GCC Code Coverage Report

Directory: ./
File: tasks/klimovich_v_crs_complex_mat_mul/stl/src/ops_stl.cpp
Date: 2026-06-04 20:25:32
Exec Total Coverage
Lines: 63 64 98.4%
Functions: 9 9 100.0%
Branches: 50 68 73.5%
Line Branch Exec Source
1 #include "klimovich_v_crs_complex_mat_mul/stl/include/ops_stl.hpp"
2
3 #include <algorithm>
4 #include <cmath>
5 #include <cstddef>
6 #include <thread>
7 #include <vector>
8
9 #include "klimovich_v_crs_complex_mat_mul/common/include/common.hpp"
10 #include "util/include/util.hpp"
11
12 namespace klimovich_v_crs_complex_mat_mul {
13 namespace {
14
15 struct RowStage {
16 std::vector<int> cols;
17 std::vector<Cplx> vals;
18 };
19
20
2/2
✓ Branch 0 taken 248 times.
✓ Branch 1 taken 208 times.
 456 void GustavsonRow(const CrsMatrix &lhs, const CrsMatrix &rhs, int row, std::vector<Cplx> &spa,
21 std::vector<int> &touched_by_row, std::vector<int> &touched_cols, RowStage &stage) {
22 touched_cols.clear();
23
24
2/2
✓ Branch 0 taken 672 times.
✓ Branch 1 taken 456 times.
 1128 for (int lp = lhs.row_offsets[row]; lp < lhs.row_offsets[row + 1]; ++lp) {
25 672 const int k = lhs.col_indices[lp];
26 672 const Cplx a_ik = lhs.data[lp];
27
2/2
✓ Branch 0 taken 1208 times.
✓ Branch 1 taken 672 times.
 1880 for (int rq = rhs.row_offsets[k]; rq < rhs.row_offsets[k + 1]; ++rq) {
28
2/2
✓ Branch 0 taken 672 times.
✓ Branch 1 taken 536 times.
 1208 const int j = rhs.col_indices[rq];
29
2/2
✓ Branch 0 taken 672 times.
✓ Branch 1 taken 536 times.
 1208 if (touched_by_row[j] != row) {
30
1/2
✓ Branch 0 taken 672 times.
✗ Branch 1 not taken.
 672 touched_by_row[j] = row;
31 touched_cols.push_back(j);
32 672 spa[j] = a_ik * rhs.data[rq];
33 } else {
34 spa[j] += a_ik * rhs.data[rq];
35 }
36 }
37 }
38
39 std::ranges::sort(touched_cols);
40
41 stage.cols.clear();
42 stage.vals.clear();
43 456 stage.cols.reserve(touched_cols.size());
44 456 stage.vals.reserve(touched_cols.size());
45
46
2/2
✓ Branch 0 taken 672 times.
✓ Branch 1 taken 456 times.
 1128 for (const int j : touched_cols) {
47
2/2
✓ Branch 0 taken 32 times.
✓ Branch 1 taken 640 times.
 672 const Cplx v = spa[j];
48
2/2
✓ Branch 0 taken 32 times.
✓ Branch 1 taken 640 times.
 672 spa[j] = Cplx(0.0, 0.0);
49
3/4
✓ Branch 0 taken 32 times.
✓ Branch 1 taken 640 times.
✓ Branch 2 taken 32 times.
✗ Branch 3 not taken.
 672 if (std::abs(v.real()) > kZeroDropTol || std::abs(v.imag()) > kZeroDropTol) {
50 stage.cols.push_back(j);
51 stage.vals.push_back(v);
52 }
53 }
54 456 }
55
56 80 CrsMatrix Assemble(int rows, int cols, const std::vector<RowStage> &per_row) {
57 CrsMatrix out(rows, cols);
58
2/2
✓ Branch 0 taken 456 times.
✓ Branch 1 taken 80 times.
 536 for (int i = 0; i < rows; ++i) {
59 456 out.row_offsets[i + 1] = out.row_offsets[i] + static_cast<int>(per_row[i].cols.size());
60 }
61
1/2
✓ Branch 1 taken 80 times.
✗ Branch 2 not taken.
 80 out.col_indices.reserve(static_cast<std::size_t>(out.row_offsets[rows]));
62
1/2
✓ Branch 1 taken 80 times.
✗ Branch 2 not taken.
 80 out.data.reserve(static_cast<std::size_t>(out.row_offsets[rows]));
63
2/2
✓ Branch 0 taken 456 times.
✓ Branch 1 taken 80 times.
 536 for (int i = 0; i < rows; ++i) {
64
2/4
✓ Branch 1 taken 456 times.
✗ Branch 2 not taken.
✓ Branch 4 taken 456 times.
✗ Branch 5 not taken.
