ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merged common texture cache code from video plugins into VideoCommon. (DX11 native mipmaps currently broken, disabled) Hopefully everything else should still be working. 

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6288 8ced0084-cf51-0410-be5f-012b33b47a6e
This commit is contained in:
Jordan Woyak 2010-10-19 22:24:27 +00:00
parent 16ff21b0d5
commit cf05cca7a6
25 changed files with 1485 additions and 1922 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Source/Core/VideoCommon/Src/HiresTextures.cpp
View File

@ -95,7 +95,7 @@ void Shutdown()
textureMap.clear(); textureMap.clear();
} }
PC_TexFormat GetHiresTex(const char *fileName, int *pWidth, int *pHeight, int texformat, u8 *data) PC_TexFormat GetHiresTex(const char *fileName, unsigned int *pWidth, unsigned int *pHeight, int texformat, u8 *data)
{ {
std::string key(fileName); std::string key(fileName);

2
Source/Core/VideoCommon/Src/HiresTextures.h
View File

@ -26,7 +26,7 @@ namespace HiresTextures
{ {
void Init(const char *gameCode); void Init(const char *gameCode);
void Shutdown(); void Shutdown();
PC_TexFormat GetHiresTex(const char *fileName, int *pWidth, int *pHeight, int texformat, u8 *data); PC_TexFormat GetHiresTex(const char *fileName, unsigned int *pWidth, unsigned int *pHeight, int texformat, u8 *data);
}; };

1
Source/Core/VideoCommon/Src/SConscript
View File

@ -30,6 +30,7 @@ files = [
'TextureConversionShader.cpp', 'TextureConversionShader.cpp',
'ImageWrite.cpp', 'ImageWrite.cpp',
'VertexManagerBase.cpp', 'VertexManagerBase.cpp',
'TextureCacheBase.cpp',
'Statistics.cpp', 'Statistics.cpp',
'Fifo.cpp', 'Fifo.cpp',
'VideoState.cpp', 'VideoState.cpp',

659
Source/Core/VideoCommon/Src/TextureCacheBase.cpp Normal file
View File

@ -0,0 +1,659 @@
#include "MemoryUtil.h"
#include "VideoConfig.h"
#include "Statistics.h"
#include "HiresTextures.h"
#include "Render.h"
#include "FileUtil.h"
#include "Profiler.h"
#include "PluginSpecs.h"
#include "TextureCacheBase.h"
// ugly
extern int frameCount;
enum
{
TEMP_SIZE = (1024 * 1024 * 4),
TEXTURE_KILL_THRESHOLD = 200,
};
TextureCache *g_texture_cache;
u8 *TextureCache::temp;
TextureCache::TexCache TextureCache::textures;
// returns the exponent of the smallest power of two which is greater than val
unsigned int GetPow2(unsigned int val)
{
unsigned int ret = 0;
for (; val; val >>= 1)
++ret;
return ret;
}
TextureCache::TCacheEntryBase::~TCacheEntryBase()
{
if (0 == addr)
return;
if (!isRenderTarget && !g_ActiveConfig.bSafeTextureCache)
{
u32 *const ptr = (u32*)g_VideoInitialize.pGetMemoryPointer(addr);
if (ptr && *ptr == hash)
*ptr = oldpixel;
}
}
// TODO: uglyness
extern PLUGIN_GLOBALS *globals;
TextureCache::TextureCache()
{
temp = (u8*)AllocateMemoryPages(TEMP_SIZE);
TexDecoder_SetTexFmtOverlayOptions(g_ActiveConfig.bTexFmtOverlayEnable, g_ActiveConfig.bTexFmtOverlayCenter);
HiresTextures::Init(globals->unique_id);
}
void TextureCache::Invalidate(bool shutdown)
{
TexCache::iterator
iter = textures.begin(),
tcend = textures.end();
for (; iter != tcend; ++iter)
{
if (shutdown)
iter->second->addr = 0;
delete iter->second;
}
textures.clear();
HiresTextures::Shutdown();
}
TextureCache::~TextureCache()
{
Invalidate(true);
FreeMemoryPages(temp, TEMP_SIZE);
temp = NULL;
}
void TextureCache::Cleanup()
{
TexCache::iterator
iter = textures.begin(),
tcend = textures.end();
while (iter != tcend)
{
if (frameCount > TEXTURE_KILL_THRESHOLD + iter->second->frameCount)
{
delete iter->second;
textures.erase(iter++);
}
else
++iter;
}
}
void TextureCache::InvalidateRange(u32 start_address, u32 size)
{
TexCache::iterator
iter = textures.begin(),
tcend = textures.end();
while (iter != tcend)
{
const int rangePosition = iter->second->IntersectsMemoryRange(start_address, size);
if (0 == rangePosition)
{
delete iter->second;
textures.erase(iter++);
}
else
++iter;
}
}
void TextureCache::MakeRangeDynamic(u32 start_address, u32 size)
{
TexCache::iterator
iter = textures.begin(),
tcend = textures.end();
for (; iter != tcend; ++iter)
{
const int rangePosition = iter->second->IntersectsMemoryRange(start_address, size);
if (0 == rangePosition)
{
iter->second->hash = 0;
}
}
}
int TextureCache::TCacheEntryBase::IntersectsMemoryRange(u32 range_address, u32 range_size) const
{
if (addr + size_in_bytes < range_address)
return -1;
if (addr >= range_address + range_size)
return 1;
return 0;
}
void TextureCache::ClearRenderTargets()
{
TexCache::iterator
iter = textures.begin(),
tcend = textures.end();
for (; iter!=tcend; ++iter)
iter->second->isRenderTarget = false;
}
TextureCache::TCacheEntryBase* TextureCache::Load(unsigned int stage,
u32 address, unsigned int width, unsigned int height, int texformat,
unsigned int tlutaddr, int tlutfmt, bool UseNativeMips, unsigned int maxlevel)
{
// necessary?
if (0 == address)
return NULL;
u8* ptr = g_VideoInitialize.pGetMemoryPointer(address);
// TexelSizeInNibbles(format)*width*height/16;
const unsigned int bsw = TexDecoder_GetBlockWidthInTexels(texformat) - 1;
const unsigned int bsh = TexDecoder_GetBlockHeightInTexels(texformat) - 1;
unsigned int expandedWidth = (width + bsw) & (~bsw);
unsigned int expandedHeight = (height + bsh) & (~bsh);
u64 hash_value = 0;
u64 texHash = 0;
u32 texID = address;
u32 full_format = texformat;
const u32 texture_size = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, texformat);
const u32 palette_size = TexDecoder_GetPaletteSize(texformat);
bool texture_is_dynamic = false;
// someone who understands this var could rename it :p
const bool isC4_C8_C14X2 = (texformat == GX_TF_C4 || texformat == GX_TF_C8 || texformat == GX_TF_C14X2);
if (isC4_C8_C14X2)
full_format = texformat | (tlutfmt << 16);
// hires texture loading and texture dumping require accurate hashes
if (g_ActiveConfig.bSafeTextureCache || g_ActiveConfig.bHiresTextures || g_ActiveConfig.bDumpTextures)
{
texHash = GetHash64(ptr, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
if (isC4_C8_C14X2)
{
// WARNING! texID != address now => may break CopyRenderTargetToTexture (cf. TODO up)
// tlut size can be up to 32768B (GX_TF_C14X2) but Safer == Slower.
// This trick (to change the texID depending on the TLUT addr) is a trick to get around
// an issue with metroid prime's fonts, where it has multiple sets of fonts on top of
// each other stored in a single texture, and uses the palette to make different characters
// visible or invisible. Thus, unless we want to recreate the textures for every drawn character,
// we must make sure that texture with different tluts get different IDs.
const u64 tlutHash = GetHash64(texMem + tlutaddr, palette_size,
g_ActiveConfig.iSafeTextureCache_ColorSamples);
texHash ^= tlutHash;
if (g_ActiveConfig.bSafeTextureCache)
texID ^= ((u32)tlutHash) ^ (tlutHash >> 32);
}
if (g_ActiveConfig.bSafeTextureCache)
hash_value = texHash;
}
TCacheEntryBase *entry = textures[texID];
if (entry)
{
if (false == g_ActiveConfig.bSafeTextureCache)
{
if (entry->isRenderTarget || entry->isDynamic)
{
if (false == g_ActiveConfig.bCopyEFBToTexture)
{
hash_value = GetHash64(ptr, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
if (isC4_C8_C14X2)
{
hash_value ^= GetHash64(&texMem[tlutaddr], palette_size,
g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
}
else
{
hash_value = 0;
}
}
else
{
hash_value = *(u32*)ptr;
}
}
else if ((entry->isRenderTarget || entry->isDynamic) && g_ActiveConfig.bCopyEFBToTexture)
{
hash_value = 0;
}
// TODO: Is the mipLevels check needed?
if (((entry->isRenderTarget || entry->isDynamic) && hash_value == entry->hash && address == entry->addr)
|| ((address == entry->addr) && (hash_value == entry->hash) && full_format == entry->format && entry->mipLevels == maxlevel))
{
entry->isDynamic = false;
goto return_entry;
}
else
{
// Let's reload the new texture data into the same texture,
// instead of destroying it and having to create a new one.
// Might speed up movie playback very, very slightly.
texture_is_dynamic = (entry->isRenderTarget || entry->isDynamic) && !g_ActiveConfig.bCopyEFBToTexture;
// TODO: Is the mipLevels check needed?
if (!entry->isRenderTarget &&
((!entry->isDynamic && width == entry->w && height == entry->h && full_format == entry->format && entry->mipLevels == maxlevel)
|| (entry->isDynamic && entry->w == width && entry->h == height)))
{
// reuse the texture
}
else
{
// delete the texture and make a new one
delete entry;
entry = NULL;
}
}
}
const unsigned int nativeW = width;
const unsigned int nativeH = height;
PC_TexFormat pcfmt = PC_TEX_FMT_NONE;
if (g_ActiveConfig.bHiresTextures)
{
// Load Custom textures
char texPathTemp[MAX_PATH];
unsigned int newWidth = width;
unsigned int newHeight = height;
sprintf(texPathTemp, "%s_%08x_%i", globals->unique_id, texHash, texformat);
pcfmt = HiresTextures::GetHiresTex(texPathTemp, &newWidth, &newHeight, texformat, temp);
if (pcfmt != PC_TEX_FMT_NONE)
{
expandedWidth = width = newWidth;
expandedHeight = height = newHeight;
// TODO: shouldn't generating mips be forced on now?
// native mips with a custom texture wouldn't make sense
}
}
if (pcfmt == PC_TEX_FMT_NONE)
pcfmt = TexDecoder_Decode(temp, ptr, expandedWidth,
expandedHeight, texformat, tlutaddr, tlutfmt, !g_texture_cache->isOGL());
const bool isPow2 = !((width & (width - 1)) || (height & (height - 1)));
unsigned int texLevels = (isPow2 && UseNativeMips && maxlevel) ?
GetPow2(std::max(width, height)) : !isPow2;
if ((texLevels > (maxlevel + 1)) && maxlevel)
texLevels = maxlevel + 1;
// create the entry/texture
if (NULL == entry)
{
textures[texID] = entry = g_texture_cache->CreateTexture(width, height, expandedWidth, texLevels, pcfmt);
entry->addr = address;
entry->format = full_format;
entry->mipLevels = maxlevel;
entry->size_in_bytes = texture_size;
entry->scaledW = entry->w = width;
entry->scaledH = entry->h = height;
entry->nativeH = nativeH;
entry->nativeW = nativeW;
entry->isRenderTarget = false;
entry->isNonPow2 = false;
entry->isDynamic = texture_is_dynamic;
entry->oldpixel = *(u32*)ptr;
if (g_ActiveConfig.bSafeTextureCache || entry->isDynamic)
entry->hash = hash_value;
else
// don't like rand() here
entry->hash = *(u32*)ptr = (u32)(((double)rand() / RAND_MAX) * 0xFFFFFFFF);
}
// load texture
entry->Load(width, height, expandedWidth, 0);
// load mips
if (texLevels > 1 && pcfmt != PC_TEX_FMT_NONE)
{
const unsigned int bsdepth = TexDecoder_GetTexelSizeInNibbles(texformat);
unsigned int level = 1;
unsigned int mipWidth = (width + 1) >> 1;
unsigned int mipHeight = (height + 1) >> 1;
ptr += texture_size;
while ((mipHeight || mipWidth) && (level < texLevels))
{
const unsigned int currentWidth = (mipWidth > 0) ? mipWidth : 1;
const unsigned int currentHeight = (mipHeight > 0) ? mipHeight : 1;
expandedWidth = (currentWidth + bsw) & (~bsw);
expandedHeight = (currentHeight + bsh) & (~bsh);
TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, texformat, tlutaddr, tlutfmt, !g_texture_cache->isOGL());
entry->Load(currentWidth, currentHeight, expandedWidth, level);
ptr += ((std::max(mipWidth, bsw) * std::max(mipHeight, bsh) * bsdepth) >> 1);
mipWidth >>= 1;
mipHeight >>= 1;
++level;
}
}
// TODO: won't this cause loaded hires textures to be dumped as well?
// dump texture to file
if (g_ActiveConfig.bDumpTextures)
{
char szTemp[MAX_PATH];
char szDir[MAX_PATH];
// make sure that the directory exists
sprintf(szDir, "%s%s", File::GetUserPath(D_DUMPTEXTURES_IDX), globals->unique_id);
if (false == File::Exists(szDir) || false == File::IsDirectory(szDir))
File::CreateDir(szDir);
sprintf(szTemp, "%s/%s_%08x_%i.png", szDir, globals->unique_id, texHash, texformat);
if (false == File::Exists(szTemp))
entry->Save(szTemp);
}
INCSTAT(stats.numTexturesCreated);
SETSTAT(stats.numTexturesAlive, textures.size());
return_entry:
entry->frameCount = frameCount;
entry->Bind(stage);
return entry;
}
void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer,
bool bIsIntensityFmt, u32 copyfmt, bool bScaleByHalf, const EFBRectangle &source_rect)
{
DVSTARTPROFILE();
float colmat[20] = {};
// last four floats of colmat for fConstAdd
float *const fConstAdd = colmat + 16;
unsigned int cbufid = -1;
if (bFromZBuffer)
{
// TODO: these values differ slightly from the DX9/11 values,
// do they need to? or can this be removed
if (g_texture_cache->isOGL())
{
switch(copyfmt)
{
case 0: // Z4
case 1: // Z8
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1;
break;
case 3: // Z16 //?
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1;
break;
case 11: // Z16 (reverse order)
colmat[2] = colmat[6] = colmat[10] = colmat[13] = 1;
break;
case 6: // Z24X8
colmat[2] = colmat[5] = colmat[8] = colmat[15] = 1;
break;
case 9: // Z8M
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1;
break;
case 10: // Z8L
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
break;
case 12: // Z16L
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy zbuf format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
else
{
switch (copyfmt)
{
case 0: // Z4
case 1: // Z8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1.0f;
cbufid = 12;
break;
case 3: // Z16 //?
colmat[1] = colmat[5] = colmat[9] = colmat[12] = 1.0f;
cbufid = 13;
break;
case 11: // Z16 (reverse order)
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
cbufid = 14;
break;
case 6: // Z24X8
colmat[0] = colmat[5] = colmat[10] = 1.0f;
cbufid = 15;
break;
case 9: // Z8M
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
cbufid = 16;
break;
case 10: // Z8L
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
cbufid = 17;
break;
case 12: // Z16L
colmat[2] = colmat[6] = colmat[10] = colmat[13] = 1.0f;
cbufid = 18;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy zbuf format: 0x%x", copyfmt);
colmat[2] = colmat[5] = colmat[8] = 1.0f;
cbufid = 19;
break;
}
}
}
else if (bIsIntensityFmt)
{
fConstAdd[0] = fConstAdd[1] = fConstAdd[2] = 16.0f/255.0f;
switch (copyfmt)
{
case 0: // I4
case 1: // I8
case 2: // IA4
case 3: // IA8
// TODO - verify these coefficients
colmat[0] = 0.257f; colmat[1] = 0.504f; colmat[2] = 0.098f;
colmat[4] = 0.257f; colmat[5] = 0.504f; colmat[6] = 0.098f;
colmat[8] = 0.257f; colmat[9] = 0.504f; colmat[10] = 0.098f;
if (copyfmt < 2)
{
fConstAdd[3] = 16.0f / 255.0f;
colmat[12] = 0.257f; colmat[13] = 0.504f; colmat[14] = 0.098f;
cbufid = 0;
}
else// alpha
{
colmat[15] = 1;
cbufid = 1;
}
break;
default:
ERROR_LOG(VIDEO, "Unknown copy intensity format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
else
{
switch (copyfmt)
{
case 0: // R4
case 8: // R8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
cbufid = 2;
break;
case 2: // RA4
case 3: // RA8
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1;
cbufid = 3;
break;
case 7: // A8
colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1;
cbufid = 4;
break;
case 9: // G8
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1;
cbufid = 5;
break;
case 10: // B8
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1;
cbufid = 6;
break;
case 11: // RG8
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1;
cbufid = 7;
break;
case 12: // GB8
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1;
cbufid = 8;
break;
case 4: // RGB565
colmat[0] = colmat[5] = colmat[10] = 1;
fConstAdd[3] = 1; // set alpha to 1
cbufid = 9;
break;
case 5: // RGB5A3
case 6: // RGBA8
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
cbufid = 10;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy color format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
cbufid = 11;
break;
}
}
const unsigned int tex_w = (abs(source_rect.GetWidth()) >> (int)bScaleByHalf);
const unsigned int tex_h = (abs(source_rect.GetHeight()) >> (int)bScaleByHalf);
const float xScale = Renderer::GetTargetScaleX();
const float yScale = Renderer::GetTargetScaleY();
unsigned int scaled_tex_w = g_ActiveConfig.bCopyEFBScaled ? (int)(tex_w * xScale) : tex_w;
unsigned int scaled_tex_h = g_ActiveConfig.bCopyEFBScaled ? (int)(tex_h * yScale) : tex_h;
bool texture_is_dynamic = false;
TCacheEntryBase *entry = textures[address];
if (entry)
{
if ((entry->isRenderTarget && entry->scaledW == scaled_tex_w && entry->scaledH == scaled_tex_h)
|| (entry->isDynamic && entry->w == tex_w && entry->h == tex_h))
{
texture_is_dynamic = entry->isDynamic;
}
else
{
// remove it and recreate it as a render target
delete entry;
entry = NULL;
}
}
if (texture_is_dynamic)
{
scaled_tex_w = tex_w;
scaled_tex_h = tex_h;
}
if (NULL == entry)
{
// create the texture
textures[address] = entry = g_texture_cache->CreateRenderTargetTexture(scaled_tex_w, scaled_tex_h);
entry->addr = address;
entry->hash = 0;
entry->w = entry->nativeW = tex_w;
entry->h = entry->nativeH = tex_h;
entry->scaledW = scaled_tex_w;
entry->scaledH = scaled_tex_h;
entry->format = copyfmt;
entry->mipLevels = 0;
entry->isRenderTarget = true;
entry->isNonPow2 = true;
entry->isDynamic = false;
}
entry->frameCount = frameCount;
Renderer::ResetAPIState(); // reset any game specific settings
entry->FromRenderTarget(bFromZBuffer, bScaleByHalf, cbufid, colmat, source_rect, bIsIntensityFmt, copyfmt);
Renderer::RestoreAPIState();
}

