Impeller causes severe performance regression on older Adreno GPUs (e.g. Adreno 506) #187009

Activity

1. puneetkukreja98

   added

   in triagePresently being triaged by the triage teamPresently being triaged by the triage team

   on May 25, 2026
2. puneetkukreja98 commented on May 25, 2026

   

   puneetkukreja98

   on May 25, 2026

   Contributor

   More actions

   @andrea75 Thanks for reporting the issue! Could you please try updating to the latest stable version (3.44.0) and check if the issue still persists? Also, if it does, is this performance regression something you only started noticing on this specific version (3.41.9), or did it occur on older versions as well?

   Similar issues:
   #162312
   #180958
   #169423
3. puneetkukreja98

   added

   waiting for responseThe Flutter team cannot make further progress on this issue until the original reporter respondsThe Flutter team cannot make further progress on this issue until the original reporter responds

   and removed

   in triagePresently being triaged by the triage teamPresently being triaged by the triage team

   on May 25, 2026
4. andrea75 commented on May 25, 2026

   

   andrea75

   on May 25, 2026

   Last edited by andrea75

   Author

   More actions

   @puneetkukreja98 thank you for your answer.

   Tested on Flutter 3.44.0 — issue persists.

   Device: Xiaomi Redmi 5 Plus (Adreno 506, Android 8.1)

   With Impeller enabled (opt-out removed):

   32 FPS average, both idle and during scroll
   100% of frames flagged as jank during scroll
   Build phase reaching 35ms on idle frames
   Repeated FlutterImageDecoderImplDefault: invalid input errors (image decode failures)
   With io.flutter.embedding.android.EnableImpeller = false:

   ~54 FPS, smooth scroll, no jank
   The runtime GPU blocklist from #160041 does not appear to cover the Adreno 506 in Flutter 3.44.0. The manual opt-out remains necessary.

   

   👍React with 👍1Reacted by Puneet kukreja
5. github-actions

   removed

   waiting for responseThe Flutter team cannot make further progress on this issue until the original reporter respondsThe Flutter team cannot make further progress on this issue until the original reporter responds

   on May 25, 2026
6. puneetkukreja98

   added

   in triagePresently being triaged by the triage teamPresently being triaged by the triage team

   engineflutter/engine related. See also e: labels.flutter/engine related. See also e: labels.

   e: impellerImpeller rendering backend issues and features requestsImpeller rendering backend issues and features requests

   team-engineOwned by Engine teamOwned by Engine team

   platform-androidAndroid applications specificallyAndroid applications specifically

   and removed

   in triagePresently being triaged by the triage teamPresently being triaged by the triage team

   on May 26, 2026
7. puneetkukreja98 commented on May 26, 2026

   

   puneetkukreja98

   on May 26, 2026

   Contributor

   More actions

   Routing this to team-engine for further investigation.

   👍React with 👍1Reacted by andrea75
8. gaaclarke commented on Jun 1, 2026

   

   gaaclarke

   on Jun 1, 2026

   Member

   More actions

   Are these devices using the vulkan or opengles backend? Did you attempt to bisect the issue a bit to see if there was a regression at some point in the performance?
9. gaaclarke

   added

   P2Important issues not at the top of the work listImportant issues not at the top of the work list

   triaged-engineTriaged by Engine teamTriaged by Engine team

   on Jun 1, 2026
10. andrea75 commented on Jun 1, 2026

    

    andrea75

    on Jun 1, 2026

    Author

    More actions

    Are these devices using the vulkan or opengles backend? Did you attempt to bisect the issue a bit to see if there was a regression at some point in the performance?

    Test results on Flutter 3.44.0:

    1. Impeller + OpenGL ES (EnableImpeller=true + EnableVulkan=false)
       → 31 FPS average, UI Jank + Raster Jank detected, Build phase 34.8 ms

    

    2. Impeller disabled (Skia) (EnableImpeller=false)
       → 55 FPS average, no jank, clean frames

    

    The issue is not Vulkan-specific — OpenGL ES backend also causes significant performance degradation on this device. Impeller itself (regardless of backend) appears to be the regression source on Adreno 506 with Flutter 3.44.0.

