deck.gl-geotiff: Use `boundless: true` when fetching COG tiles in `COGLayer`?

[kylebarron]

Currently we use boundless: false when fetching COG tiles:

deck.gl-raster/packages/deck.gl-geotiff/src/geotiff/render-pipeline.ts

Lines 157 to 161 in 72ca8ae

	const tile = await image.fetchTile(x, y, {
	boundless: false,
	pool,
	signal,
	});

and then we use clipToImageBounds to clip on CPU to the valid region of the image

deck.gl-raster/packages/geotiff/src/fetch.ts

Lines 388 to 456 in 47afad5

	/**
	* Clip a decoded tile array to the valid image bounds.
	*
	* Edge tiles in a COG are always encoded at the full tile size, with the
	* out-of-bounds region zero-padded. When `boundless=false` is requested, this
	* function copies only the valid pixel sub-rectangle into a new typed array,
	* returning a `RasterArray` whose `width`/`height` match the actual image
	* content rather than the tile dimensions.
	*
	* Interior tiles (where the tile fits entirely within the image) are returned
	* unchanged.
	*/
	function clipToImageBounds(
	self: HasTiffReference,
	x: number,
	y: number,
	array: RasterArray,
	): RasterArray {
	const { width: clippedWidth, height: clippedHeight } =
	self.image.getTileBounds(x, y);
	// Interior tile — nothing to clip.
	if (clippedWidth === self.tileWidth && clippedHeight === self.tileHeight) {
	return array;
	}
	const clippedMask = array.mask
	? clipRows(array.mask, self.tileWidth, clippedWidth, clippedHeight, 1)
	: array.mask;
	if (array.layout === "pixel-interleaved") {
	const { count, data } = array;
	const clipped = clipRows(
	data,
	self.tileWidth,
	clippedWidth,
	clippedHeight,
	count,
	);
	return {
	...array,
	width: clippedWidth,
	height: clippedHeight,
	data: clipped as typeof data,
	mask: clippedMask,
	};
	}
	// band-separate
	const { bands } = array;
	const clippedBands = bands.map(
	(band) =>
	clipRows(
	band,
	self.tileWidth,
	clippedWidth,
	clippedHeight,
	1,
	) as typeof band,
	);
	return {
	...array,
	width: clippedWidth,
	height: clippedHeight,
	bands: clippedBands,
	mask: clippedMask,
	};
	}

I think we should probably (? maybe?) avoid image clipping on CPU and upload the raw tiles as-is. Then we can use existing nodata-masking or mask-array masking to discard these pixels from rendering.

The tradeoff is that using boundless: true and removing CPU clipping:

• Should be faster by having less CPU overhead
• Should use slightly more GPU memory for the regions of the texture that are known to be nodata

In the MultiCOGLayer we always use boundless: true because we need the uvTransform to be consistent across all tiles. This is what produced the warped screenshots in #410, which was fixed in #411

[kylebarron]

This could perhaps also make it easier to implement #228 (prototyped in #305),