| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "cheremkhin_a_matr_mult_cannon_alg/seq/include/ops_seq.hpp" | ||
| 2 | |||
| 3 | #include <cmath> | ||
| 4 | #include <cstddef> | ||
| 5 | #include <utility> | ||
| 6 | #include <vector> | ||
| 7 | |||
| 8 | #include "cheremkhin_a_matr_mult_cannon_alg/common/include/common.hpp" | ||
| 9 | |||
| 10 | namespace cheremkhin_a_matr_mult_cannon_alg { | ||
| 11 | |||
| 12 | namespace { | ||
| 13 | |||
| 14 | inline std::size_t Idx(std::size_t n, std::size_t r, std::size_t c) { | ||
| 15 | 6464 | return (r * n) + c; | |
| 16 | } | ||
| 17 | |||
| 18 | std::size_t ChooseQ(std::size_t n) { | ||
| 19 | 56 | if (n <= 1) { | |
| 20 | return 1; | ||
| 21 | } | ||
| 22 | 48 | const auto root = static_cast<std::size_t>(std::sqrt(static_cast<double>(n))); | |
| 23 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 48 times.
|
48 | return (root == 0) ? 1 : root; |
| 24 | } | ||
| 25 | |||
| 26 | std::size_t CeilDiv(std::size_t a, std::size_t b) { | ||
| 27 | 56 | return (a + b - 1) / b; | |
| 28 | } | ||
| 29 | |||
| 30 | 584 | void MulAddBlock(const std::vector<double> &a, const std::vector<double> &b, std::vector<double> &c, std::size_t n, | |
| 31 | std::size_t bs, std::size_t bi, std::size_t bk, std::size_t bj) { | ||
| 32 | 584 | const std::size_t i0 = bi * bs; | |
| 33 | 584 | const std::size_t k0 = bk * bs; | |
| 34 | 584 | const std::size_t j0 = bj * bs; | |
| 35 | |||
| 36 |
2/2✓ Branch 0 taken 2376 times.
✓ Branch 1 taken 584 times.
|
2960 | for (std::size_t ii = 0; ii < bs; ++ii) { |
| 37 | 2376 | const std::size_t i = i0 + ii; | |
| 38 | 2376 | const std::size_t a_row = i * n; | |
| 39 | const std::size_t c_row = i * n; | ||
| 40 |
2/2✓ Branch 0 taken 10248 times.
✓ Branch 1 taken 2376 times.
|
12624 | for (std::size_t kk = 0; kk < bs; ++kk) { |
| 41 | 10248 | const std::size_t k = k0 + kk; | |
| 42 | 10248 | const double aik = a[a_row + k]; | |
| 43 | 10248 | const std::size_t b_row = k * n; | |
| 44 |
2/2✓ Branch 0 taken 45720 times.
✓ Branch 1 taken 10248 times.
|
55968 | for (std::size_t jj = 0; jj < bs; ++jj) { |
| 45 | 45720 | const std::size_t j = j0 + jj; | |
| 46 | 45720 | c[c_row + j] += aik * b[b_row + j]; | |
| 47 | } | ||
| 48 | } | ||
| 49 | } | ||
| 50 | 584 | } | |
| 51 | |||
| 52 | } // namespace | ||
| 53 | |||
| 54 |
1/2✓ Branch 1 taken 56 times.
✗ Branch 2 not taken.
|
56 | CheremkhinAMatrMultCannonAlgSEQ::CheremkhinAMatrMultCannonAlgSEQ(const InType &in) { |
| 55 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 56 | GetInput() = in; | ||
| 57 | GetOutput() = {}; | ||
| 58 | 56 | } | |
| 59 | |||
| 60 | 56 | bool CheremkhinAMatrMultCannonAlgSEQ::ValidationImpl() { | |
| 61 | 56 | const std::size_t n = std::get<0>(GetInput()); | |
| 62 | const auto &a = std::get<1>(GetInput()); | ||
| 63 | const auto &b = std::get<2>(GetInput()); | ||
| 64 |
3/6✓ Branch 0 taken 56 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 56 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 56 times.