    We did not bisect across intermediate versions (3.42.x / 3.43.x) — let us know if that would be useful and we'll try.
11. gaaclarke commented on Jun 1, 2026

    

    gaaclarke

    on Jun 1, 2026

    Member

    More actions

    Yes, trying a few versions of flutter would be helpful. If it's a regression it's much more actionable.
12. gaaclarke commented on Jun 1, 2026

    

    gaaclarke

    on Jun 1, 2026

    Member

    More actions

    @andrea75 can you verify that this is the case for the default counter app too with flutter? I wonder if you are measuring an app with platform views.
13. andrea75 commented on Jun 2, 2026

    

    andrea75

    on Jun 2, 2026

    Author

    More actions

    The measurement was taken on the main screen with no WebViews / platform views active — only standard Flutter widgets. So platform views are not the cause.

    We'll verify with the default counter app and report back. We'll also try to bisect across 3.42.x / 3.43.x to narrow down the regression.
14. MostafaAlyy commented on Sep 2, 2026

    

    MostafaAlyy

    on Sep 2, 2026

    More actions

    Adding lab and field data for the same class of device (Adreno 610), with answers to the three questions above and a per-primitive attribution.

    Backend: OpenGL ES. Adreno 610 is on the Vulkan denylist (driver_info_vk.cc, Adreno ≤ 650), so logcat prints Using the Impeller rendering backend (Vulkan) immediately followed by … (OpenGLES). Every number below is Impeller GLES vs Skia GL on the same GPU and driver (Adreno (TM) 610, OpenGL ES 3.2 [email protected], vivo V2424, Android 16, 120 Hz panel).

    Not platform views. Everything below is a plain ListView.builder of 40 Material cards — no platform view, no texture, no image decoding.

    Not a version regression as far as we can tell. We did not bisect; our field numbers were flat across three months of releases, and the only thing that moved them was switching the renderer.

    Minimal repro

    https://github.com/MostafaAlyy/impeller_gles_repro — 40 cards, decorations toggled with --dart-define=SCENE=plain|shadow|gradient|opacity|blur|shader|all. plain is a rounded Container + border + two Texts + LinearProgressIndicator + a gradient thumbnail; shadow adds three offset BoxShadows (blur 8/16/24); gradient a LinearGradient fill; opacity an Opacity(0.92) wrapper; shader a ShaderMask; blur a BackdropFilter(blur 12) on every 4th card. The list flings itself end-to-end forever; FrameTimings are aggregated per 3 s window and printed to logcat. Flutter 3.47.2 stable, engine a804b26. measure.sh in the repo runs the whole matrix; RESULTS-all.tsv has every row.

    Release mode, vivo V2424 (Adreno 610)

    p50 frame / p50 raster / share of frames over 16.7 ms:

    Scene	Impeller GLES	Skia GL
    plain	12.0 ms / 8.5 ms / 10 %	6.7 ms / 3.4 ms / 0 %
    blur	46.8 ms / 28.8 ms / 99 %	12.9 ms / 8.7 ms / 34 %
    all	71.7 ms / 41.7 ms / 100 %	15.8 ms / 10.3 ms / 24 %

    Per-primitive attribution (profile mode, same device)

    Scene	Impeller GLES	Skia GL
    plain	20.5 / 12.5 / 54 %	5.9 / 3.0 / 0 %
    shadow	12.4 / 8.6 / 15 %	8.5 / 4.3 / 1 %
    gradient	13.5 / 9.9 / 24 %	6.6 / 3.1 / 0 %
    opacity	13.4 / 10.3 / 22 %	6.3 / 2.7 / 0 %
    shader (ShaderMask)	13.8 / 10.8 / 25 %	7.4 / 3.7 / 0 %
    blur (BackdropFilter ×5)	46.3 / 28.6 / 99 %	9.5 / 6.9 / 6 %
    all	73.2 / 41.8 / 100 %	13.3 / 9.9 / 25 %
    all + forced 30 fps repaint	69.6 / 40.8 / 100 %	31.8 / 18.2 / 100 %

    The Dart VM timeline (Embedder stream) separates two costs:

    1. A flat scene is CPU-bound in GL command encoding. plain under Impeller GLES is a single render pass per frame, but RenderPassGLES::EncodeCommandsInReactor takes 6.5 ms of raster-thread time per frame (SurfaceFrame::Encode 8.5 ms, GPURasterizer::Draw p50 10.4 ms), and top -H shows the raster thread at ~45 % of a core. Skia on the same scene: GrDirectContext::flushAndSubmit 1.6 ms, Draw 3.9 ms, raster thread ~18 %. Reading render_pass_gles.cc, buffer_bindings_gles.cc and sampler_gles.cc, every command re-issues its full state; only the viewport, cull mode and winding are checked for redundancy. Per draw that is BlendFuncSeparate + enable/disable, StencilFuncSeparate/StencilOpSeparate per face, depth enable + DepthFunc + DepthMask, Scissor, per attribute EnableVertexAttribArray + VertexAttribPointer + VertexAttribDivisor, an unconditional UseProgram, the y-flip Uniform1fv, BindBufferRange per UBO, and per texture ActiveTexture + BindTexture + four to six TexParameteri (no sampler objects) + Uniform1i — roughly 25–35 GL calls per draw. Skia's GrGLGpu state cache and op batching avoid most of those on the same driver.

    2. Blur is GPU-bound on pass count. Five BackdropFilters make the frame 9.35 render passes (EncodeCommandsInReactor count per frame). Encoding is only 5.3 ms, but SurfaceFrame::Submit is 25.5 ms — about 18 ms per frame blocked in swap waiting for the GPU. all stacks both: 15.8 passes per frame, 12.7 ms of encoding, 38.7 ms Draw.

    A note for anyone measuring this: impeller_debug is on in profile engines (impeller/tools/args.gni), so profile builds also pay a PushDebugGroup/PopDebugGroup per command — plain reads 20 ms in profile and 12 ms in release on this device. The release numbers above are the honest ones.

    Same repro on the emulator (Android 16 x86_64, host GPU Intel Arc / Mesa)

    Scene	Impeller GLES	Skia GL	Impeller Vulkan (forced via ImpellerBackend)
    plain	5.1 / 3.7 / 0 %	5.0 / 3.5 / 0 %	47.9 / 27.5 / 98 %
    blur	16.9 / 12.1 / 46 %	4.9 / 3.5 / 0 %	85.1 / 47.0 / 99 %
    all	22.0 / 14.4 / 63 %	6.1 / 4.3 / 0 %	75.7 / 40.7 / 99 %

    With a desktop GPU behind the translator the encoding cost disappears (plain ties), but the blur/pass cost is still ~3.5× Skia. The forced-Vulkan column goes through the gfxstream translation layer and is not representative of a phone; it only shows why auto-select avoids it.

    Production app (same APK, renderer toggled via --enable-impeller=false)

    Screen	Device	Impeller GLES	Skia GL
    Home fling, p50 frame	Galaxy Tab A7 SM-T505N (Adreno 610, Android 12)	43–107 ms	8–10 ms
    Video page (texture-backed platform view)	vivo V2424	86 ms (24 fps)	26 ms (59 fps)
    Same page, error state, no video	vivo V2424	75 ms	13 ms

    Firebase Analytics across our Android fleet, per model, video-screen p50 frame before → after the switch to Skia: SM-T505N 52–69 ms → 23–28 ms, CPH2591 51 ms → 27 ms, SM-X216B (Adreno 619) 25 ms → 19 ms. We are shipping --enable-impeller=false on Android for now.

    If a per-pass GL state cache in RenderPassGLES (skipping redundant program / blend / stencil / depth / vertex-attribute / texture-parameter calls, and sampler objects on ES 3.0) is a direction the team would take, I'm happy to put a PR together — we have the Adreno 610 devices on the desk to measure it.
15. MostafaAlyy commented on Sep 2, 2026

    

    MostafaAlyy

    on Sep 2, 2026

    More actions

    Filed the encode-cost half of this (RenderPassGLES re-issuing all GL state per draw, raster thread CPU-bound on Adreno 610) as its own issue with the mechanism and a proposed fix: #192147. The blur/pass-count half stays with #169423.
16. andrea75 commented on Sep 2, 2026

    

    andrea75

    on Sep 2, 2026

    Author

    More actions

    Thanks to @MostafaAlyy for the detailed analysis on Adreno 610 (same GPU family, same symptoms). The root cause they identified — RenderPassGLES re-issuing all GL state per draw — matches our observations on Adreno 506 exactly.

    Tracking the proposed fix in #192147.

    ❤️React with ❤️1Reacted by Mostafa Ali