| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "muhammadkhon_i_stressen_alg/seq/include/ops_seq.hpp" | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <cstddef> | ||
| 5 | #include <stack> | ||
| 6 | #include <utility> | ||
| 7 | #include <vector> | ||
| 8 | |||
| 9 | #include "muhammadkhon_i_stressen_alg/common/include/common.hpp" | ||
| 10 | |||
| 11 | namespace muhammadkhon_i_stressen_alg { | ||
| 12 | |||
| 13 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | MuhammadkhonIStressenAlgSEQ::MuhammadkhonIStressenAlgSEQ(const InType &in) { |
| 14 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 15 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | GetInput() = in; |
| 16 | GetOutput() = {}; | ||
| 17 | 48 | } | |
| 18 | |||
| 19 | 48 | bool MuhammadkhonIStressenAlgSEQ::ValidationImpl() { | |
| 20 | const auto &in = GetInput(); | ||
| 21 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 48 times.
|
48 | return in.a_rows > 0 && in.a_cols_b_rows > 0 && in.b_cols > 0 && |
| 22 |
2/4✓ Branch 0 taken 48 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 48 times.
|
96 | in.a.size() == static_cast<size_t>(in.a_rows * in.a_cols_b_rows) && |
| 23 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
|
48 | in.b.size() == static_cast<size_t>(in.a_cols_b_rows * in.b_cols); |
| 24 | } | ||
| 25 | |||
| 26 | 48 | bool MuhammadkhonIStressenAlgSEQ::PreProcessingImpl() { | |
| 27 | GetOutput() = {}; | ||
| 28 | const auto &in = GetInput(); | ||
| 29 | 48 | a_rows_ = in.a_rows; | |
| 30 | 48 | a_cols_b_rows_ = in.a_cols_b_rows; | |
| 31 | 48 | b_cols_ = in.b_cols; | |
| 32 | |||
| 33 | 48 | size_t max_dim = std::max({a_rows_, a_cols_b_rows_, b_cols_}); | |
| 34 | 48 | padded_n_ = 1; | |
| 35 |
2/2✓ Branch 0 taken 192 times.
✓ Branch 1 taken 48 times.
|
240 | while (padded_n_ < max_dim) { |
| 36 | 192 | padded_n_ *= 2; | |
| 37 | } | ||
| 38 | |||
| 39 | 48 | padded_a_.assign(padded_n_ * padded_n_, 0.0); | |
| 40 | 48 | padded_b_.assign(padded_n_ * padded_n_, 0.0); | |
| 41 | |||
| 42 |
2/2✓ Branch 0 taken 1320 times.
✓ Branch 1 taken 48 times.
|
1368 | for (size_t i = 0; i < a_rows_; ++i) { |
| 43 |
2/2✓ Branch 0 taken 67368 times.
✓ Branch 1 taken 1320 times.
|
68688 | for (size_t j = 0; j < a_cols_b_rows_; ++j) { |
| 44 | 67368 | padded_a_[(i * padded_n_) + j] = in.a[(i * a_cols_b_rows_) + j]; | |
| 45 | } | ||
| 46 | } | ||
| 47 | |||
| 48 |
2/2✓ Branch 0 taken 1240 times.
✓ Branch 1 taken 48 times.
|
1288 | for (size_t i = 0; i < a_cols_b_rows_; ++i) { |
| 49 |
2/2✓ Branch 0 taken 67368 times.
✓ Branch 1 taken 1240 times.
|
68608 | for (size_t j = 0; j < b_cols_; ++j) { |
| 50 | 67368 | padded_b_[(i * padded_n_) + j] = in.b[(i * b_cols_) + j]; | |
| 51 | } | ||
| 52 | } | ||
| 53 | 48 | return true; | |
| 54 | } | ||
| 55 | |||
| 56 | 48 | bool MuhammadkhonIStressenAlgSEQ::RunImpl() { | |
| 57 | 96 | result_c_ = StrassenMultiply(padded_a_, padded_b_, padded_n_); | |
| 58 | |||
| 59 | auto &out = GetOutput(); | ||
| 60 | 48 | out.assign(a_rows_ * b_cols_, 0.0); | |
| 61 | |||
| 62 |
2/2✓ Branch 0 taken 1320 times.
