so it's linesize is 1080*4=4320 in this case), so when copying the data, they should use different linesize.

I've changed the code to this and it works fine on my computer.

![image](/attachments/c463033b-61e0-45c3-b474-57e8c3a2d7a4) In the code above, the rgb_frame's linesize is after aligning (which is 4352 in this case(my computer's align size is 64)), however, the data in "pixels" is generated in RE_render_result_rect_to_ibuf and hasn't been aligned, (so it's linesize is 1080*4=4320 in this case), so when copying the data, they should use different linesize. ![image](/attachments/4874b62d-3676-412e-9eac-0414a6c0c279) I've changed the code to this and it works fine on my computer.

image.png

118 KiB

image.png

53 KiB

Ruoyu-Lin commented

2024-04-02 20:39:12 +02:00

my pr here: #120184

my pr here: https://projects.blender.org/blender/blender/pulls/120184

👍 1

Thomas Dinges referenced this issue from a commit

2024-04-09 15:54:21 +02:00

Fix #120077: Video output artifacts due to threaded YUV conversion