GCC Code Coverage Report

Directory: ./
File: tasks/shilin_n_gauss_filter_vertical_split/mpi/src/ops_mpi.cpp
Date: 2026-01-27 01:59:34
Exec Total Coverage
Lines: 178 183 97.3%
Functions: 12 14 85.7%
Branches: 107 152 70.4%
Line Branch Exec Source
1 #include "shilin_n_gauss_filter_vertical_split/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 "shilin_n_gauss_filter_vertical_split/common/include/common.hpp"
12
13 namespace shilin_n_gauss_filter_vertical_split {
14
15
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 ShilinNGaussFilterVerticalSplitMPI::ShilinNGaussFilterVerticalSplitMPI(const InType &in) {
16 SetTypeOfTask(GetStaticTypeOfTask());
17 14 int rank = 0;
18
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
19
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank == 0) {
20 GetInput() = in;
21 }
22 14 GetOutput() = std::vector<uint8_t>();
23 14 }
24
25 14 bool ShilinNGaussFilterVerticalSplitMPI::ValidationImpl() {
26 14 int rank = 0;
27 14 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
28
29 14 int validation_result = 1;
30
31
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank == 0) {
32 const InType &input = GetInput();
33
3/6
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 7 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 7 times.
 7 if (input.width <= 0 || input.height <= 0 || input.channels <= 0) {
34 ✗ validation_result = 0;
35 } else {
36 7 size_t expected_size =
37
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 7 times.
 7 static_cast<size_t>(input.width) * static_cast<size_t>(input.height) * static_cast<size_t>(input.channels);
38
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 7 times.
 7 if (input.pixels.size() != expected_size) {
39 ✗ validation_result = 0;
40 }
41 }
42 }
43
44 14 MPI_Bcast(&validation_result, 1, MPI_INT, 0, MPI_COMM_WORLD);
45
46 14 return validation_result != 0;
47 }
48
49 14 bool ShilinNGaussFilterVerticalSplitMPI::PreProcessingImpl() {
50 14 GetOutput() = std::vector<uint8_t>();
51 14 return true;
52 }
53
54 14 bool ShilinNGaussFilterVerticalSplitMPI::RunImpl() {
55 14 int rank = 0;
56 14 int size = 0;
57 14 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
58 14 MPI_Comm_size(MPI_COMM_WORLD, &size);
59
60 14 int width = 0;
61 14 int height = 0;
62 14 int channels = 0;
63
64 14 std::vector<uint8_t> input_pixels;
65
66
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank == 0) {
67 const InType &input = GetInput();
68 7 width = input.width;
69 7 height = input.height;
70 7 channels = input.channels;
71
1/2
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
 7 input_pixels = input.pixels;
72 }
73
74
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 MPI_Bcast(&width, 1, MPI_INT, 0, MPI_COMM_WORLD);
75
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 MPI_Bcast(&height, 1, MPI_INT, 0, MPI_COMM_WORLD);
76
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 MPI_Bcast(&channels, 1, MPI_INT, 0, MPI_COMM_WORLD);
77
78
3/6
✓ Branch 0 taken 14 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 14 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 14 times.
✗ Branch 5 not taken.
 14 if (width <= 0 || height <= 0 || channels <= 0) {
79 return false;
80 }
81
82 14 int local_width = 0;
83 14 int local_start_col = 0;
84
85 14 std::vector<uint8_t> local_input;
86
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 DistributeVerticalStripes(input_pixels, local_input, width, height, channels, rank, size, local_width,
87 local_start_col);
88
89 14 std::vector<uint8_t> local_output(static_cast<size_t>(local_width) * static_cast<size_t>(height) *
90
1/4
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 14 static_cast<size_t>(channels));
91 14 ApplyGaussianKernelMPI(local_input, local_output, local_width, local_start_col, width, height, channels);
92
93 14 std::vector<uint8_t> output_pixels;
94
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank == 0) {
95 output_pixels =
96
1/4
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 14 std::vector<uint8_t>(static_cast<size_t>(width) * static_cast<size_t>(height) * static_cast<size_t>(channels));
97 }
98
99
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 GatherVerticalStripes(local_output, output_pixels, width, height, channels, rank, size, local_width, local_start_col);
100
101 14 const auto output_size = static_cast<size_t>(width) * static_cast<size_t>(height) * static_cast<size_t>(channels);
102
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank != 0) {
103
1/2
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
 7 output_pixels.resize(output_size);
104 }
105
106 // синхронизируем итоговое изображение между всеми процессами, чтобы тесты не падали на worker ranks
107
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 MPI_Bcast(output_pixels.data(), static_cast<int>(output_size), MPI_UNSIGNED_CHAR, 0, MPI_COMM_WORLD);
108
1/2
✓ Branch 1 taken 14 times.
