GCC Code Coverage Report

Directory: ./
File: tasks/marin_l_linear_filter_vertical/mpi/src/ops_mpi.cpp
Date: 2026-01-27 01:59:34
Exec Total Coverage
Lines: 97 97 100.0%
Functions: 9 9 100.0%
Branches: 76 100 76.0%
Line Branch Exec Source
1 #include "marin_l_linear_filter_vertical/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 "marin_l_linear_filter_vertical/common/include/common.hpp"
12
13 namespace marin_l_linear_filter_vertical {
14
15 namespace {
16 constexpr std::array<std::array<int, 3>, 3> kGaussKernel = {{{1, 2, 1}, {2, 4, 2}, {1, 2, 1}}};
17 constexpr int kKernelSum = 16;
18
19 14014 uint8_t ApplyKernelLocal(const std::vector<uint8_t> &local_input, int ext_col_count, int height, int row, int lx,
20 int left_pad) {
21 int sum = 0;
22
2/2
✓ Branch 0 taken 42042 times.
✓ Branch 1 taken 14014 times.
 56056 for (int ky = -1; ky <= 1; ++ky) {
23
2/2
✓ Branch 0 taken 126126 times.
✓ Branch 1 taken 42042 times.
 168168 for (int kx = -1; kx <= 1; ++kx) {
24 126126 int ny = row + ky;
25 126126 int local_nx = lx + left_pad + kx;
26 uint8_t pixel_value = 0;
27
4/4
✓ Branch 0 taken 124614 times.
✓ Branch 1 taken 1512 times.
✓ Branch 2 taken 123150 times.
✓ Branch 3 taken 1464 times.
 126126 if (ny >= 0 && ny < height && local_nx >= 0 && local_nx < ext_col_count) {
28 123150 pixel_value = local_input[(ny * ext_col_count) + local_nx];
29 }
30 126126 sum += pixel_value * kGaussKernel.at(ky + 1).at(kx + 1);
31 }
32 }
33
1/2
✓ Branch 0 taken 14014 times.
✗ Branch 1 not taken.
 14014 return static_cast<uint8_t>(std::clamp(sum / kKernelSum, 0, 255));
34 }
35
36 void ComputeColDistribution(int width, int size, std::vector<int> &col_counts, std::vector<int> &col_starts) {
37 24 int cols_per_proc = width / size;
38 24 int extra_cols = width % size;
39 int current_col = 0;
40
2/2
✓ Branch 0 taken 48 times.
✓ Branch 1 taken 24 times.
 72 for (int proc = 0; proc < size; ++proc) {
41
2/2
✓ Branch 0 taken 36 times.
✓ Branch 1 taken 12 times.
 48 col_starts[proc] = current_col;
42
2/2
✓ Branch 0 taken 36 times.
✓ Branch 1 taken 12 times.
 84 col_counts[proc] = cols_per_proc + ((proc < extra_cols) ? 1 : 0);
43 48 current_col += col_counts[proc];
44 }
45 }
46
47 void ExtractLocalData(const std::vector<uint8_t> &input_pixels, std::vector<uint8_t> &local_input, int width,
48 int height, int ext_start, int ext_col_count) {
49
4/4
✓ Branch 0 taken 252 times.
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 246 times.
✓ Branch 3 taken 10 times.
 520 for (int row = 0; row < height; ++row) {
50
4/4
✓ Branch 0 taken 7278 times.
✓ Branch 1 taken 252 times.
✓ Branch 2 taken 7228 times.
✓ Branch 3 taken 246 times.
 15004 for (int col = 0; col < ext_col_count; ++col) {
51 14506 local_input[(row * ext_col_count) + col] = input_pixels[(row * width) + (ext_start + col)];
52 }
53 }
54 }
55
56 void CopyLocalToOutput(std::vector<uint8_t> &output_pixels, const std::vector<uint8_t> &local_output, int width,
57 int height, int col_start, int col_count) {
58
4/4
✓ Branch 0 taken 252 times.
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 246 times.
✓ Branch 3 taken 10 times.
 520 for (int row = 0; row < height; ++row) {
59
4/4
✓ Branch 0 taken 7032 times.
✓ Branch 1 taken 252 times.
✓ Branch 2 taken 6982 times.
✓ Branch 3 taken 246 times.
