| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #include "batkov_f_contrast_enh_lin_hist_stretch/stl/include/ops_stl.hpp" | ||
| 2 | |||
| 3 | #include <algorithm> | ||
| 4 | #include <array> | ||
| 5 | #include <barrier> | ||
| 6 | #include <cstddef> | ||
| 7 | #include <cstdint> | ||
| 8 | #include <iterator> | ||
| 9 | #include <limits> | ||
| 10 | #include <thread> | ||
| 11 | #include <vector> | ||
| 12 | |||
| 13 | #include "batkov_f_contrast_enh_lin_hist_stretch/common/include/common.hpp" | ||
| 14 | #include "util/include/util.hpp" | ||
| 15 | |||
| 16 | namespace batkov_f_contrast_enh_lin_hist_stretch { | ||
| 17 | |||
| 18 | namespace { | ||
| 19 | |||
| 20 | constexpr size_t kParallelThreshold = 100000; | ||
| 21 | constexpr size_t kMinGrainSize = 1U << 16U; | ||
| 22 | |||
| 23 | struct Range { | ||
| 24 | size_t begin; | ||
| 25 | size_t end; | ||
| 26 | |||
| 27 | 30 | Range(size_t begin_value, size_t end_value) noexcept : begin(begin_value), end(end_value) {} | |
| 28 | }; | ||
| 29 | |||
| 30 | struct MinMax { | ||
| 31 | uint8_t min_val; | ||
| 32 | uint8_t max_val; | ||
| 33 | |||
| 34 | 104 | MinMax() noexcept : min_val(std::numeric_limits<uint8_t>::max()), max_val(std::numeric_limits<uint8_t>::min()) {} | |
| 35 | |||
| 36 | MinMax(uint8_t min_value, uint8_t max_value) noexcept : min_val(min_value), max_val(max_value) {} | ||
| 37 | }; | ||
| 38 | |||
| 39 | 48 | size_t ChooseThreadCount(size_t size, size_t requested_threads) noexcept { | |
| 40 |
2/2✓ Branch 0 taken 18 times.
✓ Branch 1 taken 30 times.
|
48 | if (size < kParallelThreshold || requested_threads <= 1) { |
| 41 | return 1; | ||
| 42 | } | ||
| 43 | |||
| 44 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 18 times.
|
18 | const size_t by_grain = std::max<size_t>(1, size / kMinGrainSize); |
| 45 | 36 | return std::max<size_t>(1, std::min({requested_threads, size, by_grain})); | |
| 46 | } | ||
| 47 | |||
| 48 | Range MakeRange(size_t size, size_t thread_index, size_t thread_count) noexcept { | ||
| 49 | 52 | return {(size * thread_index) / thread_count, (size * (thread_index + 1)) / thread_count}; | |
| 50 | } | ||
| 51 | |||
| 52 | std::array<uint8_t, 256> BuildStretchLut(uint8_t min_el, uint8_t max_el) { | ||
| 53 | 48 | std::array<uint8_t, 256> lut{}; | |
| 54 | |||
| 55 | 48 | const int min_value = static_cast<int>(min_el); | |
| 56 | 48 | const int max_value = static_cast<int>(max_el); | |
| 57 | 48 | const int range = max_value - min_value; | |
| 58 | |||
| 59 |
4/4✓ Branch 0 taken 4608 times.
✓ Branch 1 taken 18 times.
✓ Branch 2 taken 7680 times.
✓ Branch 3 taken 30 times.
|
12336 | for (size_t pixel = 0; pixel < lut.size(); ++pixel) { |
| 60 | 12288 | const int value = static_cast<int>(pixel); | |
| 61 | |||
| 62 |
4/4✓ Branch 0 taken 1818 times.
✓ Branch 1 taken 2790 times.
✓ Branch 2 taken 4230 times.
✓ Branch 3 taken 3450 times.
|
12288 | if (value <= min_value) { |
| 63 | 6048 | lut.at(pixel) = 0; | |
| 64 |
3/4✓ Branch 0 taken 2088 times.
✓ Branch 1 taken 702 times.
✓ Branch 2 taken 3450 times.
✗ Branch 3 not taken.
|
6240 | } else if (value >= max_value) { |
| 65 | 5538 | lut.at(pixel) = std::numeric_limits<uint8_t>::max(); | |
| 66 | } else { | ||
| 67 | 702 | lut.at(pixel) = static_cast<uint8_t>(((value - min_value) * 255) / range); | |
| 68 | } | ||
| 69 | } | ||
| 70 | |||
| 71 | return lut; | ||
| 72 | } | ||
| 73 | |||
| 74 | uint8_t GetLutValue(const std::array<uint8_t, 256> &lut, uint8_t pixel) noexcept { | ||
| 75 | 25442880 | return *std::next(lut.cbegin(), static_cast<std::ptrdiff_t>(pixel)); | |
| 76 | } | ||
| 77 | |||
| 78 | 52 | MinMax FindLocalMinMax(const InType &input, const Range &range) { | |
| 79 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 52 times.
|
52 | auto input_it = std::next(input.cbegin(), static_cast<std::ptrdiff_t>(range.begin)); |
| 80 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 52 times.
|
52 | const auto input_last = std::next(input.cbegin(), static_cast<std::ptrdiff_t>(range.end)); |
| 81 | |||
| 82 | MinMax result; | ||
| 83 |
2/2✓ Branch 0 taken 9072864 times.