✗ Branch 2 not taken.
 14 GetOutput() = output_pixels;
109
110 return true;
111 }
112
113 14 bool ShilinNGaussFilterVerticalSplitMPI::PostProcessingImpl() {
114 14 int rank = 0;
115 14 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
116
3/4
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
✓ Branch 2 taken 7 times.
✗ Branch 3 not taken.
 14 return (rank == 0) ? !GetOutput().empty() : true;
117 }
118
119 14 void ShilinNGaussFilterVerticalSplitMPI::DistributeVerticalStripes(const std::vector<uint8_t> &source_image,
120 std::vector<uint8_t> &destination_stripe, int width,
121 int height, int channels, int rank, int size,
122 int &local_width, int &local_start_col) {
123 // вертикальное разбиение: каждый процесс получает несколько столбцов
124 14 int base_cols_per_proc = width / size;
125 14 int remainder = width % size;
126
127
2/2
✓ Branch 0 taken 13 times.
✓ Branch 1 taken 1 times.
 14 local_start_col = (rank * base_cols_per_proc) + std::min(rank, remainder);
128
2/2
✓ Branch 0 taken 13 times.
✓ Branch 1 taken 1 times.
 14 local_width = base_cols_per_proc + (rank < remainder ? 1 : 0);
129
130 // для фильтра 3x3 нужны граничные пиксели слева и справа
131
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 int left_padding = (local_start_col > 0) ? 1 : 0;
132
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 int right_padding = (local_start_col + local_width < width) ? 1 : 0;
133 14 int extended_width = local_width + left_padding + right_padding;
134
135 14 size_t local_size = static_cast<size_t>(extended_width) * static_cast<size_t>(height) * static_cast<size_t>(channels);
136 14 destination_stripe = std::vector<uint8_t>(local_size);
137
138
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank == 0) {
139 // процесс 0 отправляет данные остальным процессам
140
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 for (int dest = 1; dest < size; ++dest) {
141 7 SendDataToProcess(source_image, dest, width, height, channels, base_cols_per_proc, remainder);
142 }
143
144 // процесс 0 обрабатывает свои данные
145 7 CopyLocalData(source_image, destination_stripe, local_start_col, local_width, width, height, channels);
146 } else {
147 // остальные процессы получают данные
148 7 MPI_Recv(destination_stripe.data(), static_cast<int>(destination_stripe.size()), MPI_UNSIGNED_CHAR, 0, 0,
149 MPI_COMM_WORLD, MPI_STATUS_IGNORE);
150 }
151 14 }
152
153 7 void ShilinNGaussFilterVerticalSplitMPI::SendDataToProcess(const std::vector<uint8_t> &input, int dest, int width,
154 int height, int channels, int base_cols_per_proc,
155 int remainder) {
156
1/2
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
 7 int dest_start = (dest * base_cols_per_proc) + std::min(dest, remainder);
157
1/2
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
 7 int dest_width = base_cols_per_proc + (dest < remainder ? 1 : 0);
158
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 7 times.
 7 int dest_left_padding = (dest_start > 0) ? 1 : 0;
159
1/2
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
 7 int dest_right_padding = (dest_start + dest_width < width) ? 1 : 0;
160 7 int dest_extended_width = dest_width + dest_left_padding + dest_right_padding;
161
162 7 int send_start_col = dest_start - dest_left_padding;
163 int send_width = dest_extended_width;
164
165 7 std::vector<uint8_t> send_data(static_cast<size_t>(send_width) * static_cast<size_t>(height) *
166 7 static_cast<size_t>(channels));
167
2/2
✓ Branch 0 taken 217 times.
