skia — Engineering Performance
24 engineers all time · Jan 2025 – Nov 2026 · built 2026-11-22 · GitHub
Performance snapshot
Today's rolling 90-day reading for skia, compared with the start of the series. Pick a window to move that comparison point.
Avg. perf / dev / mo
+159.8%
0.83 → 2.17 ETV
Active engineers
−38.1%
21.0 → 13.0
Features
−9.1pp
25.6% → 16.5%
vs. Google
0.59x
0.89x → 0.59x · −41% below
skia vs. Google
Per-engineer ETV for skia against Google as a whole. Both lines are 90-day rolling averages scaled to a 30-day month, so they share one axis and can be read against each other at any point. Pick a window to zoom the chart to it.
Performance Composition
Each month's output split by type of work: Features (new value), Maintenance (sustaining systems), Tests, Docs, and Fixes (rework). The yellow line is output per engineer, so when it rises each engineer is delivering more, whatever the team size did. Unit: Engineering Throughput Value (ETV).
Engineering capacity
Effective engineers behind skia, against its pre-AI baseline. Each subject has its own: skia's is 0.83 ETV / dev / mo, its first reading in Q1 2025. Per-engineer ETV divided by that gives a capacity multiple, and that multiple applied to the engineers active in the trailing 90 days turns it into engineer-equivalents. The line is the real headcount, so the gap between line and area is what the leverage is worth. Because each baseline is its own, every subject opens at 1.0x on its first day: multiples measure improvement and are not comparable between subjects.
Knowledge concentration
How dependent is this repo on a small number of engineers? Higher top-1 share = higher key-person risk.
Kaylee Lubick owns 16.4 % of commits.
Behind the numbers
Written summary of the work completed each month.
No monthly reports available yet.
Top engineers
Most impactful commits
Top 10 by ETV in the all-time window.
- 8.6ETV[skparagraph] Split SkParagraphTest.cpp by topic Almost every SkParagraph change adds its tests at the end of SkParagraphTest.cpp, so any two open CLs conflict there and each landing forces the others to rebase. Of 71 skparagraph/skunicode/skshaper commits since 2026-09-01, 56 touch this file, and 39 open CLs touch it now. Move the flags, macros, EPSILON constants, string helpers, ResourceFontCollection and TestCanvas into SkParagraphTestUtils.{h,cpp} in namespace skia::textlayout::test. Only linkage changes: the flags are no longer static, the EPSILON constants are inline constexpr, and getTypefaces is a file-static function. The header declares the flags as plain extern variables, since clang-format misparses DECLARE_string and DECLARE_bool, which have no trailing ';'. Then move all 248 tests, in their original order, into 18 files named SkParagraph<Topic>Test.cpp (Basic, Bidi, Cache, Decoration, Ellipsis, FontFallback, GetRects, GlyphInfo, HitTest, Justify, LetterSpacing, LineBreak, LineHeight, Metrics, Placeholder, Shaping, SoftHyphen, Strut). File-local helpers move with all the tests that use them; each file keeps only the includes it needs. New tests go to the matching topic file or a new one. The moved code is clang-formatted, since every line is new to Gerrit. clang-format does not break string literals and comments, so lines it leaves over 100 columns are fixed by hand: long literals are split into adjacent literals with the same value, commented-out SkDebugf calls and a stale TODO are removed, and an ASCII diagram keeps 4 of its 20 repeated groups. The moved tests do not change otherwise. Also fix two TextWrapper line break checks, with a test for each in SkParagraphLineBreakTest.cpp. lookAhead checked a cluster against the line as fWords + fClusters + cluster, while the max intrinsic width adds fWords + (fClusters + cluster); the float sums can differ in the last bit, and then layout(getMaxIntrinsicWidth()) wraps the last word (SkParagraph_MaxIntrinsicWidthKeepsOneLine). A single cluster wider than the line joined the current line even after other text, while a too long word or a placeholder moves to the next line; now it stays only on an empty line (SkParagraph_TooWideClusterStartsNextLine). In PRESUBMIT.py, exempt modules/*/tests/ alongside tests/ from the std::thread ban, and add CheckBannedAPIsTest in