C++ 缓存设计建议

Question

我有一个包含多种图像类型(RGB、灰色...)的 C++ 应用程序，每种类型都具有旋转或缩放等属性。每种图像类型都是通过其他类型的一些计算生成的。例如，旋转的 GrayImage 是通过旋转 GrayImage 生成的，而 GrayImage 又是通过“灰色化” RGBImage 生成的。

我想用 GetX(...) 方法设计一个缓存类来缓存各种图像(可能还有计算路径中的所有图像)。此类还知道如何生成每个图像，以防它不在缓存中。

类必须满足一些约束:

1. 由于我们处理的是不同类型和表现形式(RGB、GrayScale 等)的图像，因此缓存必须返回一个具体类，以便调用代码能够在不进行某种转换的情况下使用它。因此，缓存机制必须包含包含具体类型的不同缓存结构。 (如果我在这里错了，请修复我)

   map<...,RGBImage> 
   map<...,GrayImage> 
   

   例如。

2. 缓存必须能够灵活应对图像计算的变化。只要不是太大，代码更改是可以接受的。

我的当前版本为每种图像类型附加了一个 Key 结构。有GrayKey、RGBKey等。各种键包含 Scale 和 Rotation 等属性，并且可以具有特定于图像的属性(例如 GrayKey 的 toGrayConvertingMethod)。缓存包含以下形式的 map :

    map <XKey,XImage>

GetX(...) 方法接收一个 Key 结构作为请求旋转 GrayImage 的参数。但是，此实现会强制缓存应用大量逻辑来计算图像。它必须检查 GrayKey 是否请求旋转图像并采取相应行动。我想以更优雅的方式“编码”这种图像计算关系，但似乎找不到。

有什么建议吗？

非常感谢。

Answer 1

也许您可以使用 Boost.MultiIndex 做一些事情容器？它可以让你创建一个类型来存储图像数据，以及它是如何被操作的细节，然后根据你想要的任何键组合查找值。如果您以前没有使用过它，它可能看起来有点令人生畏，但我在下面附上了一个示例。

显然，我的例子只处理缓存机制的存储/检索部分，如果你把它粘在一起，如果查找失败，它可以生成图像，它应该做你想做的一切。扩展它也很容易……需要查找额外的参数吗？您只需要将另一个索引添加到 ImageCache typedef。

#include <boost/multi_index_container.hpp>
#include <boost/multi_index/composite_key.hpp>
#include <boost/multi_index/member.hpp>
#include <boost/multi_index/ordered_index.hpp>
#include <boost/multi_index/hashed_index.hpp>
#include <boost/multi_index/tag.hpp>

#include <boost/shared_array.hpp>

#include <algorithm>
#include <iostream>
#include <utility>

// A cache item, stores the image data, and any values we need to 
// index on.
struct ImageCacheItem
{
    enum RgbMode { RGB_MODE_COLOUR, RGB_MODE_GREYSCALE };

    // Im not sure how much copying goes on in the container,
    // so using a smart pointer to prevent copying large amounts
    // of data.
    boost::shared_array<char> imageBuffer;

    double  rotation;
    double  scale;
    RgbMode rgbMode;

    ImageCacheItem(double r, double s)
    : rotation(r), scale(s)
    {
    }
};

// These are "tag" structures, they are used as part of the 
// multi_index_container as a way to distinguish between indicies.
struct ByRotation {};
struct ByScale {};
struct ByRgb {};
struct ByRotationScale {};

// Typedef of the container itself.
typedef boost::multi_index_container<
    ImageCacheItem, // The data type for the container.  
                    // Note there is no "key" type, as the key values 
                    // are extracted from the data items theselves.
    boost::multi_index::indexed_by<

        // Define an index for the rotation value
        boost::multi_index::ordered_non_unique<
            boost::multi_index::tag<ByRotation>,
            BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, rotation)
        >,

        // Define an index for the scale value
        boost::multi_index::ordered_non_unique<
            boost::multi_index::tag<ByScale>,
            BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, scale)
        >,

        // Define an index for the rgb value
        boost::multi_index::hashed_non_unique<
            boost::multi_index::tag<ByRgb>,
            BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, ImageCacheItem::RgbMode, rgbMode)
        >,

        // Define an index by rotation + scale
        boost::multi_index::hashed_unique<
            boost::multi_index::tag<ByRotationScale>,
            boost::multi_index::composite_key<
                ImageCacheItem,
                BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, rotation),
                BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, scale)
            >
        >

    >
> ImageCache;

