Supported features and limitations

Supported features and limitations#

2025-12-17

2 min read time

Applies to Linux

This topic summarizes the hipCIM features and limitations.

Features#

  • Core image interface (cucim.core):

    • All primary image manipulation functions (read, write, and resample) are GPU-accelerated with CPU fallbacks.

    • Metadata operations (accessing dtype, dims, and shape) run on CPU only.

  • Image processing (cucim.skimage):

    • Nearly all transform operations (resize, rotate, and warp) are GPU-accelerated with CPU fallbacks.

    • Complete filter suite (Gaussian, median, and edge detectors) benefits from GPU acceleration.

    • Most morphological operations (erosion, dilation, and opening) are GPU-accelerated.

  • Segmentation:

    • Several advanced segmentation algorithms (felzenszwalb, quickshift, and active_contour) lack GPU acceleration.

    • Core segmentation operations such as watershed and SLIC are GPU-accelerated.

  • Color operations:

    • All color space conversions (rgb2gray, rgb2hsv, and rgb2lab) are GPU-accelerated.

    • Specialized operations for medical imaging, such as stain separation or combination, also benefit from GPU acceleration.

  • Whole slide imaging:

    • Patch extraction operations are GPU-accelerated.

    • Metadata operations run exclusively on the CPU.

  • Measurement functions:

    • Core measurement functions like region labeling are GPU-accelerated.

    • Some advanced functions like marching_cubes lack GPU acceleration.

Image support#

hipCIM supports the following image formats:

  • Single-level Aperio ScanScope Virtual Slide (SVS) with JPEG compression

  • Single-level Philips TIFF with JPEG compression

Note that the image support is limited by rocJPEG chroma subsampling and hardware capabilities.

hipCIM API mirrors scikit-image for image manipulation and OpenSlide for image loading.

Limitations#

  • Only supports features from amd-cupy 13.0 and later.

  • No Support for JPEG2K compression.

  • No support for GPU direct storage (KvikIO, cuFile).

  • No support for rocTX tracing.

  • No Dask support

  • No support for the following image processing operations:

    • affine, similarity, euclidean, threshold_niblack, threshold_sauvola, convex_hull_image, corner_fast, denoise_bilateral, denoise_wavelet, wiener, richardson_lucy, unsupervised_wiener, estimate_sigma, random_walker, felzenszwalb, slic, quickshift, watershed, active_contour, and all exposure operations.

  • Registration:

    • All registration functions (optical flow and daemons) are GPU-accelerated but typically lack CPU fallbacks.

  • Clara DL pipeline:

    • Data loading has partial GPU acceleration.

    • Most Clara transformations are GPU-accelerated with CPU fallbacks.