v0.35.0

Added

• Physically contiguous DMA-BUF on request, and a query for it (edgefirst-tensor, edgefirst-image). Capture engines without an IOMMU, such as the i.MX 8M Plus ISI, can only write into physically contiguous memory, and since 0.33 a DMA-BUF allocation silently falls back to the non-contiguous system heap when the CMA heap is full.
  • ImageDesc::with_contiguous(true) allocates from the CMA heap only, through Tensor::image_desc, TensorDyn::image_desc and ImageProcessor::create_image_desc. When there is no CMA heap or it cannot fit the image, allocation fails with an error naming /dev/dma_heap/linux,cma instead of falling back to the system heap, a PBO or host memory. Default allocations are unchanged.
  • Tensor::contiguity() and TensorDyn::contiguity() return the new Contiguity: Contiguous for CMA, NonContiguous for the system heap, and Unknown for an imported fd and every other kind of memory.
  • C: ef_tensor_image_desc_set_contiguous, ef_tensor_image_desc_contiguous, ef_tensor_contiguity and the ef_contiguity enum.
  • Python: Tensor.contiguity, and contiguous=True on Tensor.image and ImageProcessor.create_image.
• PixelFormat::is_rgb_or_hsv(). Reports whether a format is RGB or HSV, the flag Colorimetry::from_v4l2 now takes. Rgb, Rgba, Bgra, PlanarRgb and PlanarRgba are; Grey and the YUV formats are not.

Changed

• Breaking: Colorimetry::from_v4l2 takes is_rgb_or_hsv. The flag says whether the negotiated format's pixel encoding is RGB or HSV. It is the fifth argument in Rust and Python (Colorimetry.from_v4l2). In C it comes before out in ef_codec_colorimetry_from_v4l2, and ef_codec_abi_version is now 2.

Fixed

• Segmentation masks from models with uint8 or uint16 prototype outputs are no longer noise. The decoder narrowed every non-int8 prototype tensor to int8 by keeping the low byte, while still applying the tensor's own scale and zero point. QNN models converted with the logical split, which output uint16 (w8a16) or uint8 (w8a8) prototypes, lost most of their mask accuracy: yolov8n-seg on the IQ-9075 scored 11.2 mask mAP50-95 instead of 29.3. Such prototypes are now dequantized exactly. Models with int8 or float prototypes, including every TFLite model, are unaffected.
• An RGB V4L2 format with quantization = DEFAULT is full range. Colorimetry::from_v4l2 resolved a DEFAULT quantization from the colorspace alone, so an RGB camera reporting sRGB was published as limited range. It now follows the kernel's V4L2_MAP_QUANTIZATION_DEFAULT: a DEFAULT quantization is full range for an RGB/HSV format, whatever the colorspace, and for V4L2_COLORSPACE_JPEG; limited for a YUV format in any other recognised colorspace; and unset for a YUV format in a DEFAULT or unrecognised colorspace, so the at-use height heuristic decides.