#pragma once #include // 4J added - Storage for block & sky light data. Lighting data is normally stored as 4-bits per tile, in a DataLayer class of 16384 bytes ( 128 x 16 x 16 x 0.5 ) // This class provides more economical storage for such data by taking into consideration that it is quite common for large parts of the lighting data in a level to // be very compressible (large amounts of 0 for block lights, 0 and 15 for sky lights). // We are aiming here to balance performance (lighting data is accessed very frequently) against size. // Details of storage method: // 1. Lighting is split into horizontal planes, of which there are 128, and each taking up 128 bytes (16 x 16 x 0.5) // 2. Each of these layers has a permanently allocated index in this class (planeIndices). // 3. Data for allocatedPlaneCount planes worth of data is allocated in the data array ( allocatedPlaneCount * 128 bytes ) // 4. If a plane index for a layer is < 128, then the data for that layer is at data[ index * 128 ] // 5. If a plane index for a layer is 128, then all values for that plane are 0 // 6. If a plane index for a layer is 129, then all values for that plane are 15 // This class needs to be thread safe as there are times where chunk (and light) data are shared between server & main threads. Light values are queried // very regularly so this needs to be as light-weight as possible. // To meet these requirements, this class is now implemented using a lock-free system, implemented using a read-copy-update (RCU) type algorithm. Some details... // (1) The storage details for the class are now packed into a single int64_t, which contains both a pointer to the data that is required and a count of how many planes worth // of storage are allocated. This allows the full storage to be updated atomically using compare and exchange operations (implemented with InterlockedCompareExchangeRelease64). // (2) The data pointer referenced in this int64_t points to an area of memory which is 128 + 128 * plane_count bytes long, where the first 128 bytes stoere the plane indices, and // the rest of the data is variable in size to accomodate however many planes are required to be stored // (3) The RCU bit of the algorithm means that any read operations don't need to do any checks or locks at all. When the data needs to be updated, a copy of it is made and updated, // then an attempt is made to swap the new data in - if this succeeds then the old data pointer is deleted later at some point where we know nothing will be reading from it anymore. // This is achieved by putting the delete request in a queue which means it won't actually get deleted until 2 game ticks after the last time its reference existed, which should give // us a large margin of safety. If the attempt to swap the new data in fails, then the whole write operation has to be attempted again - this is the only time there is really a // high cost for this algorithm and such write collisions should be rare. //#define LIGHT_COMPRESSION_STATS class SparseLightStorage_SPU { private: // unsigned char planeIndices[128]; unsigned char* m_pData; // unsigned char *data; // unsigned int allocatedPlaneCount; static const int ALL_0_INDEX = 128; static const int ALL_15_INDEX = 129; public: SparseLightStorage_SPU(unsigned char* data) : m_pData(data) {} unsigned char* getDataPtr() { return m_pData; } inline int get(int x, int y, int z) // Get an individual lighting value { unsigned char *planeIndices, *data; getPlaneIndicesAndData(&planeIndices, &data); if( planeIndices[y] == ALL_0_INDEX ) { return 0; } else if ( planeIndices[y] == ALL_15_INDEX ) { return 15; } else { int planeIndex = x * 16 + z; // Index within this xz plane int byteIndex = planeIndex / 2; // Byte index within the plane (2 tiles stored per byte) int shift = ( planeIndex & 1 ) * 4; // Bit shift within the byte int retval = ( data[ planeIndices[y] * 128 + byteIndex ] >> shift ) & 15; return retval; } } inline void getPlaneIndicesAndData(unsigned char **planeIndices, unsigned char **data) { *planeIndices = m_pData; *data = m_pData + 128; } };