✓ Branch 1 taken 7 times.
 224 for (int row = 0; row < height; ++row) {
168
2/2
✓ Branch 0 taken 7179 times.
✓ Branch 1 taken 217 times.
 7396 for (int col = 0; col < send_width; ++col) {
169 7179 int src_x = send_start_col + col;
170
1/2
✓ Branch 0 taken 7179 times.
✗ Branch 1 not taken.
 7179 if (src_x >= 0 && src_x < width) {
171
2/2
✓ Branch 0 taken 21537 times.
✓ Branch 1 taken 7179 times.
 28716 for (int ch = 0; ch < channels; ++ch) {
172 21537 size_t src_idx = (static_cast<size_t>(row) * static_cast<size_t>(width) * static_cast<size_t>(channels)) +
173 21537 (static_cast<size_t>(src_x) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
174 21537 size_t dst_idx =
175 21537 (static_cast<size_t>(row) * static_cast<size_t>(send_width) * static_cast<size_t>(channels)) +
176 21537 (static_cast<size_t>(col) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
177 21537 send_data[dst_idx] = input[src_idx];
178 }
179 }
180 }
181 }
182
183
1/2
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
 7 MPI_Send(send_data.data(), static_cast<int>(send_data.size()), MPI_UNSIGNED_CHAR, dest, 0, MPI_COMM_WORLD);
184 7 }
185
186 7 void ShilinNGaussFilterVerticalSplitMPI::CopyLocalData(const std::vector<uint8_t> &input,
187 std::vector<uint8_t> &local_data, int local_start_col,
188 int local_width, int width, int height, int channels) {
189
1/2
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
 7 int left_padding = (local_start_col > 0) ? 1 : 0;
190
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 7 times.
 7 int right_padding = (local_start_col + local_width < width) ? 1 : 0;
191 7 int extended_width = local_width + left_padding + right_padding;
192
193
2/2
✓ Branch 0 taken 217 times.
✓ Branch 1 taken 7 times.
 224 for (int row = 0; row < height; ++row) {
194
2/2
✓ Branch 0 taken 7184 times.
✓ Branch 1 taken 217 times.
 7401 for (int col = -left_padding; col < local_width + right_padding; ++col) {
195 7184 int src_x = local_start_col + col;
196
1/2
✓ Branch 0 taken 7184 times.
✗ Branch 1 not taken.
 7184 if (src_x >= 0 && src_x < width) {
197
2/2
✓ Branch 0 taken 21552 times.
✓ Branch 1 taken 7184 times.
 28736 for (int ch = 0; ch < channels; ++ch) {
198 21552 size_t src_idx = (static_cast<size_t>(row) * static_cast<size_t>(width) * static_cast<size_t>(channels)) +
199 21552 (static_cast<size_t>(src_x) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
200 21552 size_t dst_idx =
201 21552 (static_cast<size_t>(row) * static_cast<size_t>(extended_width) * static_cast<size_t>(channels)) +
202 21552 (static_cast<size_t>(col + left_padding) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
203 21552 local_data[dst_idx] = input[src_idx];
204 }
205 }
206 }
207 }
208 7 }
209
210 14 void ShilinNGaussFilterVerticalSplitMPI::ApplyGaussianKernelMPI(const std::vector<uint8_t> &local_input,
211 std::vector<uint8_t> &local_output, int local_width,
212 int local_start_col, int width, int height,
213 int channels) {
214
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 int left_padding = (local_start_col > 0) ? 1 : 0;
215
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 int right_padding = (local_start_col + local_width < width) ? 1 : 0;
216 14 int extended_width = local_width + left_padding + right_padding;
217
218
2/2
✓ Branch 0 taken 434 times.
✓ Branch 1 taken 14 times.
 448 for (int row = 0; row < height; ++row) {
219
2/2
✓ Branch 0 taken 13929 times.
✓ Branch 1 taken 434 times.
