| 23 | |
| 24 | template<typename T> |
| 25 | inline void Stack(nvbench::state &state, nvbench::type_list<T>) |
| 26 | try |
| 27 | { |
| 28 | long3 shape = benchutils::GetShape<3>(state.get_string("shape")); |
| 29 | |
| 30 | using BT = typename nvcv::cuda::BaseType<T>; |
| 31 | |
| 32 | int ch = nvcv::cuda::NumElements<T>; |
| 33 | |
| 34 | state.add_global_memory_reads(shape.x * shape.y * shape.z * sizeof(T)); |
| 35 | state.add_global_memory_writes(shape.x * shape.y * shape.z * sizeof(T)); |
| 36 | |
| 37 | cvcuda::Stack op; |
| 38 | |
| 39 | // clang-format off |
| 40 | |
| 41 | nvcv::TensorBatch src(nvcv::TensorBatch::CalcRequirements(shape.x)); |
| 42 | nvcv::Tensor dst({{shape.x, shape.y, shape.z, ch}, "NHWC"}, benchutils::GetDataType<BT>()); |
| 43 | |
| 44 | for (int i = 0 ; i < shape.x; i++) |
| 45 | { |
| 46 | nvcv::Tensor srcIn({{shape.y, shape.z, ch}, "HWC"}, benchutils::GetDataType<BT>()); |
| 47 | benchutils::FillTensor<BT>(srcIn, benchutils::RandomValues<BT>()); |
| 48 | src.pushBack(srcIn); |
| 49 | } |
| 50 | |
| 51 | state.exec(nvbench::exec_tag::sync, [&op, &src, &dst](nvbench::launch &launch) |
| 52 | { |
| 53 | op(launch.get_stream(), src, dst); |
| 54 | }); |
| 55 | |
| 56 | } |
| 57 | catch (const std::exception &err) |
| 58 | { |
| 59 | state.skip(err.what()); |
| 60 | } |
| 61 | |
| 62 | // clang-format on |
| 63 | |