✓ Branch 1 taken 48 times.
|
1368 | for (size_t i = 0; i < a_rows_; ++i) { |
| 63 |
2/2✓ Branch 0 taken 68568 times.
✓ Branch 1 taken 1320 times.
|
69888 | for (size_t j = 0; j < b_cols_; ++j) { |
| 64 | 68568 | out[(i * b_cols_) + j] = result_c_[(i * padded_n_) + j]; | |
| 65 | } | ||
| 66 | } | ||
| 67 | 48 | return true; | |
| 68 | } | ||
| 69 | |||
| 70 | 48 | bool MuhammadkhonIStressenAlgSEQ::PostProcessingImpl() { | |
| 71 | 48 | return true; | |
| 72 | } | ||
| 73 | |||
| 74 | 96 | std::vector<double> MuhammadkhonIStressenAlgSEQ::Add(const std::vector<double> &mat_a, | |
| 75 | const std::vector<double> &mat_b) { | ||
| 76 | 96 | std::vector<double> res(mat_a.size()); | |
| 77 |
2/2✓ Branch 0 taken 98304 times.
✓ Branch 1 taken 96 times.
|
98400 | for (size_t i = 0; i < mat_a.size(); ++i) { |
| 78 | 98304 | res[i] = mat_a[i] + mat_b[i]; | |
| 79 | } | ||
| 80 | 96 | return res; | |
| 81 | } | ||
| 82 | |||
| 83 | 64 | std::vector<double> MuhammadkhonIStressenAlgSEQ::Subtract(const std::vector<double> &mat_a, | |
| 84 | const std::vector<double> &mat_b) { | ||
| 85 | 64 | std::vector<double> res(mat_a.size()); | |
| 86 |
2/2✓ Branch 0 taken 65536 times.
✓ Branch 1 taken 64 times.
|
65600 | for (size_t i = 0; i < mat_a.size(); ++i) { |
| 87 | 65536 | res[i] = mat_a[i] - mat_b[i]; | |
| 88 | } | ||
| 89 | 64 | return res; | |
| 90 | } | ||
| 91 | |||
| 92 | 144 | std::vector<double> MuhammadkhonIStressenAlgSEQ::BaseMultiply(const std::vector<double> &mat_a, | |
| 93 | const std::vector<double> &mat_b, size_t n) { | ||
| 94 | 144 | std::vector<double> res(n * n, 0.0); | |
| 95 |
2/2✓ Branch 0 taken 3904 times.
✓ Branch 1 taken 144 times.
|
4048 | for (size_t i = 0; i < n; ++i) { |
| 96 |
2/2✓ Branch 0 taken 119040 times.
✓ Branch 1 taken 3904 times.
|
122944 | for (size_t k = 0; k < n; ++k) { |
| 97 |
2/2✓ Branch 0 taken 3736576 times.
✓ Branch 1 taken 119040 times.
|
3855616 | for (size_t j = 0; j < n; ++j) { |
| 98 | 3736576 | res[(i * n) + j] += mat_a[(i * n) + k] * mat_b[(k * n) + j]; | |
| 99 | } | ||
| 100 | } | ||
| 101 | } | ||
| 102 | 144 | return res; | |
| 103 | } | ||
| 104 | |||
| 105 | 16 | void MuhammadkhonIStressenAlgSEQ::PushStrassenSubtasks(std::stack<StrassenFrame> &frames, | |
| 106 | const std::vector<double> &mat_a, | ||
| 107 | const std::vector<double> &mat_b, size_t n) { | ||
| 108 | 16 | size_t h = n / 2; | |
| 109 | 16 | size_t sz = h * h; | |
| 110 | 16 | std::vector<double> a11(sz); | |
| 111 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> a12(sz); |
| 112 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> a21(sz); |
| 113 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> a22(sz); |
| 114 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> b11(sz); |
| 115 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> b12(sz); |
| 116 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> b21(sz); |
| 117 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> b22(sz); |
| 118 | |||
| 119 |
2/2✓ Branch 0 taken 512 times.
