GCC Code Coverage Report

Directory: ./
File: tasks/klimovich_v_crs_complex_mat_mul/tbb/src/ops_tbb.cpp
Date: 2026-06-04 20:25:32
Exec Total Coverage
Lines: 54 55 98.2%
Functions: 10 10 100.0%
Branches: 37 50 74.0%
Line Branch Exec Source
1 #include "klimovich_v_crs_complex_mat_mul/tbb/include/ops_tbb.hpp"
2
3 #include <algorithm>
4 #include <cmath>
5 #include <cstddef>
6 #include <vector>
7
8 #include "klimovich_v_crs_complex_mat_mul/common/include/common.hpp"
9 #include "oneapi/tbb/blocked_range.h"
10 #include "oneapi/tbb/enumerable_thread_specific.h"
11 #include "oneapi/tbb/parallel_for.h"
12
13 namespace klimovich_v_crs_complex_mat_mul {
14 namespace {
15
16 struct RowStage {
17 std::vector<int> cols;
18 std::vector<Cplx> vals;
19 };
20
21 struct ThreadCtx {
22 std::vector<Cplx> spa;
23 std::vector<int> touched_by_row;
24 std::vector<int> touched_cols;
25 };
26
27 228 void GustavsonRow(const CrsMatrix &lhs, const CrsMatrix &rhs, int row, ThreadCtx &ctx, RowStage &stage) {
28 auto &spa = ctx.spa;
29 auto &touched_by_row = ctx.touched_by_row;
30
2/2
✓ Branch 0 taken 168 times.
✓ Branch 1 taken 60 times.
 228 auto &touched_cols = ctx.touched_cols;
31 touched_cols.clear();
32
33
2/2
✓ Branch 0 taken 336 times.
✓ Branch 1 taken 228 times.
 564 for (int lp = lhs.row_offsets[row]; lp < lhs.row_offsets[row + 1]; ++lp) {
34 336 const int k = lhs.col_indices[lp];
35 336 const Cplx a_ik = lhs.data[lp];
36
2/2
✓ Branch 0 taken 604 times.
✓ Branch 1 taken 336 times.
 940 for (int rq = rhs.row_offsets[k]; rq < rhs.row_offsets[k + 1]; ++rq) {
37
2/2
✓ Branch 0 taken 336 times.
✓ Branch 1 taken 268 times.
 604 const int j = rhs.col_indices[rq];
38
2/2
✓ Branch 0 taken 336 times.
✓ Branch 1 taken 268 times.
 604 if (touched_by_row[j] != row) {
39
1/2
✓ Branch 0 taken 336 times.
✗ Branch 1 not taken.
 336 touched_by_row[j] = row;
40 touched_cols.push_back(j);
41 336 spa[j] = a_ik * rhs.data[rq];
42 } else {
43 spa[j] += a_ik * rhs.data[rq];
44 }
45 }
46 }
47
48 std::ranges::sort(touched_cols);
49
50 stage.cols.clear();
51 stage.vals.clear();
52 228 stage.cols.reserve(touched_cols.size());
53 228 stage.vals.reserve(touched_cols.size());
54
55
2/2
✓ Branch 0 taken 336 times.
✓ Branch 1 taken 228 times.
 564 for (const int j : touched_cols) {
56
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 320 times.
 336 const Cplx v = spa[j];
57
2/2
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 320 times.
 336 spa[j] = Cplx(0.0, 0.0);
58
3/4
✓ Branch 0 taken 16 times.
✓ Branch 1 taken 320 times.
✓ Branch 2 taken 16 times.
✗ Branch 3 not taken.
 336 if (std::abs(v.real()) > kZeroDropTol || std::abs(v.imag()) > kZeroDropTol) {
59 stage.cols.push_back(j);
60 stage.vals.push_back(v);
61 }
62 }
63 228 }
64
65 40 CrsMatrix Assemble(int rows, int cols, const std::vector<RowStage> &per_row) {
66 CrsMatrix out(rows, cols);
67
2/2
✓ Branch 0 taken 228 times.