99
Source/Core/VideoCommon/Src/TextureCacheBase.h Normal file
View File

@ -0,0 +1,99 @@
#ifndef _TEXTURECACHEBASE_H
#define _TEXTURECACHEBASE_H
#include <map>
//#include "VideoCommon.h"
#include "TextureDecoder.h"
#include "BPMemory.h"
#include "CommonTypes.h"
class TextureCache
{
public:
struct TCacheEntryBase
{
// TODO: organize
u32 addr;
u32 size_in_bytes;
u64 hash;
//u32 paletteHash;
u32 oldpixel;
u32 format;
int frameCount;
unsigned int w, h, mipLevels;
// TODO: it looks like scaledW/H can be removed and w/h can be used in their place
unsigned int scaledW, scaledH, nativeW, nativeH;
bool isRenderTarget;
bool isDynamic; // mofified from cpu
bool isNonPow2; // doesn't seem to be used anywhere
//TCacheEntryBase()
//{
// // TODO: remove these
// isRenderTarget = 0;
// hash = 0;
// //paletteHash = 0;
// oldpixel = 0;
// addr = 0;
// size_in_bytes = 0;
// frameCount = 0;
// isNonPow2 = true;
// w = 0;
// h = 0;
// scaledW = 0;
// scaledH = 0;
//}
virtual ~TCacheEntryBase();
virtual void Bind(unsigned int stage) = 0;
virtual bool Save(const char filename[]) = 0;
virtual void Load(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int level) = 0;
virtual void FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
unsigned int cbufid, const float *colmat, const EFBRectangle &source_rect,
bool bIsIntensityFmt, u32 copyfmt) = 0;
int IntersectsMemoryRange(u32 range_address, u32 range_size) const;
};
virtual ~TextureCache(); // needs virtual for DX11 dtor
static void Cleanup();
static void Invalidate(bool shutdown);
static void InvalidateRange(u32 start_address, u32 size);
static void MakeRangeDynamic(u32 start_address, u32 size);
static void ClearRenderTargets(); // currently only used by OGL
virtual TCacheEntryBase* CreateTexture(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int tex_levels, PC_TexFormat pcfmt) = 0;
virtual TCacheEntryBase* CreateRenderTargetTexture(unsigned int scaled_tex_w, unsigned int scaled_tex_h) = 0;
static TCacheEntryBase* Load(unsigned int stage, u32 address, unsigned int width, unsigned int height,
int format, unsigned int tlutaddr, int tlutfmt, bool UseNativeMips, unsigned int maxlevel);
static void CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt,
u32 copyfmt, bool bScaleByHalf, const EFBRectangle &source_rect);
protected:
TextureCache();
static u8 *temp;
private:
typedef std::map<u32, TCacheEntryBase*> TexCache;
static TexCache textures;
virtual bool isOGL() { return false; } // Hacks for TextureDecode_real support
};
extern TextureCache *g_texture_cache;
#endif

