Change default pipeline options to match production draw calls by walley892 · Pull Request #194597 · flutter/flutter

walley892 · 2026-10-16T21:26:58Z

#193319 revealed that in our golden test suite, 0% of draw calls were handled by a prewarmed pipeline variant.

This causes a first-frame shader cache miss 100% of the time - causing in the worst case a full re-compilation + link.

This change changes our defaults to better match what we actually request at draw time.

Running impeller_golden_tests with --shader-report before this change:

Draws covered by warmed variants: 0 of 620,533 (0.0%)

Running impeller_golden_tests with --shader-report after this change:

Draws covered by warmed variants: 610,441 of 620,533 (98.4%)

Without compiling additional variants, we can't get full coverage in our golden suite, but this change is directionally correct toward the goal of eliminating first-frame shader cache misses.

Fixes #194595

Pre-launch Checklist

walley892 · 2026-10-16T21:27:38Z

/gemini review

gemini-code-assist

Code Review

This pull request refactors pipeline prewarming configurations in ContentContext by introducing more granular options tailored to specific draw characteristics, such as triangle lists, triangle strips, opaque content, and subpasses without MSAA or depth/stencil. The feedback suggests changing the prewarming configuration for the fast_gradient pipeline from options_triangle to options_trianglestrip to prevent potential shader compilation misses at draw time when filling common geometries.

gemini-code-assist · 2026-10-16T21:29:01Z

+    pipelines_->solid_fill.CreateDefault(*context_,
+                                         options_opaque_trianglestrip);
+    pipelines_->texture.CreateDefault(*context_, options_trianglestrip);
+    pipelines_->fast_gradient.CreateDefault(*context_, options_triangle);


The fast_gradient pipeline is warmed up here using options_triangle. However, like all other gradient pipelines (e.g., linear_gradient_ssbo_fill, radial_gradient_ssbo_fill), fast gradients are primarily used to fill geometries such as rectangles, rounded rectangles, and circles, which are drawn as triangle strips (PrimitiveType::kTriangleStrip). Using options_triangle here will cause a shader compilation miss at draw time when a fast gradient is used with these common geometries. It should be changed to options_trianglestrip to match the other gradient pipelines and ensure proper prewarming coverage.

Suggested change

pipelines_->fast_gradient.CreateDefault(*context_, options_triangle);

pipelines_->fast_gradient.CreateDefault(*context_, options_trianglestrip);

The only use of the FastGradientPipeline pipeline is in LinearGradientContents, which uses triangles

flutter/engine/src/flutter/impeller/entity/contents/linear_gradient_contents.cc

Line 161 in 15a5a74

return GeometryResult{

gaaclarke

#193319 revealed that in our golden test suite, 0% of draw calls were handled by a prewarmed pipeline variant.

My concern here is about over optimizing to match the a test runner's needs instead of what real world apps need. Did you look at this for doing something like running gallery or wonderous?

gemini-code-assist

Code Review

This pull request refactors pipeline initialization in ContentContext by introducing and applying more specific ContentContextOptions configurations, such as triangle versus triangle strip primitives, opaque variants, and configurations without MSAA or depth/stencil attachments. Unit tests are updated to align with these changes. Review feedback recommends resetting stencil_compare to CompareFunction::kAlways to prevent pipeline variants from inheriting non-default stencil modes, and initializing the fast_gradient pipeline with options_trianglestrip instead of options_triangle to prevent runtime shader cache misses.

gemini-code-assist · 2026-10-16T22:06:36Z

+    front_stencil.stencil_failure = StencilOperation::kKeep;
+    front_stencil.depth_stencil_pass = StencilOperation::kKeep;


When creating pipeline variants, they are initialized from the default variant's descriptor. If the default variant uses a non-default stencil mode (such as kCoverCompareInverted used by clips), any stencil fields modified by that mode will be inherited by variants that use StencilMode::kIgnore because kIgnore does not overwrite them. To prevent this, stencil_compare should be reset to CompareFunction::kAlways along with the other stencil operations.

