Skip to content

Further performance improvements for FP math#905

Merged
isaacbrodsky merged 6 commits into
uber:masterfrom
heshpdx:master
Sep 19, 2024
Merged

Further performance improvements for FP math#905
isaacbrodsky merged 6 commits into
uber:masterfrom
heshpdx:master

Conversation

@heshpdx

@heshpdx heshpdx commented Sep 9, 2024

Copy link
Copy Markdown
Contributor

I was requested to dig further in the spirit of !852 so I spent time inspecting other parts of the code. This PR unlocks more FDIV->FMUL opportunities. I inspected the assembly and confirmed that each of the changes did result in a transformation of FDIV to FMUL.

Here are the perf outliers, with the majority of other tests in the noise (since likely they do not exercise the changed code).

Grid path cells shows 4.3% upside, as measured on a Ampere Altra in the OCI Cloud, using gcc-10 out of the box. Regular h3 cmake build.
mainline:

$ ./build/bin/benchmarkGridPathCells 
        -- gridPathCellsNear: 39.136827 microseconds per iteration (10000 iterations)
        -- gridPathCellsFar: 1799.592192 microseconds per iteration (1000 iterations)

branch:

$ ./build_fast/bin/benchmarkGridPathCells 
        -- gridPathCellsNear: 37.556445 microseconds per iteration (10000 iterations)
        -- gridPathCellsFar: 1724.603344 microseconds per iteration (1000 iterations)

The experimental polygon to cells overlapping is also an outlier, 1-2%:

mainline:

        -- polygonToCellsSF_Overlapping: 5817.474534 microseconds per iteration (500 iterations)
        -- polygonToCellsAlameda_Overlapping: 13169.533630 microseconds per iteration (500 iterations)
        -- polygonToCellsSouthernExpansion_Overlapping: 381815.775600 microseconds per iteration (10 iterations)

branch:

        -- polygonToCellsSF_Overlapping: 5773.028512 microseconds per iteration (500 iterations)
        -- polygonToCellsAlameda_Overlapping: 12886.969474 microseconds per iteration (500 iterations)
        -- polygonToCellsSouthernExpansion_Overlapping: 377545.843300 microseconds per iteration (10 iterations)

@heshpdx
Further performance improvements for FP math
4604d1f 
More FDIV->FMUL opportunities unlocked, following in the
spirit of #852
@coveralls

coveralls commented Sep 9, 2024 •
edited
Loading

Copy link
Copy Markdown

Coverage Status

coverage: 98.824% (-0.002%) from 98.826%
when pulling b7542c3 on heshpdx:master
into a1037eb on uber:master.

nrabinowitz
nrabinowitz approved these changes Sep 10, 2024
View reviewed changes

@nrabinowitz nrabinowitz left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for these optimizations!

Comment thread src/h3lib/lib/localij.c Outdated
@heshpdx @nrabinowitz
Update src/h3lib/lib/localij.c
dd45fbe 
Co-authored-by: Nick Rabinowitz <public@nickrabinowitz.com>
@heshpdx

heshpdx commented Sep 18, 2024

Copy link
Copy Markdown
Contributor Author

Looks like the CIFuzz is broken. Do we need to update that runner?

@isaacbrodsky

isaacbrodsky commented Sep 18, 2024

Copy link
Copy Markdown
Collaborator

Looks like the CIFuzz is broken. Do we need to update that runner?

I'm addressing it in #907. You may need to rebase or merge that change in too

@isaacbrodsky
isaacbrodsky merged commit 4899d29 into uber:master Sep 19, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@nrabinowitz nrabinowitz nrabinowitz approved these changes
@isaacbrodsky isaacbrodsky isaacbrodsky approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@heshpdx @coveralls @isaacbrodsky @nrabinowitz