gsla/source/lzb.cpp

//
// LZB Encode / Decode
// 
#include "lzb.h"

#include <stdio.h>
#include <string.h>

#include "bctypes.h"

#define MAX_DICTIONARY_SIZE (32 * 1024)
//
// Yes This is a 32K Buffer, of bytes, with no structure to it
//
static unsigned char *pDictionary = nullptr;

struct DataString {
	// Information about the data we're trying to match
	int size;
	unsigned char *pData;
};

static int AddDictionary(const DataString& data, int dictionarySize);
static int EmitLiteral(unsigned char *pDest, DataString& data);
static int ConcatLiteral(unsigned char *pDest, DataString& data);
static int EmitReference(unsigned char *pDest, int dictionaryOffset, DataString& data);
static int DictionaryMatch(const DataString& data, int dictionarySize);

// Stuff I need for a faster version
static DataString LongestMatch(const DataString& data, const DataString& dictionary);

//
//  New Version, still Brute Force, but not as many times
//
int LZB_Compress(unsigned char* pDest, unsigned char* pSource, int sourceSize)
{
	printf("LZB Compress %d bytes\n", sourceSize);

	unsigned char *pOriginalDest = pDest;

	DataString sourceData;
	DataString dictionaryData;
	DataString candidateData;

	// Source Data Stream - will compress until the size is zero
	sourceData.pData = pSource;
	sourceData.size  = sourceSize;

	// Remember, this eventually will point at the frame buffer
	dictionaryData.pData = pSource;
	dictionaryData.size = 0;

	// dumb last emit is a literal stuff
	bool bLastEmitIsLiteral = false;
	unsigned char* pLastLiteralDest = nullptr;

	while (sourceData.size > 0)
	{
		candidateData = LongestMatch(sourceData, dictionaryData);

		if (0 == candidateData.size)
		{
			candidateData.size = 1;
			candidateData.pData = sourceData.pData;
		}

		// Adjust source stream
		sourceData.pData += candidateData.size;
		sourceData.size  -= candidateData.size;

		dictionaryData.size = AddDictionary(candidateData, dictionaryData.size);

		if (candidateData.size > 3)
		{
			// Emit a dictionary reference
			pDest += (int)EmitReference(pDest, (int)(candidateData.pData - dictionaryData.pData), candidateData);
			bLastEmitIsLiteral = false;
		}
		else if (bLastEmitIsLiteral)
		{
			// Concatenate this literal onto the previous literal
			pDest += ConcatLiteral(pLastLiteralDest, candidateData);
		}
		else
		{
			// Emit a new literal
			pLastLiteralDest = pDest;
			bLastEmitIsLiteral = true;
			pDest += EmitLiteral(pDest, candidateData);
		}
	}

	return (int)(pDest - pOriginalDest);
}


//
// This works, but it's stupidly slow, because it uses brute force, and
// because the brute force starts over everytime I grow the data string
//
int old_LZB_Compress(unsigned char* pDest, unsigned char* pSource, int sourceSize)
{
	printf("LZB_Compress %d bytes\n", sourceSize);

	// Initialize Dictionary
	int bytesInDictionary = 0;		// eventually add the ability to start with the dictionary filled
	pDictionary = pSource;

	int processedBytes = 0;
	int bytesEmitted = 0;

	// dumb last emit is a literal stuff
	bool bLastEmitIsLiteral = false;
	int  lastEmittedLiteralOffset = 0;

	DataString candidate_data;
	candidate_data.pData = pSource;
	candidate_data.size = 0;

	int MatchOffset = -1;
	int PreviousMatchOffset = -1;

	while (processedBytes < sourceSize)
	{
		// Add a byte to the candidate_data, also tally number of processed
		processedBytes++;
		candidate_data.size++;  

		// Basic Flow Idea Here
		// If there's a match, then add to the candidate data, and see if
		// there's a bigger match (use previous result to speed up search)
		// else
		// if there's a previous match, and it's large enough, emit that
		// else emit what we have as a literal


		// (KMP is probably the last planned optmization here)
		PreviousMatchOffset = MatchOffset; 

		MatchOffset = DictionaryMatch(candidate_data, bytesInDictionary);

		// The dictionary only contains bytes that have been emitted, so we
		// can't add this byte until we've emitted it?
		if (MatchOffset < 0)
		{
			// Was there a dictionary match

			// Previous Data, we can't get here with candidate_data.size == 0
			// this is an opportunity to use an assert
			candidate_data.size--;