158
Source/Core/VideoCommon/Src/VertexManagerBase.cpp
View File

@ -4,6 +4,12 @@
#include "Statistics.h" #include "Statistics.h"
#include "OpcodeDecoding.h" #include "OpcodeDecoding.h"
#include "IndexGenerator.h" #include "IndexGenerator.h"
#include "VertexShaderManager.h"
#include "PixelShaderManager.h"
#include "NativeVertexFormat.h"
#include "TextureCacheBase.h"
#include "Render.h"
#include "Profiler.h"
#include "VertexManagerBase.h" #include "VertexManagerBase.h"
@ -151,8 +157,158 @@ void VertexManager::AddVertices(int primitive, int numVertices)
AddIndices(primitive, numVertices); AddIndices(primitive, numVertices);
} }
// TODO: merge this func, (need to merge TextureCache)
void VertexManager::Flush() void VertexManager::Flush()
{ {
g_vertex_manager->vFlush(); g_vertex_manager->vFlush();
} }
// TODO: need to merge more stuff into VideoCommon to use this
#if (0)
void VertexManager::Flush()
{
if (LocalVBuffer == s_pCurBufferPointer || Flushed)
return;
Flushed = true;
VideoFifo_CheckEFBAccess();
#if defined(_DEBUG) || defined(DEBUGFAST)
PRIM_LOG("frame%d:\n texgen=%d, numchan=%d, dualtex=%d, ztex=%d, cole=%d, alpe=%d, ze=%d", g_ActiveConfig.iSaveTargetId, xfregs.numTexGens,
xfregs.nNumChans, (int)xfregs.bEnableDualTexTransform, bpmem.ztex2.op,
bpmem.blendmode.colorupdate, bpmem.blendmode.alphaupdate, bpmem.zmode.updateenable);
for (int i = 0; i < xfregs.nNumChans; ++i)
{
LitChannel* ch = &xfregs.colChans[i].color;
PRIM_LOG("colchan%d: matsrc=%d, light=0x%x, ambsrc=%d, diffunc=%d, attfunc=%d", i, ch->matsource, ch->GetFullLightMask(), ch->ambsource, ch->diffusefunc, ch->attnfunc);
ch = &xfregs.colChans[i].alpha;
PRIM_LOG("alpchan%d: matsrc=%d, light=0x%x, ambsrc=%d, diffunc=%d, attfunc=%d", i, ch->matsource, ch->GetFullLightMask(), ch->ambsource, ch->diffusefunc, ch->attnfunc);
}
for (int i = 0; i < xfregs.numTexGens; ++i)
{
TexMtxInfo tinfo = xfregs.texcoords[i].texmtxinfo;
if (tinfo.texgentype != XF_TEXGEN_EMBOSS_MAP) tinfo.hex &= 0x7ff;
if (tinfo.texgentype != XF_TEXGEN_REGULAR) tinfo.projection = 0;
PRIM_LOG("txgen%d: proj=%d, input=%d, gentype=%d, srcrow=%d, embsrc=%d, emblght=%d, postmtx=%d, postnorm=%d",
i, tinfo.projection, tinfo.inputform, tinfo.texgentype, tinfo.sourcerow, tinfo.embosssourceshift, tinfo.embosslightshift,
xfregs.texcoords[i].postmtxinfo.index, xfregs.texcoords[i].postmtxinfo.normalize);
}
PRIM_LOG("pixel: tev=%d, ind=%d, texgen=%d, dstalpha=%d, alphafunc=0x%x", bpmem.genMode.numtevstages+1, bpmem.genMode.numindstages,
bpmem.genMode.numtexgens, (u32)bpmem.dstalpha.enable, (bpmem.alphaFunc.hex>>16)&0xff);
#endif
DVSTARTPROFILE();
// set the textures
DVSTARTSUBPROFILE("VertexManager::Flush:textures");
u32 usedtextures = 0;
for (u32 i = 0; i < (u32)bpmem.genMode.numtevstages + 1; ++i)
if (bpmem.tevorders[i / 2].getEnable(i & 1))
usedtextures |= 1 << bpmem.tevorders[i/2].getTexMap(i & 1);
if (bpmem.genMode.numindstages > 0)
for (u32 i = 0; i < (u32)bpmem.genMode.numtevstages + 1; ++i)
if (bpmem.tevind[i].IsActive() && bpmem.tevind[i].bt < bpmem.genMode.numindstages)
usedtextures |= 1 << bpmem.tevindref.getTexMap(bpmem.tevind[i].bt);
for (u32 i = 0; i < 8; ++i)
{
if (usedtextures & (1 << i))
{
// TODO:
//glActiveTexture(GL_TEXTURE0 + i);
Renderer::SetSamplerState(i & 3, i >> 2);
FourTexUnits &tex = bpmem.tex[i >> 2];
TCacheEntry::TCacheEntryBase* tentry = TextureCache::Load(i,
(tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,
tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1,
tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9,
tex.texTlut[i&3].tlut_format,
(tex.texMode0[i&3].min_filter & 3) && (tex.texMode0[i&3].min_filter != 8) && g_ActiveConfig.bUseNativeMips,
(tex.texMode1[i&3].max_lod >> 4));
if (tentry)
{
// 0s are probably for no manual wrapping needed.
PixelShaderManager::SetTexDims(i, tentry->nativeW, tentry->nativeH, 0, 0);
// TODO:
//if (g_ActiveConfig.iLog & CONF_SAVETEXTURES)
//{
// // save the textures
// char strfile[255];
// sprintf(strfile, "%stex%.3d_%d.tga", File::GetUserPath(D_DUMPFRAMES_IDX), g_Config.iSaveTargetId, i);
// SaveTexture(strfile, GL_TEXTURE_2D, tentry->texture, tentry->w, tentry->h);
//}
}
else
ERROR_LOG(VIDEO, "error loading texture");
}
}
// set global constants
VertexShaderManager::SetConstants();
PixelShaderManager::SetConstants();
// finally bind
if (false == PixelShaderCache::SetShader(false, g_nativeVertexFmt->m_components))
goto shader_fail;
if (false == VertexShaderCache::SetShader(g_nativeVertexFmt->m_components))
goto shader_fail;
const int stride = g_nativeVertexFmt->GetVertexStride();
//if (g_nativeVertexFmt)
g_nativeVertexFmt->SetupVertexPointers();
g_vertex_manager->Draw(stride, false);
// run through vertex groups again to set alpha
if (false == g_ActiveConfig.bDstAlphaPass && bpmem.dstalpha.enable && bpmem.blendmode.alphaupdate)
{
if (false == PixelShaderCache::SetShader(true, g_nativeVertexFmt->m_components))
goto shader_fail;
g_vertex_manager->Draw(stride, true);
}
// TODO:
//IndexGenerator::Start(TIBuffer, LIBuffer, PIBuffer);
#if defined(_DEBUG) || defined(DEBUGFAST)
if (g_ActiveConfig.iLog & CONF_SAVESHADERS)
{
// save the shaders
char strfile[255];
sprintf(strfile, "%sps%.3d.txt", File::GetUserPath(D_DUMPFRAMES_IDX), g_ActiveConfig.iSaveTargetId);
std::ofstream fps(strfile);
fps << ps->strprog.c_str();
sprintf(strfile, "%svs%.3d.txt", File::GetUserPath(D_DUMPFRAMES_IDX), g_ActiveConfig.iSaveTargetId);
std::ofstream fvs(strfile);
fvs << vs->strprog.c_str();
}
if (g_ActiveConfig.iLog & CONF_SAVETARGETS)
{
char str[128];
sprintf(str, "%starg%.3d.tga", File::GetUserPath(D_DUMPFRAMES_IDX), g_ActiveConfig.iSaveTargetId);
TargetRectangle tr;
tr.left = 0;
tr.right = Renderer::GetTargetWidth();
tr.top = 0;
tr.bottom = Renderer::GetTargetHeight();
Renderer::SaveRenderTarget(str, tr);
}
#endif
++g_Config.iSaveTargetId;
shader_fail:
ResetBuffer();
}
#endif

4
Source/Core/VideoCommon/Src/VertexManagerBase.h
View File

@ -48,8 +48,10 @@ protected:
private: private:
static void AddIndices(int primitive, int numVertices); static void AddIndices(int primitive, int numVertices);
// temporary //virtual void Draw(u32 stride, bool alphapass) = 0;
// temp
virtual void vFlush() = 0; virtual void vFlush() = 0;
}; };
extern VertexManager *g_vertex_manager; extern VertexManager *g_vertex_manager;

8
Source/Core/VideoCommon/VideoCommon.vcproj
View File

@ -730,6 +730,14 @@
<Filter <Filter
Name="Base" Name="Base"
> >
<File
RelativePath=".\Src\TextureCacheBase.cpp"
>
</File>
<File
RelativePath=".\Src\TextureCacheBase.h"
>
</File>
<File <File
RelativePath=".\Src\VertexManagerBase.cpp" RelativePath=".\Src\VertexManagerBase.cpp"
> >

2
Source/Plugins/Plugin_VideoDX11/Src/BPFunctions.cpp
View File

@ -80,7 +80,7 @@ void CopyEFB(const BPCmd &bp, const EFBRectangle &rc, const u32 &address, const
// bpmem.zcontrol.pixel_format to PIXELFMT_Z24 is when the game wants to copy from ZBuffer (Zbuffer uses 24-bit Format) // bpmem.zcontrol.pixel_format to PIXELFMT_Z24 is when the game wants to copy from ZBuffer (Zbuffer uses 24-bit Format)
if (!g_ActiveConfig.bEFBCopyDisable) if (!g_ActiveConfig.bEFBCopyDisable)
{ {
TextureCache::CopyRenderTargetToTexture(address, fromZBuffer, isIntensityFmt, copyfmt, scaleByHalf, rc); TextureCache::CopyRenderTargetToTexture(address, fromZBuffer, isIntensityFmt, copyfmt, !!scaleByHalf, rc);
} }
} }