✓ Branch 1 taken 52 times.
|
9072916 | for (; input_it != input_last; ++input_it) { |
| 84 |
2/2✓ Branch 0 taken 204 times.
✓ Branch 1 taken 9072660 times.
|
9072864 | result.min_val = std::min(result.min_val, *input_it); |
| 85 | 9072864 | result.max_val = std::max(result.max_val, *input_it); | |
| 86 | } | ||
| 87 | |||
| 88 | 52 | return result; | |
| 89 | } | ||
| 90 | |||
| 91 | MinMax CombineLocalMinMax(const std::vector<MinMax> &locals) noexcept { | ||
| 92 | MinMax result; | ||
| 93 | |||
| 94 |
2/2✓ Branch 0 taken 52 times.
✓ Branch 1 taken 18 times.
|
70 | for (const auto &local : locals) { |
| 95 | result.min_val = std::min(result.min_val, local.min_val); | ||
| 96 | result.max_val = std::max(result.max_val, local.max_val); | ||
| 97 | } | ||
| 98 | |||
| 99 | return result; | ||
| 100 | } | ||
| 101 | |||
| 102 | ✗ | void CopyRange(const InType &input, OutType &output, const Range &range) { | |
| 103 | ✗ | const auto input_first = std::next(input.cbegin(), static_cast<std::ptrdiff_t>(range.begin)); | |
| 104 | ✗ | const auto input_last = std::next(input.cbegin(), static_cast<std::ptrdiff_t>(range.end)); | |
| 105 | const auto output_first = std::next(output.begin(), static_cast<std::ptrdiff_t>(range.begin)); | ||
| 106 | |||
| 107 | ✗ | std::copy(input_first, input_last, output_first); | |
| 108 | ✗ | } | |
| 109 | |||
| 110 | 82 | void ApplyLutRange(const InType &input, OutType &output, const Range &range, const std::array<uint8_t, 256> &lut) { | |
| 111 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 82 times.
|
82 | auto input_it = std::next(input.cbegin(), static_cast<std::ptrdiff_t>(range.begin)); |
| 112 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 82 times.
|
82 | const auto input_last = std::next(input.cbegin(), static_cast<std::ptrdiff_t>(range.end)); |
| 113 | auto output_it = std::next(output.begin(), static_cast<std::ptrdiff_t>(range.begin)); | ||
| 114 | |||
| 115 |
2/2✓ Branch 0 taken 12721440 times.
✓ Branch 1 taken 82 times.
|
12721522 | for (; input_it != input_last; ++input_it, ++output_it) { |
| 116 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 12721440 times.
|
25442880 | *output_it = GetLutValue(lut, *input_it); |
| 117 | } | ||
| 118 | 82 | } | |
| 119 | |||
| 120 | 52 | void ApplyResultRange(const InType &input, OutType &output, const Range &range, bool copy_only, | |
| 121 | const std::array<uint8_t, 256> &lut) { | ||
| 122 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 52 times.
|
52 | if (copy_only) { |
| 123 | ✗ | CopyRange(input, output, range); | |
| 124 | ✗ | return; | |
| 125 | } | ||
| 126 | |||
| 127 | 52 | ApplyLutRange(input, output, range, lut); | |
| 128 | } | ||
| 129 | |||
| 130 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | void StretchSequential(const InType &input, OutType &output) { |
| 131 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | const auto minmax = std::ranges::minmax_element(input.cbegin(), input.cend()); |
| 132 | 30 | const uint8_t min_el = *minmax.max; | |
| 133 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | const uint8_t max_el = *minmax.min; |
| 134 | const Range full_range(0, input.size()); | ||
| 135 | |||
| 136 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
|
30 | if (min_el == max_el) { |
| 137 | ✗ | CopyRange(input, output, full_range); | |
| 138 | ✗ | return; | |
| 139 | } | ||
| 140 | |||
| 141 | const auto lut = BuildStretchLut(min_el, max_el); | ||
| 142 | 30 | ApplyLutRange(input, output, full_range, lut); | |
| 143 | } | ||
| 144 | |||
| 145 | 18 | void StretchParallel(const InType &input, OutType &output, size_t thread_count) { | |
| 146 | 18 | const size_t size = input.size(); | |
| 147 | 18 | std::vector<MinMax> locals(thread_count); | |
| 148 | 18 | std::array<uint8_t, 256> lut{}; | |
| 149 | 18 | bool copy_only = false; | |
| 150 | |||
| 151 | 18 | auto completion = [&locals, &lut, ©_only]() noexcept { | |
| 152 | 18 | const MinMax global = CombineLocalMinMax(locals); | |
| 153 | 18 | copy_only = global.min_val == global.max_val; | |
| 154 | |||
| 155 |
1/2✓ Branch 0 taken 18 times.
