Worked on accelerating CPU-bound image interpolation in the ROCm/pytorch and pytorch/pytorch repositories, focusing on performance and maintainability. Developed a NEON-optimized implementation of torch.nn.functional.interpolate for RGB images in ChannelsLast format, achieving 3x-6x speedups while ensuring bitwise-equivalent outputs and antialiasing support. Refactored upsampling kernel dispatch logic in C++ to unify channels-last and separable paths, improving code clarity and maintainability. Introduced a NEON 'block of 4' optimization for F.interpolate, resulting in 20-30% speedups for bilinear and bicubic modes. Validated changes with comprehensive testing and benchmarking, emphasizing robust performance and correctness across common image processing workflows.