676
Source/Plugins/Plugin_VideoDX11/Src/TextureCache.cpp
View File

@ -40,39 +40,143 @@
#include "TextureCache.h" #include "TextureCache.h"
#include "HiresTextures.h" #include "HiresTextures.h"
namespace DX11
{
ID3D11BlendState* efbcopyblendstate = NULL; ID3D11BlendState* efbcopyblendstate = NULL;
ID3D11RasterizerState* efbcopyraststate = NULL; ID3D11RasterizerState* efbcopyraststate = NULL;
ID3D11DepthStencilState* efbcopydepthstate = NULL; ID3D11DepthStencilState* efbcopydepthstate = NULL;
ID3D11Buffer* efbcopycbuf[20] = { NULL }; ID3D11Buffer* efbcopycbuf[20] = {};
u8* TextureCache::temp = NULL; TextureCache::TCacheEntry::~TCacheEntry()
TextureCache::TexCache TextureCache::textures;
extern int frameCount;
#define TEMP_SIZE (2048*2048*4)
#define TEXTURE_KILL_THRESHOLD 200
void TextureCache::TCacheEntry::Destroy(bool shutdown)
{ {
SAFE_RELEASE(texture); texture->Release();
if (!isRenderTarget && !shutdown && !g_ActiveConfig.bSafeTextureCache)
{
u32 *ptr = (u32*)g_VideoInitialize.pGetMemoryPointer(addr);
if (ptr && *ptr == hash)
*ptr = oldpixel;
}
} }
void TextureCache::Init() void TextureCache::TCacheEntry::Bind(unsigned int stage)
{
D3D::gfxstate->SetShaderResource(stage, texture->GetSRV());
}
bool TextureCache::TCacheEntry::Save(const char filename[])
{
return SUCCEEDED(PD3DX11SaveTextureToFileA(D3D::context, texture->GetTex(), D3DX11_IFF_PNG, filename));
}
void TextureCache::TCacheEntry::Load(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int level)
{
// TODO: remove hax
if (level != 0) return;
D3D::ReplaceRGBATexture2D(texture->GetTex(), TextureCache::temp, width, height, expanded_width, level, usage);
}
TextureCache::TCacheEntryBase* TextureCache::CreateTexture(unsigned int width,
unsigned int height, unsigned int expanded_width,
unsigned int tex_levels, PC_TexFormat pcfmt)
{
D3D11_USAGE usage = D3D11_USAGE_DEFAULT;
D3D11_CPU_ACCESS_FLAG cpu_access = (D3D11_CPU_ACCESS_FLAG)0;
D3D11_SUBRESOURCE_DATA srdata, *data = NULL;
// TODO: remove hax
tex_levels = 1;
if (1 == tex_levels)
{
usage = D3D11_USAGE_DYNAMIC;
cpu_access = D3D11_CPU_ACCESS_WRITE;
srdata.pSysMem = TextureCache::temp;
srdata.SysMemPitch = 4 * expanded_width;
// testing
//data = &srdata;
}
const D3D11_TEXTURE2D_DESC texdesc = CD3D11_TEXTURE2D_DESC(DXGI_FORMAT_R8G8B8A8_UNORM,
width, height, 1, tex_levels, D3D11_BIND_SHADER_RESOURCE, usage, cpu_access);
ID3D11Texture2D *pTexture;
const HRESULT hr = D3D::device->CreateTexture2D(&texdesc, data, &pTexture);
CHECK(SUCCEEDED(hr), "Create texture of the TextureCache");
TCacheEntry* const entry = new TCacheEntry(new D3DTexture2D(pTexture, D3D11_BIND_SHADER_RESOURCE));
entry->usage = usage;
// TODO: better debug names
D3D::SetDebugObjectName((ID3D11DeviceChild*)entry->texture->GetTex(), "a texture of the TextureCache");
D3D::SetDebugObjectName((ID3D11DeviceChild*)entry->texture->GetSRV(), "shader resource view of a texture of the TextureCache");
SAFE_RELEASE(pTexture);
if (0 == tex_levels)
PD3DX11FilterTexture(D3D::context, entry->texture->GetTex(), 0, D3DX11_DEFAULT);
return entry;
}
void TextureCache::TCacheEntry::FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
unsigned int cbufid, const float colmat[], const EFBRectangle &source_rect,
bool bIsIntensityFmt, u32 copyfmt)
{
// stretch picture with increased internal resolution
const D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, (float)scaledW, (float)scaledH);
D3D::context->RSSetViewports(1, &vp);
// set transformation
if (NULL == efbcopycbuf[cbufid])
{
const D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(20 * sizeof(float), D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DEFAULT);
D3D11_SUBRESOURCE_DATA data;
data.pSysMem = colmat;
HRESULT hr = D3D::device->CreateBuffer(&cbdesc, &data, &efbcopycbuf[cbufid]);
CHECK(SUCCEEDED(hr), "Create efb copy constant buffer %d", cbufid);
D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopycbuf[cbufid], "a constant buffer used in TextureCache::CopyRenderTargetToTexture");
}
D3D::context->PSSetConstantBuffers(0, 1, &efbcopycbuf[cbufid]);
const TargetRectangle targetSource = Renderer::ConvertEFBRectangle(source_rect);
// TODO: try targetSource.asRECT();
const D3D11_RECT sourcerect = CD3D11_RECT(targetSource.left, targetSource.top, targetSource.right, targetSource.bottom);
// Use linear filtering if (bScaleByHalf), use point filtering otherwise
if (bScaleByHalf)
D3D::SetLinearCopySampler();
else
D3D::SetPointCopySampler();
D3D::stateman->PushBlendState(efbcopyblendstate);
D3D::stateman->PushRasterizerState(efbcopyraststate);
D3D::stateman->PushDepthState(efbcopydepthstate);
D3D::context->OMSetRenderTargets(1, &texture->GetRTV(), NULL);
D3D::drawShadedTexQuad(
(bFromZBuffer) ? g_framebufferManager.GetEFBDepthTexture()->GetSRV() : g_framebufferManager.GetEFBColorTexture()->GetSRV(),
&sourcerect, Renderer::GetFullTargetWidth(), Renderer::GetFullTargetHeight(),
(bFromZBuffer) ? PixelShaderCache::GetDepthMatrixProgram() : PixelShaderCache::GetColorMatrixProgram(),
VertexShaderCache::GetSimpleVertexShader(), VertexShaderCache::GetSimpleInputLayout());
D3D::context->OMSetRenderTargets(1, &g_framebufferManager.GetEFBColorTexture()->GetRTV(), g_framebufferManager.GetEFBDepthTexture()->GetDSV());
D3D::stateman->PopBlendState();
D3D::stateman->PopDepthState();
D3D::stateman->PopRasterizerState();
}
TextureCache::TCacheEntryBase* TextureCache::CreateRenderTargetTexture(
unsigned int scaled_tex_w, unsigned int scaled_tex_h)
{
return new TCacheEntry(D3DTexture2D::Create(scaled_tex_w, scaled_tex_h,
(D3D11_BIND_FLAG)((int)D3D11_BIND_RENDER_TARGET | (int)D3D11_BIND_SHADER_RESOURCE),
D3D11_USAGE_DEFAULT, DXGI_FORMAT_R8G8B8A8_UNORM));
}
TextureCache::TextureCache()
{ {
HRESULT hr; HRESULT hr;
temp = (u8*)AllocateMemoryPages(TEMP_SIZE);
TexDecoder_SetTexFmtOverlayOptions(g_ActiveConfig.bTexFmtOverlayEnable, g_ActiveConfig.bTexFmtOverlayCenter);
HiresTextures::Init(globals->unique_id);
D3D11_BLEND_DESC blenddesc; D3D11_BLEND_DESC blenddesc;
blenddesc.AlphaToCoverageEnable = FALSE; blenddesc.AlphaToCoverageEnable = FALSE;
blenddesc.IndependentBlendEnable = FALSE; blenddesc.IndependentBlendEnable = FALSE;
@ -85,8 +189,8 @@ void TextureCache::Init()
blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_ZERO; blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_ZERO;
blenddesc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD; blenddesc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
hr = D3D::device->CreateBlendState(&blenddesc, &efbcopyblendstate); hr = D3D::device->CreateBlendState(&blenddesc, &efbcopyblendstate);
CHECK(hr==S_OK, "Create blend state for TextureCache::CopyRenderTargetToTexture"); CHECK(hr==S_OK, "Create blend state for CopyRenderTargetToTexture");
D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopyblendstate, "blend state used in TextureCache::CopyRenderTargetToTexture"); D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopyblendstate, "blend state used in CopyRenderTargetToTexture");
D3D11_DEPTH_STENCIL_DESC depthdesc; D3D11_DEPTH_STENCIL_DESC depthdesc;
depthdesc.DepthEnable = FALSE; depthdesc.DepthEnable = FALSE;
@ -96,8 +200,8 @@ void TextureCache::Init()
depthdesc.StencilReadMask = D3D11_DEFAULT_STENCIL_READ_MASK; depthdesc.StencilReadMask = D3D11_DEFAULT_STENCIL_READ_MASK;
depthdesc.StencilWriteMask = D3D11_DEFAULT_STENCIL_WRITE_MASK; depthdesc.StencilWriteMask = D3D11_DEFAULT_STENCIL_WRITE_MASK;
hr = D3D::device->CreateDepthStencilState(&depthdesc, &efbcopydepthstate); hr = D3D::device->CreateDepthStencilState(&depthdesc, &efbcopydepthstate);
CHECK(hr==S_OK, "Create depth state for TextureCache::CopyRenderTargetToTexture"); CHECK(hr==S_OK, "Create depth state for CopyRenderTargetToTexture");
D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopydepthstate, "depth stencil state used in TextureCache::CopyRenderTargetToTexture"); D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopydepthstate, "depth stencil state used in CopyRenderTargetToTexture");
D3D11_RASTERIZER_DESC rastdesc; D3D11_RASTERIZER_DESC rastdesc;
rastdesc.CullMode = D3D11_CULL_NONE; rastdesc.CullMode = D3D11_CULL_NONE;
@ -111,528 +215,18 @@ void TextureCache::Init()
rastdesc.MultisampleEnable = false; rastdesc.MultisampleEnable = false;
rastdesc.AntialiasedLineEnable = false; rastdesc.AntialiasedLineEnable = false;
hr = D3D::device->CreateRasterizerState(&rastdesc, &efbcopyraststate); hr = D3D::device->CreateRasterizerState(&rastdesc, &efbcopyraststate);
CHECK(hr==S_OK, "Create rasterizer state for TextureCache::CopyRenderTargetToTexture"); CHECK(hr==S_OK, "Create rasterizer state for CopyRenderTargetToTexture");
D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopyraststate, "rasterizer state used in TextureCache::CopyRenderTargetToTexture"); D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopyraststate, "rasterizer state used in CopyRenderTargetToTexture");
} }
void TextureCache::Invalidate(bool shutdown) TextureCache::~TextureCache()
{ {
for (TexCache::iterator iter = textures.begin(); iter != textures.end(); ++iter)
iter->second.Destroy(shutdown);
textures.clear();
HiresTextures::Shutdown();
}
void TextureCache::Shutdown()
{
Invalidate(true);
FreeMemoryPages(temp, TEMP_SIZE);
temp = NULL;
SAFE_RELEASE(efbcopyblendstate); SAFE_RELEASE(efbcopyblendstate);
SAFE_RELEASE(efbcopyraststate); SAFE_RELEASE(efbcopyraststate);
SAFE_RELEASE(efbcopydepthstate); SAFE_RELEASE(efbcopydepthstate);
for (unsigned int k = 0; k < 20;k++) for (unsigned int k = 0; k < 20; ++k)
SAFE_RELEASE(efbcopycbuf[k]); SAFE_RELEASE(efbcopycbuf[k]);
} }
void TextureCache::Cleanup()
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
if (frameCount > TEXTURE_KILL_THRESHOLD + iter->second.frameCount)
{
iter->second.Destroy(false);
iter = textures.erase(iter);
}
else
{
++iter;
}
}
}
void TextureCache::InvalidateRange(u32 start_address, u32 size)
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
if (iter->second.IntersectsMemoryRange(start_address, size))
{
iter->second.Destroy(false);
textures.erase(iter++);
}
else
{
++iter;
}
}
}
bool TextureCache::TCacheEntry::IntersectsMemoryRange(u32 range_address, u32 range_size)
{
if (addr + size_in_bytes < range_address) return false;
if (addr >= range_address + range_size) return false;
return true;
}
// returns the exponent of the smallest power of two which is greater than val
unsigned int GetPow2(unsigned int val)
{
unsigned int ret = 0;
for (;val;val>>=1) ret++;
return ret;
}
TextureCache::TCacheEntry* TextureCache::Load(unsigned int stage, u32 address, unsigned int width, unsigned int height, unsigned int tex_format, unsigned int tlutaddr, unsigned int tlutfmt, bool UseNativeMips, unsigned int maxlevel)
{
// notes (about "UNsafe texture cache"):
// Have to be removed soon.
// But we keep it until the "safe" way became rock solid
// pros: it has an unique ID held by the texture data itself (@address) once cached.
// cons: it writes this unique ID in the gc RAM <- very dangerous (break MP1) and ugly
// notes (about "safe texture cache"):
// Metroids text issue (character table):
// Same addr, same GX_TF_C4 texture data but different TLUT (hence different outputs).
// That's why we have to hash the TLUT too for TLUT tex_format dependent textures (ie. GX_TF_C4, GX_TF_C8, GX_TF_C14X2).
// And since the address and tex data don't change, the key index in the cacheEntry map can't be the address but
// have to be a real unique ID.
// DONE but not satifiying yet -> may break copyEFBToTexture sometimes.
// Pokemon Colosseum text issue (plain text):
// Use a GX_TF_I4 512x512 text-flush-texture at a const address.
// The problem here was just the sparse hash on the texture. This texture is partly overwrited (what is needed only)
// so lot's of remaning old text. Thin white chars on black bg too.
// TODO: - clean this up when ready to kill old "unsafe texture cache"
// - fix the key index situation with CopyRenderTargetToTexture.
// Could happen only for GX_TF_C4, GX_TF_C8 and GX_TF_C14X2 fmt.
// Wonder if we can't use tex width&height to know if EFB might be copied to it...
// raw idea: TOCHECK if addresses are aligned we have few bits left...
if (address == 0)
return NULL;
u8 *ptr = g_VideoInitialize.pGetMemoryPointer(address);
unsigned int bsw = TexDecoder_GetBlockWidthInTexels(tex_format) - 1; // TexelSizeInNibbles(format)*width*height/16;
unsigned int bsh = TexDecoder_GetBlockHeightInTexels(tex_format) - 1; // TexelSizeInNibbles(format)*width*height/16;
unsigned int bsdepth = TexDecoder_GetTexelSizeInNibbles(tex_format);
unsigned int expandedWidth = (width + bsw) & (~bsw);
unsigned int expandedHeight = (height + bsh) & (~bsh);
u64 hash_value = 0;
u32 texID = address;
u64 texHash = 0;
u32 FullFormat = tex_format;
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
FullFormat = (tex_format | (tlutfmt << 16));
// hires texture loading and texture dumping require accurate hashes
if (g_ActiveConfig.bSafeTextureCache || g_ActiveConfig.bHiresTextures || g_ActiveConfig.bDumpTextures)