PRESUBMIT_test.py: moving SkParagraph_GetRectsForRangeConcurrently into SkParagraphCacheTest.cpp marks its four pre-existing std::thread lines as new. Without the PRESUBMIT.py change, CheckBannedAPIsTest.testStdThreadAllowedInTestsAndModuleTests fails. Tested: - dm --src tests --match SkParagraph, Linux x86-64 Debug, clang + lld, FreeType/fontconfig, ICU, --nativeFonts --paragraph_fonts extra_fonts --norun_paragraph_tests_needing_system_fonts: 242/242 pass before and after the split (the other 6 are disabled or ICU4X-only); with the TextWrapper fix, 244/244 pass, and both new tests fail without it - PRESUBMIT_test.py (40 tests) and CheckChangeOnCommit presubmit, Linux x86-64, Python 3: pass - Script check: before formatting, every line below the shared prelude and all 248 test names appear exactly once across the new files; clang-format leaves every file equal modulo whitespace; after the manual fixes, every new file is clang-format clean, no line is over 100 columns, and with comments dropped and adjacent literals joined, every test body has the same tokens as before - Trybots on PS8 (split only), all green: Housekeeper-OnDemand-Presubmit Test-Ubuntu24.04-Clang-GCE-CPU-AVX2-x86_64-Debug-All-NativeFonts_FreeType_ASAN_Upload Test-Mac14-Clang-MacMini8.1-CPU-AVX2-x86_64-Debug-All-NativeFonts_Fontations Test-Mac14-Clang-MacMini9.1-CPU-AppleM1-arm64-Debug-All BazelBuild-core-release-linux_x64 BazelBuild-full_library-release-linux_x64 Housekeeper-PerCommit-CheckGeneratedFiles Not tested: Windows, Android, google3 Assisted-by: Claude Opus 5.5 (Medium) Change-Id: I9aa9cea84e135b764b2ce3d3c333531393692117 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1430718 Reviewed-by: Alexis Cruz-Ayala <alexisdavidc@google.com> Commit-Queue: Julia Lavrova <jlavrova@google.com>Julia Lavrova · 56d35df3 · 2026-11-17
- 4.2ETV[ganesh] Low-level failures during flush are now made visible to clients Prior to this CL, Ganesh could silently drop draws due to low-level failures. This could lead to security issues in Chrome. This CL attempts to surface some of those errors all the way to the top of Skia's API - allowing clients to handle the failure themselves. Note, this CL does not cover every possible failure. Follow-up CL will be required to handle more failure modes. Bug: 584700430 Bug: 576134673 Change-Id: I3afaa53221b057d4815792779857f02a6f46597c Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1408317 Reviewed-by: Michael Ludwig <michaelludwig@google.com> Commit-Queue: Robert Phillips <robertphillips@google.com>Robert Phillips · 241a6ae1 · 2026-11-12
- 3.3ETV[rust jpeg] Initial implementation of rust jpeg Add a CXX FFI bridge connecting zune-jpeg (decoder) and jpeg-encoder (encoder) crates to Skia, following the same pattern as rust_bmp and rust_png. Decoder features: - Baseline, progressive, grayscale, RGB, and CMYK/YCCK JPEG decoding - ICC profile extraction (multi-segment APP2) and validation - EXIF orientation parsing (via shared SkExif::Parse) - Color space transforms (via skcms) - Incremental decode with rewind support - Gainmap extraction (ISO 21496-1, Adobe, Apple) via MPF - OSS-Fuzz integration Encoder features: - RGB, RGBA, BGRA, Grayscale, kRGB_888x pixel formats - Alpha blending (ignore or blend-on-black) - XMP metadata insertion Rust-side JPEG segment scanning, ICC/EXIF metadata extraction, TIFF IFD parsing, and MPF (Multi-Picture Format) parsing are all implemented in memory-safe Rust. Higher-level gainmap detection (XMP, ISO 21496-1) reuses existing C++ SkJpegMetadataDecoder. Known gaps vs SkJpegCodec: - No true incremental decoding — requires zune-jpeg support - No YUV plane decode (onQueryYUVAInfo) — requires zune-jpeg raw planes - No scanline decoding - requires zune-jpeg support - No getSampler (subset decoding) - No Container XMP gainmap fallback (requires EOI offset) Bug: 493315750 Change-Id: Ib96f274241dbfda36a752c350b2fbb439403d206 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1189316 Reviewed-by: Florin Malita <fmalita@google.com> Reviewed-by: Łukasz Anforowicz <lukasza@google.com> Commit-Queue: Florin Malita <fmalita@google.com>Sergio Gonzalez Martin · 5dac00a4 · 2026-08-05
