| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "krykov_e_simple_iterations/mpi/include/ops_mpi.hpp" | ||
| 2 | |||
| 3 | #include <mpi.h> | ||
| 4 | |||
| 5 | #include <cctype> | ||
| 6 | #include <cmath> | ||
| 7 | #include <cstddef> | ||
| 8 | #include <ranges> | ||
| 9 | #include <utility> | ||
| 10 | #include <vector> | ||
| 11 | |||
| 12 | #include "krykov_e_simple_iterations/common/include/common.hpp" | ||
| 13 | |||
| 14 | namespace krykov_e_simple_iterations { | ||
| 15 | namespace { | ||
| 16 | |||
| 17 | constexpr double kEps = 1e-5; | ||
| 18 | constexpr int kMaxIter = 10000; | ||
| 19 | |||
| 20 | 92 | void CalculateLocalXNew(int start, int count, size_t n, const std::vector<double> &local_a, | |
| 21 | const std::vector<double> &local_b, const std::vector<double> &x, | ||
| 22 | std::vector<double> &local_x_new) { | ||
| 23 |
2/2✓ Branch 0 taken 123 times.
✓ Branch 1 taken 92 times.
|
215 | for (int i = 0; i < count; ++i) { |
| 24 | 123 | int global_i = start + i; | |
| 25 | double sum = 0.0; | ||
| 26 |
2/2✓ Branch 0 taken 367 times.
✓ Branch 1 taken 123 times.
|
490 | for (size_t j = 0; j < n; ++j) { |
| 27 | if (std::cmp_not_equal(j, global_i)) { | ||
| 28 | 244 | sum += local_a[(i * n) + j] * x[j]; | |
| 29 | } | ||
| 30 | } | ||
| 31 | 123 | local_x_new[i] = (local_b[i] - sum) / local_a[(i * n) + global_i]; | |
| 32 | } | ||
| 33 | 92 | } | |
| 34 | |||
| 35 | double CalculateLocalNorm(int start, int count, const std::vector<double> &x_new, const std::vector<double> &x) { | ||
| 36 | double local_norm = 0.0; | ||
| 37 |
2/2✓ Branch 0 taken 123 times.
✓ Branch 1 taken 92 times.
|
215 | for (int i = 0; i < count; ++i) { |
| 38 | 123 | int gi = start + i; | |
| 39 | 123 | double diff = x_new[gi] - x[gi]; | |
| 40 | 123 | local_norm += diff * diff; | |
| 41 | } | ||
| 42 | return local_norm; | ||
| 43 | } | ||
| 44 | |||
| 45 |
1/2✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
|
20 | void CalculateChunkSizesAndDispls(int size, int n, std::vector<int> &chunk_sizes, std::vector<int> &displs) { |
| 46 |
1/2✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
|
20 | if (displs.empty()) { |
| 47 | return; | ||
| 48 | } | ||
| 49 | 20 | int base = n / size; | |
| 50 | 20 | int rem = n % size; | |
| 51 | |||
| 52 | 20 | displs[0] = 0; | |
| 53 |
2/2✓ Branch 0 taken 40 times.
✓ Branch 1 taken 20 times.
|
60 | for (int i = 0; i < size; ++i) { |
| 54 |
4/4✓ Branch 0 taken 32 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 20 times.
✓ Branch 3 taken 20 times.
|
72 | chunk_sizes[i] = base + (i < rem ? 1 : 0); |
| 55 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 20 times.
|
40 | if (i > 0) { |
| 56 | 20 | displs[i] = displs[i - 1] + chunk_sizes[i - 1]; | |
| 57 | } | ||
| 58 | } | ||
| 59 | } | ||
| 60 | |||
| 61 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
10 | void CalculateMatrixChunkSizesAndDispls(int size, int n, std::vector<int> &matrix_chunk_sizes, |
| 62 | std::vector<int> &matrix_displs) { | ||
| 63 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
10 | if (matrix_displs.empty()) { |
| 64 | return; | ||
| 65 | } | ||
| 66 | 10 | int base = n / size; | |
| 67 | 10 | int rem = n % size; | |
| 68 | |||
| 69 | 10 | matrix_displs[0] = 0; | |
| 70 |
2/2✓ Branch 0 taken 20 times.
✓ Branch 1 taken 10 times.
|
30 | for (int i = 0; i < size; ++i) { |
| 71 |
2/2✓ Branch 0 taken 16 times.
✓ Branch 1 taken 4 times.
|
20 | int rows = base + (i < rem ? 1 : 0); |
| 72 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
20 | matrix_chunk_sizes[i] = rows * n; |
| 73 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
20 | if (i > 0) { |
| 74 | 10 | matrix_displs[i] = matrix_displs[i - 1] + matrix_chunk_sizes[i - 1]; | |
| 75 | } | ||
| 76 | } | ||
| 77 | } | ||
| 78 | |||
| 79 | } // namespace | ||
| 80 | |||
| 81 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | KrykovESimpleIterationsMPI::KrykovESimpleIterationsMPI(const InType &in) { |
| 82 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 83 | GetInput() = in; | ||
| 84 | 10 | } | |
| 85 | |||
| 86 | 10 | bool KrykovESimpleIterationsMPI::ValidationImpl() { | |
| 87 | const auto &[n, a, b] = GetInput(); | ||
| 88 |
3/6✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 10 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 10 times.
|
10 | return n > 0 && a.size() == n * n && b.size() == n; |
| 89 | } | ||
| 90 | |||
| 91 | 10 | bool KrykovESimpleIterationsMPI::PreProcessingImpl() { | |
| 92 | 10 | return true; | |
| 93 | } | ||
| 94 | |||
| 95 | 10 | bool KrykovESimpleIterationsMPI::RunImpl() { | |
| 96 | 10 | int size = 0; | |
| 97 | 10 | int rank = 0; | |
| 98 | 10 | MPI_Comm_size(MPI_COMM_WORLD, &size); | |
| 99 | 10 | MPI_Comm_rank(MPI_COMM_WORLD, &rank); | |
| 100 | |||
| 101 | 10 | size_t n = 0; | |
| 102 | 10 | std::vector<double> a; | |
| 103 | 10 | std::vector<double> b; | |
| 104 | |||
| 105 |
2/2✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
|
10 | if (rank == 0) { |
| 106 | const auto &input = GetInput(); | ||
| 107 |
1/2✓ Branch 1 taken 5 times.