{
texHash = GetHash64(ptr,TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
{
// WARNING! texID != address now => may break CopyRenderTargetToTexture (cf. TODO up)
// tlut size can be up to 32768B (GX_TF_C14X2) but Safer == Slower.
// This trick (to change the texID depending on the TLUT addr) is a trick to get around
// an issue with metroid prime's fonts, where it has multiple sets of fonts on top of
// each other stored in a single texture, and uses the palette to make different characters
// visible or invisible. Thus, unless we want to recreate the textures for every drawn character,
// we must make sure that texture with different tluts get different IDs.
u64 tlutHash = GetHash64(&texMem[tlutaddr], TexDecoder_GetPaletteSize(tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
texHash ^= tlutHash;
if (g_ActiveConfig.bSafeTextureCache)
{
texID = texID ^ ((u32)(tlutHash & 0xFFFFFFFF)) ^ ((u32)((tlutHash >> 32) & 0xFFFFFFFF));
}
}
if (g_ActiveConfig.bSafeTextureCache)
hash_value = texHash;
}
bool skip_texture_create = false;
TexCache::iterator iter = textures.find(texID);
if (iter != textures.end())
{
TCacheEntry &entry = iter->second;
if (!g_ActiveConfig.bSafeTextureCache)
hash_value = ((u32*)ptr)[0];
// TODO: Is the (entry.MipLevels == maxlevel) check needed?
if (entry.isRenderTarget || ((address == entry.addr) && (hash_value == entry.hash) && FullFormat == entry.fmt && entry.MipLevels == maxlevel))
{
entry.frameCount = frameCount;
D3D::gfxstate->SetShaderResource(stage, entry.texture->GetSRV());
return &entry;
}
else
{
// Let's reload the new texture data into the same texture,
// instead of destroying it and having to create a new one.
// Might speed up movie playback very, very slightly.
// TODO: Is the (entry.MipLevels < maxlevel) check needed?
if (width == entry.w && height==entry.h && FullFormat == entry.fmt && entry.MipLevels < maxlevel)
{
skip_texture_create = true;
}
else
{
entry.Destroy(false);
textures.erase(iter);
}
}
}
// Make an entry in the table
TCacheEntry& entry = textures[texID];
entry.Realw = width;
entry.Realh = height;
PC_TexFormat pcfmt = PC_TEX_FMT_NONE;
pcfmt = TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt, true);
if (g_ActiveConfig.bHiresTextures)
{
// Load Custom textures
char texPathTemp[MAX_PATH];
int newWidth = width;
int newHeight = height;
sprintf(texPathTemp, "%s_%08x_%i", globals->unique_id, texHash, tex_format);
pcfmt = HiresTextures::GetHiresTex(texPathTemp, &newWidth, &newHeight, GX_TF_RGBA8, temp);
if (pcfmt != PC_TEX_FMT_NONE)
{
expandedWidth = width = newWidth;
expandedHeight = height = newHeight;
}
}
if (pcfmt == PC_TEX_FMT_NONE)
pcfmt = TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt, true);
entry.oldpixel = ((u32*)ptr)[0];
if (g_ActiveConfig.bSafeTextureCache) entry.hash = hash_value;
else entry.hash = ((u32*)ptr)[0] = (u32)(((double)rand() / RAND_MAX) * 0xFFFFFFFF);
bool isPow2 = !((width & (width - 1)) || (height & (height - 1)));
unsigned int TexLevels = (isPow2 && UseNativeMips && maxlevel) ? GetPow2(max(width, height)) : ((isPow2)? 0 : 1);
if (TexLevels > (maxlevel + 1) && maxlevel)
TexLevels = maxlevel + 1;
D3D11_USAGE usage = (TexLevels == 1) ? D3D11_USAGE_DYNAMIC : D3D11_USAGE_DEFAULT;
if (!skip_texture_create)
{
// TODO: A little more verbosity in the debug names would be quite helpful..
D3D11_CPU_ACCESS_FLAG cpu_access = (TexLevels == 1) ? D3D11_CPU_ACCESS_WRITE : (D3D11_CPU_ACCESS_FLAG)0;
ID3D11Texture2D* pTexture = NULL;
HRESULT hr;
D3D11_SUBRESOURCE_DATA data;
data.pSysMem = temp;
data.SysMemPitch = 4*expandedWidth;
D3D11_TEXTURE2D_DESC texdesc = CD3D11_TEXTURE2D_DESC(DXGI_FORMAT_R8G8B8A8_UNORM, width, height, 1, TexLevels, D3D11_BIND_SHADER_RESOURCE, usage, cpu_access);
hr = D3D::device->CreateTexture2D(&texdesc, (TexLevels==1)?&data:NULL, &pTexture);
CHECK(hr==S_OK, "Create texture of the TextureCache");
entry.texture = new D3DTexture2D(pTexture, D3D11_BIND_SHADER_RESOURCE);
CHECK(entry.texture!=NULL, "Create texture of the TextureCache");
D3D::SetDebugObjectName((ID3D11DeviceChild*)entry.texture->GetTex(), "a texture of the TextureCache");
D3D::SetDebugObjectName((ID3D11DeviceChild*)entry.texture->GetSRV(), "shader resource view of a texture of the TextureCache");
SAFE_RELEASE(pTexture);
// if (TexLevels == 1), we already loaded the data into our texture upon creation
if (TexLevels != 1) D3D::ReplaceRGBATexture2D(entry.texture->GetTex(), temp, width, height, expandedWidth, 0, usage);
}
else
{
D3D::ReplaceRGBATexture2D(entry.texture->GetTex(), temp, width, height, expandedWidth, 0, usage);
}
entry.addr = address;
entry.size_in_bytes = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format);
entry.isRenderTarget = false;
entry.isNonPow2 = false;
entry.MipLevels = maxlevel;
if (TexLevels == 0) PD3DX11FilterTexture(D3D::context, entry.texture->GetTex(), 0, D3DX11_DEFAULT);
else if (TexLevels > 1 && pcfmt != PC_TEX_FMT_NONE)
{
unsigned int level = 1;
unsigned int mipWidth = (width + 1) >> 1;
unsigned int mipHeight = (height + 1) >> 1;
ptr += entry.size_in_bytes;
while ((mipHeight || mipWidth) && (level < TexLevels))
{
unsigned int currentWidth = (mipWidth > 0) ? mipWidth : 1;
unsigned int currentHeight = (mipHeight > 0) ? mipHeight : 1;
expandedWidth = (currentWidth + bsw) & (~bsw);
expandedHeight = (currentHeight + bsh) & (~bsh);
TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt, true);
D3D::ReplaceRGBATexture2D(entry.texture->GetTex(), temp, currentWidth, currentHeight, expandedWidth, level, usage);
u32 size = (max(mipWidth, bsw) * max(mipHeight, bsh) * bsdepth) >> 1;
ptr += size;
mipWidth >>= 1;
mipHeight >>= 1;
level++;
}
}
entry.frameCount = frameCount;
entry.w = width;
entry.h = height;
entry.Scaledw = width;
entry.Scaledh = height;
entry.fmt = FullFormat;
// dump texture to file
if (g_ActiveConfig.bDumpTextures)
{
char szTemp[MAX_PATH];
char szDir[MAX_PATH];
// make sure that the directory exists
sprintf(szDir, "%s%s", File::GetUserPath(D_DUMPTEXTURES_IDX), globals->unique_id);
if (!File::Exists(szDir) || !File::IsDirectory(szDir))
File::CreateDir(szDir);
sprintf(szTemp, "%s/%s_%08x_%i.png", szDir, globals->unique_id, texHash, tex_format);
if (!File::Exists(szTemp))
if(FAILED(PD3DX11SaveTextureToFileA(D3D::context, entry.texture->GetTex(), D3DX11_IFF_PNG, szTemp))) PanicAlert("!!!");
}
INCSTAT(stats.numTexturesCreated);
SETSTAT(stats.numTexturesAlive, textures.size());
D3D::gfxstate->SetShaderResource(stage, entry.texture->GetSRV());
return &entry;
}
void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, unsigned int bScaleByHalf, const EFBRectangle &source_rect)
{
int efb_w = source_rect.GetWidth();
int efb_h = source_rect.GetHeight();
int tex_w = (abs(source_rect.GetWidth()) >> bScaleByHalf);
int tex_h = (abs(source_rect.GetHeight()) >> bScaleByHalf);
int Scaledtex_w = (g_ActiveConfig.bCopyEFBScaled) ? ((int)(Renderer::GetTargetScaleX() * tex_w)) : tex_w;
int Scaledtex_h = (g_ActiveConfig.bCopyEFBScaled) ? ((int)(Renderer::GetTargetScaleY() * tex_h)) : tex_h;
TexCache::iterator iter;
D3DTexture2D* tex = NULL;
iter = textures.find(address);
if (iter != textures.end())
{
if (iter->second.isRenderTarget && iter->second.Scaledw == Scaledtex_w && iter->second.Scaledh == Scaledtex_h)
{
tex = iter->second.texture;
iter->second.frameCount = frameCount;
}
else
{
// Remove it and recreate it as a render target
SAFE_RELEASE(iter->second.texture);
textures.erase(iter);
}
}
if (!tex)
{
TCacheEntry entry;
entry.isRenderTarget = true;
entry.hash = 0;
entry.frameCount = frameCount;
entry.w = entry.Realw = tex_w;
entry.h = entry.Realh = tex_h;
entry.Scaledw = Scaledtex_w;
entry.Scaledh = Scaledtex_h;
entry.fmt = copyfmt;
entry.isNonPow2 = true;
entry.texture = D3DTexture2D::Create(Scaledtex_w, Scaledtex_h, (D3D11_BIND_FLAG)((int)D3D11_BIND_RENDER_TARGET|(int)D3D11_BIND_SHADER_RESOURCE), D3D11_USAGE_DEFAULT, DXGI_FORMAT_R8G8B8A8_UNORM);
if (entry.texture == NULL) PanicAlert("CopyRenderTargetToTexture failed to create entry.texture at %s %d\n", __FILE__, __LINE__);
textures[address] = entry;
tex = entry.texture;
}
float colmat[20]= {0.0f}; // last four floats for fConstAdd
unsigned int cbufid = (unsigned int)-1;
// TODO: Move this to TextureCache::Init()
if (bFromZBuffer)
{
switch(copyfmt)
{
case 0: // Z4
case 1: // Z8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1.0f;
cbufid = 12;
break;
case 3: // Z16 //?
colmat[1] = colmat[5] = colmat[9] = colmat[12] = 1.0f;
cbufid = 13;
break;
case 11: // Z16 (reverse order)
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
cbufid = 14;
break;
case 6: // Z24X8
colmat[0] = colmat[5] = colmat[10] = 1.0f;
cbufid = 15;
break;
case 9: // Z8M
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
cbufid = 16;
break;
case 10: // Z8L
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
cbufid = 17;
break;
case 12: // Z16L
colmat[2] = colmat[6] = colmat[10] = colmat[13] = 1.0f;
cbufid = 18;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy zbuf format: 0x%x", copyfmt);
colmat[2] = colmat[5] = colmat[8] = 1.0f;
cbufid = 19;
break;
}
}
else if (bIsIntensityFmt)
{
colmat[16] = colmat[17] = colmat[18] = 16.0f/255.0f;
switch (copyfmt)
{
case 0: // I4
case 1: // I8
case 2: // IA4
case 3: // IA8
// TODO - verify these coefficients
colmat[0] = 0.257f; colmat[1] = 0.504f; colmat[2] = 0.098f;
colmat[4] = 0.257f; colmat[5] = 0.504f; colmat[6] = 0.098f;
colmat[8] = 0.257f; colmat[9] = 0.504f; colmat[10] = 0.098f;
if (copyfmt < 2)
{
colmat[19] = 16.0f / 255.0f;
colmat[12] = 0.257f; colmat[13] = 0.504f; colmat[14] = 0.098f;
cbufid = 0;
}
else// alpha
{
colmat[15] = 1;
cbufid = 1;
}
break;
default:
ERROR_LOG(VIDEO, "Unknown copy intensity format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
else
{
switch (copyfmt)
{
case 0: // R4
case 8: // R8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
cbufid = 2;
break;
case 2: // RA4
case 3: // RA8
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1;
cbufid = 3;
break;
case 7: // A8
colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1;
cbufid = 4;
break;
case 9: // G8
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1;
cbufid = 5;
break;
case 10: // B8
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1;
cbufid = 6;
break;
case 11: // RG8
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1;
cbufid = 7;
break;
case 12: // GB8
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1;
cbufid = 8;
break;
case 4: // RGB565
colmat[0] = colmat[5] = colmat[10] = 1;
colmat[19] = 1; // set alpha to 1
cbufid = 9;
break;
case 5: // RGB5A3
case 6: // RGBA8
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
cbufid = 10;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy color format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
cbufid = 11;
break;
}
}
Renderer::ResetAPIState(); // reset any game specific settings
// stretch picture with increased internal resolution
D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, (float)Scaledtex_w, (float)Scaledtex_h);
D3D::context->RSSetViewports(1, &vp);
D3D11_RECT destrect = CD3D11_RECT(0, 0, Scaledtex_w, Scaledtex_h);
// set transformation
if (efbcopycbuf[cbufid] == NULL)
{
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(20*sizeof(float), D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DEFAULT);
D3D11_SUBRESOURCE_DATA data;
data.pSysMem = colmat;
HRESULT hr = D3D::device->CreateBuffer(&cbdesc, &data, &efbcopycbuf[cbufid]);
CHECK(hr==S_OK, "Create efb copy constant buffer %d", cbufid);
D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopycbuf[cbufid], "a constant buffer used in TextureCache::CopyRenderTargetToTexture");
}
D3D::context->PSSetConstantBuffers(0, 1, &efbcopycbuf[cbufid]);
TargetRectangle targetSource = Renderer::ConvertEFBRectangle(source_rect);
D3D11_RECT sourcerect = CD3D11_RECT(targetSource.left, targetSource.top, targetSource.right, targetSource.bottom);
// Use linear filtering if (bScaleByHalf), use point filtering otherwise
if (bScaleByHalf) D3D::SetLinearCopySampler();
else D3D::SetPointCopySampler();
D3D::stateman->PushBlendState(efbcopyblendstate);
D3D::stateman->PushRasterizerState(efbcopyraststate);
D3D::stateman->PushDepthState(efbcopydepthstate);
D3D::context->OMSetRenderTargets(1, &tex->GetRTV(), NULL);
D3D::drawShadedTexQuad(
(bFromZBuffer) ? g_framebufferManager.GetEFBDepthTexture()->GetSRV() : g_framebufferManager.GetEFBColorTexture()->GetSRV(),
&sourcerect,
Renderer::GetFullTargetWidth(),
Renderer::GetFullTargetHeight(),
(bFromZBuffer) ? PixelShaderCache::GetDepthMatrixProgram() : PixelShaderCache::GetColorMatrixProgram(), VertexShaderCache::GetSimpleVertexShader(), VertexShaderCache::GetSimpleInputLayout());
D3D::context->OMSetRenderTargets(1, &g_framebufferManager.GetEFBColorTexture()->GetRTV(), g_framebufferManager.GetEFBDepthTexture()->GetDSV());
D3D::stateman->PopBlendState();
D3D::stateman->PopDepthState();
D3D::stateman->PopRasterizerState();
Renderer::RestoreAPIState();
} }