			MatchOffset = PreviousMatchOffset; //DictionaryMatch(candidate_data, bytesInDictionary);

			if ((MatchOffset >= 0) && candidate_data.size > 3)
			{
				processedBytes--;
				bytesInDictionary = AddDictionary(candidate_data, bytesInDictionary);
				bytesEmitted += EmitReference(pDest + bytesEmitted, MatchOffset, candidate_data);
				bLastEmitIsLiteral = false;
			}
			else
			{
				candidate_data.size++;
				// Add Dictionary
				bytesInDictionary = AddDictionary(candidate_data, bytesInDictionary);

				if (bLastEmitIsLiteral)
				{
					// If the last emit was a literal, I want to concatenate
					// this literal into the previous opcode, to save space
					bytesEmitted += ConcatLiteral(pDest + lastEmittedLiteralOffset, candidate_data);
				}
				else
				{
					lastEmittedLiteralOffset = bytesEmitted;
					bytesEmitted += EmitLiteral(pDest + bytesEmitted, candidate_data);
				}
				bLastEmitIsLiteral = true;
			}
		}
	}

	if (candidate_data.size > 0)
	{

		int MatchOffset = DictionaryMatch(candidate_data, bytesInDictionary);

		if ((MatchOffset >=0) && candidate_data.size > 2)
		{
			bytesInDictionary = AddDictionary(candidate_data, bytesInDictionary);
			bytesEmitted += EmitReference(pDest + bytesEmitted, MatchOffset, candidate_data);
		}
		else
		{
			// Add Dictionary
			bytesInDictionary = AddDictionary(candidate_data, bytesInDictionary);

			if (bLastEmitIsLiteral)
			{
				// If the last emit was a literal, I want to concatenate
				// this literal into the previous opcode, to save space
				bytesEmitted += ConcatLiteral(pDest + lastEmittedLiteralOffset, candidate_data);
			}
			else
			{
				bytesEmitted += EmitLiteral(pDest + bytesEmitted, candidate_data);
			}
		}
	}

	return bytesEmitted;
}

//------------------------------------------------------------------------------
// Return new dictionarySize
static int AddDictionary(const DataString& data, int dictionarySize)
{
	//int dataIndex = 0;
	//while ((dictionarySize < MAX_DICTIONARY_SIZE) && (dataIndex < data.size))
	//{
	//	pDictionary[ dictionarySize++ ] = data.pData[ dataIndex++ ];
	//}

	dictionarySize += data.size;

	return dictionarySize;
}

//------------------------------------------------------------------------------
//
// Return longest match of data, in dictionary
//

DataString LongestMatch(const DataString& data, const DataString& dictionary)
{
	DataString result;
	result.pData = nullptr;
	result.size = 0;

	// Find the longest matching data in the dictionary
	if ((dictionary.size > 0) && (data.size > 0))
	{
		DataString candidate;
		candidate.pData = data.pData;
		candidate.size = 0;

		// First Check for a pattern / run-length style match
		// Check the end of the dictionary, to see if this data could be a
		// pattern "run" (where we can repeat a pattern for X many times for free
		// using the memcpy with overlapping source/dest buffers)
		// (This is a dictionary based pattern run/length)
		{
			// Check for pattern sizes, start small
			int max_pattern_size = 4096;
			if (dictionary.size < max_pattern_size)  max_pattern_size = dictionary.size;
			if (data.size < max_pattern_size) max_pattern_size = data.size;

			for (int pattern_size = 1; pattern_size <= max_pattern_size; ++pattern_size)
			{
				bool bMatch = true;
				int pattern_start = dictionary.size - pattern_size;

				for (int dataIndex = 0; dataIndex < data.size; ++dataIndex)
				{
					if (data.pData[ dataIndex ] == dictionary.pData[ pattern_start + (dataIndex % pattern_size) ])
					{
						candidate.pData = dictionary.pData + pattern_start;
						candidate.size = dataIndex+1;
						continue;
					}

					bMatch = false;
					break;
				}

				if (candidate.size < pattern_size)
					break;

				if (candidate.size > result.size)
				{
					result = candidate;
				}
			}
		}

		// As an optimization
		int dictionarySize = dictionary.size - 1;	// This last string has already been checked by, the
												    // run-length matcher above

		if (dictionarySize >= candidate.size)
		{
			// Check the dictionary for a match, brute force
			for (int idx = 0; idx <= (dictionarySize-candidate.size); ++idx)
			{
				bool bMatch = true;
				for (int dataIdx = 0; dataIdx < data.size; ++dataIdx)
				{
					if (data.pData[ dataIdx ] == dictionary.pData[ idx + dataIdx ])
					{
						if (dataIdx > (candidate.size-1))
						{
							candidate.pData = dictionary.pData + idx;
							candidate.size = dataIdx - 1;
						}
						continue;
					}

					bMatch = false;
					break;
				}

				if (candidate.size > result.size)
				{
					result = candidate;
					break;
				}
			}
		}
	}

	return result;
}


//------------------------------------------------------------------------------
//
// Return offset into dictionary where the string matches
//
// -1 means, no match
//
static int DictionaryMatch(const DataString& data, int dictionarySize)
{
	if( (0 == dictionarySize ) ||
		(0 == data.size) ||
		(data.size > 16384) ) // 16384 is largest string copy we can encode
	{
		return -1;
	}