✗ Branch 2 not taken.
|
5 | n = std::get<0>(input); |
| 108 |
1/2✓ Branch 1 taken 5 times.
✗ Branch 2 not taken.
|
5 | a = std::get<1>(input); |
| 109 |
1/2✓ Branch 1 taken 5 times.
✗ Branch 2 not taken.
|
5 | b = std::get<2>(input); |
| 110 | } | ||
| 111 | |||
| 112 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | MPI_Bcast(&n, 1, MPI_UNSIGNED_LONG, 0, MPI_COMM_WORLD); |
| 113 | |||
| 114 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<int> chunk_sizes(size); |
| 115 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<int> displs(size); |
| 116 | 10 | CalculateChunkSizesAndDispls(size, static_cast<int>(n), chunk_sizes, displs); | |
| 117 | |||
| 118 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<int> matrix_chunk_sizes(size); |
| 119 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<int> matrix_displs(size); |
| 120 | 10 | CalculateMatrixChunkSizesAndDispls(size, static_cast<int>(n), matrix_chunk_sizes, matrix_displs); | |
| 121 | |||
| 122 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | int local_rows = chunk_sizes[rank]; |
| 123 | 10 | int local_matrix_size = matrix_chunk_sizes[rank]; | |
| 124 | |||
| 125 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<double> local_a(local_matrix_size); |
| 126 |
3/6✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 5 times.
✓ Branch 4 taken 5 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
|
10 | std::vector<double> local_b(local_rows); |
| 127 | |||
| 128 |
3/4✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 3 taken 10 times.
✗ Branch 4 not taken.
|
15 | MPI_Scatterv(rank == 0 ? a.data() : nullptr, matrix_chunk_sizes.data(), matrix_displs.data(), MPI_DOUBLE, |
| 129 | local_a.data(), local_matrix_size, MPI_DOUBLE, 0, MPI_COMM_WORLD); | ||
| 130 | |||
| 131 |
3/4✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 3 taken 10 times.
✗ Branch 4 not taken.
|
15 | MPI_Scatterv(rank == 0 ? b.data() : nullptr, chunk_sizes.data(), displs.data(), MPI_DOUBLE, local_b.data(), |
| 132 | local_rows, MPI_DOUBLE, 0, MPI_COMM_WORLD); | ||
| 133 | |||
| 134 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | int start = displs[rank]; |
| 135 | int count = local_rows; | ||
| 136 | |||
| 137 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<double> x(n, 0.0); |
| 138 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<double> x_new(n, 0.0); |
| 139 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<double> local_x_new(count, 0.0); |
| 140 | |||
| 141 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<int> recv_counts(size); |
| 142 |
1/4✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
10 | std::vector<int> allgather_displs(size); |
| 143 | 10 | CalculateChunkSizesAndDispls(size, static_cast<int>(n), recv_counts, allgather_displs); | |
| 144 | |||
| 145 |
1/2✓ Branch 0 taken 92 times.
✗ Branch 1 not taken.
|
92 | for (int iter = 0; iter < kMaxIter; ++iter) { |
| 146 | 92 | CalculateLocalXNew(start, count, n, local_a, local_b, x, local_x_new); | |
| 147 | |||
| 148 |
1/2✓ Branch 1 taken 92 times.
✗ Branch 2 not taken.
|
92 | MPI_Allgatherv(local_x_new.data(), count, MPI_DOUBLE, x_new.data(), recv_counts.data(), allgather_displs.data(), |
| 149 | MPI_DOUBLE, MPI_COMM_WORLD); | ||
| 150 | |||
| 151 | 92 | double local_norm = CalculateLocalNorm(start, count, x_new, x); | |
| 152 | 92 | double global_norm = 0.0; | |
| 153 |
1/2✓ Branch 1 taken 92 times.
✗ Branch 2 not taken.
|
92 | MPI_Allreduce(&local_norm, &global_norm, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD); |
| 154 | |||
| 155 |
1/2✓ Branch 1 taken 92 times.
✗ Branch 2 not taken.
|
92 | x = x_new; |
| 156 | |||
| 157 |
2/2✓ Branch 0 taken 82 times.
✓ Branch 1 taken 10 times.
|
92 | if (std::sqrt(global_norm) < kEps) { |
| 158 | break; | ||
| 159 | } | ||
| 160 | } | ||
| 161 | |||
| 162 |
1/2✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
|
10 | GetOutput() = x; |
| 163 | |||
| 164 | 10 | return true; | |
| 165 | } | ||
| 166 | |||
| 167 | 10 | bool KrykovESimpleIterationsMPI::PostProcessingImpl() { | |
| 168 | 10 | return true; | |
| 169 | } | ||
| 170 | |||
| 171 | } // namespace krykov_e_simple_iterations | ||
| 172 |