 14512 for (int col = 0; col < col_count; ++col) {
60 14014 output_pixels[(row * width) + (col_start + col)] = local_output[(row * col_count) + col];
61 }
62 }
63 }
64
65 int ComputePadding(int col_start, int col_count, int width, bool is_left) {
66 if (is_left) {
67
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 10 times.
 22 return (col_start > 0) ? 1 : 0;
68 }
69
1/2
✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
 10 return ((col_start + col_count) < width) ? 1 : 0;
70 }
71
72 12 void SendDataToWorkers(const std::vector<uint8_t> &input_pixels, const std::vector<int> &col_counts,
73 const std::vector<int> &col_starts, int width, int height, int size) {
74
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
 24 for (int proc = 1; proc < size; ++proc) {
75
2/2
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 10 times.
 12 if (col_counts[proc] == 0) {
76 2 continue;
77 }
78
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
 10 int p_col_start = col_starts[proc];
79 int p_col_count = col_counts[proc];
80 int p_left_pad = ComputePadding(p_col_start, p_col_count, width, true);
81 int p_right_pad = ComputePadding(p_col_start, p_col_count, width, false);
82 10 int p_ext_col_count = p_col_count + p_left_pad + p_right_pad;
83 10 int p_ext_start = p_col_start - p_left_pad;
84
85 10 std::vector<uint8_t> send_buf(static_cast<size_t>(p_ext_col_count) * static_cast<size_t>(height));
86 ExtractLocalData(input_pixels, send_buf, width, height, p_ext_start, p_ext_col_count);
87
1/2
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
 10 MPI_Send(send_buf.data(), static_cast<int>(send_buf.size()), MPI_UNSIGNED_CHAR, proc, 0, MPI_COMM_WORLD);
88 }
89 12 }
90
91 12 void ReceiveResultsFromWorkers(std::vector<uint8_t> &output_pixels, const std::vector<int> &col_counts,
92 const std::vector<int> &col_starts, int width, int height, int size) {
93
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
 24 for (int proc = 1; proc < size; ++proc) {
94
2/2
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 10 times.
 12 if (col_counts[proc] == 0) {
95 2 continue;
96 }
97 10 int p_col_start = col_starts[proc];
98 int p_col_count = col_counts[proc];
99
1/2
✓ Branch 2 taken 10 times.
✗ Branch 3 not taken.
 10 std::vector<uint8_t> recv_buf(static_cast<size_t>(p_col_count) * static_cast<size_t>(height));
100
1/2
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
 10 MPI_Recv(recv_buf.data(), static_cast<int>(recv_buf.size()), MPI_UNSIGNED_CHAR, proc, 1, MPI_COMM_WORLD,
101 MPI_STATUS_IGNORE);
102 CopyLocalToOutput(output_pixels, recv_buf, width, height, p_col_start, p_col_count);
103 }
104 12 }
105
106 22 void ApplyFilterToLocalData(const std::vector<uint8_t> &local_input, std::vector<uint8_t> &local_output,
107 int ext_col_count, int local_col_count, int height, int left_pad) {
108
2/2
✓ Branch 0 taken 498 times.
✓ Branch 1 taken 22 times.
 520 for (int row = 0; row < height; ++row) {
109
2/2
✓ Branch 0 taken 14014 times.
✓ Branch 1 taken 498 times.
 14512 for (int lx = 0; lx < local_col_count; ++lx) {
110 14014 local_output[(row * local_col_count) + lx] =
111 14014 ApplyKernelLocal(local_input, ext_col_count, height, row, lx, left_pad);
112 }
113 }
114 22 }
115 } // namespace
116
117
1/2
✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
 24 MarinLLinearFilterVerticalMPI::MarinLLinearFilterVerticalMPI(const InType &in) {
118 SetTypeOfTask(GetStaticTypeOfTask());
119 GetInput() = in;
120 GetOutput() = {};
121 24 }
122
123 24 bool MarinLLinearFilterVerticalMPI::ValidationImpl() {
124 24 int rank = 0;
125 24 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
126
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
 24 if (rank == 0) {
127 const auto &input = GetInput();
128
2/4
✓ Branch 0 taken 12 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 12 times.
✗ Branch 3 not taken.
 12 if (input.width <= 0 || input.height <= 0) {
129 return false;
130 }
131 12 auto expected_size = static_cast<size_t>(input.width) * static_cast<size_t>(input.height);
132 12 return input.pixels.size() == expected_size;
133 }
134 return true;
135 }
136
137 24 bool MarinLLinearFilterVerticalMPI::PreProcessingImpl() {
138 24 int rank = 0;
139 24 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
140
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 12 times.
