Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

vvc_deblock.asm: reduce stack usage #275

Open
wants to merge 10 commits into
base: deblock_asm
Choose a base branch
from
Open

vvc_deblock.asm: reduce stack usage #275

wants to merge 10 commits into from

Conversation

stone-d-chen
Copy link
Collaborator

Since we're storing the lines of the CTU in registers, we can use the CTU as additional storage (it gets overwritten at the very end with the correct values).

@stone-d-chen
vvc_deblock.asm: reduce stack usage
a441dd5 
Since we're storing the lines of the CTU in registers, we can use the CTU as additional storage (it gets overwritten at the very end with the correct values).
@stone-d-chen
Copy link
Collaborator Author

@nuomi2021, I think something like this can work to reduce the stack usage.

@stone-d-chen
vvc_deblock.asm: remove [rsp + 32] stack usage
bdf22b0 
remove [rsp + 32] stack usage
@stone-d-chen
vvc_deblock.asm: rename r3strideq to srcstrideq; eliminate rsp + 16 use
8805c18 
First we rename vertical chroma's parameter r3strideq to src3strideq, what the horizontal calls it. This was we can use src3strideq in the macros. Eliminate [rsp + 16] usage by storing in the deblock unit instead.
@stone-d-chen
vvc_deblock.asm: remove [rsp] usage
997466c 
Use pixq + stride and pixq + 2 * stride to store the masks. For some reason movh [pix], is causing a mismatch in CodingToolsSets_B_Tencent_2.bit for 8 bit.
@stone-d-chen
vvc_deblock.asm: remove extraneous mask loading in chroma strong
8b0b158 
Additionally reuses the new freed up pix0q
@stone-d-chen
vvc_deblock.asm: add LOAD_TC macro and use it in weak chroma
fcf65c6 
Instead of storing the tc values in tcptr, just reload it each time.
@stone-d-chen
vvc_deblock.asm: remove spatial_maskq stack use
700948d
@stone-d-chen
vvc_deblock.asm: update function signatures
404b679 
Since we've removed stack usage and some GPRs, update the function signature

Auto stash before checking out "HEAD"
@stone-d-chen
vvc_deblock: remove extraneous instructions
4163124 
e.g. pxor m10, m10, when still cleared
@stone-d-chen
Copy link
Collaborator Author

stone-d-chen commented Jan 3, 2025 •
edited
Loading

@nuomi2021, I've eliminated stack use.

Benchmarks

Anything using the weak path does not get too much performance gain unfortunately, too much overhead.

vvc_h_loop_filter_chroma_8_mix_no-shift_c: 289.3
vvc_h_loop_filter_chroma_8_mix_no-shift_avx: 119.3

vvc_h_loop_filter_chroma_8_mix_shift_c: 859.3
vvc_h_loop_filter_chroma_8_mix_shift_avx: 199.3

vvc_h_loop_filter_chroma_8_one-side_no-shift_c: 669.3
vvc_h_loop_filter_chroma_8_one-side_no-shift_avx: 199.3

vvc_h_loop_filter_chroma_8_one-side_shift_c: 869.3
vvc_h_loop_filter_chroma_8_one-side_shift_avx: 159.3

vvc_h_loop_filter_chroma_8_strong_no-shift_c: 639.3
vvc_h_loop_filter_chroma_8_strong_no-shift_avx: 199.3

vvc_h_loop_filter_chroma_8_strong_shift_c: 1109.3
vvc_h_loop_filter_chroma_8_strong_shift_avx: 169.3

vvc_h_loop_filter_chroma_8_weak_no-shift_c: 309.3
vvc_h_loop_filter_chroma_8_weak_no-shift_avx: 129.3

vvc_h_loop_filter_chroma_8_weak_shift_c: 369.3
vvc_h_loop_filter_chroma_8_weak_shift_avx: 109.3

vvc_h_loop_filter_chroma_10_mix_no-shift_c: 439.3
vvc_h_loop_filter_chroma_10_mix_no-shift_avx: 169.3

vvc_h_loop_filter_chroma_10_mix_shift_c: 489.3
vvc_h_loop_filter_chroma_10_mix_shift_avx: 89.3

vvc_h_loop_filter_chroma_10_one-side_no-shift_c: 679.3
vvc_h_loop_filter_chroma_10_one-side_no-shift_avx: 149.3

vvc_h_loop_filter_chroma_10_one-side_shift_c: 909.3
vvc_h_loop_filter_chroma_10_one-side_shift_avx: 149.3

vvc_h_loop_filter_chroma_10_strong_no-shift_c: 909.3
vvc_h_loop_filter_chroma_10_strong_no-shift_avx: 169.3

vvc_h_loop_filter_chroma_10_strong_shift_c: 1129.3
vvc_h_loop_filter_chroma_10_strong_shift_avx: 159.3

vvc_h_loop_filter_chroma_10_weak_no-shift_c: 209.3
vvc_h_loop_filter_chroma_10_weak_no-shift_avx: 99.3

vvc_h_loop_filter_chroma_10_weak_shift_c: 259.3
vvc_h_loop_filter_chroma_10_weak_shift_avx: 89.3

vvc_h_loop_filter_chroma_12_mix_no-shift_c: 559.3
vvc_h_loop_filter_chroma_12_mix_no-shift_avx: 99.3

vvc_h_loop_filter_chroma_12_mix_shift_c: 799.3
vvc_h_loop_filter_chroma_12_mix_shift_avx: 209.3

