The existing specification requires that low dynamic range (LDR) ASTC
textures are decompressed to FP16 values per component. In many cases,
decompressing LDR textures to a lower precision intermediate result
gives acceptable image quality. Source material for LDR textures is
typically authored as 8-bit UNORM values, so decoding to FP16 values
adds little value. On the other hand, reducing precision of the decoded
result reduces the size of the decompressed data, potentially improving
texture cache performance and saving power.
The goal of this extension is to enable this efficiency gain on existing
ASTC texture data. This is achieved by giving the application the
ability to select the intermediate decoding precision.
Three decoding options are provided:
Decode to FORMAT_R16G16B16A16_SFLOAT
precision: This is the default, and matches the required behavior in
the core API.
Decode to FORMAT_R8G8B8A8_UNORM
precision: This is provided as an option in LDR mode.
Decode to FORMAT_E5B9G9R9_UFLOAT_PACK32
precision: This is provided as an option in both LDR and HDR mode.
In this mode, negative values cannot be represented and are clamped
to zero. The alpha component is ignored, and the results are as if
alpha was 1.0. This decode mode is optional and support can be
queried via the physical device properties.
1) Are implementations allowed to decode at a higher precision than what
is requested?
RESOLUTION: No.
If we allow this, then this extension could be exposed on all
implementations that support ASTC.
But developers would have no way of knowing what precision was actually
used, and thus whether the image quality is sufficient at reduced
precision.
2) Should the decode mode be image view state and/or sampler state?
RESOLUTION: Image view state only.
Some implementations treat the different decode modes as different
texture formats.