TileGen.cpp

/**
 ** Supermodel
 ** A Sega Model 3 Arcade Emulator.
 ** Copyright 2011-2012 Bart Trzynadlowski, Nik Henson 
 **
 ** This file is part of Supermodel.
 **
 ** Supermodel is free software: you can redistribute it and/or modify it under
 ** the terms of the GNU General Public License as published by the Free 
 ** Software Foundation, either version 3 of the License, or (at your option)
 ** any later version.
 **
 ** Supermodel is distributed in the hope that it will be useful, but WITHOUT
 ** ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 ** FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 ** more details.
 **
 ** You should have received a copy of the GNU General Public License along
 ** with Supermodel.  If not, see <http://www.gnu.org/licenses/>.
 **/
 
/*
 * TileGen.cpp
 * 
 * Implementation of the CTileGen class: 2D tile generator. Palette decoding 
 * and synchronization with the renderer (which may run in a separate thread)
 * are performed here as well. For a description of the tile generator
 * hardware, please refer to the 2D rendering engine source code.
 *
 * Palettes
 * --------
 *
 * Multiple copies of the 32K-color palette data are maintained. The first is
 * the raw data as written to the VRAM. Two copies are computed, one for layers
 * A/A' and the other for layers B/B'. These pairs of layers have independent 
 * color offset registers associated with them. The renderer uses these 
 * "computed" palettes.
 *
 * The computed palettes are updated whenever the real palette is modified, a
 * single color entry at a time. If the color register is modified, the entire
 * palette has to be recomputed accordingly.
 *
 * The read-only copy of the palette, which is generated for the renderer, only
 * stores the two computed palettes.
 *
 * TO-DO List:
 * -----------
 * - For consistency, the registers should probably be byte reversed (this is a
 *   little endian device), forcing the Model3 Read32/Write32 handlers to
 *   manually reverse the data. This keeps with the convention for VRAM.
 */

#include <cstring>
#include "Supermodel.h"

// Macros that divide memory regions into pages and mark them as dirty when they are written to
#define PAGE_WIDTH 10
#define PAGE_SIZE (1<<PAGE_WIDTH)
#define DIRTY_SIZE(arraySize) (1+(arraySize-1)/(8*PAGE_SIZE))
#define MARK_DIRTY(dirtyArray, addr) dirtyArray[addr>>(PAGE_WIDTH+3)] |= 1<<((addr>>PAGE_WIDTH)&7)

// Offsets of memory regions within TileGen memory pool
#define OFFSET_VRAM         0x000000	// VRAM and palette data
#define OFFSET_PAL_A        0x120000	// computed A/A' palette
#define OFFSET_PAL_B		0x140000	// computed B/B' palette 
#define MEM_POOL_SIZE_RW    (0x120000+0x040000)

#define OFFSET_VRAM_RO      0x160000   // [read-only snapshot]
#define OFFSET_PAL_RO_A     0x280000   // [read-only snapshot]
#define OFFSET_PAL_RO_B		0x2A0000
#define MEM_POOL_SIZE_RO    (0x120000+0x040000)

#define OFFSET_VRAM_DIRTY   0x2C0000
#define OFFSET_PAL_A_DIRTY  (OFFSET_VRAM_DIRTY+DIRTY_SIZE(0x120000))
#define OFFSET_PAL_B_DIRTY	(OFFSET_PAL_A_DIRTY+DIRTY_SIZE(0x20000))
#define MEM_POOL_SIZE_DIRTY (DIRTY_SIZE(0x120000)+2*DIRTY_SIZE(0x20000))	// VRAM + 2 palette dirty buffers

#define MEMORY_POOL_SIZE	(MEM_POOL_SIZE_RW+MEM_POOL_SIZE_RO+MEM_POOL_SIZE_DIRTY)


/******************************************************************************
 Save States
******************************************************************************/

void CTileGen::SaveState(CBlockFile *SaveState)
{
	SaveState->NewBlock("Tile Generator", __FILE__);
	SaveState->Write(vram, 0x120000); // Don't write out palette, read-only snapshots or dirty page arrays, just VRAM
	SaveState->Write(regs, sizeof(regs));
}

void CTileGen::LoadState(CBlockFile *SaveState)
{
	if (OKAY != SaveState->FindBlock("Tile Generator"))
	{
		ErrorLog("Unable to load tile generator state. Save state file is corrupt.");
		return;
	}
	
	// Load memory one word at a time
	for (int i = 0; i < 0x120000; i += 4)
	{
		UINT32 data;
	