vvc_h_loop_filter_chroma_12_one-side_no-shift_c: 679.3
vvc_h_loop_filter_chroma_12_one-side_no-shift_avx: 149.3

vvc_h_loop_filter_chroma_12_one-side_shift_c: 919.3
vvc_h_loop_filter_chroma_12_one-side_shift_avx: 139.3

vvc_h_loop_filter_chroma_12_strong_no-shift_c: 919.3
vvc_h_loop_filter_chroma_12_strong_no-shift_avx: 169.3

vvc_h_loop_filter_chroma_12_strong_shift_c: 1119.3
vvc_h_loop_filter_chroma_12_strong_shift_avx: 159.3

vvc_h_loop_filter_chroma_12_weak_no-shift_c: 219.3
vvc_h_loop_filter_chroma_12_weak_no-shift_avx: 99.3

vvc_h_loop_filter_chroma_12_weak_shift_c: 249.3
vvc_h_loop_filter_chroma_12_weak_shift_avx: 89.3

vvc_v_loop_filter_chroma_8_mix_no-shift_c: 439.3
vvc_v_loop_filter_chroma_8_mix_no-shift_avx: 329.3

vvc_v_loop_filter_chroma_8_mix_shift_c: 1799.3
vvc_v_loop_filter_chroma_8_mix_shift_avx: 369.3

vvc_v_loop_filter_chroma_8_one-side_no-shift_c: 709.3
vvc_v_loop_filter_chroma_8_one-side_no-shift_avx: 369.3

vvc_v_loop_filter_chroma_8_one-side_shift_c: 889.3
vvc_v_loop_filter_chroma_8_one-side_shift_avx: 369.3

vvc_v_loop_filter_chroma_8_strong_no-shift_c: 919.3
vvc_v_loop_filter_chroma_8_strong_no-shift_avx: 369.3

vvc_v_loop_filter_chroma_8_strong_shift_c: 1139.3
vvc_v_loop_filter_chroma_8_strong_shift_avx: 369.3

vvc_v_loop_filter_chroma_8_weak_no-shift_c: 429.3
vvc_v_loop_filter_chroma_8_weak_no-shift_avx: 339.3

vvc_v_loop_filter_chroma_8_weak_shift_c: 459.3
vvc_v_loop_filter_chroma_8_weak_shift_avx: 319.3

vvc_v_loop_filter_chroma_10_mix_no-shift_c: 629.3
vvc_v_loop_filter_chroma_10_mix_no-shift_avx: 319.3

vvc_v_loop_filter_chroma_10_mix_shift_c: 399.3
vvc_v_loop_filter_chroma_10_mix_shift_avx: 359.3

vvc_v_loop_filter_chroma_10_one-side_no-shift_c: 749.3
vvc_v_loop_filter_chroma_10_one-side_no-shift_avx: 339.3

vvc_v_loop_filter_chroma_10_one-side_shift_c: 959.3
vvc_v_loop_filter_chroma_10_one-side_shift_avx: 339.3

vvc_v_loop_filter_chroma_10_strong_no-shift_c: 949.3
vvc_v_loop_filter_chroma_10_strong_no-shift_avx: 339.3

vvc_v_loop_filter_chroma_10_strong_shift_c: 1179.3
vvc_v_loop_filter_chroma_10_strong_shift_avx: 329.3

vvc_v_loop_filter_chroma_10_weak_no-shift_c: 339.3
vvc_v_loop_filter_chroma_10_weak_no-shift_avx: 319.3

vvc_v_loop_filter_chroma_10_weak_shift_c: 439.3
vvc_v_loop_filter_chroma_10_weak_shift_avx: 309.3

vvc_v_loop_filter_chroma_12_mix_no-shift_c: 859.3
vvc_v_loop_filter_chroma_12_mix_no-shift_avx: 339.3

vvc_v_loop_filter_chroma_12_mix_shift_c: 729.3
vvc_v_loop_filter_chroma_12_mix_shift_avx: 379.3

vvc_v_loop_filter_chroma_12_one-side_no-shift_c: 739.3
vvc_v_loop_filter_chroma_12_one-side_no-shift_avx: 349.3

vvc_v_loop_filter_chroma_12_one-side_shift_c: 949.3
vvc_v_loop_filter_chroma_12_one-side_shift_avx: 339.3

vvc_v_loop_filter_chroma_12_strong_no-shift_c: 989.3
vvc_v_loop_filter_chroma_12_strong_no-shift_avx: 339.3

vvc_v_loop_filter_chroma_12_strong_shift_c: 1179.3
vvc_v_loop_filter_chroma_12_strong_shift_avx: 329.3

vvc_v_loop_filter_chroma_12_weak_no-shift_c: 359.3
vvc_v_loop_filter_chroma_12_weak_no-shift_avx: 329.3

vvc_v_loop_filter_chroma_12_weak_shift_c: 399.3
vvc_v_loop_filter_chroma_12_weak_shift_avx: 309.3

@stone-d-chen stone-d-chen requested a review from nuomi2021 January 3, 2025 22:09
@stone-d-chen
vvcdsp_init: call DEBLOCK_INIT for AVX2
fbab6e0 
The deblocking code only uses AVX so it can be initialized for both AVX and AVX2
@stone-d-chen stone-d-chen marked this pull request as ready for review January 4, 2025 20:06
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@nuomi2021 nuomi2021 Awaiting requested review from nuomi2021

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@stone-d-chen