v0.11.2

What's Changed

Changed

• edgefirst-hal bumped from 0.24.0 to 0.24.2 (transitive update via
  the workspace's edgefirst-hal = "0.24" semver constraint). The 0.24.2
  release fixes a GL_TEXTURE_SWIZZLE_R leak in the GL backend's
  RGBA → PlanarRgb conversion path
  (EdgeFirstAI/hal#84) —
  the swizzle was left at GL_BLUE after the last iteration of the
  per-channel loop, then inherited by the next draw_decoded_masks
  call's bg blit and observed as canvas.R := src.B across the entire
  overlay on NXP Vivante GC7000 and Mali Valhall targets. The yolov8
  example's saved overlay JPEG now renders in natural colours on
  imx8mp-frdm and imx95-frdm.

Added

• The yolov8 example now synthesises a canonical Ultralytics dshape
  per output role (Detection, Boxes, Scores, MaskCoefficients,
  Protos) when the .dvm file does not ship an edgefirst.json
  metadata block — for example the official Kinara 1.2.1 exports
  (yolov8n-kinara-1.2.1.dvm, yolov8n-seg-kinara-1.2.1.dvm) which
  have no zip footer. Without this fallback, the seg path errored at
  mask materialisation with
  mask_coefficients [N, 32] incompatible with protos [32, 160, 160] (expected [N, 160]) because the HAL decoder fell back to "shape is
  already canonical" and read the NCHW proto tensor as if it were
  NHWC. Verified end-to-end with both detection and segmentation Kinara
  exports on imx8mp-frdm and imx95-frdm.

---

crates.io: ara2 | ara2-sys
PyPI: edgefirst-ara2