✓ Branch 1 taken 16 times.
|
528 | for (size_t i = 0; i < h; ++i) { |
| 120 |
2/2✓ Branch 0 taken 16384 times.
✓ Branch 1 taken 512 times.
|
16896 | for (size_t j = 0; j < h; ++j) { |
| 121 | 16384 | size_t idx_src = (i * n) + j; | |
| 122 | 16384 | size_t idx_dst = (i * h) + j; | |
| 123 | 16384 | a11[idx_dst] = mat_a[idx_src]; | |
| 124 | 16384 | a12[idx_dst] = mat_a[idx_src + h]; | |
| 125 | 16384 | a21[idx_dst] = mat_a[idx_src + (h * n)]; | |
| 126 | 16384 | a22[idx_dst] = mat_a[idx_src + (h * n) + h]; | |
| 127 | |||
| 128 | 16384 | b11[idx_dst] = mat_b[idx_src]; | |
| 129 | 16384 | b12[idx_dst] = mat_b[idx_src + h]; | |
| 130 | 16384 | b21[idx_dst] = mat_b[idx_src + (h * n)]; | |
| 131 | 16384 | b22[idx_dst] = mat_b[idx_src + (h * n) + h]; | |
| 132 | } | ||
| 133 | } | ||
| 134 | |||
| 135 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
32 | frames.push({{}, {}, n, 1}); |
| 136 | |||
| 137 | 16 | frames.push({Subtract(a12, a22), Add(b21, b22), h, 0}); | |
| 138 | 16 | frames.push({Subtract(a21, a11), Add(b11, b12), h, 0}); | |
| 139 | 16 | frames.push({Add(a11, a12), b22, h, 0}); | |
| 140 | 16 | frames.push({a22, Subtract(b21, b11), h, 0}); | |
| 141 | 16 | frames.push({a11, Subtract(b12, b22), h, 0}); | |
| 142 | 16 | frames.push({Add(a21, a22), b11, h, 0}); | |
| 143 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
|
16 | frames.push({Add(a11, a22), Add(b11, b22), h, 0}); |
| 144 |
21/42✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✓ Branch 4 taken 16 times.
✗ Branch 5 not taken.
✓ Branch 7 taken 16 times.
✗ Branch 8 not taken.
✓ Branch 10 taken 16 times.
✗ Branch 11 not taken.
✓ Branch 13 taken 16 times.
✗ Branch 14 not taken.
✓ Branch 16 taken 16 times.
✗ Branch 17 not taken.
✓ Branch 19 taken 16 times.
✗ Branch 20 not taken.
✓ Branch 22 taken 16 times.
✗ Branch 23 not taken.
✓ Branch 25 taken 16 times.
✗ Branch 26 not taken.
✓ Branch 28 taken 16 times.
✗ Branch 29 not taken.
✓ Branch 31 taken 16 times.
✗ Branch 32 not taken.
✓ Branch 34 taken 16 times.
✗ Branch 35 not taken.
✓ Branch 37 taken 16 times.
✗ Branch 38 not taken.
✓ Branch 40 taken 16 times.
✗ Branch 41 not taken.
✓ Branch 43 taken 16 times.
✗ Branch 44 not taken.
✓ Branch 46 taken 16 times.
✗ Branch 47 not taken.
✓ Branch 49 taken 16 times.
✗ Branch 50 not taken.
✓ Branch 52 taken 16 times.
✗ Branch 53 not taken.
✓ Branch 55 taken 16 times.
✗ Branch 56 not taken.
✓ Branch 58 taken 16 times.
✗ Branch 59 not taken.
✓ Branch 61 taken 16 times.
