onnxruntime
onnxruntime copied to clipboard
webgpu: add FormatTransform kernel and tests
- add the WebGPU FormatTransform kernel, headers, and WGSL template, supporting Plain <-> nChw4c/ABcd16a4b conversions
- register the FormatTransform schema in the internal NHWC domain with padding-aware shape inference
- hook the kernel into the WebGPU execution provider and add WebGPU tests covering both blocked formats, padding cases, and round trips