GCC Code Coverage Report

Directory: ./
File: tasks/gasenin_l_image_smooth/mpi/src/ops_mpi.cpp
Date: 2026-01-10 02:40:41
Exec Total Coverage
Lines: 121 121 100.0%
Functions: 10 10 100.0%
Branches: 69 82 84.1%
Line Branch Exec Source
1 #include "gasenin_l_image_smooth/mpi/include/ops_mpi.hpp"
2
3 #include <mpi.h>
4
5 #include <algorithm>
6 #include <array>
7 #include <cstddef>
8 #include <cstdint>
9 #include <vector>
10
11 #include "gasenin_l_image_smooth/common/include/common.hpp"
12
13 namespace gasenin_l_image_smooth {
14
15
1/2
✓ Branch 1 taken 12 times.
✗ Branch 2 not taken.
 12 GaseninLImageSmoothMPI::GaseninLImageSmoothMPI(const InType &in) {
16 SetTypeOfTask(GetStaticTypeOfTask());
17 GetInput() = in;
18 12 }
19
20 12 bool GaseninLImageSmoothMPI::ValidationImpl() {
21 12 int rank = 0;
22 12 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
23
24 12 std::array<int, 4> params = {0, 0, 0, 0};
25
26
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 6 times.
 12 if (rank == 0) {
27 6 params[1] = GetInput().width;
28 6 params[2] = GetInput().height;
29 6 params[3] = GetInput().kernel_size;
30
31
3/6
✓ Branch 0 taken 6 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 6 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 6 times.
✗ Branch 5 not taken.
 6 if (params[1] > 0 && params[2] > 0 && params[3] > 0) {
32 6 params[0] = 1;
33 } else {
34 params[0] = 0;
35 }
36 }
37
38 12 MPI_Bcast(params.data(), 4, MPI_INT, 0, MPI_COMM_WORLD);
39
40
1/2
✓ Branch 0 taken 12 times.
✗ Branch 1 not taken.
 12 if (params[0] == 0) {
41 return false;
42 }
43
44 12 GetInput().width = params[1];
45 12 GetInput().height = params[2];
46 12 GetInput().kernel_size = params[3];
47
48 12 return true;
49 }
50
51 12 bool GaseninLImageSmoothMPI::PreProcessingImpl() {
52 12 int rank = 0;
53 12 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
54
55
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 6 times.
 12 if (rank == 0) {
56 GetOutput() = GetInput();
57 std::fill(GetOutput().data.begin(), GetOutput().data.end(), 0);
58 } else {
59 6 GetOutput().width = GetInput().width;
60 6 GetOutput().height = GetInput().height;
61 6 GetOutput().kernel_size = GetInput().kernel_size;
62 }
63 12 return true;
64 }
65
66 namespace {
67
68 void CalculateRowDistribution(int rank, int size, int height, int &start_row, int &end_row, int &local_rows) {
69 12 const int base_rows = height / size;
70 12 const int extra_rows = height % size;
71
72 12 start_row = (rank * base_rows) + std::min(rank, extra_rows);
73
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 end_row = start_row + base_rows + (rank < extra_rows ? 1 : 0);
74 12 local_rows = end_row - start_row;
75 }
76
77 11 void PrepareScatterParameters(int rank, int size, int width, int height, int kernel_radius, int base_rows,
78 int extra_rows, std::vector<int> &sendcounts, std::vector<int> &displs) {
79
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 5 times.
 11 if (rank == 0) {
80 6 sendcounts.resize(size);
81 6 displs.resize(size);
82
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 6 times.
 18 for (int process_idx = 0; process_idx < size; ++process_idx) {
83
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 const int row_start = (process_idx * base_rows) + std::min(process_idx, extra_rows);
84
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 const int row_end = row_start + base_rows + (process_idx < extra_rows ? 1 : 0);
85
86
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 if (row_end > row_start) {
87 11 const int actual_start = std::max(0, row_start - kernel_radius);
88 11 const int actual_end = std::min(height, row_end + kernel_radius);
89 11 sendcounts[process_idx] = (actual_end - actual_start) * width;
90 11 displs[process_idx] = actual_start * width;
91 } else {
92 1 sendcounts[process_idx] = 0;
93 1 displs[process_idx] = 0;
94 }
95 }
96 }
97 11 }
98
99 10824 void ProcessInteriorPixel(int buffer_y, int col, int width, int kernel_size, const std::vector<uint8_t> &local_data,
100 std::vector<uint8_t> &local_result, int local_row_idx) {
101 10824 const int kernel_sq = kernel_size * kernel_size;
102 10824 const int radius = kernel_size / 2;
103
104 10824 const auto *row_ptr = local_data.data() + (static_cast<ptrdiff_t>(buffer_y - radius) * width) + (col - radius);
105 int sum = 0;
106
107
2/2
✓ Branch 0 taken 53096 times.