✗ Branch 62 not taken.
|
128 | } |
| 145 | |||
| 146 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 16 times.
|
16 | std::vector<double> MuhammadkhonIStressenAlgSEQ::CombineStrassenResults(std::stack<std::vector<double>> &results, |
| 147 | size_t n) { | ||
| 148 | auto p7 = std::move(results.top()); | ||
| 149 | results.pop(); | ||
| 150 | auto p6 = std::move(results.top()); | ||
| 151 | results.pop(); | ||
| 152 | auto p5 = std::move(results.top()); | ||
| 153 | results.pop(); | ||
| 154 | auto p4 = std::move(results.top()); | ||
| 155 | results.pop(); | ||
| 156 | auto p3 = std::move(results.top()); | ||
| 157 | results.pop(); | ||
| 158 | auto p2 = std::move(results.top()); | ||
| 159 | results.pop(); | ||
| 160 | auto p1 = std::move(results.top()); | ||
| 161 | results.pop(); | ||
| 162 | |||
| 163 | 16 | size_t h = n / 2; | |
| 164 |
1/4✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
16 | std::vector<double> res(n * n); |
| 165 |
2/2✓ Branch 0 taken 512 times.
✓ Branch 1 taken 16 times.
|
528 | for (size_t i = 0; i < h; ++i) { |
| 166 |
2/2✓ Branch 0 taken 16384 times.
✓ Branch 1 taken 512 times.
|
16896 | for (size_t j = 0; j < h; ++j) { |
| 167 | 16384 | size_t idx = (i * h) + j; | |
| 168 | 16384 | res[(i * n) + j] = p1[idx] + p4[idx] - p5[idx] + p7[idx]; | |
| 169 | 16384 | res[(i * n) + j + h] = p3[idx] + p5[idx]; | |
| 170 | 16384 | res[((i + h) * n) + j] = p2[idx] + p4[idx]; | |
| 171 | 16384 | res[((i + h) * n) + j + h] = p1[idx] - p2[idx] + p3[idx] + p6[idx]; | |
| 172 | } | ||
| 173 | } | ||
| 174 | 16 | return res; | |
| 175 | } | ||
| 176 | |||
| 177 | 48 | std::vector<double> MuhammadkhonIStressenAlgSEQ::StrassenMultiply(const std::vector<double> &mat_a, | |
| 178 | const std::vector<double> &mat_b, size_t n) { | ||
| 179 | std::stack<StrassenFrame> frames; | ||
| 180 | std::stack<std::vector<double>> results; | ||
| 181 | |||
| 182 | 48 | frames.push({mat_a, mat_b, n, 0}); | |
| 183 | |||
| 184 |
2/2✓ Branch 0 taken 176 times.
✓ Branch 1 taken 48 times.
|
224 | while (!frames.empty()) { |
| 185 | StrassenFrame current = std::move(frames.top()); | ||
| 186 | frames.pop(); | ||
| 187 | |||
| 188 |
2/2✓ Branch 0 taken 160 times.
✓ Branch 1 taken 16 times.
|
176 | if (current.stage == 0) { |
| 189 |
2/2✓ Branch 0 taken 144 times.
✓ Branch 1 taken 16 times.
|
160 | if (current.n <= 32) { |
| 190 |
1/2✓ Branch 1 taken 144 times.
✗ Branch 2 not taken.
|
288 | results.push(BaseMultiply(current.mat_a, current.mat_b, current.n)); |
| 191 | } else { | ||
| 192 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
16 | PushStrassenSubtasks(frames, current.mat_a, current.mat_b, current.n); |
| 193 | } | ||
| 194 | } else { | ||
| 195 |
1/2✓ Branch 1 taken 16 times.
✗ Branch 2 not taken.
|
32 | results.push(CombineStrassenResults(results, current.n)); |
| 196 | } | ||
| 197 | 176 | } | |
| 198 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
96 | return results.top(); |
| 199 |
3/6✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
✓ Branch 4 taken 48 times.
✗ Branch 5 not taken.
✓ Branch 7 taken 48 times.
✗ Branch 8 not taken.
|
48 | } |
| 200 | } // namespace muhammadkhon_i_stressen_alg | ||
| 201 |