- 2.8ETVValidate direct-mask Slug data during deserialization Direct-mask glyphs are packed into the glyph atlas without padding when atlas bilerp is unsupported, so they are only safe to draw at integer device positions, with an integer translation of their creation matrix. Slug deserialization did not enforce this: a crafted Slug could supply fractional or non-finite glyph positions, or a direct-mask creation matrix that differs from the Slug's initial position matrix (bypassing SlugImpl::canDraw), leading to out-of-bounds atlas sampling under MSAA. Unbounded offsets could also overflow the float-to-int conversion used when clipping direct-mask quads. Deserialization now requires: * direct-mask glyph positions to be finite integers within +/-2^21; * direct-mask creation matrices to equal the Slug's initial position matrix, and canDrawDirect to match the SubRun type; * finite bounds, origins, matrices, scales and SDFT ranges, and invertible, non-perspective creation matrices; * all glyph positions to be finite and within 1e14; * boolean flags to be 0 or 1 (read with readBool; the wire format is unchanged). At draw time, VertexFiller::CanUseDirect also requires the offset between the creation matrix and the position matrix to be a finite integer within +/-2^21. The position matrix is computed exactly as at creation (single-precision preTranslate), so drawing with an unchanged matrix always yields a zero offset. SlugImpl::canDraw now only gates Slugs that contain direct-mask SubRuns; every other SubRun type is padded in the atlas (entries are keyed by padding) or doesn't use it. Creation and deserialization share the same checks, so every Slug that is created deserializes. SlugImpl::Make returns null if the position matrix overflows or the bounds or origin are invalid. In SubRunContainer::MakeInAlloc, which is used by all GPU text, not just Slugs: * direct masks are only used if the position matrix is finite, invertible, and has a translation within 1e14. Otherwise the glyphs fall back to drawables or paths; * direct-mask glyphs whose device position exceeds 2^21 fall back to the drawable/path/transformed-mask stages; * SDFT and transformed-mask SubRuns are only created if their creation matrix (and SDFT matrix range) is valid. For example, an infinite fGlyphsAsPathsFontSize gives large text an infinite SDFT range, so that text now uses direct masks or paths instead; * glyphs with positions beyond 1e14 are rejected. Bug: 577341943 Change-Id: Ic81bc3ba9ea0af3a213047b3abdf40950a991cca Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1407596 Reviewed-by: Thomas Smith <thomsmit@google.com> Reviewed-by: Nathan Sanchez <nathanasanchez@google.com> Commit-Queue: Greg Daniel <egdaniel@google.com>Greg Daniel · 65fc4569 · 2026-10-30
- 2.5ETV[wgsl] Apply array polyfills in synthetic sksl tests correctly This refactors `writeUniformsAndBuffers`'s body that handled interface blocks into a helper `writeInterfaceBlock` (logic unchanged) so that it could be called by `writeNonBlockUniformsForTests` in place of it generating bespoke WGSL that was 90% the same as the default generated interface WGSL. This allows the synthetic global uniform block to participate in the array polyfills and matrix polyfills correctly, although there was a minor impact on generated test wgsl where @group and @binding changed order. The array/matrix polyfill types also had to be updated to use @align(16) instead of @size(16). The latter adjusts the size but does not propagate the stricter alignment into any wrapping type. Bug: b/465408252 Change-Id: I41077f77ff8bd752f20c8e0f7603bfd4023a45cf Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1121416 Commit-Queue: Michael Ludwig <michaelludwig@google.com> Reviewed-by: Thomas Smith <thomsmit@google.com>Michael Ludwig · 0d60c01b · 2026-01-14
- 2.2ETVReconstruct subRun bounds from glyphs * Reconstruct the bounds of a subRun after deserialization instead of packaging onto the VertexFiller. * An attacker could create a VertexFiller with creation bounds that did not contain its glyphs but were entirely contained within the current clip, enabling to the glyphs to ignore the creation bounds clip and sample from stale scratch textures Bug: b/513948227 Change-Id: Ib4902657e6a50dd5675db4d73a1576b77c4ce88e Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1239916 Commit-Queue: Thomas Smith <thomsmit@google.com> Reviewed-by: Michael Ludwig <michaelludwig@google.com>Thomas Smith · f93ed13d · 2026-05-28
