| 23 | |
| 24 | template<typename ST> |
| 25 | inline void PairwiseMatcher(nvbench::state &state, nvbench::type_list<ST>) |
| 26 | try |
| 27 | { |
| 28 | long3 shape = benchutils::GetShape<3>(state.get_string("shape")); |
| 29 | |
| 30 | int matchesPerPoint = static_cast<int>(state.get_int64("matchesPerPoint")); |
| 31 | |
| 32 | bool crossCheck = state.get_string("crossCheck") == "T"; |
| 33 | bool readNumSets = state.get_string("readNumSets") == "T"; |
| 34 | bool writeDistances = state.get_string("writeDistances") == "T"; |
| 35 | |
| 36 | NVCVNormType normType = benchutils::GetNormType(state.get_string("normType")); |
| 37 | |
| 38 | NVCVPairwiseMatcherType algoChoice; |
| 39 | |
| 40 | if (state.get_string("algoChoice") == "BRUTE_FORCE") |
| 41 | { |
| 42 | algoChoice = NVCV_BRUTE_FORCE; |
| 43 | } |
| 44 | else |
| 45 | { |
| 46 | throw std::invalid_argument("Unexpected algorithm choice = " + state.get_string("algoChoice")); |
| 47 | } |
| 48 | |
| 49 | int maxMatches = shape.y * matchesPerPoint; |
| 50 | |
| 51 | cvcuda::PairwiseMatcher op(algoChoice); |
| 52 | |
| 53 | state.add_global_memory_reads((crossCheck ? 3 : 2) * shape.x * shape.y * shape.z * sizeof(ST)); |
| 54 | state.add_global_memory_writes(shape.x * (sizeof(int) + maxMatches * (2 * sizeof(int) + sizeof(float)))); |
| 55 | |
| 56 | // clang-format off |
| 57 | |
| 58 | nvcv::Tensor set1({{shape.x, shape.y, shape.z}, "NMD"}, benchutils::GetDataType<ST>()); |
| 59 | nvcv::Tensor set2({{shape.x, shape.y, shape.z}, "NMD"}, benchutils::GetDataType<ST>()); |
| 60 | |
| 61 | nvcv::Tensor matches({{shape.x, maxMatches, 2}, "NMD"}, nvcv::TYPE_S32); |
| 62 | |
| 63 | nvcv::Tensor numMatches({{shape.x}, "N"}, nvcv::TYPE_S32); |
| 64 | |
| 65 | nvcv::Tensor numSet1, numSet2, distances; |
| 66 | |
| 67 | if (readNumSets) |
| 68 | { |
| 69 | numSet1 = nvcv::Tensor({{shape.x}, "N"}, nvcv::TYPE_S32); |
| 70 | numSet2 = nvcv::Tensor({{shape.x}, "N"}, nvcv::TYPE_S32); |
| 71 | |
| 72 | benchutils::FillTensor<int>(numSet1, [&shape](auto &){ return shape.y; }); |
| 73 | benchutils::FillTensor<int>(numSet2, [&shape](auto &){ return shape.y; }); |
| 74 | } |
| 75 | if (writeDistances) |
| 76 | { |
| 77 | distances = nvcv::Tensor({{shape.x, maxMatches}, "NM"}, nvcv::TYPE_F32); |
| 78 | } |
| 79 | |
| 80 | benchutils::FillTensor<ST>(set1, benchutils::RandomValues<ST>()); |
| 81 | benchutils::FillTensor<ST>(set2, benchutils::RandomValues<ST>()); |
| 82 |
nothing calls this directly
no test coverage detected