v0.19.0

Release Highlights

CV-CUDA v0.19.0-beta includes the following changes.

New Features and Enhancements

Added cvcuda.tv, a torchvision-compatible functional API for geometry,
color, photometric and tensor transforms, without runtime dependencies on
PyTorch or torchvision. Functions accept torch.Tensor and
cvcuda.Tensor inputs, including erase's fill value and elastic's
displacement. As in torchvision, displacement is relative to the identity
sampling grid.
cvcuda.tv.Compose (which requires PyTorch and TorchVision) runs
TorchVision v2 transform pipelines, including random and AutoAugment-family
transforms, as one CV-CUDA chain with a single import and export;
see the supported-transform table.
Added TvResize with nearest-exact, nearest-floor, linear and cubic
sampling, optional antialiasing, and torchvision-style Python size/max-size
handling. Supports U8/U16/S16/F16/F32 HWC/NHWC and CHW/NCHW tensors, and
uniform single-plane C1/3/4 variable-shape batches. Input and output scalar
types must match; integer results round ties-to-even and saturate. The C/C++
API requires caller-owned workspace.
Added Split and Merge for tensors and variable-shape image batches.
Added free-threaded CPython 3.14 support, including concurrent operator
submission, shared-buffer dependency tracking and safe batch mutation.
See Free-threaded Python for build instructions and concurrency semantics.
Expanded data-type support:
- CopyMakeBorder: S8/S32/S64/F64 on tensor, var-shape and planar paths.
- CustomCrop: all uniformly packed 8-, 16-, 32- and 64-bit signed,
  unsigned and floating-point channel types.
- Invert: S16/S32/S64/F64; Solarize: F64; AutoContrast: S16;
  BrightnessContrast: S64/F64.
- Normalize: F64 on tensor, var-shape, scalar and planar paths, using
  double-precision arithmetic for images. Parameters remain F32.
- Posterize: F16/F32/F64 images, quantizing normalized [0, 1] values
  into 2 ** bits levels with bits in [0, 8].
Extended HistogramEq with F16/F32/F64 images and a Pillow-compatible
lookup-table method. Floating-point images are binned to 8 bits, equalized
and scaled back. Select the method with NVCVHistogramEqMethod or
Python's method= keyword: NVCV_HISTOGRAM_EQ_PILLOW matches Pillow's
table bit-exactly; the default NVCV_HISTOGRAM_EQ_OPENCV and existing
entry points are unchanged.
Added CvtColor grayscale codes RGB2GRAY_TV and BGR2GRAY_TV,
bit-exact with torchvision.transforms.v2.functional.rgb_to_grayscale.
They use BT.601 weights 0.2989 / 0.587 / 0.114 (as in MATLAB
rgb2gray and TensorFlow), evaluated in float32 and truncated for integer
outputs. Unlike RGB2GRAY (0.299, rounded to nearest), the weights
sum to 0.9999, so white maps to 254 for U8. RGBA2GRAY_TV and
BGRA2GRAY_TV ignore alpha.
Added AdvCvtColor support for the existing full-range BT.601, BT.709
and BT.2020 NVCV_COLOR_SPEC_*_ER specifications, also exposed through
Python's cvcuda.ColorSpec. Each uses the same full-range U8 arithmetic
as its base variant
(issue #289).
Expanded interpolation and border options:
- Resize accepts NVCV_INTERP_NEAREST_FLOOR (Python
  cvcuda.Interp.NEAREST_FLOOR) for PyTorch/torchvision nearest
  alignment. The existing NVCV_INTERP_NEAREST center-based convention
  is unchanged.
- WarpAffine, WarpPerspective and Remap tensor paths accept
  NVCV_BORDER_CONSTANT_COVERAGE. It blends the border value according
  to the interpolation footprint's out-of-bounds coverage, at full
  precision before rounding. Var-shape paths do not support this mode.
- Remap accepts non-uniform per-channel constant border values for
  planar tensors and planar var-shape batches with up to four channels,
  including constant-coverage mode for tensors. Wider planar tensors still
  require a uniform value. Interleaved F16/F32 inputs now support three
  and four channels.
Added rounding options:
- AdjustContrast and by-value BrightnessContrast accept
  NVCV_ROUND_TRUNCATE through new C/C++ submit entry points and
  Python's round= keyword. It truncates integer outputs to match
  torchvision; floating-point outputs are unchanged. Nearest rounding
  remains the default and the behavior of existing entry points.
- ConvertTo accepts NVCV_ROUND_NEAREST_WRAP and
  NVCV_ROUND_TRUNCATE_WRAP (Python cvcuda.Round.NEAREST_WRAP and
  TRUNCATE_WRAP) for 8- and 16-bit integer destinations. These modes
  wrap instead of saturating; negative intermediates still map to zero
  for unsigned outputs. 32-bit integer destinations and other operators
  do not support them.

