Skip to content

revert commit 1d842ee6 to fix #1979

Cidana Developers requested to merge Cidana-Developers/SVT-AV1:remove_stack2heap into master

Description

This is merge request of revertion the commit 1d842ee6 that causes speed regression on windows.

After this revertion, the MacOS stack overflow issue can not be re-produced in CI progress. So we do not apply the option 2 in this MR. Another MR is prepared here, the test shows similar performance.

If the option2 is required, we will close this MR and create another MR includes option2

Issue

Closes Issue 1979

Author(s)

@Cidana-Developers

Performance impact

  • quality
  • memory
  • speed
  • 8 bit
  • 10 bit
  • N/A
Test Case:
AMD Ryzen 7 3700X 8-Core Processor (8C16T)  @ 3.60 GHz + 32GB DRAM
Windows10 Pro 21H1
SvtAv1EncApp.exe -i tf2_lossless_10bit.yuv  -w 1920 -h 1080 --input-depth 10 --preset 7 --crf 14 --tune 0
Runs: 10
commit d92858be is the bad commit on master for base
commit 91a3aa39 only reserts the patch of moving memory from stack to heap in MR1947
Test NO, master commit d92858be FPS revertion commit 91a3aa39 FPS
1 11.037 13.978
2 10.892 13.909
3 10.887 13.909
4 10.929 13.899
5 10.869 13.936
6 10.896 13.879
7 10.953 13.920
8 10.878 13.945
9 10.917 13.911
10 10.968 13.889
AVG. 10.923 13.918
Deviation 27.42%

Test set

  • obj-1-fast can be found here
  • other
  • N/A
Table 1 10-bit ref scale off revert version vs. master PSNR SSIM VMAF AVG PSNR/SSIM/VMAF Cycles Dev Max Value Memory Deviation Abs Max Clip Memory Deviation
ALL svt_M5_91a3aa39_off_10bit vs. svt_M5_master_off_10bit 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 1.08% -0.11% 0.27%
ALL svt_M8_91a3aa39_off_10bit vs. svt_M8_master_off_10bit 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.97% 0.17% -1.32%
ALL svt_M12_91a3aa39_off_10bit vs. svt_M12_master_off_10bit 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.98% 0.10% 0.28%
Table 2 10-bit ref scale random mode revert version vs. master PSNR SSIM VMAF AVG PSNR/SSIM/VMAF Cycles Dev Max Value Memory Deviation Abs Max Clip Memory Deviation
ALL svt_M5_91a3aa39_resize_random_10bit vs. svt_M5_master_resize_random_10bit 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% -0.28% 0.68% 1.96%
ALL svt_M8_91a3aa39_resize_random_10bit vs. svt_M8_master_resize_random_10bit 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% -0.61% -1.12% 2.33%
ALL svt_M12_91a3aa39_resize_random_10bit vs. svt_M12_master_resize_random_10bit 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% -0.79% 0.37% -2.31%

Table 1 and Table 2 show lossless in both default case and resize random mode case.

Merge method

  • Allow the maintainer to squash and merge when PR is ready to create a 1-commit to the master branch. The maintainer will be able to fix typos / combine commit messages to create a more readable 1-commit message or use whatever is stated in the 'Description' section
  • I will clean up my commits and the maintainer shall use 'rebase and merge' to the master branch

Merge request reports