Repository navigation
Incorrect bit packing in GPU-encoded BC7 #67
Description
Activity
Thanks for the report. I'll take a look now that I'm back in the office.
Does this look like the correct fix to you?
block.w = (get_color_index(15) << 29) | (get_color_index(14) << 26) | ( get_color_index(13) << 23) | (get_color_index(12) << 20) | (get_color_index(11) << 17) | (get_color_index(10) << 14) | (get_color_index(9) << 11) | (get_color_index(8) << 8) | (get_color_index(7) << 5) | (get_color_index(6) << 3) | | get_color_index(5);Yeah that looks right.
Do you have a test image that uses this particular codepath you can share?
^ Encode this image with texconv -f BC7_UNORM -m 1 -y defect.png
The first byte of the output block (offset 148 in the DDS) is 0xbe, a.k.a. mode 1 partition 47.
The middle 2 pixels in the second row should be the same in the encoded result because they're in the same subset and have the same input value, but they wind up being different because the bits are in the wrong place.
Awesome. Thanks for the specific repro...
Fixed in this commit
Fixed in the February 2018 release
https://github.com/Microsoft/DirectXTex/blob/master/DirectXTex/Shaders/BC7Encode.hlsl
In block_package1, when fixup index is 6:
Indexes 6 and 7 are shifted 1 bit higher than they should be. 6 and 7 should be 2 bits apart, and all others 3 bits apart, but index 5 is given 4 bits and index 7 is given 2.