| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "bortsova_a_integrals_rectangle/all/include/ops_all.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | #include <omp.h> | ||
| 5 | |||
| 6 | #include <algorithm> | ||
| 7 | #include <cstdint> | ||
| 8 | #include <vector> | ||
| 9 | |||
| 10 | #include "bortsova_a_integrals_rectangle/common/include/common.hpp" | ||
| 11 | #include "util/include/util.hpp" | ||
| 12 | |||
| 13 | namespace bortsova_a_integrals_rectangle { | ||
| 14 | |||
| 15 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | BortsovaAIntegralsRectangleALL::BortsovaAIntegralsRectangleALL(const InType &in) { |
| 16 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 17 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | GetInput() = in; |
| 18 | 20 | GetOutput() = 0.0; | |
| 19 | 20 | } | |
| 20 | |||
| 21 |
1/2✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
|
20 | bool BortsovaAIntegralsRectangleALL::ValidationImpl() { |
| 22 | const auto &input = GetInput(); | ||
| 23 |
3/6✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 20 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 20 times.
|
20 | return input.func && !input.lower_bounds.empty() && input.lower_bounds.size() == input.upper_bounds.size() && |
| 24 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20 times.
|
20 | input.num_steps > 0; |
| 25 | } | ||
| 26 | |||
| 27 | 20 | bool BortsovaAIntegralsRectangleALL::PreProcessingImpl() { | |
| 28 | const auto &input = GetInput(); | ||
| 29 | 20 | func_ = input.func; | |
| 30 | 20 | num_steps_ = input.num_steps; | |
| 31 | 20 | dims_ = static_cast<int>(input.lower_bounds.size()); | |
| 32 | |||
| 33 | 20 | midpoints_.resize(dims_); | |
| 34 | 20 | volume_ = 1.0; | |
| 35 | 20 | total_points_ = 1; | |
| 36 | |||
| 37 |
2/2✓ Branch 0 taken 34 times.
✓ Branch 1 taken 20 times.
|
54 | for (int di = 0; di < dims_; di++) { |
| 38 | 34 | double step = (input.upper_bounds[di] - input.lower_bounds[di]) / static_cast<double>(num_steps_); | |
| 39 | 34 | volume_ *= step; | |
| 40 | 34 | total_points_ *= num_steps_; | |
| 41 | |||
| 42 | 34 | midpoints_[di].resize(num_steps_); | |
| 43 |
2/2✓ Branch 0 taken 5842 times.
✓ Branch 1 taken 34 times.
|
5876 | for (int si = 0; si < num_steps_; si++) { |
| 44 | 5842 | midpoints_[di][si] = input.lower_bounds[di] + ((si + 0.5) * step); | |
| 45 | } | ||
| 46 | } | ||
| 47 | |||
| 48 | 20 | return true; | |
| 49 | } | ||
| 50 | |||
| 51 | 40 | double BortsovaAIntegralsRectangleALL::ComputePartialSum(int64_t begin, int64_t end) { | |
| 52 | 40 | std::vector<int> indices(dims_, 0); | |
| 53 |
1/4✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
40 | std::vector<double> point(dims_); |
| 54 | |||
| 55 | int64_t temp = begin; | ||
| 56 |
2/2✓ Branch 0 taken 68 times.
✓ Branch 1 taken 40 times.
|
108 | for (int di = dims_ - 1; di >= 0; di--) { |
| 57 | 68 | indices[di] = static_cast<int>(temp % num_steps_); | |
| 58 | 68 | temp /= num_steps_; | |
| 59 | } | ||
| 60 | |||
| 61 | double local_sum = 0.0; | ||
| 62 |
2/2✓ Branch 0 taken 48201 times.
✓ Branch 1 taken 40 times.
|
48241 | for (int64_t pt = begin; pt < end; pt++) { |
| 63 |
2/2✓ Branch 0 taken 110201 times.