✓ Branch 1 taken 10824 times.
 63920 for (int kernel_y = 0; kernel_y < kernel_size; ++kernel_y) {
108
2/2
✓ Branch 0 taken 289448 times.
✓ Branch 1 taken 53096 times.
 342544 for (int kernel_x = 0; kernel_x < kernel_size; ++kernel_x) {
109 289448 sum += row_ptr[kernel_x];
110 }
111 53096 row_ptr += width;
112 }
113 10824 local_result[(local_row_idx * width) + col] = static_cast<uint8_t>(sum / kernel_sq);
114 10824 }
115
116 1505 void ProcessBorderPixel(int global_y, int col, int width, int height, int kernel_radius, int actual_start,
117 const std::vector<uint8_t> &local_data, std::vector<uint8_t> &local_result, int local_row_idx) {
118 int sum = 0;
119 int count = 0;
120
121
2/2
✓ Branch 0 taken 8435 times.
✓ Branch 1 taken 1505 times.
 9940 for (int kernel_y = -kernel_radius; kernel_y <= kernel_radius; ++kernel_y) {
122 8435 const int global_y_offset = global_y + kernel_y;
123
2/2
✓ Branch 0 taken 7780 times.
✓ Branch 1 taken 655 times.
 8435 const int clamped_y = Clamp(global_y_offset, 0, height - 1);
124 8435 const int buffer_y = clamped_y - actual_start;
125
126
2/2
✓ Branch 0 taken 50761 times.
✓ Branch 1 taken 8435 times.
 59196 for (int kernel_x = -kernel_radius; kernel_x <= kernel_radius; ++kernel_x) {
127
2/2
✓ Branch 0 taken 46912 times.
✓ Branch 1 taken 3849 times.
 50761 const int global_x_offset = Clamp(col + kernel_x, 0, width - 1);
128 50761 sum += local_data[(buffer_y * width) + global_x_offset];
129 50761 ++count;
130 }
131 }
132
133 1505 const int index = (local_row_idx * width) + col;
134
1/2
✓ Branch 0 taken 1505 times.
✗ Branch 1 not taken.
 1505 local_result[index] = (count > 0) ? static_cast<uint8_t>(sum / count) : 0;
135 1505 }
136
137 11 void ProcessLocalRows(int start_row, int local_rows, int width, int height, int kernel_size, int kernel_radius,
138 int actual_start, const std::vector<uint8_t> &local_data, std::vector<uint8_t> &local_result) {
139 11 const int row_offset_in_buffer = start_row - actual_start;
140
141
2/2
✓ Branch 0 taken 199 times.
✓ Branch 1 taken 11 times.
 210 for (int local_row_idx = 0; local_row_idx < local_rows; ++local_row_idx) {
142 199 const int global_y = start_row + local_row_idx;
143 199 const int buffer_y = row_offset_in_buffer + local_row_idx;
144
4/4
✓ Branch 0 taken 191 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 7 times.
✓ Branch 3 taken 184 times.
 199 const bool is_global_border_y = (global_y < kernel_radius) || (global_y >= height - kernel_radius);
145
146
2/2
✓ Branch 0 taken 12329 times.
✓ Branch 1 taken 199 times.
 12528 for (int col = 0; col < width; ++col) {
147
4/4
✓ Branch 0 taken 11942 times.
✓ Branch 1 taken 387 times.
✓ Branch 2 taken 11555 times.
✓ Branch 3 taken 387 times.
 12329 const bool is_border_col = (col < kernel_radius) || (col >= width - kernel_radius);
148
149
2/2
✓ Branch 0 taken 10824 times.
✓ Branch 1 taken 731 times.