 24 if (rank == 0) {
141 const auto &input = GetInput();
142 12 width_ = input.width;
143 12 height_ = input.height;
144 12 input_pixels_ = input.pixels;
145 }
146 24 return true;
147 }
148
149 24 bool MarinLLinearFilterVerticalMPI::RunImpl() {
150 24 int rank = 0;
151 24 int size = 0;
152 24 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
153 24 MPI_Comm_size(MPI_COMM_WORLD, &size);
154
155 24 MPI_Bcast(&width_, 1, MPI_INT, 0, MPI_COMM_WORLD);
156 24 MPI_Bcast(&height_, 1, MPI_INT, 0, MPI_COMM_WORLD);
157
158 24 output_pixels_.resize(static_cast<size_t>(width_) * static_cast<size_t>(height_));
159
160 24 std::vector<int> col_counts(size);
161
1/4
✓ Branch 1 taken 24 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 24 std::vector<int> col_starts(size);
162 24 ComputeColDistribution(width_, size, col_counts, col_starts);
163
164
2/2
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 22 times.
 24 int local_col_start = col_starts[rank];
165 24 int local_col_count = col_counts[rank];
166
167
2/2
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 22 times.
 24 if (local_col_count == 0) {
168
1/2
✓ Branch 1 taken 2 times.
✗ Branch 2 not taken.
 2 MPI_Bcast(output_pixels_.data(), static_cast<int>(output_pixels_.size()), MPI_UNSIGNED_CHAR, 0, MPI_COMM_WORLD);
169 return true;
170 }
171
172 int left_pad = ComputePadding(local_col_start, local_col_count, width_, true);
173
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 10 times.
 22 int right_pad = ComputePadding(local_col_start, local_col_count, width_, false);
174 22 int extended_col_count = local_col_count + left_pad + right_pad;
175
176
1/4
✓ Branch 1 taken 22 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 22 std::vector<uint8_t> local_input(static_cast<size_t>(extended_col_count) * static_cast<size_t>(height_));
177
178
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 10 times.
 22 if (rank == 0) {
179 12 int ext_start = local_col_start - left_pad;
180 12 ExtractLocalData(input_pixels_, local_input, width_, height_, ext_start, extended_col_count);
181
1/2
✓ Branch 1 taken 12 times.
✗ Branch 2 not taken.
 12 SendDataToWorkers(input_pixels_, col_counts, col_starts, width_, height_, size);
182 } else {
183
1/2
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
 10 MPI_Recv(local_input.data(), static_cast<int>(local_input.size()), MPI_UNSIGNED_CHAR, 0, 0, MPI_COMM_WORLD,
184 MPI_STATUS_IGNORE);
185 }
186
187
1/4
✓ Branch 1 taken 22 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
 22 std::vector<uint8_t> local_output(static_cast<size_t>(local_col_count) * static_cast<size_t>(height_));
188 22 ApplyFilterToLocalData(local_input, local_output, extended_col_count, local_col_count, height_, left_pad);
189
190
2/2
✓ Branch 0 taken 12 times.
✓ Branch 1 taken 10 times.
 22 if (rank == 0) {
191 12 CopyLocalToOutput(output_pixels_, local_output, width_, height_, local_col_start, local_col_count);
192
1/2
✓ Branch 1 taken 12 times.
✗ Branch 2 not taken.
 12 ReceiveResultsFromWorkers(output_pixels_, col_counts, col_starts, width_, height_, size);
193 } else {
194
1/2
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
 10 MPI_Send(local_output.data(), static_cast<int>(local_output.size()), MPI_UNSIGNED_CHAR, 0, 1, MPI_COMM_WORLD);
195 }
196
197
1/2
✓ Branch 1 taken 22 times.
✗ Branch 2 not taken.
 22 MPI_Bcast(output_pixels_.data(), static_cast<int>(output_pixels_.size()), MPI_UNSIGNED_CHAR, 0, MPI_COMM_WORLD);
198
199 return true;
200 }
201
202 24 bool MarinLLinearFilterVerticalMPI::PostProcessingImpl() {
203 24 GetOutput() = output_pixels_;
204 24 return true;
205 }
206
207 } // namespace marin_l_linear_filter_vertical
208