		SaveState->Read(&data, sizeof(data));
		WriteRAM(i, data);
	}	
	SaveState->Read(regs, sizeof(regs));
	
	// Because regs were read after palette, must recompute
	RecomputePalettes();
	
	// If multi-threaded, update read-only snapshots too
	if (g_Config.gpuMultiThreaded)
		UpdateSnapshots(true);
}


/******************************************************************************
 Rendering
******************************************************************************/

void CTileGen::BeginVBlank(void)
{
/*
	printf("08: %X\n", regs[0x08/4]);
	printf("0C: %X\n", regs[0x0C/4]);
	printf("20: %X\n", regs[0x20/4]);
	printf("40: %X\n", regs[0x40/4]);
	printf("44: %X\n", regs[0x44/4]);
	printf("60: %08X\n", regs[0x60/4]);
	printf("64: %08X\n", regs[0x64/4]);
	printf("68: %08X\n", regs[0x68/4]);
	printf("6C: %08X\n", regs[0x6C/4]);
	printf("\n");
*/
}

void CTileGen::EndVBlank(void)
{
	//
}

void CTileGen::RecomputePalettes(void)
{
	// Writing the colors forces palettes to be computed
	if (g_Config.gpuMultiThreaded)
	{
		for (unsigned colorAddr = 0; colorAddr < 32768*4; colorAddr += 4 )
		{
			MARK_DIRTY(palDirty[0], colorAddr);
			MARK_DIRTY(palDirty[1], colorAddr);
			WritePalette(colorAddr/4, *(UINT32 *) &vram[0x100000+colorAddr]);
		}
	}
	else
	{
		for (unsigned colorAddr = 0; colorAddr < 32768*4; colorAddr += 4 )
			WritePalette(colorAddr/4, *(UINT32 *) &vram[0x100000+colorAddr]);
	}
}

UINT32 CTileGen::SyncSnapshots(void)
{
	// Good time to recompute the palettes
	if (recomputePalettes)
	{
		RecomputePalettes();
		recomputePalettes = false;
	}
	
	if (!g_Config.gpuMultiThreaded)
		return 0;
	
	// Update read-only snapshots
	return UpdateSnapshots(false);
}

UINT32 CTileGen::UpdateSnapshot(bool copyWhole, UINT8 *src, UINT8 *dst, unsigned size, UINT8 *dirty)
{
	unsigned dirtySize = DIRTY_SIZE(size);
	if (copyWhole)
	{
		// If updating whole region, then just copy all data in one go
		memcpy(dst, src, size);
		memset(dirty, 0, dirtySize);
		return size;
	}
	else
	{
		// Otherwise, loop through dirty pages array to find out what needs to be updated and copy only those parts
		UINT32 copied = 0;
		UINT8 *pSrc = src;
		UINT8 *pDst = dst;
		for (unsigned i = 0; i < dirtySize; i++)
		{
			UINT8 d = dirty[i];
			if (d)
			{
				for (unsigned j = 0; j < 8; j++)
				{
					if (d&1)
					{
						// If not at very end of region, then copy an extra 4 bytes to allow for a possible 32-bit overlap
						UINT32 toCopy = (i < dirtySize - 1 || j < 7 ? PAGE_SIZE + 4 : PAGE_SIZE);
						memcpy(pDst, pSrc, toCopy);
						copied += toCopy;
					}
					d >>= 1;
					pSrc += PAGE_SIZE;	
					pDst += PAGE_SIZE;
				}
				dirty[i] = 0;
			}
			else
			{
				pSrc += 8 * PAGE_SIZE;	
				pDst += 8 * PAGE_SIZE;
			}
		}
		return copied;
	}
}

UINT32 CTileGen::UpdateSnapshots(bool copyWhole)
{
	// Update all memory region snapshots
	UINT32 palACopied  = UpdateSnapshot(copyWhole, (UINT8*)pal[0],  (UINT8*)palRO[0],  0x020000, palDirty[0]);
	UINT32 palBCopied  = UpdateSnapshot(copyWhole, (UINT8*)pal[1],  (UINT8*)palRO[1],  0x020000, palDirty[1]);
	UINT32 vramCopied = UpdateSnapshot(copyWhole, (UINT8*)vram, (UINT8*)vramRO, 0x120000, vramDirty);
	memcpy(regsRO, regs, sizeof(regs)); // Always copy whole of regs buffer
	//printf("TileGen copied - palA:%4uK, palB:%4uK, vram:%4uK, regs:%uK\n", palACopied / 1024, palBCopied / 1024, vramCopied / 1024, sizeof(regs) / 1024);
	return palACopied + palBCopied + vramCopied + sizeof(regs);
}

void CTileGen::BeginFrame(void)
{
	// NOTE: Render2D->WriteVRAM(addr, data) is no longer being called for RAM addresses that are written
	// to and instead this class relies upon the fact that Render2D currently marks everything as dirty
	// with every frame.  If this were to change in the future then code to handle marking the correct
	// parts of the renderer as dirty would need to be added here.
	
