WIP: DOC: Document CUDA acceleration

docs/methods.md

@@ -26,6 +26,20 @@ intensity images.
 
 The default method.
 
+To use an NVIDIA CUDA GPU-accelerated version, install the
+`itkwasm-downsample-cucim` package:
+
+```sh
+pip install itkwasm-downsample-cucim
+
+# Install dependencies for cucim (assuming that CUDA 11.x is used for CuPy)
+pip install scipy scikit-image cupy-cuda11x
+# or via conda / mamba / micromamba
+conda create -n cucim -c rapidsai -c conda-forge cucim cudatoolkit=<CUDA version>
+```
+
+And GPU-accelerated filtering is applied by default after installation.
+
 ## `ITKWASM_BIN_SHRINK`
 
 Uses the [local mean] for the output value. [WebAssembly] build.
@@ -34,6 +48,9 @@ Fast but generates more artifacts than gaussian-based methods.
 
 Appropriate for intensity images.
 
+An NVIDIA CUDA GPU-accelerated version can be installed similar to
+`ITKWASM_GAUSSIAN` above.
+
 ## `ITKWASM_LABEL_IMAGE`
 
 A sample is the mode of the linearly weighted [local labels] in the image.
@@ -56,7 +73,8 @@ To use a GPU-accelerated version, install the `itk-vkfft` package:
 pip install itk-vkfft
 ```
 
-And GPU-accelerated, FFT-based filtering is applied by default after installation.
+And GPU-accelerated, FFT-based filtering is applied by default after
+installation.
 
 ## `ITK_BIN_SHRINK`
 
@@ -105,7 +123,6 @@ Install required dependencies with:
 pip install "ngff-zarr[dask-image]"
 ```
 
-
 [aliasing artifacts]:
   https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem
 [dask-image]: https://image.dask.org/