Function CopyMakeBorder

bench/BenchCopyMakeBorder.cpp:25–86

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

    NVCVBorderType borderType = benchutils::GetBorderType(state.get_string("border"));

    float4 borderValue{0.f, 0.f, 0.f, 0.f};

    int top  = srcShape.y / 2;
    int left = srcShape.z / 2;

    long3 dstShape{srcShape.x, top + srcShape.y, left + srcShape.z};

    state.add_global_memory_reads(srcShape.x * srcShape.y * srcShape.z * sizeof(T));
    state.add_global_memory_writes(dstShape.x * dstShape.y * dstShape.z * sizeof(T));

    cvcuda::CopyMakeBorder op;

    // clang-format off

    if (varShape < 0) // negative var shape means use Tensor
    {
        nvcv::Tensor src({{srcShape.x, srcShape.y, srcShape.z, 1}, "NHWC"}, benchutils::GetDataType<T>());
        nvcv::Tensor dst({{dstShape.x, dstShape.y, dstShape.z, 1}, "NHWC"}, benchutils::GetDataType<T>());

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

        state.exec(nvbench::exec_tag::sync,
                   [&op, &src, &dst, &top, &left, &borderType, &borderValue](nvbench::launch &launch)
        {
            op(launch.get_stream(), src, dst, top, left, borderType, borderValue);
        });
    }
    else // zero and positive var shape means use ImageBatchVarShape
    {
        nvcv::ImageBatchVarShape src(srcShape.x);
        nvcv::ImageBatchVarShape dst(dstShape.x);

        benchutils::FillImageBatch<T>(src, long2{srcShape.z, srcShape.y}, long2{varShape, varShape},
                                      benchutils::RandomValues<T>());
        benchutils::FillImageBatch<T>(dst, long2{dstShape.z, dstShape.y}, long2{varShape, varShape},
                                      benchutils::RandomValues<T>());

        nvcv::Tensor topTensor({{srcShape.x, 1, 1, 1}, "NHWC"}, nvcv::TYPE_S32);
        nvcv::Tensor leftTensor({{srcShape.x, 1, 1, 1}, "NHWC"}, nvcv::TYPE_S32);

        benchutils::FillTensor<int>(topTensor, [&top](const long4_16a &){ return top; });
        benchutils::FillTensor<int>(leftTensor, [&left](const long4_16a &){ return left; });

        state.exec(nvbench::exec_tag::sync,
                   [&op, &src, &dst, &topTensor, &leftTensor, &borderType, &borderValue](nvbench::launch &launch)
        {
            op(launch.get_stream(), src, dst, topTensor, leftTensor, borderType, borderValue);
        });
    }
}