Function testConvert_OutBitDepth

tests/cpu/AVX512_tests.cpp:78–156

template<OCIO::BitDepth inBD, OCIO::BitDepth outBD>
void testConvert_OutBitDepth()
{
    typedef typename OCIO::BitDepthInfo<inBD>::Type InType;
    typedef typename OCIO::BitDepthInfo<outBD>::Type OutType;

    size_t maxValue = OCIO::BitDepthInfo<inBD>::maxValue + 1;

    if (OCIO::BitDepthInfo<inBD>::isFloat)
        maxValue = 65536;

    std::vector<InType> inImage(maxValue);
    std::vector<OutType> outImage(maxValue);

    for (unsigned i = 0; i < maxValue; i++)
    {
        inImage[i] = scale_unsigned<inBD>(i);
    }

    float scale = (float)OCIO::BitDepthInfo<outBD>::maxValue / (float)OCIO::BitDepthInfo<inBD>::maxValue;
    __m512 s = _mm512_set1_ps(scale);

    for (unsigned i = 0; i < inImage.size(); i += 64)
    {
        __m512 r, g, b, a;
        OCIO::AVX512RGBAPack<inBD>::Load(&inImage[i], r, g, b, a);
        r = _mm512_mul_ps(r, s);
        g = _mm512_mul_ps(g, s);
        b = _mm512_mul_ps(b, s);
        a = _mm512_mul_ps(a, s);
        OCIO::AVX512RGBAPack<outBD>::Store(&outImage[i], r, g, b, a);
    }
    for (unsigned i = 0; i < outImage.size(); i++)
    {
        float v = (float)inImage[i] * scale;

        if (OCIO::BitDepthInfo<outBD>::isFloat)
            v = (OutType)v; // casts to half if format is half
        else
            v = rintf(v);

        OCIO_CHECK_ASSERT_MESSAGE(!OCIO::FloatsDiffer(v, (float)outImage[i], 0, false),
                                  GetErrorMessage(v, (float)outImage[i], inBD, outBD));
    }

    // Test Load/Store Masked
    for (unsigned pixel_count = 0; pixel_count <= 16; pixel_count++)
    {
        __m512 r, g, b, a;
        // reset all values to zero
        for (unsigned i = 0; i < outImage.size(); i++)
        {
            outImage[i] = 0;
        }

        OCIO::AVX512RGBAPack<inBD>::LoadMasked(&inImage[0], r, g, b, a, pixel_count);
        r = _mm512_mul_ps(r, s);
        g = _mm512_mul_ps(g, s);
        b = _mm512_mul_ps(b, s);