Suggested change

front_stencil.stencil_failure = StencilOperation::kKeep;

front_stencil.depth_stencil_pass = StencilOperation::kKeep;

front_stencil.stencil_compare = CompareFunction::kAlways;

front_stencil.stencil_failure = StencilOperation::kKeep;

front_stencil.depth_stencil_pass = StencilOperation::kKeep;

walley892 · 2026-10-16T23:11:25Z

My concern here is about over optimizing to match the a test runner's needs instead of what real world apps need. Did you look at this for doing something like running gallery or wonderous?

You're right that the test runner is a proxy/imperfect signal, and it's what flagged out these defaults as bad initially. But, looking at the callstack, you can see that the defaults we used were in principle never able to be used.

For example, the CirclePipeline geometry is always specified as kTriangeStrip and never as kTrainge (the default before this change)

flutter/engine/src/flutter/impeller/tessellator/tessellator.cc

Line 496 in 3d4dafa

PrimitiveType::kTriangleStrip, 4,

and the depth mode is always set as kGreaterEqual and never as kAlways (the default before this change)

flutter/engine/src/flutter/impeller/entity/contents/contents.cc

Line 29 in 3d4dafa

opts.depth_compare = CompareFunction::kGreaterEqual;

Because of the defaults before, no default pipelines matched the requested variant - so this change should be an improvement across all apps.

I ran an ad-hoc test on Wondrous:

Draws hitting a warmed variant: 0.0% before -> 76.2% with these changes

gaaclarke

Okay, sounds good. Code is looking good. Let's get an automated devicelab test that reflects the thing you are optimizing and let's land it before this so we can get the measurement of the improvment. It's probably easiest to make a variant of the new_gallery test and measure the number of hit warmed up shaders.

gaaclarke · 2026-10-16T23:26:11Z

Otherwise there's no way to make sure we maintain this if it is something we care about.

Although, are you expecting this to show up in some of the existing time benchmarks in a meaningful way?

walley892 · 2026-10-16T23:35:35Z

Although, are you expecting this to show up in some of the existing time benchmarks in a meaningful way?

Yeah, this should reduce the time to first frame rasterized metrics.

It's probably easiest to make a variant of the new_gallery test and measure the number of hit warmed up shaders

This is a good idea to prevent this happening again when we add new shaders. Will do

gaaclarke · 2026-10-16T23:52:52Z

Yea, the nice thing about counting them directly is that it's less noisy.

…riants

walley892 · 2026-10-27T18:06:04Z

Benchmark has landed, showing in luci

https://flutter-flutter-perf.luci.app/e?begin=1793036891&enable_chart_tooltip=true&end=1793123291&plotSummary=true&queries=sub_result%3Dshader_warmed_draw_percent&use_titles=true

gaaclarke

lgtm!

auto-submit · 2026-10-27T18:59:22Z

autosubmit label was removed for flutter/flutter/194597, because - The status or check suite Presubmit has failed. Please fix the issues identified (or deflake) before re-applying this label.