	Render2D->BeginFrame();
}

void CTileGen::EndFrame(void)
{
	Render2D->EndFrame();
}

/******************************************************************************
 Emulation Functions
******************************************************************************/

UINT32 CTileGen::ReadRAM(unsigned addr)
{
	return *(UINT32 *) &vram[addr];
}

void CTileGen::WriteRAM(unsigned addr, UINT32 data)
{
	if (g_Config.gpuMultiThreaded)
		MARK_DIRTY(vramDirty, addr);
	*(UINT32 *) &vram[addr] = data;
		
	// Update palette if required
	if (addr >= 0x100000)
    {
		addr -= 0x100000;
		unsigned color = addr/4;	// color index
		
		// Same address in both palettes must be marked dirty
		if (g_Config.gpuMultiThreaded)
		{
			MARK_DIRTY(palDirty[0], addr);
			MARK_DIRTY(palDirty[1], addr);
		}
			
		// Both palettes will be modified simultaneously
        WritePalette(color, data);
    }
}

void CTileGen::InitPalette(void)
{
	for (int i = 0; i < 0x20000/4; i++)
	{
		WritePalette(i, *(UINT32 *) &vram[0x100000 + i*4]);
		if (g_Config.gpuMultiThreaded)
		{
			palRO[0][i] = pal[0][i];
			palRO[1][i] = pal[1][i];
		}
	}
}

static inline UINT32 AddColorOffset(UINT8 r, UINT8 g, UINT8 b, UINT8 a, UINT32 offsetReg)
{
	INT32	ir, ig, ib;
	
	/*
	 * Color offsets are signed but I'm not sure whether or not their range is 
	 * merely [-128,+127], which would mean adding to a 0 component would not 
	 * result full intensity (only +127 at most). Alternatively, the signed 
	 * value might have to be multiplied by 2. That is assumed here. In either 
	 * case, the signed addition should be saturated.
	 */
	ib = (INT32) (INT8)((offsetReg>>16)&0xFF);
	ig = (INT32) (INT8)((offsetReg>>8)&0xFF);
	ir = (INT32) (INT8)((offsetReg>>0)&0xFF);
	ib *= 2;
	ig *= 2;
	ir *= 2;
	
	// Add with saturation
	ib += (INT32) (UINT32) b;
	if (ib < 0)			ib = 0;
	else if (ib > 0xFF)	ib = 0xFF;
	ig += (INT32) (UINT32) g;
	if (ig < 0)			ig = 0;
	else if (ig > 0xFF)	ig = 0xFF;
	ir += (INT32) (UINT32) r;
	if (ir < 0)			ir = 0;
	else if (ir > 0xFF)	ir = 0xFF;
	
	// Construct the final 32-bit ABGR-format color
	r = (UINT8) ir;
	g = (UINT8) ig;
	b = (UINT8) ib;
	return ((UINT32)a<<24)|((UINT32)b<<16)|((UINT32)g<<8)|(UINT32)r;
}

void CTileGen::WritePalette(unsigned color, UINT32 data)
{
	UINT8		r, g, b, a;
	
	a = 0xFF * ((data>>15)&1); 	// decode the RGBA (make alpha 0xFF or 0x00)
    a = ~a;                  	// invert it (set on Model 3 means clear pixel)
	
	if ((data&0x8000))
    	r = g = b = 0;
	else
    {
    	b = (data>>7)&0xF8;
        g = (data>>2)&0xF8;
        r = (data<<3)&0xF8;
	}
	
	pal[0][color] = AddColorOffset(r, g, b, a, regs[0x40/4]);	// A/A'
	pal[1][color] = AddColorOffset(r, g, b, a, regs[0x44/4]);	// B/B'
}

void CTileGen::WriteRegister(unsigned reg, UINT32 data)
{
	reg &= 0xFF;
		
	switch (reg)
	{
	case 0x08:
	case 0x0C:
	case 0x20:
	case 0x60:
	case 0x64:
	case 0x68:
	case 0x6C:
		break;
	case 0x40:	// layer A/A' color offset
	case 0x44:	// layer B/B' color offset
		// We only have a mechanism to recompute both palettes simultaneously.
		// These regs are often written together in the same frame. To avoid
		// needlessly recomputing both palettes twice, we defer the operation.
		if (regs[reg/4] != data)	// only if changed
			recomputePalettes = true;
		break;
	case 0x10:	// IRQ acknowledge
		IRQ->Deassert(data&0xFF);
		break;
	default:
		DebugLog("Tile Generator reg %02X = %08X\n", reg, data);
		//printf("%02X = %08X\n", reg, data);
		break;
	}
	
