BC7 color ramp interpolation

[lexaknyazev]

Generation of color ramp uses different math than MSDN example.

Compare:
https://github.com/GPUOpen-Tools/Compressonator/blob/70d7d78f309f762b3c65fd112c2aa06e68ac06c6/Compressonator/Source/Codec/BC7/BC7_utils.cpp#L131-L142

and (from here):

UINT8 interpolate(UINT8 e0, UINT8 e1, UINT8 index, UINT8 indexprecision)
{
    if(indexprecision == 2)
        return (UINT8) (((64 - aWeights2[index])*UINT16(e0) + aWeights2[index]*UINT16(e1) + 32) >> 6);
    else if(indexprecision == 3)
        return (UINT8) (((64 - aWeights3[index])*UINT16(e0) + aWeights3[index]*UINT16(e1) + 32) >> 6);
    else // indexprecision == 4
        return (UINT8) (((64 - aWeights4[index])*UINT16(e0) + aWeights4[index]*UINT16(e1) + 32) >> 6);
}

The difference is that MSDN approach avoids floating-point calculations by using right-shift at the end.

This gives different results, e.g.:

mode 6:
color_bits = 7 + 1
ep0 = 1
ep1 = 82

index_bits = 4
index = 7 (30/64)

MSDN:
((64 - 30) * 1 + 30 * 82) >> 6 == 38

Compressonator:
1.0 * (1.0 - 0.46875) + 82.0 * 0.46875 + 0.5 == 39.46875

[NPCompress]

Hi @lexaknyazev
Thanks for bringing this to our attention. The Ramp Decode for BC7 will be investigated to determine its accuracy with our CPU decoder implementation vs that with GPU decode results using a number of sample blocks.

Once completed the results and any changes will be posted alongside this issue report.

[xooiamd]

hi @lexaknyazev , using MSDN interpolation by bit shifting will affect the CPU decoder (as shown in the snapshot CPU_Decode_View image attached) and SSIM droped to 0.63. However, GPU decoder is fine ( as shown in the GPU_Decode_View image attached) and SSIM is 0.99.


Therefore, definition "USE_HIGH_PRECISION_INTERPOLATION_BC7" has been added to BC7_Definitions.h file in the latest commit in master branch. You can uncomment this definition to use the bit shifting MSDN defined interpolation if you are only interested in GPU Decoder. Thanks!

[lexaknyazev]

Hmmm,.. That looks strange.
First of all, my example above is incorrect since it lacks + 32 operation. So, results are actually the same (after floor):

MSDN:
((64 - 30) * 1 + 30 * 82 + 32) >> 6 == 39

Compressonator:
1.0 * (1.0 - 0.46875) + 82.0 * 0.46875 + 0.5 == 39.46875

Moreover, I couldn't find any other example that shows any difference between these formulas, so it's a bit suspicious that decoded images are that different.

[xooiamd]

hi @lexaknyazev, the BC7 ramp calculation either by division and floor VS bit shifting produce same result. My attachment of the snapshot above are incorrect as I forgot the loop for cluster when calculating the ramp with bit shifting as shown by msdn. Both methods produce same result. Please see the latest commit 7ecb80f on the update. Thanks!