✗ Branch 1 not taken.
|
18 | if (!copy_only) { |
| 156 | 36 | lut = BuildStretchLut(global.min_val, global.max_val); | |
| 157 | } | ||
| 158 | 18 | }; | |
| 159 | |||
| 160 |
1/2✓ Branch 1 taken 18 times.
✗ Branch 2 not taken.
|
18 | std::barrier<decltype(completion)> barrier(static_cast<std::ptrdiff_t>(thread_count), completion); |
| 161 | |||
| 162 | 18 | std::vector<std::thread> threads; | |
| 163 |
1/2✓ Branch 1 taken 18 times.
✗ Branch 2 not taken.
|
18 | threads.reserve(thread_count); |
| 164 | |||
| 165 |
2/2✓ Branch 0 taken 52 times.
✓ Branch 1 taken 18 times.
|
70 | for (size_t thread_index = 0; thread_index < thread_count; ++thread_index) { |
| 166 |
1/2✓ Branch 1 taken 52 times.
✗ Branch 2 not taken.
|
52 | threads.emplace_back([&, thread_index]() { |
| 167 | 52 | const Range range = MakeRange(size, thread_index, thread_count); | |
| 168 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 52 times.
|
52 | locals.at(thread_index) = FindLocalMinMax(input, range); |
| 169 | 52 | barrier.arrive_and_wait(); | |
| 170 | 52 | ApplyResultRange(input, output, range, copy_only, lut); | |
| 171 | 52 | }); | |
| 172 | } | ||
| 173 | |||
| 174 |
2/2✓ Branch 0 taken 52 times.
✓ Branch 1 taken 18 times.
|
70 | for (auto &thread : threads) { |
| 175 |
1/2✓ Branch 1 taken 52 times.
✗ Branch 2 not taken.
|
52 | thread.join(); |
| 176 | } | ||
| 177 | 36 | } | |
| 178 | |||
| 179 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 18 times.
|
48 | void StretchContrast(const InType &input, OutType &output, size_t requested_threads) { |
| 180 | 48 | const size_t thread_count = ChooseThreadCount(input.size(), requested_threads); | |
| 181 | |||
| 182 |
2/2✓ Branch 0 taken 30 times.
✓ Branch 1 taken 18 times.
|
48 | if (thread_count == 1) { |
| 183 | 30 | StretchSequential(input, output); | |
| 184 | 30 | return; | |
| 185 | } | ||
| 186 | |||
| 187 | 18 | StretchParallel(input, output, thread_count); | |
| 188 | } | ||
| 189 | |||
| 190 | } // namespace | ||
| 191 | |||
| 192 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | BatkovFContrastEnhLinHistStretchSTL::BatkovFContrastEnhLinHistStretchSTL(const InType &in) { |
| 193 | SetTypeOfTask(GetStaticTypeOfTask()); | ||
| 194 |
1/2✓ Branch 1 taken 48 times.
✗ Branch 2 not taken.
|
48 | GetInput() = in; |
| 195 | 48 | } | |
| 196 | |||
| 197 | 48 | bool BatkovFContrastEnhLinHistStretchSTL::ValidationImpl() { | |
| 198 | 48 | return !GetInput().empty(); | |
| 199 | } | ||
| 200 | |||
| 201 | 48 | bool BatkovFContrastEnhLinHistStretchSTL::PreProcessingImpl() { | |
| 202 | 48 | GetOutput().resize(GetInput().size()); | |
| 203 | 48 | return true; | |
| 204 | } | ||
| 205 | |||
| 206 | 48 | bool BatkovFContrastEnhLinHistStretchSTL::RunImpl() { | |
| 207 | const auto &input = GetInput(); | ||
| 208 | auto &output = GetOutput(); | ||
| 209 | |||
| 210 | 48 | const size_t requested_threads = static_cast<size_t>(std::max(1, ppc::util::GetNumThreads())); | |
| 211 | 48 | StretchContrast(input, output, requested_threads); | |
| 212 | |||
| 213 | 48 | return true; | |
| 214 | } | ||
| 215 | |||
| 216 | 48 | bool BatkovFContrastEnhLinHistStretchSTL::PostProcessingImpl() { | |
| 217 | 48 | return !GetOutput().empty(); | |
| 218 | } | ||
| 219 | |||
| 220 | } // namespace batkov_f_contrast_enh_lin_hist_stretch | ||
| 221 |