sRGB decompression is adding a non-1 alpha. about astc-encoder

[solidpixel]

Triage notes:

• Bug isn't a regression; it's always been broken
• Only impacts sRGB constant color blocks

Cause is this bit of code:

int mask = (decode_mode == DECODE_LDR_SRGB) ? 0xFF00 : 0xFFFF;
...
alpha = sf16_to_float(unorm16_to_sf16(scb->constant_color[3] & mask));

... applying the 0xFF00 mask results in a 0.996 (254) floating point value. In real hardware decode this isn't a bug - the top 8 bits are sent off to a sRGB-to-linear conversion unit which would handle this correctly. However, the reference compressor doesn't linearize (we want to write sRGB encoded data back into the output texture) so when using the entire 16-bit result the loss of precision seeps though into the output.

For textures compressed with astcenc itself we'd be safe just removing the masking, but we've got no guarantee that another compressor writes anything sensible into those bottom 8 bits. The correct fix is therefore to take the top 8 bits and rescale them to fill the whole 16-bit range (replicate the top 8 bits in the bottom 8).