 14363 for (int local_col = 0; local_col < local_width; ++local_col) {
220 13929 ProcessPixelWithKernel(local_input, local_output, row, local_col, local_width, left_padding, extended_width,
221 height, channels);
222 }
223 }
224 14 }
225
226 13929 void ShilinNGaussFilterVerticalSplitMPI::ProcessPixelWithKernel(const std::vector<uint8_t> &local_input,
227 std::vector<uint8_t> &local_output, int row,
228 int local_col, int local_width, int left_padding,
229 int extended_width, int height, int channels) {
230 // ядро гаусса 3x3
231 13929 constexpr std::array<std::array<double, 3>, 3> kKernel = {{{{1.0 / 16.0, 2.0 / 16.0, 1.0 / 16.0}},
232 {{2.0 / 16.0, 4.0 / 16.0, 2.0 / 16.0}},
233 {{1.0 / 16.0, 2.0 / 16.0, 1.0 / 16.0}}}};
234
235 13929 int col_in_extended = local_col + left_padding;
236
237
2/2
✓ Branch 0 taken 41787 times.
✓ Branch 1 taken 13929 times.
 55716 for (int ch = 0; ch < channels; ++ch) {
238 41787 double sum = 0.0;
239
240
2/2
✓ Branch 0 taken 125361 times.
✓ Branch 1 taken 41787 times.
 167148 for (int ky = -1; ky <= 1; ++ky) {
241
2/2
✓ Branch 0 taken 376083 times.
✓ Branch 1 taken 125361 times.
 501444 for (int kx = -1; kx <= 1; ++kx) {
242 376083 int px = col_in_extended + kx;
243 376083 int py = row + ky;
244
245 double pixel_val = 0.0;
246
4/4
✓ Branch 0 taken 372177 times.
✓ Branch 1 taken 3906 times.
✓ Branch 2 taken 368355 times.
✓ Branch 3 taken 3822 times.
 376083 if (px >= 0 && px < extended_width && py >= 0 && py < height) {
247 368355 size_t idx = (static_cast<size_t>(py) * static_cast<size_t>(extended_width) * static_cast<size_t>(channels)) +
248 368355 (static_cast<size_t>(px) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
249 368355 pixel_val = static_cast<double>(local_input[idx]);
250 }
251 376083 const int kernel_y_idx = ky + 1;
252 376083 const int kernel_x_idx = kx + 1;
253 376083 const auto kernel_y = static_cast<size_t>(kernel_y_idx);
254 376083 const auto kernel_x = static_cast<size_t>(kernel_x_idx);
255 376083 sum += pixel_val * kKernel.at(kernel_y).at(kernel_x);
256 }
257 }
258
259 41787 size_t out_idx = (static_cast<size_t>(row) * static_cast<size_t>(local_width) * static_cast<size_t>(channels)) +
260 41787 (static_cast<size_t>(local_col) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
261 41787 local_output[out_idx] = static_cast<uint8_t>(std::clamp(sum, 0.0, 255.0));
262 }
263 13929 }
264
265 14 void ShilinNGaussFilterVerticalSplitMPI::GatherVerticalStripes(const std::vector<uint8_t> &local_stripe,
266 std::vector<uint8_t> &final_image, int width, int height,
267 int channels, int rank, int size, int local_width,
268 int local_start_col) {
269 14 int base_cols_per_proc = width / size;
270 14 int remainder = width % size;
271
272
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 if (rank == 0) {
273 int src_start = (0 * base_cols_per_proc) + std::min(0, remainder);
274
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 1 times.
 7 int src_width = base_cols_per_proc + (0 < remainder ? 1 : 0);
275 GatherFromRank0(local_stripe, final_image, width, height, channels, size, local_width, src_start, src_width);
276 7 GatherFromOtherRanks(final_image, width, height, channels, size, base_cols_per_proc, remainder);
277 } else {
278 SendUnpaddedData(local_stripe, local_width, local_start_col, width, height, channels);
279 }
280 14 }
281
282 ✗ void ShilinNGaussFilterVerticalSplitMPI::GatherFromRank0(const std::vector<uint8_t> &local_stripe,
283 std::vector<uint8_t> &final_image, int width, int height,
284 int channels, int /* size */, int local_width, int src_start,
285 int src_width) {
286
2/4
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✓ Branch 2 taken 217 times.
✓ Branch 3 taken 7 times.
 224 for (int row = 0; row < height; ++row) {
287
2/4
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✓ Branch 2 taken 6967 times.