// Utility typedefs, you'll want these to shorten the iterator
// data types when you're looking things up (see main).
typedef ImageCache::index<ByRotation>::type      ImageCacheByRotation;
typedef ImageCache::index<ByScale>::type         ImageCacheByScale;
typedef ImageCache::index<ByRgb>::type           ImageCacheByRgb;
typedef ImageCache::index<ByRotationScale>::type ImageCacheByRotationScale;

int main()
{
    // Create the cache and add time "images" to it.
    ImageCache cache;
    cache.insert(ImageCacheItem(10, 10));
    cache.insert(ImageCacheItem(10, 20));
    cache.insert(ImageCacheItem(20, 20));

    // look up the images with scale of 20.
    typedef ImageCacheByScale::iterator ScaleIter;

    std::pair<ScaleIter, ScaleIter> scaleResult = cache.get<ByScale>().equal_range(20);
    std::cout << "Found " << std::distance(scaleResult.first, scaleResult.second) << " Results" << std::endl;

    // look up the image with rotation = 10 && scale = 20.
    ImageCacheByRotationScale::iterator rsResult = cache.get<ByRotationScale>().find(boost::make_tuple(10, 20));
    std::cout << "Found " << (rsResult != cache.get<ByRotationScale>().end() ? 1 : 0) << " Results" << std::endl;

    return 0;
}

编辑:它很大...

我试过扩展上面的示例以在缓存中找到与您搜索的内容最接近但有偏差的图像，所以如果您想要旋转 45，并且比例为 10，如果没有找到完全匹配，它会倾向于一个属性相同而另一个属性为 0 的结果(即比例为 10，但旋转为 0，因此您需要做的就是旋转) .

代码被注释以解释其作用，但基本上，它使用模板递归按顺序搜索索引，一旦索引找到一些匹配项，它就会尝试按相关性顺序对它们进行排序，并返回最佳匹配。要添加另一个属性，您需要执行以下操作:

1. 将属性添加到 ImageCacheItem
2. 将属性比较添加到 ImageCacheSimilarity
3. (可选)将另一个与其匹配的索引添加到 ImageCache typedef

它可能不是最佳解决方案，但我认为它涵盖了您在评论中提到的用例。

#include <boost/multi_index_container.hpp>
#include <boost/multi_index/composite_key.hpp>
#include <boost/multi_index/member.hpp>
#include <boost/multi_index/ordered_index.hpp>
#include <boost/multi_index/hashed_index.hpp>
#include <boost/multi_index/tag.hpp>

#include <boost/mpl/list.hpp>
#include <boost/optional.hpp>
#include <boost/ref.hpp>
#include <boost/shared_array.hpp>

#include <algorithm>
#include <cmath>
#include <iostream>
#include <utility>
#include <vector>
#include <typeinfo>

// A cache item, stores the image data, and any values we need to 
// index on.
struct ImageCacheItem
{
    enum RgbMode { RGB_MODE_COLOUR, RGB_MODE_GREYSCALE };

    // Im not sure how much copying goes on in the container,
    // so using a smart pointer to prevent copying large amounts
    // of data.
    boost::shared_array<char> imageBuffer;

    double  rotation;
    double  scale;
    RgbMode rgbMode;

    ImageCacheItem(double r, double s)
    : rotation(r), scale(s)
    {
    }
};

// Calculates the similarity between two ImageCacheItem objects.
int ImageCacheSimilarity(const ImageCacheItem& item, const ImageCacheItem& target)
{
    const double EPSILON = 0.0000001;

    int score = 0;

    // 2 points for an exact match
    // 1 point if the value is 0 (e.g. not rotated, so can be used as a starting point)
    // -1 point otherwise
    score += (std::fabs(item.rotation - target.rotation) < EPSILON) 
        ? 2 
        : ((std::fabs(item.rotation) < EPSILON) ? 1 : -1);

    score += (std::fabs(item.scale - target.scale) < EPSILON) 
        ? 2 
        : ((std::fabs(item.scale) < EPSILON) ? 1 : -1);

    score += (item.rgbMode == target.rgbMode) ? 2 : 0;

    return score;
}

// Orders ImageCacheItem objects based on their similarity to a target value.
struct ImageCacheCmp
{
    const ImageCacheItem& target;

    ImageCacheCmp(const ImageCacheItem& t)
    : target(t)
    {
    }

    bool operator()(const ImageCacheItem& a, const ImageCacheItem& b)
    {
        return (ImageCacheSimilarity(a, target) > ImageCacheSimilarity(b, target));
    }
};

// These are "tag" structures, they are used as part of the 
// multi_index_container as a way to distinguish between indicies.
struct ByRotation {};
struct ByScale {};
struct ByRgb {};
struct ByRotationScale {};

// Typedef of the container itself.
typedef boost::multi_index_container<
    ImageCacheItem, // The data type for the container.  
                    // Note there is no "key" type, as the key values 
                    // are extracted from the data items theselves.
    boost::multi_index::indexed_by<

        // The order of indicies here will affect performance, put the 
        // ones that match against the most fields first.  Its not required
        // to make it work, but it will reduce the number of matches to 
        // compare against later on.

