ChartGPU v0.3.5

Compare: v0.3.4…v0.3.5

Scope: 239 files, ~42k insertions / ~17k deletions, 65+ commits, PRs #166–#177.

This release is primarily a performance and architecture update, driven by work against the SciChart.js JavaScript Chart Performance Suite (ingestion rate and FPS, production dist). It also includes streaming/time-axis correctness fixes, logarithmic axes, candlestick price-axis UI, and site updates.

---

Performance and architecture (#166)

Upload / data residency

• Ranged append for full-raw and GPU-decimation line buffers
• Fixed-capacity ring FIFO: appendData(..., { maxPoints })
• Equal-N y-only rewrite paths (line and scatter)
• Dual-store hit-test skip when tooltips are off
• Axes-only setOption reuse when series element identities are stable
• Domain-space bar packing for axes-only redraws
• Unbounded append near device storage limits auto-windows like maxPoints (avoids domain/GPU desync past maxStorageBufferBindingSize)

Frame graph

compute (scatter density + line decimation)
  → main 4× MSAA
  → optional dense-hairline sampleCount-1 pass
  → overlay 4× MSAA
  → swapchain

• Dense hairline LOD for high point counts / multi-series segment budgets (draw-only)
• Multi-chart submit batching on a shared GPUDevice (microtask-coalesced queue.submit)
• Sticky auto-range domains for streaming (Y headroom; X tracks data max with headroom 0)

Coordinator structure

• Thin createRenderCoordinator shell
• Composition root and domain modules under src/core/renderCoordinator/
• Docs/architecture updated for the current pass graph

Suite-related work in this PR covered axes-only redraws, ring FIFO + GPU decimation, scatter/y-update/unsorted-line rewrites, FIFO large-N paths, N×M line series, and multi-chart batching.

---

Streaming and correctness (#167, [#168])

• Time-axis packing affine so epoch-ms domains stay on-screen
• Pure type: 'area' through DataStore with ranged append and ring/staging geometry
• Full-width unbounded streaming (no empty-right grow/reset from X headroom)
• Modular ring raw lines: shaders use ringStart / ringCapacity so wrap order is logical
• CPU LTTB keeps domain x in Float64 through index selection (avoids epoch-ms collapse)
• Dual-store / tooltip-off resync and related hit-test ownership fixes
• performance.lod: 'auto' | 'strict'
• auto (default): dense hairline, scatter radius compaction, equal-N LTTB sticky indices
• strict: honor stroke width / marker size; full LTTB recompute on equal-N y change

---

Other fixes

PR	Change
#169	Time-axis precision and streaming zoom (#161)
#170	Chart sizing under page zoom and CSS zoom (#155)
#171	Area fill disconnects on null gaps instead of fanning to origin (#153)
#172	Line null gaps preserved under zoom; legend above axis labels

---

Features

Logarithmic axes (#174 / [#135])

• AxisType: 'log', optional logBase
• Shader-side log projection; DataStore stays linear
• Dual logBaseX / logBaseY, log-space affines, tick-aligned grids
• Sticky log domains, positive-only auto-bounds, area baseline guards on log Y
• Candlestick streaming: instance re-upload on append / forming-bar updates

Candlestick price axis (#175, [#176])

• Candle-primary charts: first Y-axis defaults to the right (dual-Y-safe gutters)
• Series priceLabel: last-price DOM badge, optional price line, optional bar-close countdown
• Axis header top-rail label

Site (#176)

The project homepage on GitHub Pages is now the live landing site (previously separate from the Pages homepage). It includes live WebGPU demos (multi-series hero stream, multi-chart dashboard wall, candlestick price-axis demo), install CTA, and basic project links. Served from the repo’s Pages setup; source lives with the dev-site landing work in this PR.

URL: https://chartgpu.github.io/ChartGPU/

Version (#177)

• Package version bumped to 0.3.5

---

Consumer-facing notes

:::ts
chart.appendData(seriesIndex, newPoints, { maxPoints: 1_000_000 });

ChartGPU.create(el, {
  performance: { lod: 'auto' }, // or 'strict'
});

// log axis
yAxis: { type: 'log', logBase: 10 }

// candlestick
series: [{ type: 'candlestick', data: ohlc, priceLabel: true }]

Topic	Behavior
Streaming	Prefer appendData + maxPoints over sliding-window full setOption
Axes-only updates	Keep series object identity stable for the reuse path
LOD	auto may hairline dense lines / compact dense scatter; use strict for exact width/markers
Storage	Very long unbounded streams may auto-window near device storage caps
Tooltips	Off skips dual-store hit-test work on streaming frames

See docs/performance.md (github.com).

---

Pull requests

PR	Title
#166	perf: upload architecture, frame graph, and modular coordinator
#167	fix: time-axis affine, area streaming, and full-width append
#168	fix: close PR 167 correctness blockers (ring, LTTB, dual-store, LOD)
#169	fix: improve time-axis precision and streaming zoom (#161)
#170	fix: correct chart sizing under page zoom and CSS zoom (#155)
#171	fix: area fill null gaps fan to origin (#153)
#172	fix: preserve line null gaps under zoom; stack legend above axis labels
#174	feat: logarithmic axes (#135) and streaming candlestick re-upload
#175	feat: candlestick exchange-style Y-axis (right rail + last-price badge)
#176	Candlestick price axis, live landing site, and GitHub Pages homepage
#177	chore: bump version to 0.3.5

Full changelog: https://github.com/ChartGPU/ChartGPU/compare/v0.3.4...v0.3.5

:::bash
npm install @chartgpu/chartgpu@0.3.5