69
Source/Plugins/Plugin_VideoDX11/Src/TextureCache.h
View File

@ -24,45 +24,42 @@
#include "VideoCommon.h" #include "VideoCommon.h"
#include "BPMemory.h" #include "BPMemory.h"
class TextureCache #include "TextureCacheBase.h"
namespace DX11
{
class TextureCache : public ::TextureCache
{ {
public: public:
struct TCacheEntry TextureCache();
{ ~TextureCache();
D3DTexture2D* texture;
u32 addr;
u32 size_in_bytes;
u64 hash;
u32 paletteHash;
u32 oldpixel;
int frameCount;
unsigned int w, h, fmt, MipLevels;
int Realw, Realh, Scaledw, Scaledh;
bool isRenderTarget;
bool isNonPow2;
TCacheEntry() : texture(NULL), addr(0), size_in_bytes(0), hash(0), paletteHash(0), oldpixel(0),
frameCount(0), w(0), h(0), fmt(0), MipLevels(0), Realw(0), Realh(0), Scaledw(0), Scaledh(0),
isRenderTarget(false), isNonPow2(true) {}
void Destroy(bool shutdown);
bool IntersectsMemoryRange(u32 range_address, u32 range_size);
};
static void Init();
static void Cleanup();
static void Shutdown();
static void Invalidate(bool shutdown);
static void InvalidateRange(u32 start_address, u32 size);
static TCacheEntry* TextureCache::Load(unsigned int stage, u32 address, unsigned int width, unsigned int height, unsigned int tex_format, unsigned int tlutaddr, unsigned int tlutfmt, bool UseNativeMips, unsigned int maxlevel);
static void CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, unsigned int bScaleByHalf, const EFBRectangle &source_rect);
private: private:
typedef std::map<u32,TCacheEntry> TexCache; struct TCacheEntry : TCacheEntryBase
{
D3DTexture2D *const texture;
static u8* temp; D3D11_USAGE usage;
static TexCache textures;
TCacheEntry(D3DTexture2D *_tex) : texture(_tex) {}
~TCacheEntry();
void Load(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int levels);
void FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
unsigned int cbufid, const float* colmat, const EFBRectangle &source_rect,
bool bIsIntensityFmt, u32 copyfmt);
void Bind(unsigned int stage);
bool Save(const char filename[]);
};
TCacheEntryBase* CreateTexture(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int tex_levels, PC_TexFormat pcfmt);
TCacheEntryBase* CreateRenderTargetTexture(unsigned int scaled_tex_w, unsigned int scaled_tex_h);
}; };
}

4
Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
View File

@ -214,7 +214,7 @@ void VertexManager::vFlush()
{ {
Renderer::SetSamplerState(i & 3, i >> 2); Renderer::SetSamplerState(i & 3, i >> 2);
FourTexUnits &tex = bpmem.tex[i >> 2]; FourTexUnits &tex = bpmem.tex[i >> 2];
TextureCache::TCacheEntry* tentry = TextureCache::Load(i, TextureCache::TCacheEntryBase* tentry = TextureCache::Load(i,
(tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5, (tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,
tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1, tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1,
tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9, tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9,
@ -225,7 +225,7 @@ void VertexManager::vFlush()
if (tentry) if (tentry)
{ {
// 0s are probably for no manual wrapping needed. // 0s are probably for no manual wrapping needed.
PixelShaderManager::SetTexDims(i, tentry->Realw, tentry->Realh, 0, 0); PixelShaderManager::SetTexDims(i, tentry->nativeW, tentry->nativeH, 0, 0);
} }
else else
ERROR_LOG(VIDEO, "error loading texture"); ERROR_LOG(VIDEO, "error loading texture");

4
Source/Plugins/Plugin_VideoDX11/Src/main.cpp
View File

@ -215,7 +215,7 @@ void Video_Prepare()
// internal interfaces // internal interfaces
Renderer::Init(); Renderer::Init();
TextureCache::Init(); g_texture_cache = new DX11::TextureCache;
g_vertex_manager = new DX11::VertexManager; g_vertex_manager = new DX11::VertexManager;
VertexShaderCache::Init(); VertexShaderCache::Init();
PixelShaderCache::Init(); PixelShaderCache::Init();
@ -258,7 +258,7 @@ void Shutdown()
PixelShaderCache::Shutdown(); PixelShaderCache::Shutdown();
VertexShaderCache::Shutdown(); VertexShaderCache::Shutdown();
delete g_vertex_manager; delete g_vertex_manager;
TextureCache::Shutdown(); delete g_texture_cache;
Renderer::Shutdown(); Renderer::Shutdown();
EmuWindow::Close(); EmuWindow::Close();

2
Source/Plugins/Plugin_VideoDX9/Src/BPFunctions.cpp
View File

@ -85,7 +85,7 @@ void CopyEFB(const BPCmd &bp, const EFBRectangle &rc, const u32 &address, const
// bpmem.zcontrol.pixel_format to PIXELFMT_Z24 is when the game wants to copy from ZBuffer (Zbuffer uses 24-bit Format) // bpmem.zcontrol.pixel_format to PIXELFMT_Z24 is when the game wants to copy from ZBuffer (Zbuffer uses 24-bit Format)
if (!g_ActiveConfig.bEFBCopyDisable) if (!g_ActiveConfig.bEFBCopyDisable)
{ {
TextureCache::CopyRenderTargetToTexture(address, fromZBuffer, isIntensityFmt, copyfmt, scaleByHalf, rc); TextureCache::CopyRenderTargetToTexture(address, fromZBuffer, isIntensityFmt, copyfmt, !!scaleByHalf, rc);
} }
} }

