Skip to content

Memory Management

WhatsCanvas bounds its main retained CPU and GPU caches and exposes their estimated footprint through Canvas::RenderStats. These limits control long-running, content-heavy applications without requiring callers to manage backend resources directly.

Retained resource budgets

Resource Default bound Policy
Fill tessellation cache 8 MiB, 256 entries LRU by retained vector capacity
Fill anti-alias mesh cache 8 MiB, 64 entries LRU by retained vertex and coverage capacity
Stroke tessellation cache 8 MiB, 256 entries LRU by retained vector capacity
GPU bitmap-text cache 32 MiB, 256 entries LRU by uploaded RGBA byte count
Render-target pool 32 MiB, 12 targets Oldest idle target is evicted first
Portable glyph atlas 2048 x 2048 initially, up to 4096 x 4096 Grows on demand and rebuilds retained glyphs

An oversized tessellation can remain as the cache's sole entry because the current draw needs a stable reference. A bitmap-text texture or render target larger than its entire budget is used but not retained. The render-target byte estimate is conservative across backends because attachment formats and driver allocation details differ.

Runtime diagnostics

Call Canvas::getRenderStats() after a frame. The memory-related fields are:

  • glyphAtlasTextureCount and glyphAtlasTextureBytes
  • pooledRenderTargetCount and pooledRenderTargetBytes
  • renderTargetPoolReuseCount, renderTargetPoolAllocationCount, and renderTargetPoolEvictionCount
  • tessellationCacheBytes (fill triangulation plus anti-alias meshes) and strokeCacheBytes
  • bitmapTextCacheSize and bitmapTextCacheBytes
  • trackedResourceBytes, the convenience sum of the byte fields above

The values describe resources retained by WhatsCanvas. They are not process RSS or exact driver heap measurements. Use platform profilers for allocator overhead, driver-internal allocations, mapped memory, and fragmentation.

Temporary working memory

The software Gaussian blur uses two floating-point working images. The vertical pass writes directly to the final RGBA8 destination, avoiding a third full-resolution floating-point image. Inner-shadow filtering follows the same two-buffer strategy for alpha coverage.

OpenGL and Vulkan image filters reuse temporary render targets. The pool keeps targets for 64 usage cycles so multiple filters in one frame do not age each other out, while the byte and target-count budgets remain hard upper bounds. Pooled targets are cleared before the backend device or graphics context is released.

Measuring changes

Build and run the end-to-end filter benchmark with identical backend, build type, dimensions, warmup, and frame count:

cmake --build build --target WhatsCanvasImageFilterBenchmarks --config Release
./build/WhatsCanvasImageFilterBenchmarks \
  --backend software --warmup 3 --frames 10 --width 1920 --height 1080

Each FILTER_BENCHMARK line includes a deterministic output hash, pooled render-target diagnostics, and retained cache estimates. Compare the hash before comparing timing or process memory so a visual regression cannot be mistaken for an optimization.