	// Check the end of the dictionary, to see if this data could be a
	// pattern "run" (where we can repeat a pattern for X many times for free
	// using the memcpy with overlapping source/dest buffers)
	// (This is a dictionary based pattern run/length)

	{
		// Check for pattern sizes, start small
		int max_pattern_size = 256;
		if (dictionarySize < max_pattern_size)  max_pattern_size = dictionarySize;
		if (data.size < max_pattern_size) max_pattern_size = data.size;

		for (int pattern_size = 1; pattern_size <= max_pattern_size; ++pattern_size)
		{
			bool bMatch = true;
			int pattern_start = dictionarySize - pattern_size;

			for (int dataIndex = 0; dataIndex < data.size; ++dataIndex)
			{
				if (data.pData[ dataIndex ] == pDictionary[ pattern_start + (dataIndex % pattern_size) ])
					continue;

				bMatch = false;
				break;
			}

			if (bMatch)
			{
				// Return a RLE Style match result
				return pattern_start;
			}
		}
	}

	// As an optimization
	dictionarySize -= 1;	// This last string has already been checked by, the
						    // run-length matcher above

	if (dictionarySize < data.size)
	{
		return -1;
	}

	int result = -1;

	// Check the dictionary for a match, brute force
	for (int idx = 0; idx <= (dictionarySize-data.size); ++idx)
	{
		bool bMatch = true;
		for (int dataIdx = 0; dataIdx < data.size; ++dataIdx)
		{
			if (data.pData[ dataIdx ] == pDictionary[ idx + dataIdx ])
				continue;

			bMatch = false;
			break;
		}

		if (bMatch)
		{
			result = idx;
			break;
		}
	}

	return result;
}

//------------------------------------------------------------------------------
//
// Emit a literal, that appends itself to an existing literal
//
static int ConcatLiteral(unsigned char *pDest, DataString& data)
{
	// Return Size
	int outSize = (int)data.size;

	int opCode  = pDest[0];
	    opCode |= (int)(((pDest[1])&0x7F)<<8);

    int skip = opCode;
	opCode += outSize;

	// Opcode
	*pDest++ = (unsigned char)(opCode & 0xFF);
	*pDest++ = (unsigned char)((opCode >> 8) & 0x7F);

	pDest += skip;

	// Literal Data
	for (int idx = 0; idx < data.size; ++idx)
	{
		*pDest++ = data.pData[ idx ];
	}

	// Clear
	data.pData += data.size;
	data.size = 0;

	return outSize;
}

//------------------------------------------------------------------------------

static int EmitLiteral(unsigned char *pDest, DataString& data)
{
	// Return Size
	int outSize = 2 + (int)data.size;

	// Opcode
	*pDest++ = (unsigned char)(data.size & 0xFF);
	*pDest++ = (unsigned char)((data.size >> 8) & 0x7F);

	// Literal Data
	for (int idx = 0; idx < data.size; ++idx)
	{
		*pDest++ = data.pData[ idx ];
	}

	// Clear
	data.pData += data.size;
	data.size = 0;

	return outSize;
}

//------------------------------------------------------------------------------

static int EmitReference(unsigned char *pDest, int dictionaryOffset, DataString& data)
{
	// Return Size
	int outSize = 2 + 2;

	// Opcode
	*pDest++ = (unsigned char)(data.size & 0xFF);
	*pDest++ = (unsigned char)((data.size >> 8) & 0x7F) | 0x80;

	*pDest++ = (unsigned char)(dictionaryOffset & 0xFF);
	*pDest++ = (unsigned char)((dictionaryOffset>>8) & 0xFF);

	// Clear
	data.pData += data.size;
	data.size = 0;

	return outSize;
}

//------------------------------------------------------------------------------
//
// Std C memcpy seems to be stopping the copy from happening, when I overlap
// the buffer to get a pattern run copy (overlapped buffers)
//
static void my_memcpy(u8* pDest, u8* pSrc, int length)
{
	while (length-- > 0)
	{
		*pDest++ = *pSrc++;
	}
}

//------------------------------------------------------------------------------
//
//  Simple Decompress, for validation
//
void LZB_Decompress(unsigned char* pDest, unsigned char* pSource, int destSize)
{
	int decompressedBytes = 0;

	while (decompressedBytes < destSize)
	{
		u16 opcode  = *pSource++;
		    opcode |= ((u16)(*pSource++))<<8;

		if (opcode & 0x8000)
		{
			// Dictionary
			opcode &= 0x7FFF;

			// Dictionary Copy from the output stream
			u16 offset  = *pSource++;
			    offset |= ((u16)(*pSource++))<<8;

			my_memcpy(&pDest[ decompressedBytes ], &pDest[ offset ], opcode);
			decompressedBytes += opcode;
		}
		else
		{
			// Literal Copy, from compressed stream
			memcpy(&pDest[ decompressedBytes ], pSource, opcode);
			decompressedBytes += opcode;
			pSource += opcode;
		}
	}
}

//------------------------------------------------------------------------------