| 23 | |
| 24 | template<typename T> |
| 25 | inline void MinMaxLoc(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 | long maxLocs = state.get_int64("maxLocations"); |
| 31 | |
| 32 | // clang-format off |
| 33 | |
| 34 | nvcv::Tensor minVal({{shape.x}, "N"}, nvcv::TYPE_U32); |
| 35 | nvcv::Tensor minLoc({{shape.x, maxLocs}, "NM"}, nvcv::TYPE_2S32); |
| 36 | nvcv::Tensor numMin({{shape.x}, "N"}, nvcv::TYPE_S32); |
| 37 | |
| 38 | nvcv::Tensor maxVal({{shape.x}, "N"}, nvcv::TYPE_U32); |
| 39 | nvcv::Tensor maxLoc({{shape.x, maxLocs}, "NM"}, nvcv::TYPE_2S32); |
| 40 | nvcv::Tensor numMax({{shape.x}, "N"}, nvcv::TYPE_S32); |
| 41 | |
| 42 | // clang-format on |
| 43 | |
| 44 | // R/W bandwidth rationale: |
| 45 | // 1 read to find min/max + 1 read to collect their locations |
| 46 | // 2 writes of min/max values (U32), locations (2S32) and quantity (S32) |
| 47 | state.add_global_memory_reads(shape.x * shape.y * shape.z * sizeof(T) * 2); |
| 48 | state.add_global_memory_writes(shape.x * (sizeof(uint32_t) + maxLocs * sizeof(int2) + sizeof(int)) * 2); |
| 49 | |
| 50 | cvcuda::MinMaxLoc op; |
| 51 | |
| 52 | // clang-format off |
| 53 | |
| 54 | if (varShape < 0) // negative var shape means use Tensor |
| 55 | { |
| 56 | nvcv::Tensor src({{shape.x, shape.y, shape.z, 1}, "NHWC"}, benchutils::GetDataType<T>()); |
| 57 | |
| 58 | benchutils::FillTensorWithMinMax<T>(src, maxLocs); |
| 59 | |
| 60 | state.exec(nvbench::exec_tag::sync, |
| 61 | [&op, &src, &minVal, &minLoc, &numMin, &maxVal, &maxLoc, &numMax](nvbench::launch &launch) |
| 62 | { |
| 63 | op(launch.get_stream(), src, minVal, minLoc, numMin, maxVal, maxLoc, numMax); |
| 64 | }); |
| 65 | } |
| 66 | else // zero and positive var shape means use ImageBatchVarShape |
| 67 | { |
| 68 | nvcv::ImageBatchVarShape src(shape.x); |
| 69 | |
| 70 | benchutils::FillImageBatchWithMinMax<T>(src, long2{shape.z, shape.y}, long2{varShape, varShape}, maxLocs); |
| 71 | |
| 72 | state.exec(nvbench::exec_tag::sync, |
| 73 | [&op, &src, &minVal, &minLoc, &numMin, &maxVal, &maxLoc, &numMax](nvbench::launch &launch) |
| 74 | { |
| 75 | op(launch.get_stream(), src, minVal, minLoc, numMin, maxVal, maxLoc, numMax); |
| 76 | }); |
| 77 | } |
| 78 | } |
| 79 | catch (const std::exception &err) |
| 80 | { |
| 81 | state.skip(err.what()); |
| 82 | } |