Loongarch： multiple SIMD optimization functions are added

Multiple LSX optimizations and a few LASX optimizations were added. On the loongarch platform， the decoding performance has also been greatly improved. Welcome reviews. @jbk @jamrial @gramner

added 1 commit

• 7e0e8fee - loongarch: refactor loopfilter

Compare with previous version

Hi Hecai, thanks for your contribution. Could you provide some overall performance numbers? would be nice to know where it stands ATM.

added 1 commit

• bd2825e7 - LoongArch: Add save_tmvs_lsx

Compare with previous version

approved this merge request

@HecaiYuan could you please click "Edit" (top/right) and then select "Allow commits from members who can merge to the target branch." (at the bottom)? That way I can rebase & merge for you.

(Or rebase manually, if that was already selected.)

I tried to turn this option on and found that it was not selectable. I'm not sure what caused this. @rbultje

I looked at the revised code and found it more comfortable and reasonable. Our plan is to continually maintain community work and deliver high quality code to the best of our ability. In addition, the modified code will be in another patch.

resolved all threads

added 2 commits

• 0f618ba1 - loongarch: rewrite optimization functions in loongarch/itx.S
• 6973f5fc - loongarch: minor improvement on decode_symbol_adapt

Compare with previous version

added 2 commits

• 489225ba - loongarch: rewrite optimization functions in loongarch/itx.S
• 55c54ad3 - loongarch: minor improvement on decode_symbol_adapt

Compare with previous version

Relevant functions in loongarch/itx.S have been adjusted. The number of lines of code is significantly reduced. @rbultje @jbk

Please rebase this set on top of master.

added 56 commits

• 55c54ad3...f2c3ccd6 - 21 commits from branch videolan:master
• f2c3ccd6...7c63bb1b - 25 earlier commits
• 411fc219 - Loongarch: Optimized ipred_z1 8bpc functions by LSX
• 90a9549b - Loongarch: Optimized load_tmvs_c function by LSX
• af11a10a - loongarch: add lasx implementation of wiener filter for 8 bpc
• b9e9a0ef - loongarch: Refine prep_8tap_8bpc_lasx
• 96d6e472 - loongarch: rewirte warp_8x8/8x8t_lsx for 8 bpc
• 70582027 - loongarch: add lasx implementation of sgr_3x3 for 8 bpc
• 3d96175d - loongarch: refactor loopfilter
• 757f294a - LoongArch: Add save_tmvs_lsx
• 62a51df1 - loongarch: rewrite optimization functions in loongarch/itx.S
• ed004fe9 - loongarch: minor improvement on decode_symbol_adapt

Compare with previous version

merged

Thanks @HecaiYuan & team!

added loongarch performance labels

changed milestone to %1.5.0

@HecaiYuan sorry to poke you, but would it be possible for you to run the argon test suite with these optimizations merged? These samples test things like clipping behaviour in the inverse transforms or potential overflows in MC better than the standard conformance samples. There's a script in the repo to run it. Thanks!

In particular, these tests currently do fail, after this MR:

$ ../tests/dav1d_argon.bash -f
Mismatch in profile0_core/streams/test10571_10597_10562.obu                

This breakage seems to have started since commit 13a857d0.

The suite itself is here

mentioned in issue #448 (closed)

In issue  #448 (closed), the problem has been solved. @mstorsjo Thank you for suite. @lu_zero