trunk/515f9f99f8799bf9dcc71d8b740c093f67310135: Add Lanczos interpolation mode for CPU images (#177320)
New CPU-native Lanczos mode in PyTorch matches PIL-SIMD quality while dramatically accelerating image preprocessing.
The PyTorch development team has merged a significant pull request (#177320) that introduces native Lanczos interpolation support to the framework's core image processing functions. The new `mode="lanczos"` parameter for `torch.nn.functional.interpolate` delivers mathematically identical results to the industry-standard PIL-SIMD library when processing uint8 images, while achieving performance gains of 2X to 10X in relevant CPU-based scenarios. This implementation leverages existing separable interpolation infrastructure to automatically provide optimized AVX2 and NEON SIMD paths for both float32 and uint8 data types, with full backward pass support for gradient computation.
The addition specifically targets CPU-based image preprocessing pipelines common in machine learning workflows. Lanczos interpolation is widely regarded as superior to bicubic methods for quality-critical applications, and its absence in PyTorch previously forced developers to maintain external PIL dependencies. The implementation currently supports image batches with antialiasing enabled, while excluding 1D/3D resampling and GPU backends like CUDA and MPS for now. Extensive testing confirms bitwise equality with PIL-SIMD outputs and establishes the same error tolerances as PyTorch's existing bicubic implementation when comparing against standard PIL.
- Bitwise-exact equivalence with PIL-SIMD on uint8 data, ensuring no quality regression
- 2X-10X performance improvement over standard PIL for CPU-based image resizing
- Supports AVX2/NEON SIMD, float32/uint8 dtypes, antialiasing, and backward passes for gradients
Why It Matters
Eliminates PIL dependency for high-quality image preprocessing in ML pipelines, significantly accelerating data loading and augmentation on CPU.