674
Source/Plugins/Plugin_VideoDX9/Src/TextureCache.cpp
View File

@ -43,570 +43,44 @@
#include "debugger/debugger.h" #include "debugger/debugger.h"
u8 *TextureCache::temp = NULL;
TextureCache::TexCache TextureCache::textures;
extern int frameCount; extern int frameCount;
#define TEMP_SIZE (2048*2048*4) namespace DX9
#define TEXTURE_KILL_THRESHOLD 200
void TextureCache::TCacheEntry::Destroy(bool shutdown)
{ {
if (texture)
texture->Release(); TextureCache::TCacheEntry::~TCacheEntry()
texture = 0; {
if (!isRenderTarget && !shutdown && !g_ActiveConfig.bSafeTextureCache) texture->Release();
{
u32 *ptr = (u32*)g_VideoInitialize.pGetMemoryPointer(addr);
if (ptr && *ptr == hash)
*ptr = oldpixel;
}
} }
void TextureCache::Init() void TextureCache::TCacheEntry::Bind(unsigned int stage)
{ {
temp = (u8*)AllocateMemoryPages(TEMP_SIZE); D3D::SetTexture(stage, texture);
TexDecoder_SetTexFmtOverlayOptions(g_ActiveConfig.bTexFmtOverlayEnable, g_ActiveConfig.bTexFmtOverlayCenter);
HiresTextures::Init(globals->unique_id);
} }
void TextureCache::Invalidate(bool shutdown) bool TextureCache::TCacheEntry::Save(const char filename[])
{ {
for (TexCache::iterator iter = textures.begin(); iter != textures.end(); ++iter) return SUCCEEDED(PD3DXSaveTextureToFileA(filename, D3DXIFF_PNG, texture, 0));
iter->second.Destroy(shutdown);
textures.clear();
HiresTextures::Shutdown();
} }
void TextureCache::Shutdown() void TextureCache::TCacheEntry::Load(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int level)
{ {
Invalidate(true); D3D::ReplaceTexture2D(texture, temp, width, height, expanded_width, d3d_fmt, swap_r_b, level);
FreeMemoryPages(temp, TEMP_SIZE);
temp = NULL;
} }
void TextureCache::Cleanup() void TextureCache::TCacheEntry::FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
unsigned int cbufid, const float *colmat, const EFBRectangle &source_rect,
bool bIsIntensityFmt, u32 copyfmt)
{ {
TexCache::iterator iter = textures.begin(); const LPDIRECT3DTEXTURE9 read_texture = bFromZBuffer ?
while (iter != textures.end()) g_framebufferManager.GetEFBDepthTexture(source_rect) :
g_framebufferManager.GetEFBColorTexture(source_rect);
if (!isDynamic || g_ActiveConfig.bCopyEFBToTexture)
{ {
if (frameCount > TEXTURE_KILL_THRESHOLD + iter->second.frameCount)
{
iter->second.Destroy(false);
iter = textures.erase(iter);
}
else
{
++iter;
}
}
}
void TextureCache::InvalidateRange(u32 start_address, u32 size)
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
int rangePosition = iter->second.IntersectsMemoryRange(start_address, size);
if (rangePosition == 0)
{
iter->second.Destroy(false);
textures.erase(iter++);
}
else
{
++iter;
}
}
}
void TextureCache::MakeRangeDynamic(u32 start_address, u32 size)
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
int rangePosition = iter->second.IntersectsMemoryRange(start_address, size);
if ( rangePosition == 0)
{
iter->second.hash = 0;
}
++iter;
}
}
int TextureCache::TCacheEntry::IntersectsMemoryRange(u32 range_address, u32 range_size)
{
if (addr + size_in_bytes < range_address)
return -1;
if (addr >= range_address + range_size)
return 1;
return 0;
}
TextureCache::TCacheEntry *TextureCache::Load(int stage, u32 address, int width, int height, int tex_format, int tlutaddr, int tlutfmt,bool UseNativeMips, int maxlevel)
{
// notes (about "UNsafe texture cache"):
// Have to be removed soon.
// But we keep it until the "safe" way became rock solid
// pros: it has an unique ID held by the texture data itself (@address) once cached.
// cons: it writes this unique ID in the gc RAM <- very dangerous (break MP1) and ugly
// notes (about "safe texture cache"):
// Metroids text issue (character table):
// Same addr, same GX_TF_C4 texture data but different TLUT (hence different outputs).
// That's why we have to hash the TLUT too for TLUT tex_format dependent textures (ie. GX_TF_C4, GX_TF_C8, GX_TF_C14X2).
// And since the address and tex data don't change, the key index in the cacheEntry map can't be the address but
// have to be a real unique ID.
// DONE but not satifiying yet -> may break copyEFBToTexture sometimes.
// Pokemon Colosseum text issue (plain text):
// Use a GX_TF_I4 512x512 text-flush-texture at a const address.
// The problem here was just the sparse hash on the texture. This texture is partly overwrited (what is needed only)
// so lot's of remaning old text. Thin white chars on black bg too.
// TODO: - clean this up when ready to kill old "unsafe texture cache"
// - fix the key index situation with CopyRenderTargetToTexture.
// Could happen only for GX_TF_C4, GX_TF_C8 and GX_TF_C14X2 fmt.
// Wonder if we can't use tex width&height to know if EFB might be copied to it...
// raw idea: TOCHECK if addresses are aligned we have few bits left...
if (address == 0)
return NULL;
u8 *ptr = g_VideoInitialize.pGetMemoryPointer(address);
int bsw = TexDecoder_GetBlockWidthInTexels(tex_format) - 1; // TexelSizeInNibbles(format)*width*height/16;
int bsh = TexDecoder_GetBlockHeightInTexels(tex_format) - 1; // TexelSizeInNibbles(format)*width*height/16;
int bsdepth = TexDecoder_GetTexelSizeInNibbles(tex_format);
int expandedWidth = (width + bsw) & (~bsw);
int expandedHeight = (height + bsh) & (~bsh);
u64 hash_value = 0;
u32 texID = address;
u64 texHash = 0;
u32 FullFormat = tex_format;
bool TextureisDynamic = false;
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
FullFormat = (tex_format | (tlutfmt << 16));
if (g_ActiveConfig.bSafeTextureCache || g_ActiveConfig.bHiresTextures || g_ActiveConfig.bDumpTextures)
{
texHash = GetHash64(ptr,TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
{
// WARNING! texID != address now => may break CopyRenderTargetToTexture (cf. TODO up)
// tlut size can be up to 32768B (GX_TF_C14X2) but Safer == Slower.
// This trick (to change the texID depending on the TLUT addr) is a trick to get around
// an issue with metroid prime's fonts, where it has multiple sets of fonts on top of
// each other stored in a single texture, and uses the palette to make different characters
// visible or invisible. Thus, unless we want to recreate the textures for every drawn character,
// we must make sure that texture with different tluts get different IDs.
u64 tlutHash = GetHash64(&texMem[tlutaddr], TexDecoder_GetPaletteSize(tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
texHash ^= tlutHash;
if (g_ActiveConfig.bSafeTextureCache)
{
texID = texID ^ ((u32)(tlutHash & 0xFFFFFFFF)) ^ ((u32)((tlutHash >> 32) & 0xFFFFFFFF));
}
}
if (g_ActiveConfig.bSafeTextureCache)
hash_value = texHash;
}
bool skip_texture_create = false;
TexCache::iterator iter = textures.find(texID);
if (iter != textures.end())
{
TCacheEntry &entry = iter->second;
if (!g_ActiveConfig.bSafeTextureCache)
{
if(entry.isRenderTarget || entry.isDynamic)
{
if(!g_ActiveConfig.bCopyEFBToTexture)
{
hash_value = GetHash64(ptr,TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
{
hash_value ^= GetHash64(&texMem[tlutaddr], TexDecoder_GetPaletteSize(tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
}
else
{
hash_value = 0;
}
}
else
{
hash_value = ((u32 *)ptr)[0];
}
}
else
{
if(entry.isRenderTarget || entry.isDynamic)
{
if(g_ActiveConfig.bCopyEFBToTexture)
{
hash_value = 0;
}
}
}
if (((entry.isRenderTarget || entry.isDynamic) && hash_value == entry.hash && address == entry.addr)
|| ((address == entry.addr) && (hash_value == entry.hash) && FullFormat == entry.fmt/* && entry.MipLevels == maxlevel*/))
{
entry.frameCount = frameCount;
entry.isDynamic = false;
D3D::SetTexture(stage, entry.texture);
return &entry;
}
else
{
// Let's reload the new texture data into the same texture,
// instead of destroying it and having to create a new one.
// Might speed up movie playback very, very slightly.
TextureisDynamic = (entry.isRenderTarget || entry.isDynamic) && !g_ActiveConfig.bCopyEFBToTexture;
if (!entry.isRenderTarget &&
((!entry.isDynamic && width == entry.w && height==entry.h && FullFormat == entry.fmt /* && entry.MipLevels < maxlevel*/)
|| (entry.isDynamic && entry.w == width && entry.h == height)))
{
skip_texture_create = true;
}
else
{
entry.Destroy(false);
textures.erase(iter);
}
}
}
// Make an entry in the table
TCacheEntry& entry = textures[texID];
entry.isDynamic = TextureisDynamic;
entry.Realw = width;
entry.Realh = height;
PC_TexFormat pcfmt = PC_TEX_FMT_NONE;
if (g_ActiveConfig.bHiresTextures)
{
// Load Custom textures
char texPathTemp[MAX_PATH];
int newWidth = width;
int newHeight = height;
sprintf(texPathTemp, "%s_%08x_%i", globals->unique_id, texHash, tex_format);
pcfmt = HiresTextures::GetHiresTex(texPathTemp, &newWidth, &newHeight, tex_format, temp);
if (pcfmt != PC_TEX_FMT_NONE)
{
expandedWidth = width = newWidth;
expandedHeight = height = newHeight;
}
}
if (pcfmt == PC_TEX_FMT_NONE)
pcfmt = TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt);
entry.oldpixel = ((u32 *)ptr)[0];
if (g_ActiveConfig.bSafeTextureCache || entry.isDynamic)
entry.hash = hash_value;
else
{
entry.hash = (u32)(((double)rand() / RAND_MAX) * 0xFFFFFFFF);
((u32 *)ptr)[0] = entry.hash;
}
entry.addr = address;
entry.size_in_bytes = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format);
entry.isRenderTarget = false;
bool isPow2 = !((width & (width - 1)) || (height & (height - 1)));
entry.isNonPow2 = false;
int TexLevels = (width > height)?width:height;
TexLevels = (isPow2 && UseNativeMips && (maxlevel > 0)) ? (int)(log((double)TexLevels)/log((double)2)) + 1 : (isPow2? 0 : 1);
if(TexLevels > (maxlevel + 1) && maxlevel > 0)
TexLevels = (maxlevel + 1);
entry.MipLevels = maxlevel;
D3DFORMAT d3d_fmt;
bool swap_r_b = false;
switch (pcfmt)
{
case PC_TEX_FMT_BGRA32:
d3d_fmt = D3DFMT_A8R8G8B8;
break;
case PC_TEX_FMT_RGBA32:
d3d_fmt = D3DFMT_A8R8G8B8;
swap_r_b = true;
break;
case PC_TEX_FMT_RGB565:
d3d_fmt = D3DFMT_R5G6B5;
break;
case PC_TEX_FMT_IA4_AS_IA8:
d3d_fmt = D3DFMT_A8L8;
break;
case PC_TEX_FMT_I8:
case PC_TEX_FMT_I4_AS_I8:
// A hack which means the format is a packed
// 8-bit intensity texture. It is unpacked
// to A8L8 in D3DTexture.cpp
d3d_fmt = D3DFMT_A8P8;
break;
case PC_TEX_FMT_IA8:
d3d_fmt = D3DFMT_A8L8;
break;
case PC_TEX_FMT_DXT1:
d3d_fmt = D3DFMT_DXT1;
break;
}
if (!skip_texture_create)
{
entry.texture = D3D::CreateTexture2D((BYTE*)temp, width, height, expandedWidth, d3d_fmt, swap_r_b, TexLevels);
}
else
{
D3D::ReplaceTexture2D(entry.texture, (BYTE*)temp, width, height, expandedWidth, d3d_fmt, swap_r_b, 0);
}
if(TexLevels > 1 && pcfmt != PC_TEX_FMT_NONE)
{
int level = 1;
int mipWidth = (width + 1) >> 1;
int mipHeight = (height + 1) >> 1;
ptr += entry.size_in_bytes;
while((mipHeight || mipWidth) && (level < TexLevels))
{
u32 currentWidth = (mipWidth > 0)? mipWidth : 1;
u32 currentHeight = (mipHeight > 0)? mipHeight : 1;
expandedWidth = (currentWidth + bsw) & (~bsw);
expandedHeight = (currentHeight + bsh) & (~bsh);
TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt);
D3D::ReplaceTexture2D(entry.texture, (BYTE*)temp, currentWidth, currentHeight, expandedWidth, d3d_fmt, swap_r_b,level);
u32 size = (max(mipWidth, bsw) * max(mipHeight, bsh) * bsdepth) >> 1;
ptr += size;
mipWidth >>= 1;
mipHeight >>= 1;
level++;
}
}
entry.frameCount = frameCount;
entry.w = width;
entry.h = height;
entry.Scaledw = width;
entry.Scaledh = height;
entry.fmt = FullFormat;
if (g_ActiveConfig.bDumpTextures)
{
// dump texture to file
char szTemp[MAX_PATH];
char szDir[MAX_PATH];
const char* uniqueId = globals->unique_id;
static bool bCheckedDumpDir = false;
sprintf(szDir, "%s%s", File::GetUserPath(D_DUMPTEXTURES_IDX), uniqueId);
if (!bCheckedDumpDir)
{
if (!File::Exists(szDir) || !File::IsDirectory(szDir))
File::CreateDir(szDir);
bCheckedDumpDir = true;
}
sprintf(szTemp, "%s/%s_%08x_%i.png", szDir, uniqueId, texHash, tex_format);
if (!File::Exists(szTemp))
PD3DXSaveTextureToFileA(szTemp,D3DXIFF_PNG,entry.texture,0);
}
INCSTAT(stats.numTexturesCreated);
SETSTAT(stats.numTexturesAlive, textures.size());
//Set the texture!
D3D::SetTexture(stage, entry.texture);
DEBUGGER_PAUSE_LOG_AT(NEXT_NEW_TEXTURE,true,{printf("A new texture (%d x %d) is loaded", width, height);});
return &entry;
}
void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle &source_rect)
{
int tex_w = (abs(source_rect.GetWidth()) >> bScaleByHalf);
int tex_h = (abs(source_rect.GetHeight()) >> bScaleByHalf);
//compensate the texture grow if supersampling is enabled to conserve memory usage
float SuperSampleCompensation = g_ActiveConfig.iMultisampleMode + 1;
SuperSampleCompensation = 1.0f / SuperSampleCompensation;
float xScale = Renderer::GetTargetScaleX();
float yScale = Renderer::GetTargetScaleY();
int Scaledtex_w = (g_ActiveConfig.bCopyEFBScaled)?((int)(xScale * SuperSampleCompensation * tex_w)):tex_w;
int Scaledtex_h = (g_ActiveConfig.bCopyEFBScaled)?((int)(yScale * SuperSampleCompensation * tex_h)):tex_h;
TexCache::iterator iter;
LPDIRECT3DTEXTURE9 tex = NULL;
iter = textures.find(address);
bool TextureisDynamic = false;
if (iter != textures.end())
{
if ((iter->second.isRenderTarget && iter->second.Scaledw == Scaledtex_w && iter->second.Scaledh == Scaledtex_h)
|| (iter->second.isDynamic && iter->second.w == tex_w && iter->second.h == tex_h))
{
tex = iter->second.texture;
TextureisDynamic = iter->second.isDynamic;
iter->second.frameCount = frameCount;
}
else
{
// Remove it and recreate it as a render target
if(iter->second.texture)
iter->second.texture->Release();
iter->second.texture = 0;
textures.erase(iter);
}
}
if (TextureisDynamic)
{
Scaledtex_w = tex_w;
Scaledtex_h = tex_h;
}
if (!tex)
{
TCacheEntry entry;
entry.addr = address;
entry.isRenderTarget = true;
entry.hash = 0;
entry.frameCount = frameCount;
entry.w = entry.Realw = tex_w;
entry.h = entry.Realh = tex_h;
entry.Scaledw = Scaledtex_w;
entry.Scaledh = Scaledtex_h;
entry.fmt = copyfmt;
entry.isNonPow2 = true;
entry.isDynamic = false;
D3D::dev->CreateTexture(Scaledtex_w, Scaledtex_h, 1, D3DUSAGE_RENDERTARGET, D3DFMT_A8R8G8B8, D3DPOOL_DEFAULT, &entry.texture, 0);
textures[address] = entry;
tex = entry.texture;
}
// Make sure to resolve anything we need to read from.
LPDIRECT3DTEXTURE9 read_texture = bFromZBuffer ? g_framebufferManager.GetEFBDepthTexture(source_rect) : g_framebufferManager.GetEFBColorTexture(source_rect);
// We have to run a pixel shader, for color conversion.
Renderer::ResetAPIState(); // reset any game specific settings
if(!TextureisDynamic || g_ActiveConfig.bCopyEFBToTexture)
{
float colmat[16]= {0.0f};
float fConstAdd[4] = {0.0f};
if (bFromZBuffer)
{
switch(copyfmt)
{
case 0: // Z4
case 1: // Z8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1.0f;
break;
case 3: // Z16 //?
colmat[1] = colmat[5] = colmat[9] = colmat[12] = 1.0f;
break;
case 11: // Z16 (reverse order)
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
break;
case 6: // Z24X8
colmat[0] = colmat[5] = colmat[10] = 1.0f;
break;
case 9: // Z8M
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
break;
case 10: // Z8L
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
break;
case 12: // Z16L
colmat[2] = colmat[6] = colmat[10] = colmat[13] = 1.0f;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy zbuf format: 0x%x", copyfmt);
colmat[2] = colmat[5] = colmat[8] = 1.0f;
break;
}
}
else if (bIsIntensityFmt)
{
fConstAdd[0] = fConstAdd[1] = fConstAdd[2] = 16.0f/255.0f;
switch (copyfmt)
{
case 0: // I4
case 1: // I8
case 2: // IA4
case 3: // IA8
colmat[0] = 0.257f; colmat[1] = 0.504f; colmat[2] = 0.098f;
colmat[4] = 0.257f; colmat[5] = 0.504f; colmat[6] = 0.098f;
colmat[8] = 0.257f; colmat[9] = 0.504f; colmat[10] = 0.098f;
if (copyfmt < 2)
{
fConstAdd[3] = 16.0f / 255.0f;
colmat[12] = 0.257f; colmat[13] = 0.504f; colmat[14] = 0.098f;
}
else// alpha
colmat[15] = 1;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy intensity format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
else
{
switch (copyfmt)
{
case 0: // R4
case 8: // R8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
break;
case 2: // RA4
case 3: // RA8
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1;
break;
case 7: // A8
colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1;
break;
case 9: // G8
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1;
break;
case 10: // B8
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1;
break;
case 11: // RG8
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1;
break;
case 12: // GB8
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1;
break;
case 4: // RGB565
colmat[0] = colmat[5] = colmat[10] = 1;
fConstAdd[3] = 1; // set alpha to 1
break;
case 5: // RGB5A3
case 6: // RGBA8
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy color format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
LPDIRECT3DSURFACE9 Rendersurf = NULL; LPDIRECT3DSURFACE9 Rendersurf = NULL;
tex->GetSurfaceLevel(0,&Rendersurf); texture->GetSurfaceLevel(0, &Rendersurf);
D3D::dev->SetDepthStencilSurface(NULL); D3D::dev->SetDepthStencilSurface(NULL);
D3D::dev->SetRenderTarget(0, Rendersurf); D3D::dev->SetRenderTarget(0, Rendersurf);
@ -615,18 +89,18 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
// Stretch picture with increased internal resolution // Stretch picture with increased internal resolution
vp.X = 0; vp.X = 0;
vp.Y = 0; vp.Y = 0;
vp.Width = Scaledtex_w; vp.Width = scaledW;
vp.Height = Scaledtex_h; vp.Height = scaledH;
vp.MinZ = 0.0f; vp.MinZ = 0.0f;
vp.MaxZ = 1.0f; vp.MaxZ = 1.0f;
D3D::dev->SetViewport(&vp); D3D::dev->SetViewport(&vp);
RECT destrect; RECT destrect;
destrect.bottom = Scaledtex_h; destrect.bottom = scaledH;
destrect.left = 0; destrect.left = 0;
destrect.right = Scaledtex_w; destrect.right = scaledW;
destrect.top = 0; destrect.top = 0;
const float* const fConstAdd = colmat + 16; // fConstAdd is the last 4 floats of colmat
PixelShaderManager::SetColorMatrix(colmat, fConstAdd); // set transformation PixelShaderManager::SetColorMatrix(colmat, fConstAdd); // set transformation
TargetRectangle targetSource = Renderer::ConvertEFBRectangle(source_rect); TargetRectangle targetSource = Renderer::ConvertEFBRectangle(source_rect);
RECT sourcerect; RECT sourcerect;
@ -635,10 +109,9 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
sourcerect.right = targetSource.right; sourcerect.right = targetSource.right;
sourcerect.top = targetSource.top; sourcerect.top = targetSource.top;
if (bFromZBuffer)
if(bFromZBuffer)
{ {
if(bScaleByHalf || g_ActiveConfig.iMultisampleMode) if (bScaleByHalf || g_ActiveConfig.iMultisampleMode)
{ {
D3D::ChangeSamplerState(0, D3DSAMP_MINFILTER, D3DTEXF_LINEAR); D3D::ChangeSamplerState(0, D3DSAMP_MINFILTER, D3DTEXF_LINEAR);
D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR); D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR);
@ -655,30 +128,29 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR); D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR);
} }
D3DFORMAT bformat = g_framebufferManager.GetEFBDepthRTSurfaceFormat(); D3DFORMAT bformat = g_framebufferManager.GetEFBDepthRTSurfaceFormat();
int SSAAMode = g_ActiveConfig.iMultisampleMode; int SSAAMode = g_ActiveConfig.iMultisampleMode;
D3D::drawShadedTexQuad(
read_texture, D3D::drawShadedTexQuad(read_texture, &sourcerect,
&sourcerect, Renderer::GetFullTargetWidth(), Renderer::GetFullTargetHeight(),
Renderer::GetFullTargetWidth() , scaledW, scaledH,
Renderer::GetFullTargetHeight(), ((bformat != FOURCC_RAWZ && bformat != D3DFMT_D24X8) && bFromZBuffer) ?
Scaledtex_w, PixelShaderCache::GetDepthMatrixProgram(SSAAMode) :
Scaledtex_h, PixelShaderCache::GetColorMatrixProgram(SSAAMode),
((bformat != FOURCC_RAWZ && bformat != D3DFMT_D24X8) && bFromZBuffer)? PixelShaderCache::GetDepthMatrixProgram(SSAAMode): PixelShaderCache::GetColorMatrixProgram(SSAAMode),
VertexShaderCache::GetSimpleVertexShader(SSAAMode)); VertexShaderCache::GetSimpleVertexShader(SSAAMode));
Rendersurf->Release(); Rendersurf->Release();
} }
if(!g_ActiveConfig.bCopyEFBToTexture) if (!g_ActiveConfig.bCopyEFBToTexture)
{ {
textures[address].hash = TextureConverter::EncodeToRamFromTexture( hash = TextureConverter::EncodeToRamFromTexture(
address, addr,
read_texture, read_texture,
Renderer::GetFullTargetWidth(), Renderer::GetFullTargetWidth(),
Renderer::GetFullTargetHeight(), Renderer::GetFullTargetHeight(),
xScale, Renderer::GetTargetScaleX(),
yScale, Renderer::GetTargetScaleY(),
(float)((Renderer::GetFullTargetWidth() - Renderer::GetTargetWidth()) / 2), (float)((Renderer::GetFullTargetWidth() - Renderer::GetTargetWidth()) / 2),
(float)((Renderer::GetFullTargetHeight() - Renderer::GetTargetHeight()) / 2) , (float)((Renderer::GetFullTargetHeight() - Renderer::GetTargetHeight()) / 2) ,
bFromZBuffer, bFromZBuffer,
@ -690,8 +162,68 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
D3D::RefreshSamplerState(0, D3DSAMP_MINFILTER); D3D::RefreshSamplerState(0, D3DSAMP_MINFILTER);
D3D::RefreshSamplerState(0, D3DSAMP_MAGFILTER); D3D::RefreshSamplerState(0, D3DSAMP_MAGFILTER);
D3D::SetTexture(0,NULL); D3D::SetTexture(0, NULL);
D3D::dev->SetRenderTarget(0, g_framebufferManager.GetEFBColorRTSurface()); D3D::dev->SetRenderTarget(0, g_framebufferManager.GetEFBColorRTSurface());
D3D::dev->SetDepthStencilSurface(g_framebufferManager.GetEFBDepthRTSurface()); D3D::dev->SetDepthStencilSurface(g_framebufferManager.GetEFBDepthRTSurface());
Renderer::RestoreAPIState(); }
TextureCache::TCacheEntryBase* TextureCache::CreateTexture(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int tex_levels, PC_TexFormat pcfmt)
{
D3DFORMAT d3d_fmt;
bool swap_r_b = false;
switch (pcfmt)
{
case PC_TEX_FMT_BGRA32:
d3d_fmt = D3DFMT_A8R8G8B8;
break;
case PC_TEX_FMT_RGBA32:
d3d_fmt = D3DFMT_A8R8G8B8;
swap_r_b = true;
break;
case PC_TEX_FMT_RGB565:
d3d_fmt = D3DFMT_R5G6B5;
break;
case PC_TEX_FMT_IA4_AS_IA8:
d3d_fmt = D3DFMT_A8L8;
break;
case PC_TEX_FMT_I8:
case PC_TEX_FMT_I4_AS_I8:
// A hack which means the format is a packed
// 8-bit intensity texture. It is unpacked
// to A8L8 in D3DTexture.cpp
d3d_fmt = D3DFMT_A8P8;
break;
case PC_TEX_FMT_IA8:
d3d_fmt = D3DFMT_A8L8;
break;
case PC_TEX_FMT_DXT1:
d3d_fmt = D3DFMT_DXT1;
break;
}
TCacheEntry* entry = new TCacheEntry(D3D::CreateTexture2D(temp, width, height, expanded_width, d3d_fmt, swap_r_b, tex_levels));
entry->swap_r_b = swap_r_b;
entry->d3d_fmt = d3d_fmt;
return entry;
}
TextureCache::TCacheEntryBase* TextureCache::CreateRenderTargetTexture(
unsigned int scaled_tex_w, unsigned int scaled_tex_h)
{
LPDIRECT3DTEXTURE9 texture;
D3D::dev->CreateTexture(scaled_tex_w, scaled_tex_h, 1, D3DUSAGE_RENDERTARGET,
D3DFMT_A8R8G8B8, D3DPOOL_DEFAULT, &texture, 0);
return new TCacheEntry(texture);
}
} }

