| 23 | |
| 24 | template<typename T> |
| 25 | inline void CropFlipNormalizeReformat(nvbench::state &state, nvbench::type_list<T>) |
| 26 | try |
| 27 | { |
| 28 | long3 srcShape = benchutils::GetShape<3>(state.get_string("shape")); |
| 29 | long varShape = state.get_int64("varShape"); |
| 30 | long3 dstShape = srcShape; |
| 31 | |
| 32 | NVCVBorderType borderType = benchutils::GetBorderType(state.get_string("border")); |
| 33 | |
| 34 | float borderValue{0.f}; |
| 35 | |
| 36 | float globalScale = 1.234f; |
| 37 | float globalShift = 2.345f; |
| 38 | float epsilon = 12.34f; |
| 39 | uint32_t flags = 0; |
| 40 | |
| 41 | long3 baseShape{srcShape.x, 1, 1}; |
| 42 | long3 scaleShape{srcShape.x, 1, 1}; |
| 43 | long3 cropShape{srcShape.x, 1, 1}; |
| 44 | |
| 45 | state.add_global_memory_reads(srcShape.x * srcShape.y * srcShape.z * sizeof(T) |
| 46 | + baseShape.x * baseShape.y * baseShape.z * sizeof(float) |
| 47 | + scaleShape.x * scaleShape.y * scaleShape.z * sizeof(float) |
| 48 | + cropShape.x * cropShape.y * cropShape.z * sizeof(int) * 4); |
| 49 | state.add_global_memory_writes(dstShape.x * dstShape.y * dstShape.z * sizeof(T)); |
| 50 | |
| 51 | cvcuda::CropFlipNormalizeReformat op; |
| 52 | |
| 53 | // clang-format off |
| 54 | |
| 55 | nvcv::Tensor dst({{dstShape.x, dstShape.y, dstShape.z, 1}, "NHWC"}, benchutils::GetDataType<T>()); |
| 56 | |
| 57 | nvcv::Tensor flipCode({{srcShape.x}, "N"}, nvcv::TYPE_S32); |
| 58 | |
| 59 | nvcv::Tensor base({{baseShape.x, baseShape.y, baseShape.z, 1}, "NHWC"}, nvcv::TYPE_F32); |
| 60 | nvcv::Tensor scale({{scaleShape.x, scaleShape.y, scaleShape.z, 1}, "NHWC"}, nvcv::TYPE_F32); |
| 61 | |
| 62 | nvcv::Tensor crop({{cropShape.x, cropShape.y, cropShape.z, 4}, "NHWC"}, nvcv::TYPE_S32); |
| 63 | |
| 64 | benchutils::FillTensor<int>(flipCode, [](auto &){ return -1; }); |
| 65 | |
| 66 | benchutils::FillTensor<float>(base, benchutils::RandomValues<T>()); |
| 67 | benchutils::FillTensor<float>(scale, benchutils::RandomValues<float>(0.f, 1.f)); |
| 68 | |
| 69 | // Always crop entire source image for easy bandwidth calculations |
| 70 | benchutils::FillTensor<int>(crop, [&srcShape](const long4_16a &c) |
| 71 | { |
| 72 | if (c.w == 2) |
| 73 | { |
| 74 | return (int)srcShape.z; |
| 75 | } |
| 76 | else if (c.w == 3) |
| 77 | { |
| 78 | return (int)srcShape.y; |
| 79 | } |
| 80 | return 0; |
| 81 | }); |
| 82 |
nothing calls this directly
no test coverage detected