	// Modify register
	regs[reg/4] = data;
}

void CTileGen::Reset(void)
{
	unsigned memSize = (g_Config.gpuMultiThreaded ? MEMORY_POOL_SIZE : MEM_POOL_SIZE_RW);
	memset(memoryPool, 0, memSize);
	memset(regs, 0, sizeof(regs));
	memset(regsRO, 0, sizeof(regsRO));
	
	InitPalette();
	recomputePalettes = false;

	DebugLog("Tile Generator reset\n");
}


/******************************************************************************
 Configuration, Initialization, and Shutdown
******************************************************************************/

void CTileGen::AttachRenderer(CRender2D *Render2DPtr)
{
	Render2D = Render2DPtr;

	// If multi-threaded, attach read-only snapshots to renderer instead of real ones
	if (g_Config.gpuMultiThreaded)
	{
		Render2D->AttachVRAM(vramRO);
		Render2D->AttachPalette((const UINT32 **)palRO);
		Render2D->AttachRegisters(regsRO);
	}
	else
	{
		Render2D->AttachVRAM(vram);
		Render2D->AttachPalette((const UINT32 **)pal);
		Render2D->AttachRegisters(regs);
	}

	DebugLog("Tile Generator attached a Render2D object\n");
}


bool CTileGen::Init(CIRQ *IRQObjectPtr)
{
	unsigned memSize   = (g_Config.gpuMultiThreaded ? MEMORY_POOL_SIZE : MEM_POOL_SIZE_RW);
	float	 memSizeMB = (float)memSize/(float)0x100000;
	
	// Allocate all memory for all TileGen RAM regions
	memoryPool = new(std::nothrow) UINT8[memSize];
	if (NULL == memoryPool)
		return ErrorLog("Insufficient memory for tile generator object (needs %1.1f MB).", memSizeMB);
	
	// Set up main pointers
	vram = (UINT8 *) &memoryPool[OFFSET_VRAM];
	pal[0] = (UINT32 *) &memoryPool[OFFSET_PAL_A];
	pal[1] = (UINT32 *) &memoryPool[OFFSET_PAL_B];

	// If multi-threaded, set up pointers for read-only snapshots and dirty page arrays too
	if (g_Config.gpuMultiThreaded)
	{
		vramRO = (UINT8 *) &memoryPool[OFFSET_VRAM_RO];
		palRO[0] = (UINT32 *) &memoryPool[OFFSET_PAL_RO_A];
		palRO[1] = (UINT32 *) &memoryPool[OFFSET_PAL_RO_B];
		vramDirty = (UINT8 *) &memoryPool[OFFSET_VRAM_DIRTY];
		palDirty[0] = (UINT8 *) &memoryPool[OFFSET_PAL_A_DIRTY];
		palDirty[1] = (UINT8 *) &memoryPool[OFFSET_PAL_B_DIRTY];
	}

	// Hook up the IRQ controller
	IRQ = IRQObjectPtr;
	
	DebugLog("Initialized Tile Generator (allocated %1.1f MB and connected to IRQ controller)\n", memSizeMB);
	return OKAY;
}

CTileGen::CTileGen(void)
{
	IRQ = NULL;
	memoryPool = NULL;
	DebugLog("Built Tile Generator\n");
}

CTileGen::~CTileGen(void)
{
	// Dump tile generator RAM
#if 0
	FILE *fp;
	fp = fopen("tileram", "wb");
	if (NULL != fp)
	{
		fwrite(memoryPool, sizeof(UINT8), 0x120000, fp);
		fclose(fp);
		printf("dumped %s\n", "tileram");
	}
	else
		printf("unable to dump %s\n", "tileram");
#endif
		
	IRQ = NULL;
	if (memoryPool != NULL)
	{
		delete [] memoryPool;
		memoryPool = NULL;
	}
	DebugLog("Destroyed Tile Generator\n");
}