Function CvtColor

bench/BenchCvtColor.cpp:103–174

template<typename BT>
inline void CvtColor(nvbench::state &state, nvbench::type_list<BT>)
try
{
    long3 shape    = benchutils::GetShape<3>(state.get_string("shape"));
    long  varShape = state.get_int64("varShape");

    ConvCodeToFormat formats = str2Frmt(state.get_string("code"));

    NVCVColorConversionCode code = std::get<0>(formats);
    nvcv::ImageFormat       inFormat{std::get<1>(formats)};
    nvcv::ImageFormat       outFormat{std::get<2>(formats)};

    state.add_global_memory_reads(shape.x * shape.y * shape.z * bytesPerPixel<BT>(inFormat));
    state.add_global_memory_writes(shape.x * shape.y * shape.z * bytesPerPixel<BT>(outFormat));

    cvcuda::CvtColor op;

    if (varShape < 0) // negative var shape means use Tensor
    {
        nvcv::Tensor src = CreateTensor(shape.x, shape.z, shape.y, inFormat);
        nvcv::Tensor dst = CreateTensor(shape.x, shape.z, shape.y, outFormat);

        benchutils::FillTensor<BT>(src, benchutils::RandomValues<BT>());

        state.exec(nvbench::exec_tag::sync,
                   [&op, &src, &dst, &code](nvbench::launch &launch) { op(launch.get_stream(), src, dst, code); });
    }
    else // zero and positive var shape means use ImageBatchVarShape
    {
        if (inFormat.chromaSubsampling() != nvcv::ChromaSubsampling::CSS_444
            || outFormat.chromaSubsampling() != nvcv::ChromaSubsampling::CSS_444)
        {
            state.skip("Skipping formats that have subsampled planes for the varshape benchmark");
        }

        std::vector<nvcv::Image>          imgSrc;
        std::vector<nvcv::Image>          imgDst;
        nvcv::ImageBatchVarShape          src(shape.x);
        nvcv::ImageBatchVarShape          dst(shape.x);
        std::vector<std::vector<uint8_t>> srcVec(shape.x);

        auto randomValuesU8 = benchutils::RandomValues<uint8_t>();

        for (int i = 0; i < shape.x; i++)
        {
            imgSrc.emplace_back(nvcv::Size2D{(int)shape.z, (int)shape.y}, inFormat);
            imgDst.emplace_back(nvcv::Size2D{(int)shape.z, (int)shape.y}, outFormat);

            int srcRowStride = imgSrc[i].size().w * inFormat.planePixelStrideBytes(0);
            int srcBufSize   = imgSrc[i].size().h * srcRowStride;
            srcVec[i].resize(srcBufSize);
            for (int idx = 0; idx < srcBufSize; idx++)
            {
                srcVec[i][idx] = randomValuesU8();
            }

            auto imgData = imgSrc[i].exportData<nvcv::ImageDataStridedCuda>();
            CUDA_CHECK_ERROR(cudaMemcpy2D(imgData->plane(0).basePtr, imgData->plane(0).rowStride, srcVec[i].data(),