| 24 | |
| 25 | template<typename T> |
| 26 | inline void OSD(nvbench::state &state, nvbench::type_list<T>) |
| 27 | try |
| 28 | { |
| 29 | long3 shape = benchutils::GetShape<3>(state.get_string("shape")); |
| 30 | long varShape = state.get_int64("varShape"); |
| 31 | int numElem = static_cast<int>(state.get_int64("numElem")); |
| 32 | |
| 33 | int ch = nvcv::cuda::NumElements<T>; |
| 34 | |
| 35 | using BT = nvcv::cuda::BaseType<T>; |
| 36 | |
| 37 | std::vector<std::vector<std::shared_ptr<cvcuda::priv::NVCVElement>>> elementVec; |
| 38 | |
| 39 | for (int n = 0; n < (int)shape.x; n++) |
| 40 | { |
| 41 | std::vector<std::shared_ptr<cvcuda::priv::NVCVElement>> curVec; |
| 42 | for (int i = 0; i < numElem; i++) |
| 43 | { |
| 44 | NVCVPoint point; |
| 45 | point.centerPos.x = shape.z / 2; |
| 46 | point.centerPos.y = shape.y / 2; |
| 47 | point.radius = std::min(shape.z, shape.y) / 2; |
| 48 | point.color = {0, 0, 0, 255}; |
| 49 | auto element = std::make_shared<cvcuda::priv::NVCVElement>(NVCVOSDType::NVCV_OSD_POINT, &point); |
| 50 | curVec.push_back(element); |
| 51 | } |
| 52 | elementVec.push_back(curVec); |
| 53 | } |
| 54 | |
| 55 | std::shared_ptr<cvcuda::priv::NVCVElementsImpl> ctx = std::make_shared<cvcuda::priv::NVCVElementsImpl>(elementVec); |
| 56 | |
| 57 | state.add_global_memory_reads(shape.x * shape.y * shape.z * sizeof(T) + numElem * sizeof(int) * 16); |
| 58 | state.add_global_memory_writes(shape.x * shape.y * shape.z * sizeof(T)); |
| 59 | |
| 60 | cvcuda::OSD op; |
| 61 | |
| 62 | // clang-format off |
| 63 | |
| 64 | if (varShape < 0) // negative var shape means use Tensor |
| 65 | { |
| 66 | nvcv::Tensor src({{shape.x, shape.y, shape.z, ch}, "NHWC"}, benchutils::GetDataType<BT>()); |
| 67 | nvcv::Tensor dst({{shape.x, shape.y, shape.z, ch}, "NHWC"}, benchutils::GetDataType<BT>()); |
| 68 | |
| 69 | benchutils::FillTensor<BT>(src, benchutils::RandomValues<BT>()); |
| 70 | |
| 71 | state.exec(nvbench::exec_tag::sync, [&op, &src, &dst, &ctx](nvbench::launch &launch) |
| 72 | { |
| 73 | op(launch.get_stream(), src, dst, (NVCVElements)ctx.get()); |
| 74 | }); |
| 75 | } |
| 76 | else // zero and positive var shape means use ImageBatchVarShape |
| 77 | { |
| 78 | throw std::invalid_argument("ImageBatchVarShape not implemented for this operator"); |
| 79 | } |
| 80 | } |
| 81 | catch (const std::exception &err) |
| 82 | { |
| 83 | state.skip(err.what()); |