 456 out.col_indices.insert(out.col_indices.end(), per_row[i].cols.begin(), per_row[i].cols.end());
65 456 out.data.insert(out.data.end(), per_row[i].vals.begin(), per_row[i].vals.end());
66 }
67 80 return out;
68 ✗ }
69
70 } // namespace
71
72
1/2
✓ Branch 2 taken 80 times.
✗ Branch 3 not taken.
 80 KlimovichVCrsComplexMatMulStl::KlimovichVCrsComplexMatMulStl(const InType &in) {
73 SetTypeOfTask(GetStaticTypeOfTask());
74 GetInput() = in;
75 80 GetOutput() = CrsMatrix();
76 80 }
77
78 80 bool KlimovichVCrsComplexMatMulStl::ValidationImpl() {
79 const auto &lhs = std::get<0>(GetInput());
80 const auto &rhs = std::get<1>(GetInput());
81 80 return lhs.n_cols == rhs.n_rows;
82 }
83
84 80 bool KlimovichVCrsComplexMatMulStl::PreProcessingImpl() {
85 80 return true;
86 }
87
88 80 CrsMatrix KlimovichVCrsComplexMatMulStl::MultiplyCrs(const CrsMatrix &lhs, const CrsMatrix &rhs) {
89 80 std::vector<RowStage> per_row(static_cast<std::size_t>(lhs.n_rows));
90
91
1/2
✓ Branch 1 taken 80 times.
✗ Branch 2 not taken.
 80 const int requested_threads = std::max(1, ppc::util::GetNumThreads());
92
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 68 times.
 80 const int total_rows = lhs.n_rows;
93
2/2
✓ Branch 0 taken 54 times.
✓ Branch 1 taken 26 times.
 80 const int worker_count = std::max(1, std::min(requested_threads, total_rows));
94
95 180 auto worker = [&](int begin, int end) {
96 180 std::vector<Cplx> spa(static_cast<std::size_t>(rhs.n_cols), Cplx(0.0, 0.0));
97
1/4
✓ Branch 1 taken 180 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 180 std::vector<int> touched_by_row(static_cast<std::size_t>(rhs.n_cols), -1);
98 180 std::vector<int> touched_cols;
99
1/2
✓ Branch 1 taken 180 times.
✗ Branch 2 not taken.
 180 touched_cols.reserve(static_cast<std::size_t>(rhs.n_cols));
100
101
2/2
✓ Branch 0 taken 456 times.
✓ Branch 1 taken 180 times.
 636 for (int i = begin; i < end; ++i) {
102
1/2
✓ Branch 1 taken 456 times.
✗ Branch 2 not taken.
 456 GustavsonRow(lhs, rhs, i, spa, touched_by_row, touched_cols, per_row[i]);
103 }
104 180 };
105
106 80 const int chunk = total_rows / worker_count;
107 80 const int leftover = total_rows % worker_count;
108
109 80 std::vector<std::thread> pool;
110
1/2
✓ Branch 1 taken 80 times.
✗ Branch 2 not taken.
 80 pool.reserve(static_cast<std::size_t>(worker_count));
111
112 int cursor = 0;
113
2/2
✓ Branch 0 taken 180 times.
✓ Branch 1 taken 80 times.
 260 for (int worker_index = 0; worker_index < worker_count; ++worker_index) {
114 180 const int begin = cursor;
115
2/2
✓ Branch 0 taken 132 times.
✓ Branch 1 taken 48 times.
 180 const int end = begin + chunk + (worker_index < leftover ? 1 : 0);
116 cursor = end;
117
1/2
✓ Branch 0 taken 180 times.
✗ Branch 1 not taken.
 180 if (begin < end) {
118
1/2
✓ Branch 1 taken 180 times.
✗ Branch 2 not taken.
 180 pool.emplace_back(worker, begin, end);
119 }
120 }
121
2/2
✓ Branch 0 taken 180 times.
✓ Branch 1 taken 80 times.
 260 for (auto &th : pool) {
122
1/2
✓ Branch 1 taken 180 times.
✗ Branch 2 not taken.
 180 th.join();
123 }
124
125
1/2
✓ Branch 1 taken 80 times.
✗ Branch 2 not taken.
 160 return Assemble(lhs.n_rows, rhs.n_cols, per_row);
126 80 }
127
128 80 bool KlimovichVCrsComplexMatMulStl::RunImpl() {
129 const auto &lhs = std::get<0>(GetInput());
130 const auto &rhs = std::get<1>(GetInput());
131 80 GetOutput() = MultiplyCrs(lhs, rhs);
132 80 return true;
133 }
134
135 80 bool KlimovichVCrsComplexMatMulStl::PostProcessingImpl() {
136 80 return true;
137 }
138
139 } // namespace klimovich_v_crs_complex_mat_mul
140