walley892 · 2026-10-27T21:15:46Z

Oops. Test failure from the rebase on the benchmark PR

flutter/flutter@d4c8577...0c2473c 2026-10-28 [email protected] Roll Packages from a8549ae to 13d38e0 (9 revisions) (flutter/flutter#195488) 2026-10-28 [email protected] Use int.oneBitCount in PersistentHashMap and scope 53-bit caps to JS (flutter/flutter#195005) 2026-10-28 [email protected] Roll pub packages (flutter/flutter#195483) 2026-10-28 [email protected] [iPhone Duo] Use the correct UIScreen to report the display size and update when fold/unfold (flutter/flutter#195199) 2026-10-28 [email protected] [fuchsia] Fix portable_ui_test.h includes for newer Fuchsia SDKs (flutter/flutter#195420) 2026-10-28 [email protected] Change default pipeline options to match production draw calls (flutter/flutter#194597) 2026-10-28 [email protected] Roll Fuchsia Test Scripts from 2T18KC7-LDQtmQiLP... to xl1LWBuppzu-BnOzV... (flutter/flutter#195450) 2026-10-27 [email protected] [flutter_tools] Inject dependencies into Darwin build targets, native assets, and Xcode build settings (flutter/flutter#194052) 2026-10-27 [email protected] Revert: Reland square corners fix (flutter/flutter#195436) 2026-10-27 [email protected] [flutter_tools] Add DeviceService app launch and flutter run support for tool extensions (flutter/flutter#194589) 2026-10-27 [email protected] [engine] Honor targetWidth/targetHeight for animated images (flutter/flutter#194615) 2026-10-27 [email protected] [Android][a11y] Set visibleToUser=false for hidden semantics nodes (flutter/flutter#194411) 2026-10-27 [email protected] Fix repo deletion test flake (flutter/flutter#195349) 2026-10-27 [email protected] Roll Packages from 23a05b6 to a8549ae (10 revisions) (flutter/flutter#195434) 2026-10-27 [email protected] [web] Capture pointers to the Flutter view so drags survive platform view iframes (flutter/flutter#195187) 2026-10-27 [email protected] [web] Add opt-in raw samples, measured sample count and warm-up phase hook to the web benchmark recorder (flutter/flutter#195118) 2026-10-27 [email protected] Link migration guide in deprecation messages (flutter/flutter#195361) 2026-10-27 [email protected] Pin basic_material_app_macos__compile to macOS 26 (flutter/flutter#195429) 2026-10-27 [email protected] Add benchmark for warmed pipeline usage (flutter/flutter#195001) 2026-10-27 [email protected] [flutter_tools] Add device log streaming support to DeviceService and ExtensionBackedDevice (flutter/flutter#194964) 2026-10-27 [email protected] Roll Dart SDK from 1d16c04dc346 to 65fc979759ca (4 revisions) (flutter/flutter#195383) If this roll has caused a breakage, revert this CL and set the roller to dry run mode using the controls here: https://autoroll.skia.org/r/flutter-packages Please CC [email protected],[email protected] on the revert to ensure that a human is aware of the problem. To file a bug in Packages: https://github.com/flutter/flutter/issues/new/choose To report a problem with the AutoRoller itself, please file a bug: https://issues.skia.org/issues/new?component=1389291&template=1850622 Documentation for the AutoRoller is here: https://skia.googlesource.com/buildbot/+doc/main/autoroll/README.md

…er#194597) flutter#193319 revealed that in our golden test suite, 0% of draw calls were handled by a prewarmed pipeline variant. This causes a first-frame shader cache miss 100% of the time - causing in the worst case a full re-compilation + link. This change changes our defaults to better match what we actually request at draw time. Running impeller_golden_tests with --shader-report before this change: ``` Draws covered by warmed variants: 0 of 620,533 (0.0%) ``` Running impeller_golden_tests with --shader-report after this change: ``` Draws covered by warmed variants: 610,441 of 620,533 (98.4%) ``` Without compiling additional variants, we can't get full coverage in our golden suite, but this change is directionally correct toward the goal of eliminating first-frame shader cache misses. Fixes flutter#194595 ## Pre-launch Checklist - [x] I read the [Contributor Guide] and followed the process outlined there for submitting PRs. - [x] I read the [AI contribution guidelines] and understand my responsibilities, or I am not using AI tools. - [x] I read the [Tree Hygiene] wiki page, which explains my responsibilities. - [x] I read and followed the [Flutter Style Guide], including [Features we expect every widget to implement]. - [x] I signed the [CLA]. - [x] I listed at least one issue that this PR fixes in the description above. - [x] I updated/added relevant in-code documentation (doc comments with `///`). - [x] If this PR introduces a new feature or capability, I created and linked a website documentation issue or PR in [flutter/website] (or verified none is needed). - [x] I added new tests to check the change I am making, or this PR is [test-exempt]. - [x] I followed the [breaking change policy] and added [Data Driven Fixes] where supported. - [x] All existing and new tests are passing.

