commit | 7db7d624eb69d6cc85065c84319f6726daf4a8a0 | [log] [tgz] |
---|---|---|
author | Romain Guy | Thu May 30 17:33:23 2024 -0700 |
committer | Romain Guy | Fri May 31 13:10:50 2024 -0700 |
tree | 17f5c8190cedb65387573d0abb9bc2d7ef72f4fd | |
parent | 27fe9cfbaa47a0f8ca8b0a0033990e4feca372e8 [diff] |
Optimize Matrix computations The implementation of graphics layer needs to apply a TRS matrix (translation/rotation/scale) which was previously done by computing separate transforms successively. This was not only inefficient from a CPU perspective, it also led to the allocation of 3 temporary matrices. This was the cause of thousands/dozens of thousands of allocations when scrolling lazy containers. This CL combines all the operations necessary to create a TRS matrix around a pivot point into a single method called resetToPivotedTransform(). With this change, updating a layer's matrix goes from 1062 instructions and 39 branches down to 168 instructions and 1 branch. More importantly, all allocations are eliminated. Another important optimization is the addition of array size checks all over the file. By construction we know the array always has a size of 16, but the runtime cannot guarantee this and inserts a lot of bounds checks. This is especially bad when we do multiple accesses as it generates a lot of instructions and branches we do not need. By adding size checks manually, we give a hint to ART that the check should happen only once, greatly reducing the number of extra instructions. For instance in the isIdentity() method, this removes half of the instructions and half of the branches. Additional optimizations: - Implement reset as a series of assignments. Saves 9 instructions and 6 branches - Do the same in setFrom, saving 7 branches for a few more instructions - Inline dot(), making timesAssign go from 201 instructions and 7 branches down to 189 and 4 - Rewrite isIdentity without a loop. For matrices that benefit from early returns, it runs 10% faster. For identity matrices, it runs 80% faster Test: MatrixTest, compose-ui tests RelNote: N/A Change-Id: Ia478735b8ddcc182ac1aaf82d013e4e14832a9c5
Jetpack is a suite of libraries, tools, and guidance to help developers write high-quality apps easier. These components help you follow best practices, free you from writing boilerplate code, and simplify complex tasks, so you can focus on the code you care about.
Jetpack comprises the androidx.*
package libraries, unbundled from the platform APIs. This means that it offers backward compatibility and is updated more frequently than the Android platform, making sure you always have access to the latest and greatest versions of the Jetpack components.
Our official AARs and JARs binaries are distributed through Google Maven.
You can learn more about using it from Android Jetpack landing page.
For contributions via GitHub, see the GitHub Contribution Guide.
Note: The contributions workflow via GitHub is currently experimental - only contributions to the following projects are being accepted at this time:
When contributing to Jetpack, follow the code review etiquette.
We are not currently accepting new modules.
Head over to the onboarding docs to learn more about getting set up and the development workflow!
Our continuous integration system builds all in progress (and potentially unstable) libraries as new changes are merged. You can manually download these AARs and JARs for your experimentation.
Before uploading your first contribution, you will need setup a password and agree to the contribution agreement:
Generate a HTTPS password: https://android-review.googlesource.com/new-password
Agree to the Google Contributor Licenses Agreement: https://android-review.googlesource.com/settings/new-agreement
AndroidX uses git to store all the binary Gradle dependencies. They are stored in prebuilts/androidx/internal
and prebuilts/androidx/external
directories in your checkout. All the dependencies in these directories are also available from google()
, or mavenCentral()
. We store copies of these dependencies to have hermetic builds. You can pull in a new dependency using our importMaven tool.