✓ Branch 1 taken 48201 times.
|
158402 | for (int di = 0; di < dims_; di++) { |
| 64 | 110201 | point[di] = midpoints_[di][indices[di]]; | |
| 65 | } | ||
| 66 | 48201 | local_sum += func_(point); | |
| 67 | |||
| 68 |
2/2✓ Branch 0 taken 49341 times.
✓ Branch 1 taken 10 times.
|
49351 | for (int di = dims_ - 1; di >= 0; di--) { |
| 69 |
2/2✓ Branch 0 taken 1150 times.
✓ Branch 1 taken 48191 times.
|
49341 | indices[di]++; |
| 70 |
2/2✓ Branch 0 taken 1150 times.
✓ Branch 1 taken 48191 times.
|
49341 | if (indices[di] < num_steps_) { |
| 71 | break; | ||
| 72 | } | ||
| 73 | 1150 | indices[di] = 0; | |
| 74 | } | ||
| 75 | } | ||
| 76 | 40 | return local_sum; | |
| 77 | } | ||
| 78 | |||
| 79 | 20 | bool BortsovaAIntegralsRectangleALL::RunImpl() { | |
| 80 | 20 | int rank = 0; | |
| 81 | 20 | int size = 1; | |
| 82 | 20 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 83 | 20 | MPI_Comm_size(MPI_COMM_WORLD, &size); | |
| 84 | |||
| 85 | 20 | int64_t chunk = total_points_ / size; | |
| 86 | 20 | int64_t remainder = total_points_ % size; | |
| 87 |
2/2✓ Branch 0 taken 19 times.
✓ Branch 1 taken 1 times.
|
20 | int64_t mpi_begin = (rank * chunk) + std::min(static_cast<int64_t>(rank), remainder); |
| 88 |
2/2✓ Branch 0 taken 19 times.
✓ Branch 1 taken 1 times.
|
20 | int64_t mpi_end = mpi_begin + chunk + (static_cast<int64_t>(rank) < remainder ? 1 : 0); |
| 89 | |||
| 90 | 20 | int num_threads = ppc::util::GetNumThreads(); | |
| 91 | 20 | int64_t local_count = mpi_end - mpi_begin; | |
| 92 | 20 | std::vector<double> thread_sums(num_threads, 0.0); | |
| 93 | |||
| 94 | 20 | #pragma omp parallel num_threads(num_threads) default(none) shared(thread_sums, mpi_begin, local_count, num_threads) | |
| 95 | { | ||
| 96 | int tid = omp_get_thread_num(); | ||
| 97 | int64_t th_chunk = local_count / num_threads; | ||
| 98 | int64_t th_rem = local_count % num_threads; | ||
| 99 | int64_t th_begin = mpi_begin + (tid * th_chunk) + std::min(static_cast<int64_t>(tid), th_rem); | ||
| 100 | int64_t th_end = th_begin + th_chunk + (static_cast<int64_t>(tid) < th_rem ? 1 : 0); | ||
| 101 | thread_sums[tid] = ComputePartialSum(th_begin, th_end); | ||
| 102 | } | ||
| 103 | |||
| 104 | 20 | double local_sum = 0.0; | |
| 105 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 20 times.
|
60 | for (int ti = 0; ti < num_threads; ti++) { |
| 106 | 40 | local_sum += thread_sums[ti]; | |
| 107 | } | ||
| 108 | |||
| 109 | 20 | double global_sum = 0.0; | |
| 110 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | MPI_Reduce(&local_sum, &global_sum, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD); |
| 111 |
1/2✓ Branch 1 taken 20 times.
✗ Branch 2 not taken.
|
20 | MPI_Bcast(&global_sum, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD); |
| 112 | |||
| 113 |
1/2✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
|
20 | GetOutput() = global_sum * volume_; |
| 114 | 20 | return true; | |
| 115 | } | ||
| 116 | |||
| 117 | 20 | bool BortsovaAIntegralsRectangleALL::PostProcessingImpl() { | |
| 118 | 20 | return true; | |
| 119 | } | ||
| 120 | |||
| 121 | } // namespace bortsova_a_integrals_rectangle | ||
| 122 |