Change default pipeline options to match production draw calls

4b9af9d

github-actions Bot added engine flutter/engine related. See also e: labels. e: impeller Impeller rendering backend issues and features requests labels Oct 16, 2026

gemini-code-assist Bot reviewed Oct 16, 2026

View reviewed changes

Fix golden diffs

2e61fdf

walley892 added the CICD Run CI/CD label Oct 16, 2026

walley892 requested a review from gaaclarke October 16, 2026 22:02

walley892 marked this pull request as ready for review October 16, 2026 22:02

gaaclarke requested changes Oct 16, 2026

View reviewed changes

gemini-code-assist Bot reviewed Oct 16, 2026

View reviewed changes

walley892 requested a review from gaaclarke October 16, 2026 23:13

gaaclarke requested changes Oct 16, 2026

View reviewed changes

Merge remote-tracking branch 'upstream/master' into change_default_va…

e4818ac

…riants

walley892 requested a review from gaaclarke October 27, 2026 18:06

gaaclarke previously approved these changes Oct 27, 2026

View reviewed changes

walley892 added the autosubmit Merge PR when tree becomes green via auto submit App label Oct 27, 2026

auto-submit Bot removed the autosubmit Merge PR when tree becomes green via auto submit App label Oct 27, 2026

Fix tests

007d480

walley892 dismissed gaaclarke’s stale review via 007d480 October 27, 2026 21:15

walley892 requested a review from gaaclarke October 27, 2026 21:15

gaaclarke approved these changes Oct 27, 2026

View reviewed changes

walley892 added the autosubmit Merge PR when tree becomes green via auto submit App label Oct 27, 2026

walley892 added this to Impeller Fast Start Oct 27, 2026

auto-submit Bot added this pull request to the merge queue Oct 28, 2026

Merged via the queue into flutter:master with commit 1e47dc5 Oct 28, 2026
24 checks passed

flutter-dashboard Bot removed the autosubmit Merge PR when tree becomes green via auto submit App label Oct 28, 2026

github-project-automation Bot moved this to Done in Impeller Fast Start Oct 28, 2026

This was referenced Oct 28, 2026

Roll Flutter from d4c857754c0e to 97cd226ec5ee (19 revisions) flutter/packages#13469

Closed

Roll Flutter from d4c857754c0e to 0c2473cd50c8 (21 revisions) flutter/packages#13470

Merged

	pipelines_->fast_gradient.CreateDefault(*context_, options_triangle);
	pipelines_->fast_gradient.CreateDefault(*context_, options_trianglestrip);

		front_stencil.stencil_failure = StencilOperation::kKeep;
		front_stencil.depth_stencil_pass = StencilOperation::kKeep;

Conversation

walley892 commented Oct 16, 2026

Pre-launch Checklist

Uh oh!

walley892 commented Oct 16, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Oct 16, 2026

Choose a reason for hiding this comment

Uh oh!

walley892 Oct 16, 2026

Choose a reason for hiding this comment

Uh oh!

gaaclarke left a comment

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Oct 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

walley892 commented Oct 16, 2026

Uh oh!

gaaclarke left a comment

Choose a reason for hiding this comment

Uh oh!

gaaclarke commented Oct 16, 2026

Uh oh!

walley892 commented Oct 16, 2026

Uh oh!

gaaclarke commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

walley892 commented Oct 27, 2026

Uh oh!

gaaclarke left a comment

Choose a reason for hiding this comment

Uh oh!

auto-submit Bot commented Oct 27, 2026

Uh oh!

walley892 commented Oct 27, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

gaaclarke commented Oct 16, 2026 •

edited

Loading