✓ Branch 3 taken 217 times.
 7184 for (int col = 0; col < src_width; ++col) {
288
2/4
✗ Branch 0 not taken.
✗ Branch 1 not taken.
✓ Branch 2 taken 20901 times.
✓ Branch 3 taken 6967 times.
 27868 for (int ch = 0; ch < channels; ++ch) {
289 20901 size_t local_idx =
290 20901 (static_cast<size_t>(row) * static_cast<size_t>(local_width) * static_cast<size_t>(channels)) +
291 20901 (static_cast<size_t>(col) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
292 20901 size_t global_idx = (static_cast<size_t>(row) * static_cast<size_t>(width) * static_cast<size_t>(channels)) +
293 20901 (static_cast<size_t>(src_start + col) * static_cast<size_t>(channels)) +
294 static_cast<size_t>(ch);
295 20901 final_image[global_idx] = local_stripe[local_idx];
296 }
297 }
298 }
299 ✗ }
300
301 7 void ShilinNGaussFilterVerticalSplitMPI::GatherFromOtherRanks(std::vector<uint8_t> &final_image, int width, int height,
302 int channels, int size, int base_cols_per_proc,
303 int remainder) {
304
2/2
✓ Branch 0 taken 7 times.
✓ Branch 1 taken 7 times.
 14 for (int src = 1; src < size; ++src) {
305
1/2
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
 7 int src_start = (src * base_cols_per_proc) + std::min(src, remainder);
306
1/2
✓ Branch 0 taken 7 times.
✗ Branch 1 not taken.
 7 int src_width = base_cols_per_proc + (src < remainder ? 1 : 0);
307
308 7 std::vector<uint8_t> recv_data(static_cast<size_t>(src_width) * static_cast<size_t>(height) *
309
1/2
✓ Branch 2 taken 7 times.
✗ Branch 3 not taken.
 7 static_cast<size_t>(channels));
310
311
1/2
✓ Branch 1 taken 7 times.
✗ Branch 2 not taken.
 7 MPI_Recv(recv_data.data(), static_cast<int>(recv_data.size()), MPI_UNSIGNED_CHAR, src, 0, MPI_COMM_WORLD,
312 MPI_STATUS_IGNORE);
313
314
2/2
✓ Branch 0 taken 217 times.
✓ Branch 1 taken 7 times.
 224 for (int row = 0; row < height; ++row) {
315
2/2
✓ Branch 0 taken 6962 times.
✓ Branch 1 taken 217 times.
 7179 for (int col = 0; col < src_width; ++col) {
316
2/2
✓ Branch 0 taken 20886 times.
✓ Branch 1 taken 6962 times.
 27848 for (int ch = 0; ch < channels; ++ch) {
317 20886 size_t recv_idx =
318 20886 (static_cast<size_t>(row) * static_cast<size_t>(src_width) * static_cast<size_t>(channels)) +
319 20886 (static_cast<size_t>(col) * static_cast<size_t>(channels)) + static_cast<size_t>(ch);
320 20886 size_t global_idx = (static_cast<size_t>(row) * static_cast<size_t>(width) * static_cast<size_t>(channels)) +
321 20886 (static_cast<size_t>(src_start + col) * static_cast<size_t>(channels)) +
322 static_cast<size_t>(ch);
323 20886 final_image[global_idx] = recv_data[recv_idx];
324 }
325 }
326 }
327 }
328 7 }
329
330 ✗ void ShilinNGaussFilterVerticalSplitMPI::SendUnpaddedData(const std::vector<uint8_t> &local_stripe, int local_width,
331 int /*local_start_col*/, int /*width*/, int height,
332 int channels) {
333 // local_stripe уже содержит unpadded данные (local_output после ApplyGaussianKernelMPI)
334 // размер: local_width * height * channels
335 7 size_t data_size = static_cast<size_t>(local_width) * static_cast<size_t>(height) * static_cast<size_t>(channels);
336 7 MPI_Send(local_stripe.data(), static_cast<int>(data_size), MPI_UNSIGNED_CHAR, 0, 0, MPI_COMM_WORLD);
337 7 }
338
339 } // namespace shilin_n_gauss_filter_vertical_split
340