- 2.1ETV[wgsl] Refactor const-eval workaround into automatic helper Bug: b/465408252 Change-Id: Ic8f8ee4ac16b994b92fb994d262cee25230538bc Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1118697 Reviewed-by: Thomas Smith <thomsmit@google.com> Commit-Queue: Michael Ludwig <michaelludwig@google.com>Michael Ludwig · 2b8e3730 · 2025-12-12
- 2.1ETV[ganesh] Guard GrBufferAllocPool against overflow Previously, GrBufferAllocPool::makeSpaceAtLeast checked whether the active block had sufficient space using raw unsigned addition (minSize + pad), which could wrap around size_t when minSize was near SIZE_MAX and write padding out of bounds. Check alignment rounding, padding additions, and vertex/index/indirect byte multiplications with SkSafeMath, and reject zero or negative sizes and counts. Cap single-block allocations to 1.5 GiB, make GrCpuBuffer::Make fallible via sk_malloc_canfail so oversized or OOM allocations return nullptr and skip the draw, and cap GrVertexChunkBuilder's initial heuristic preallocation to 512 KiB (growing up to 16 MiB per chunk as vertices are written) with a fallback to the default buffer size. Bug: 584003006 Change-Id: Ic6e50df23729f569b8b77bf13c926d199719c89a Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1428176 Commit-Queue: Thomas Smith <thomsmit@google.com> Reviewed-by: Nathan Sanchez <nathanasanchez@google.com> Reviewed-by: Thomas Smith <thomsmit@google.com>Greg Daniel · 774ce698 · 2026-11-13
- 2.0ETV[ganesh] More triangulator guards * more speculative fixes for triangulator dereferencing behavior Bug: chromium:473156318 Bug: chromium:470210175 Bug: chromium:546237339 Change-Id: I4d41121c08161b5eebecd46b28c7b8a641477ca2 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1323017 Reviewed-by: Robert Phillips <robertphillips@google.com> Commit-Queue: Thomas Smith <thomsmit@google.com>Thomas Smith · c12ebabc · 2026-08-19
- 2.0ETVRevert "Reland "[graphite] Extracts early in drawGeometry"" This reverts commit 81de4113e3e7cfe8ec91413fbbe51101dcb354e3. Reason for revert: Now breaking chromium roll. Original change's description: > Reland "[graphite] Extracts early in drawGeometry" > > * Reintroduce notify image in use and flush in snapDrawTask. > > * Fixes an issue where multi-draw dependencies were not correctly tracked. > > This reverts commit 1b271fd02a65ba97e12bcaa32f67afa50b5d9b52. > > > Original change's description: > > Revert "[graphite] Extracts early in drawGeometry" > > > > This reverts commit 25f00cb247f23b4a8cbe7a1245bdf609fa0be846. > > > > Reason for revert: Breaks android roll > > > > Original change's description: > > > [graphite] Extracts early in drawGeometry > > > > > > * Moves the creation of UniquePaintIDs from DrawPass::Snap to PaintParams::toKey, which is called in Device::drawGeometry > > > > > > * Moves blend mode calculations into PaintParams, and adds an enum DstUsage to DrawTypes. > > > > > > * Moves the creation of a draw pass from DrawPass::Make to DrawList::snapDrawPass. > > > > > > * Texture and uniform trackers commensurately moved to DrawList. > > > > > > Change-Id: Ie843db44bfad0cd51773ffa7e42050fdbd7c22e3 > > > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1045336 > > > Commit-Queue: Thomas Smith <thomsmit@google.com> > > > Reviewed-by: Michael Ludwig <michaelludwig@google.com> > > > > No-Presubmit: true > > No-Tree-Checks: true > > No-Try: true > > Change-Id: I19ad73d77051295e37ac9adaae77f228e4934834 > > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1052396 > > Commit-Queue: Thomas Smith <thomsmit@google.com> > > Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com> > > Change-Id: Ib8b9aa5b3ed998bdecd3b56a03ca13f189518178 > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1052657 > Reviewed-by: Michael Ludwig <michaelludwig@google.com> > Commit-Queue: Thomas Smith <thomsmit@google.com> No-Presubmit: true No-Tree-Checks: true No-Try: true Change-Id: I0132ab1e71955f6a8b35b3107afe9ae48f5654aa Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1059636 Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com> Commit-Queue: Thomas Smith <thomsmit@google.com>Thomas Smith · 5aafacb1 · 2025-09-22