|
56 | return n > 0 && a.size() == n * n && b.size() == n * n; |
| 65 | } | ||
| 66 | |||
| 67 | 56 | bool CheremkhinAMatrMultCannonAlgSEQ::PreProcessingImpl() { | |
| 68 | GetOutput() = {}; | ||
| 69 | 56 | return true; | |
| 70 | } | ||
| 71 | |||
| 72 | 56 | bool CheremkhinAMatrMultCannonAlgSEQ::RunImpl() { | |
| 73 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 48 times.
|
56 | const std::size_t n = std::get<0>(GetInput()); |
| 74 | const auto &a_in = std::get<1>(GetInput()); | ||
| 75 | const auto &b_in = std::get<2>(GetInput()); | ||
| 76 | |||
| 77 | const std::size_t q = ChooseQ(n); | ||
| 78 | const std::size_t bs = CeilDiv(n, q); | ||
| 79 | 56 | const std::size_t np = q * bs; | |
| 80 | |||
| 81 | 56 | std::vector<double> a(np * np, 0.0); | |
| 82 |
1/4✓ Branch 1 taken 56 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
56 | std::vector<double> b(np * np, 0.0); |
| 83 |
1/4✓ Branch 1 taken 56 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
56 | std::vector<double> c(np * np, 0.0); |
| 84 | |||
| 85 |
2/2✓ Branch 0 taken 336 times.
✓ Branch 1 taken 56 times.
|
392 | for (std::size_t i = 0; i < n; ++i) { |
| 86 |
2/2✓ Branch 0 taken 3232 times.
✓ Branch 1 taken 336 times.
|
3568 | for (std::size_t j = 0; j < n; ++j) { |
| 87 | 3232 | a[Idx(np, i, j)] = a_in[Idx(n, i, j)]; | |
| 88 | 3232 | b[Idx(np, i, j)] = b_in[Idx(n, i, j)]; | |
| 89 | } | ||
| 90 | } | ||
| 91 | |||
| 92 |
2/2✓ Branch 0 taken 104 times.
✓ Branch 1 taken 56 times.
|
160 | for (std::size_t bi = 0; bi < q; ++bi) { |
| 93 |
2/2✓ Branch 0 taken 232 times.
✓ Branch 1 taken 104 times.
|
336 | for (std::size_t bj = 0; bj < q; ++bj) { |
| 94 |
2/2✓ Branch 0 taken 584 times.
✓ Branch 1 taken 232 times.
|
816 | for (std::size_t step = 0; step < q; ++step) { |
| 95 | 584 | const std::size_t bk = (bj + bi + step) % q; | |
| 96 | 584 | MulAddBlock(a, b, c, np, bs, bi, bk, bj); | |
| 97 | } | ||
| 98 | } | ||
| 99 | } | ||
| 100 | |||
| 101 |
1/4✓ Branch 1 taken 56 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
|
56 | std::vector<double> out(n * n, 0.0); |
| 102 |
2/2✓ Branch 0 taken 336 times.
✓ Branch 1 taken 56 times.
|
392 | for (std::size_t i = 0; i < n; ++i) { |
| 103 |
2/2✓ Branch 0 taken 3232 times.
✓ Branch 1 taken 336 times.
|
3568 | for (std::size_t j = 0; j < n; ++j) { |
| 104 | 3232 | out[Idx(n, i, j)] = c[Idx(np, i, j)]; | |
| 105 | } | ||
| 106 | } | ||
| 107 | |||
| 108 | GetOutput() = std::move(out); | ||
| 109 | 56 | return true; | |
| 110 | } | ||
| 111 | |||
| 112 | 56 | bool CheremkhinAMatrMultCannonAlgSEQ::PostProcessingImpl() { | |
| 113 | 56 | return true; | |
| 114 | } | ||
| 115 | |||
| 116 | } // namespace cheremkhin_a_matr_mult_cannon_alg | ||
| 117 |