| 23 | |
| 24 | template<typename T> |
| 25 | inline void Gaussian(nvbench::state &state, nvbench::type_list<T>) |
| 26 | try |
| 27 | { |
| 28 | long3 shape = benchutils::GetShape<3>(state.get_string("shape")); |
| 29 | long varShape = state.get_int64("varShape"); |
| 30 | double sigma = state.get_float64("sigma"); |
| 31 | |
| 32 | NVCVBorderType borderType = benchutils::GetBorderType(state.get_string("border")); |
| 33 | |
| 34 | int kernelSize = (int)std::round(sigma * (std::is_same_v<nvcv::cuda::BaseType<T>, uint8_t> ? 3 : 4) * 2 + 1) | 1; |
| 35 | int2 ksize2{kernelSize, kernelSize}; |
| 36 | |
| 37 | nvcv::Size2D kernelSize2{kernelSize, kernelSize}; |
| 38 | double2 sigma2{sigma, sigma}; |
| 39 | |
| 40 | state.add_global_memory_reads(shape.x * shape.y * shape.z * sizeof(T)); |
| 41 | state.add_global_memory_writes(shape.x * shape.y * shape.z * sizeof(T)); |
| 42 | |
| 43 | cvcuda::Gaussian op(kernelSize2, shape.x); |
| 44 | |
| 45 | // clang-format off |
| 46 | |
| 47 | if (varShape < 0) // negative var shape means use Tensor |
| 48 | { |
| 49 | nvcv::Tensor src({{shape.x, shape.y, shape.z, 1}, "NHWC"}, benchutils::GetDataType<T>()); |
| 50 | nvcv::Tensor dst({{shape.x, shape.y, shape.z, 1}, "NHWC"}, benchutils::GetDataType<T>()); |
| 51 | |
| 52 | benchutils::FillTensor<T>(src, benchutils::RandomValues<T>()); |
| 53 | |
| 54 | state.exec(nvbench::exec_tag::sync, |
| 55 | [&op, &src, &dst, &kernelSize2, &sigma2, &borderType](nvbench::launch &launch) |
| 56 | { |
| 57 | op(launch.get_stream(), src, dst, kernelSize2, sigma2, borderType); |
| 58 | }); |
| 59 | } |
| 60 | else // zero and positive var shape means use ImageBatchVarShape |
| 61 | { |
| 62 | nvcv::ImageBatchVarShape src(shape.x); |
| 63 | nvcv::ImageBatchVarShape dst(shape.x); |
| 64 | |
| 65 | benchutils::FillImageBatch<T>(src, long2{shape.z, shape.y}, long2{varShape, varShape}, |
| 66 | benchutils::RandomValues<T>()); |
| 67 | dst.pushBack(src.begin(), src.end()); |
| 68 | |
| 69 | nvcv::Tensor kernelSizeTensor({{shape.x}, "N"}, nvcv::TYPE_2S32); |
| 70 | nvcv::Tensor sigmaTensor({{shape.x}, "N"}, nvcv::TYPE_2F64); |
| 71 | |
| 72 | benchutils::FillTensor<int2>(kernelSizeTensor, [&ksize2](const long4_16a &){ return ksize2; }); |
| 73 | benchutils::FillTensor<double2>(sigmaTensor, [&sigma2](const long4_16a &){ return sigma2; }); |
| 74 | |
| 75 | state.exec(nvbench::exec_tag::sync, |
| 76 | [&op, &src, &dst, &kernelSizeTensor, &sigmaTensor, &borderType](nvbench::launch &launch) |
| 77 | { |
| 78 | op(launch.get_stream(), src, dst, kernelSizeTensor, sigmaTensor, borderType); |
| 79 | }); |
| 80 | } |
| 81 | } |
| 82 | catch (const std::exception &err) |