✓ Branch 1 taken 40 times.
 268 for (int i = 0; i < rows; ++i) {
68 228 out.row_offsets[i + 1] = out.row_offsets[i] + static_cast<int>(per_row[i].cols.size());
69 }
70
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 out.col_indices.reserve(static_cast<std::size_t>(out.row_offsets[rows]));
71
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 out.data.reserve(static_cast<std::size_t>(out.row_offsets[rows]));
72
2/2
✓ Branch 0 taken 228 times.
✓ Branch 1 taken 40 times.
 268 for (int i = 0; i < rows; ++i) {
73
2/4
✓ Branch 1 taken 228 times.
✗ Branch 2 not taken.
✓ Branch 4 taken 228 times.
✗ Branch 5 not taken.
 228 out.col_indices.insert(out.col_indices.end(), per_row[i].cols.begin(), per_row[i].cols.end());
74 228 out.data.insert(out.data.end(), per_row[i].vals.begin(), per_row[i].vals.end());
75 }
76 40 return out;
77 ✗ }
78
79 } // namespace
80
81
1/2
✓ Branch 2 taken 40 times.
✗ Branch 3 not taken.
 40 KlimovichVCrsComplexMatMulTbb::KlimovichVCrsComplexMatMulTbb(const InType &in) {
82 SetTypeOfTask(GetStaticTypeOfTask());
83 GetInput() = in;
84 40 GetOutput() = CrsMatrix();
85 40 }
86
87 40 bool KlimovichVCrsComplexMatMulTbb::ValidationImpl() {
88 const auto &lhs = std::get<0>(GetInput());
89 const auto &rhs = std::get<1>(GetInput());
90 40 return lhs.n_cols == rhs.n_rows;
91 }
92
93 40 bool KlimovichVCrsComplexMatMulTbb::PreProcessingImpl() {
94 40 return true;
95 }
96
97 40 CrsMatrix KlimovichVCrsComplexMatMulTbb::MultiplyCrs(const CrsMatrix &lhs, const CrsMatrix &rhs) {
98 40 std::vector<RowStage> per_row(static_cast<std::size_t>(lhs.n_rows));
99
100 40 oneapi::tbb::enumerable_thread_specific<ThreadCtx> tls([&rhs] {
101 40 ThreadCtx c;
102
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 c.spa.assign(static_cast<std::size_t>(rhs.n_cols), Cplx(0.0, 0.0));
103
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 c.touched_by_row.assign(static_cast<std::size_t>(rhs.n_cols), -1);
104
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 c.touched_cols.reserve(static_cast<std::size_t>(rhs.n_cols));
105 40 return c;
106
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 });
107
108
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 40 oneapi::tbb::parallel_for(oneapi::tbb::blocked_range<int>(0, lhs.n_rows), [&](const auto &range) {
109 228 auto &ctx = tls.local();
110
2/2
✓ Branch 0 taken 228 times.
✓ Branch 1 taken 228 times.
 456 for (int i = range.begin(); i < range.end(); ++i) {
111 228 GustavsonRow(lhs, rhs, i, ctx, per_row[i]);
112 }
113 228 });
114
115
1/2
✓ Branch 1 taken 40 times.
✗ Branch 2 not taken.
 80 return Assemble(lhs.n_rows, rhs.n_cols, per_row);
116 40 }
117
118 40 bool KlimovichVCrsComplexMatMulTbb::RunImpl() {
119 const auto &lhs = std::get<0>(GetInput());
120 const auto &rhs = std::get<1>(GetInput());
121 40 GetOutput() = MultiplyCrs(lhs, rhs);
122 40 return true;
123 }
124
125 40 bool KlimovichVCrsComplexMatMulTbb::PostProcessingImpl() {
126 40 return true;
127 }
128
129 } // namespace klimovich_v_crs_complex_mat_mul
130