Performance Improvements

Optimized AverageBlur, ConvertTo, CopyMakeBorder,
CropFlipNormalizeReformat, HQResize, Invert, PadAndStack,
Posterize, RandomResizedCrop, Resize, Rotate, WarpAffine
and WarpPerspective for selected workloads.
Reduced Python operator-call and tensor/image export overhead without
changing the API. Improved cache reuse lets larger idle allocations serve
different shapes while preserving requested strides and alignment. Live
views and exported buffers prevent reuse of their storage.
Accelerated aligned, pitched single-channel U8 Reformat from NCHW to NHWC.

Bug Fixes

Fixed Python default-stream synchronization for cross-stream operator
submissions and DLPack exports. CUDA stream handles now remain valid while
dependent resources are alive, and avoidable pending-cache growth is prevented.
Hardened input validation:
- TensorShapeInfo and TensorDataAccessStrided reject dimensions
  outside the signed 32-bit range with NVCV_ERROR_OVERFLOW instead of
  silently truncating them.
- Invert, Posterize, Solarize, AdjustSharpness and
  BrightnessContrast report errors for unsupported image heights,
  tensors exceeding 32-bit addressing, padded column strides and multichannel
  tensor data types, rather than CUDA failures or incorrect results. Image
  channels must use the tensor's C dimension. BrightnessContrast
  also rejects batches larger than 65535 images.
- GammaContrast rejects non-F32 gamma tensors; MinAreaRect rejects
  mismatched totalContours counts.
Fixed shared-memory races in HistogramEq OpenCV CDF normalization.
Fixed Morphology dilation of all-negative floating-point images.
Fixed misaligned memory accesses in WarpPerspective and in Invert,
Posterize and Solarize planar var-shape images wrapping caller-owned memory.
Fixed integer size calculations, including CopyMakeBorder launch
failures for four-channel 64-bit pixels and ceiling-division precision
loss for large values.
Fixed allocation-failure reporting in AdaptiveThreshold, WarpAffine,
WarpPerspective and RandomResizedCrop, including a device-memory
leak when RandomResizedCrop host allocation fails.
Fixed the installed CMake package so find_package(cvcuda) discovers its
CUDA Toolkit and nvcv_types dependencies
(issue #295).

Compatibility Changes

Breaking ABI change: cvcudaAdjustSharpnessSubmit and
cvcudaAdjustSharpnessVarShapeSubmit now take a double factor
(<= FLT_MAX). Rebuild applications using these entry points. The
calculation of 1 - factor now rounds to float once, matching torchvision;
outputs may differ by one level.
Versioned requests for the pre-1.0 cvcuda and nvcv_types CMake
packages must match the installed major and minor. Major-only requests and
requests for a different minor are rejected. Specify the installed minor,
for example find_package(cvcuda 0.19 CONFIG REQUIRED).
Corrected AdvCvtColor BT.601/BT.709 RGB-to-YUV luma coefficients to
preserve neutral grayscale values. Output can differ from earlier releases
by up to one code value for BT.601 and two for BT.709; regenerate affected
golden buffers.
AdjustContrast rejects non-finite contrast factors and factors that
cannot be represented as float32 instead of accepting them for asynchronous
execution. ConvertTo rejects unknown NVCVRoundMode values instead
of silently using nearest rounding.

Compatibility and Known Limitations

For full details, see the main README's compatibility table and known limitations.

License

CV-CUDA is licensed under the Apache 2.0 license.

Resources

Acknowledgements

CV-CUDA originated as a collaborative effort between NVIDIA and the ByteDance Machine Learning team.

Navigation Menu

Repository navigation

CV-CUDA Release v0.19.0

Choose a tag to compare

Sorry, something went wrong.