When using GAUSSIAN_NORMALS, image::resampleImage automatically unitizes
its output. This means that clients (e.g. mipgen) need to ensure that
the source image is in [-1,+1] rather than [0,+1].
Due to an assertion in etc2comp, this unfortunately uses RGB8 instead of
RG11 but it still helps quite a bit (3.3 MB savings in the unzipped KTX
size).
This shrinks file size as follows, although it seems to darken the
rendered result.
```
1398209 Oct 2 16:22 ao.ktx
699172 Oct 2 16:20 ao_etc.ktx
1398209 Oct 2 16:22 metallic.ktx
699172 Oct 2 16:19 metallic_etc.ktx
1398209 Oct 2 16:22 roughness.ktx
699172 Oct 2 16:20 roughness_etc.ktx
```
At build time, mipgen now creates several compressed and non-compressed
variants for albedo.
For ASTC we use the ARM encoder, for S3TC we use the STB encoder. The
latter is somewhat limited so we may wish to investigate other libraries
in the future (e.g., AMD Compressenator).
At run time, we detect which of these variants to download, based on
available WebGL extensions. In practice, this means that desktop
browsers will use the S3TC variant and mobile browsers will use the ASTC
variant.
The makes unzipped albedo go from ~4 MB to 682 KB using S3TC.
Mobile support (ASTC and ETC2) is coming in a future PR, stay tuned.
The compression API now takes block size instead of bit rate. Otherwise
it is quite possible for the encoder to generate a block size that is
not one of the valid block sizes that graphics API's can actually
consume.
(1) Suzanne's normal map is actually linear but we were telling GL that
it is SRGB, so the shader was not reading (0.5,0.5,1.0) from a perfectly
flat normal map.
(2) --linear in mipgen was not honored when consuming PNG files.
(3) The build was not passing --linear to mipgen for the normal map.
The libimage API does not include the original image in the returned mip
array, which might make sense when generating individual files, but for
a KTX bundle, we want to include level 0.
I tested the validity of the resulting KTX files by using macOS Preview,
which can read some types of KTX files. (!)
This CL also adds mipgen to the WebGL build, although its outputs are
not yet used by the sample apps.
Sometimes the source image might have a useless 1-bit alpha channel
which our image-processing pipeline expands into a proper alpha channel.
This consumes memory and breaks an assertion in our encoder.
I ran into this while testing mipgen with the Suzanne textures.
* Introduce mipgen frontend.
For now this is fairly simple although it does let you choose a filter.
Another cool feature is that an HTML file is generated which makes it
easy to review the generated miplevels.
This does not expose the Boundary enum to users because it has not yet
been implemented in the core Image library.
* Fix up mipgen per code review.
Philip Rideout
Ben Doherty
cubeleo
Sebastian Hartte