Skip to content

decomp: add prefetch for matched seq on aarch64#3164

Merged
terrelln merged 1 commit intofacebook:devfrom
JunHe77:prefetch
Jul 29, 2022
Merged

decomp: add prefetch for matched seq on aarch64#3164
terrelln merged 1 commit intofacebook:devfrom
JunHe77:prefetch

Conversation

@JunHe77
Copy link
Contributor

@JunHe77 JunHe77 commented Jun 18, 2022

match is used for following sequence copy. It is
only updated when extDict is needed, which is a
low probability case. So it can be prefetched to
reduce cache miss.
The benchmarks on various Arm platforms showed
uplift from 1% ~ 14% with gcc-11/clang-14.

Signed-off-by: Jun He [email protected]
Change-Id: If201af4799d2455d74c79f8387404439d7f684ae

@JunHe77
decomp: add prefetch for matched seq on aarch64
a53c1cc 
match is used for following sequence copy. It is
only updated when extDict is needed, which is a
low probability case. So it can be prefetched to
reduce cache miss.
The benchmarks on various Arm platforms showed
uplift from 1% ~ 14% with gcc-11/clang-14.

Signed-off-by: Jun He <[email protected]>
Change-Id: If201af4799d2455d74c79f8387404439d7f684ae
@JunHe77
Copy link
Contributor Author

JunHe77 commented Jun 18, 2022

Benchmark changes on Arm N1/A72/A57 with gcc-11:

file N1/L2 N1/L8 N1/L15 A72/L2 A72/L8 A72/L15 A57/L2 A57/L8 A57/L15
dickens 3.144% 3.869% 3.967% 4.460% 4.389% 11.991% 5.341% 8.165% 12.130%
mr 1.192% 3.658% 3.606% 2.977% 6.147% 13.810% 1.826% 8.634% 12.909%
nci 0.737% 1.070% 1.811% -0.812% 1.065% 3.879% 0.164% 0.240% 2.399%
ooffice 0.134% 2.445% 2.389% -0.726% 0.080% 4.424% -0.306% 3.859% 5.377%
osdb 1.251% 2.669% 3.856% -0.125% 1.912% 6.158% 0.443% 3.182% 5.333%
reymont 2.639% 2.904% 3.641% 4.097% 3.379% 7.885% 3.649% 5.660% 9.412%
samba 1.129% 1.291% 1.393% 0.589% 1.950% 2.650% 0.758% 1.750% 1.401%
sao 0.728% 2.475% 2.633% -0.059% 2.364% 6.211% 0.633% 4.288% 6.761%
webster 2.138% 2.952% 2.483% 2.587% 5.809% 8.609% 3.530% 5.328% 7.289%
xml 0.604% 1.010% 0.934% -0.099% -0.110% 0.394% 0.497% 0.607% 1.045%
x-ray 0.102% 4.541% 4.543% -0.159% 6.198% 12.980% 0.102% 11.965% 14.064%

@JunHe77
Copy link
Contributor Author

JunHe77 commented Jun 18, 2022

Benchmark changes on Arm N1/A72/A57 with clang-14:

file N1/L2 N1/L8 N1/L15 A72/L2 A72/L8 A72/L15 A57/L2 A57/L8 A57/L15
dickens 2.755% 2.502% 2.235% 5.207% 6.836% 7.938% 5.987% 6.615% 6.607%
mozilla 0.741% 1.053% 1.032% 1.562% 2.610% 2.648% 1.066% 2.305% 2.180%
mr 2.270% 2.156% 2.110% 3.870% 4.890% 7.011% 3.409% 6.377% 6.168%
nci 0.841% 1.464% 2.739% 1.498% 2.558% 3.520% 1.122% 1.473% 2.337%
ooffice 1.103% 1.408% 1.704% 1.451% 3.761% 3.848% 1.181% 3.578% 3.808%
osdb 1.970% 2.545% 2.973% 4.665% 4.347% 5.006% 2.029% 3.241% 2.862%
reymont 2.379% 2.250% 2.490% 3.589% 5.267% 5.028% 4.517% 4.581% 4.335%
samba 1.738% 1.628% 1.501% 4.137% 3.137% 4.069% 2.723% 2.799% 3.032%
sao 0.445% 1.371% 1.314% 3.716% 4.120% 5.453% 1.299% 4.017% 4.459%
webster 2.782% 2.242% 1.893% 5.957% 5.264% 6.333% 4.216% 4.408% 3.738%
xml 1.512% 1.992% 1.557% 2.649% 1.462% 2.174% 2.044% 2.395% 2.080%
x-ray 0.363% 2.357% 2.477% 0.705% 7.155% 6.502% 0.721% 7.207% 6.986%

@JunHe77
Copy link
Contributor Author

JunHe77 commented Jun 18, 2022

This change may lead to regression on x86. My test on Xeon5218 is:

file gcc-11/L2 gcc-11/L8 gcc-11/L15 clang-14/L2 clang-14/L8 clang-14/L15
dickens -2.672% -3.085% -3.879% -0.514% -1.577% -3.390%
mozilla 0.813% 0.102% -0.066% -0.349% -0.721% -1.044%
mr -1.689% -3.238% -5.032% -0.313% -1.618% -3.501%
nci 1.174% 0.675% 1.035% -0.130% -0.362% -0.441%
ooffice -0.301% -1.163% -1.887% -0.427% -1.106% -1.597%
osdb 0.180% -1.443% 0.655% -0.138% -1.137% 1.217%
reymont -1.821% -0.914% -0.781% -0.493% -1.264% -1.666%
samba -0.278% -0.388% -0.472% -0.195% -0.410% -0.463%
sao -0.037% -2.148% -2.747% -0.184% -1.062% -1.971%
webster -1.548% -1.107% -0.989% -0.576% -1.172% -1.560%
xml 0.055% -0.056% 0.026% -0.036% -0.311% -0.431%
x-ray -0.155% -3.566% -4.536% -0.028% -2.175% -3.596%

@Cyan4973
Copy link
Contributor

Cyan4973 commented Jun 22, 2022

I believe this is a rather good policy.
It's surprising that it doesn't work for x64 though (I assume you meant x64 when you stated "x86").

There are a few quirks, such as overlapped prefetch orders when decoding in long mode, or useless prefetching when extDict is present (which might not be captured by benchmark). But these are refinements that can be worked on later on.

@terrelln terrelln merged commit 558cf20 into facebook:dev Jul 29, 2022
@Cyan4973 Cyan4973 mentioned this pull request Feb 9, 2023
Merged
@JunHe77 JunHe77 deleted the prefetch branch March 12, 2023 07:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Cyan4973 Cyan4973 Cyan4973 approved these changes

Assignees

No one assigned

Labels

CLA Signed

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@JunHe77 @Cyan4973 @facebook-github-bot @terrelln