elasota / convectionkernels Goto Github PK

thirdparty/cvtt/ConvectionKernels_ETC.cpp:3134:55: error: no member named 'sqrt' in namespace 'std'
const float lengthRatio = static_cast(std::sqrt(ca0LengthSq / ca1UNLengthSq));
~~~~~^

Please use #include <cmath>.

I'm trying to use cvtt::Internal::BC6HComputer::UnpackOne directly, to decode a BC6H (unsigned) image block by block. With various other decompression libraries (DirectXTex and others) everything seems to work, but cvtt seems to produce "something entirely else". Happens with all BC6H files I tried so far, on both Mac (clang 13) & Windows (vs2022).

Attached the source file (IndoorHDRI001.dds), what is the expected decoding result (with DirectXTex in this case), IndoorHDRI001-dxtex.hdr and what I get with cvtt, IndoorHDRI001-convection.hdr. Github does not allow attaching dds/hdr directly, so they are all in a zip file.

IndoorHDRI001.zip

My usage is basically like this:

cvtt::PixelBlockF16 rgba;
for (int i = 0; i < height; i += 4)
{
    for (int j = 0; j < width; j += 4)
    {
        cvtt::Internal::BC6HComputer::UnpackOne(rgba, src, format == DDSFormat::BC6HS);
        src += 16;
        for (int c = 0; c < 16; ++c) {
            float* dst_f = (float*)(dst + (i*width+j)*12 + width*(c/4)*12 + (c&3)*12);
            dst_f[0] = half_to_float_fast5(rgba.m_pixels[c][0]);
            dst_f[1] = half_to_float_fast5(rgba.m_pixels[c][1]);
            dst_f[2] = half_to_float_fast5(rgba.m_pixels[c][2]);
        }
    }
}

This line should probably use alphaAggError, not rgbAggError, to compute the error for the alpha channel:

(found by clang's address sanitizer, since right now the code causes a buffer overflow)