 11555 if (!is_global_border_y && !is_border_col) {
150 10824 ProcessInteriorPixel(buffer_y, col, width, kernel_size, local_data, local_result, local_row_idx);
151 } else {
152 1505 ProcessBorderPixel(global_y, col, width, height, kernel_radius, actual_start, local_data, local_result,
153 local_row_idx);
154 }
155 }
156 }
157 11 }
158
159 11 void PrepareGatherParameters(int rank, int size, int width, int height, std::vector<int> &sendcounts,
160 std::vector<int> &displs) {
161
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 5 times.
 11 if (rank == 0) {
162 6 sendcounts.resize(size);
163 6 displs.resize(size);
164
165 6 const int base_rows = height / size;
166 6 const int extra_rows = height % size;
167
168
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 6 times.
 18 for (int process_idx = 0; process_idx < size; ++process_idx) {
169
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 const int row_start = (process_idx * base_rows) + std::min(process_idx, extra_rows);
170
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 const int row_end = row_start + base_rows + (process_idx < extra_rows ? 1 : 0);
171 12 sendcounts[process_idx] = (row_end - row_start) * width;
172 12 displs[process_idx] = row_start * width;
173 }
174 }
175 11 }
176
177 } // namespace
178
179 12 bool GaseninLImageSmoothMPI::RunImpl() {
180 12 int rank = 0;
181 12 int size = 0;
182 12 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
183 12 MPI_Comm_size(MPI_COMM_WORLD, &size);
184
185 12 const int width = GetInput().width;
186 12 const int height = GetInput().height;
187 12 const int kernel_size = GetInput().kernel_size;
188 12 const int kernel_radius = kernel_size / 2;
189
190 int start_row = 0;
191 int end_row = 0;
192 int local_rows = 0;
193
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 CalculateRowDistribution(rank, size, height, start_row, end_row, local_rows);
194
195
2/2
✓ Branch 0 taken 11 times.
✓ Branch 1 taken 1 times.
 12 if (local_rows <= 0) {
196 return true;
197 }
198
199 const int overlap_top = kernel_radius;
200 const int overlap_bottom = kernel_radius;
201 11 const int actual_start = std::max(0, start_row - overlap_top);
202 11 const int actual_end = std::min(height, end_row + overlap_bottom);
203 11 const int extended_rows = actual_end - actual_start;
204 11 const int recv_count = extended_rows * width;
205
206 11 std::vector<uint8_t> local_data(recv_count);
207
208 11 const int base_rows = height / size;
209 11 const int extra_rows = height % size;
210 11 std::vector<int> sendcounts;
211 11 std::vector<int> displs;
212
213
1/2
✓ Branch 1 taken 11 times.
✗ Branch 2 not taken.
 11 PrepareScatterParameters(rank, size, width, height, kernel_radius, base_rows, extra_rows, sendcounts, displs);
214
215
3/4
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 5 times.
✓ Branch 3 taken 11 times.
✗ Branch 4 not taken.
 17 MPI_Scatterv((rank == 0 ? GetInput().data.data() : nullptr), sendcounts.data(), displs.data(), MPI_UNSIGNED_CHAR,
216 local_data.data(), recv_count, MPI_UNSIGNED_CHAR, 0, MPI_COMM_WORLD);
217
218
1/4
✓ Branch 1 taken 11 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 11 std::vector<uint8_t> local_result(static_cast<size_t>(local_rows) * width);
219
220 11 ProcessLocalRows(start_row, local_rows, width, height, kernel_size, kernel_radius, actual_start, local_data,
221 local_result);
222
223
1/2
✓ Branch 1 taken 11 times.
✗ Branch 2 not taken.
 11 PrepareGatherParameters(rank, size, width, height, sendcounts, displs);
224
225
1/2
✓ Branch 1 taken 11 times.
✗ Branch 2 not taken.
 11 MPI_Gatherv(local_result.data(), local_rows * width, MPI_UNSIGNED_CHAR,
226
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 5 times.
 11 (rank == 0 ? GetOutput().data.data() : nullptr), sendcounts.data(), displs.data(), MPI_UNSIGNED_CHAR, 0,
227 MPI_COMM_WORLD);
228
229 return true;
230 }
231
232 12 bool GaseninLImageSmoothMPI::PostProcessingImpl() {
233 12 return true;
234 }
235
236 } // namespace gasenin_l_image_smooth
237