        // Define an index by rotation + scale
        boost::multi_index::hashed_unique<
            boost::multi_index::tag<ByRotationScale>,
            boost::multi_index::composite_key<
                ImageCacheItem,
                BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, rotation),
                BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, scale)
            >
        >,

        // Define an index for the rotation value
        boost::multi_index::ordered_non_unique<
            boost::multi_index::tag<ByRotation>,
            BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, rotation)
        >,

        // Define an index for the scale value
        boost::multi_index::ordered_non_unique<
            boost::multi_index::tag<ByScale>,
            BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, double, scale)
        >,

        // Define an index for the rgb value
        boost::multi_index::hashed_non_unique<
            boost::multi_index::tag<ByRgb>,
            BOOST_MULTI_INDEX_MEMBER(ImageCacheItem, ImageCacheItem::RgbMode, rgbMode)
        >
    >
> ImageCache;

// Type of the vector used when collecting index results.  It stores
// references to the values in the cache to minimise copying.
typedef std::vector<boost::reference_wrapper<const ImageCacheItem> > ImageCacheResults;

// Utility class for overload resolution
template <int I>
struct Int2Type
{
    enum { value = I };
};

void FindMatches(
    const ImageCacheItem& item, 
    const ImageCache& cache, 
    ImageCacheResults& results,
    const Int2Type<boost::mpl::size<ImageCache::index_type_list>::type::value>&)
{
    // End of template recursion
}

template <int I>
void FindMatches(
    const ImageCacheItem& item, 
    const ImageCache& cache, 
    ImageCacheResults& results,
    const Int2Type<I>&)
{
    // Get the index being searched
    typedef typename ImageCache::nth_index<I>::type Index;

    // This type knows how to extract the relevant bits of ImageCacheItem
    // for this particular index.
    typename Index::key_from_value keyExtractor;

    // Look for matches in the index.
    std::pair<typename Index::const_iterator, typename Index::const_iterator> iter = 
        cache.get<I>().equal_range(keyExtractor(item));

    // If we found any results, add them to 'results', otherwise
    // continue to the next index.
    if (iter.first != iter.second)
    {
        results.reserve(std::distance(iter.first, iter.second));

        for ( ; iter.first != iter.second; ++iter.first)
        {
            results.push_back(boost::cref(*iter.first));
        }
    }
    else
    {
        FindMatches(item, cache, results, Int2Type<I + 1>());
    }
}

boost::optional<ImageCacheItem> FindClosestImage(const ImageCacheItem& item, const ImageCache& cache)
{
    // Find exact/partial matches according to the indicies.
    ImageCacheResults results;
    FindMatches(item, cache, results, Int2Type<0>());

    // If no matches were found, return an empty value
    if (results.empty())
    {
        return boost::optional<ImageCacheItem>();
    }

    // We got this far, so we must have some candiates, the problem is
    // we dont know which is the best match, so here we sort the results
    // based on proximity to the "item".  However, we are only interested
    // in the best match, so do a partial_sort.
    std::partial_sort(results.begin(), results.begin() + 1, results.end(), ImageCacheCmp(item));

    return results.front().get();
}

int main()
{
    // Create the cache and add some "images" to it.
    ImageCache cache;
    cache.insert(ImageCacheItem(10, 20));
    cache.insert(ImageCacheItem(10, 10));
    cache.insert(ImageCacheItem(10, 2));
    cache.insert(ImageCacheItem(20, 20));
    cache.insert(ImageCacheItem(30, 20));
    cache.insert(ImageCacheItem(30, 0));

    // Look for an image similar to rotation = 30 && scale = 2.
    boost::optional<ImageCacheItem> result = FindClosestImage(ImageCacheItem(30, 2), cache);

    // Have to check if result is value before usage, it would be empty 
    // if not match is found.
    if (result)
    {
        std::cout << "Found (" << result->rotation 
                  << ", " << result->scale << ")" 
                  << std::endl;
    }
    else
    {
        std::cout << "No Results" << std::endl;
    }

    return 0;
}