75
Source/Plugins/Plugin_VideoDX9/Src/TextureCache.h
View File

@ -25,64 +25,41 @@
#include "VideoCommon.h" #include "VideoCommon.h"
#include "BPMemory.h" #include "BPMemory.h"
class TextureCache #include "TextureCacheBase.h"
namespace DX9
{ {
public:
struct TCacheEntry class TextureCache : public ::TextureCache
{
private:
struct TCacheEntry : TCacheEntryBase
{ {
LPDIRECT3DTEXTURE9 texture; const LPDIRECT3DTEXTURE9 texture;
u32 addr; D3DFORMAT d3d_fmt;
u32 size_in_bytes; bool swap_r_b;
u64 hash;
u32 paletteHash;
u32 oldpixel;
int frameCount; TCacheEntry(LPDIRECT3DTEXTURE9 _tex) : texture(_tex) {}
int w, h, fmt,MipLevels, Realw, Realh, Scaledw, Scaledh; ~TCacheEntry();
bool isRenderTarget; void Load(unsigned int width, unsigned int height,
bool isDynamic;// mofified from cpu unsigned int expanded_width, unsigned int levels);
bool isNonPow2;
TCacheEntry() void FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
{ unsigned int cbufid, const float* colmat, const EFBRectangle &source_rect,
texture = 0; bool bIsIntensityFmt, u32 copyfmt);
isRenderTarget = 0;
hash = 0; void Bind(unsigned int stage);
paletteHash = 0; bool Save(const char filename[]);
oldpixel = 0;
addr = 0;
size_in_bytes = 0;
frameCount = 0;
isNonPow2 = true;
w = 0;
h = 0;
Realw = 0;
Realh = 0;
Scaledw = 0;
Scaledh = 0;
}
void Destroy(bool shutdown);
int IntersectsMemoryRange(u32 range_address, u32 range_size);
}; };
private: TCacheEntryBase* CreateTexture(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int tex_levels, PC_TexFormat pcfmt);
typedef std::map<u32,TCacheEntry> TexCache; TCacheEntryBase* CreateRenderTargetTexture(unsigned int scaled_tex_w, unsigned int scaled_tex_h);
static u8 *temp;
static TexCache textures;
public:
static void Init();
static void Cleanup();
static void Shutdown();
static void Invalidate(bool shutdown);
static void InvalidateRange(u32 start_address, u32 size);
static void MakeRangeDynamic(u32 start_address, u32 size);
static TCacheEntry *Load(int stage, u32 address, int width, int height, int format, int tlutaddr, int tlutfmt,bool UseNativeMips, int maxlevel);
static void CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle &source_rect);
}; };
}
#endif #endif

4
Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp
View File

@ -133,7 +133,7 @@ void VertexManager::vFlush()
{ {
Renderer::SetSamplerState(i & 3, i >> 2); Renderer::SetSamplerState(i & 3, i >> 2);
FourTexUnits &tex = bpmem.tex[i >> 2]; FourTexUnits &tex = bpmem.tex[i >> 2];
TextureCache::TCacheEntry* tentry = TextureCache::Load(i, TextureCache::TCacheEntryBase* tentry = TextureCache::Load(i,
(tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5, (tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,
tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1, tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1,
tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9, tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9,
@ -144,7 +144,7 @@ void VertexManager::vFlush()
if (tentry) if (tentry)
{ {
// 0s are probably for no manual wrapping needed. // 0s are probably for no manual wrapping needed.
PixelShaderManager::SetTexDims(i, tentry->Realw, tentry->Realh, 0, 0); PixelShaderManager::SetTexDims(i, tentry->nativeW, tentry->nativeH, 0, 0);
} }
else else
ERROR_LOG(VIDEO, "error loading texture"); ERROR_LOG(VIDEO, "error loading texture");

4
Source/Plugins/Plugin_VideoDX9/Src/main.cpp
View File

@ -238,7 +238,7 @@ void Video_Prepare()
// internal interfaces // internal interfaces
Renderer::Init(); Renderer::Init();
TextureCache::Init(); g_texture_cache = new DX9::TextureCache;
g_vertex_manager = new DX9::VertexManager; g_vertex_manager = new DX9::VertexManager;
// VideoCommon // VideoCommon
BPInit(); BPInit();
@ -276,7 +276,7 @@ void Shutdown()
PixelShaderCache::Shutdown(); PixelShaderCache::Shutdown();
VertexShaderCache::Shutdown(); VertexShaderCache::Shutdown();
delete g_vertex_manager; delete g_vertex_manager;
TextureCache::Shutdown(); delete g_texture_cache;
Renderer::Shutdown(); Renderer::Shutdown();
D3D::Shutdown(); D3D::Shutdown();
EmuWindow::Close(); EmuWindow::Close();

2
Source/Plugins/Plugin_VideoOGL/Src/BPFunctions.cpp
View File

@ -83,7 +83,7 @@ void CopyEFB(const BPCmd &bp, const EFBRectangle &rc, const u32 &address, const
// bpmem.zcontrol.pixel_format to PIXELFMT_Z24 is when the game wants to copy from ZBuffer (Zbuffer uses 24-bit Format) // bpmem.zcontrol.pixel_format to PIXELFMT_Z24 is when the game wants to copy from ZBuffer (Zbuffer uses 24-bit Format)
if (!g_ActiveConfig.bEFBCopyDisable) if (!g_ActiveConfig.bEFBCopyDisable)
{ {
TextureCache::CopyRenderTargetToTexture(address, fromZBuffer, isIntensityFmt, copyfmt, scaleByHalf, rc); TextureCache::CopyRenderTargetToTexture(address, fromZBuffer, isIntensityFmt, copyfmt, !!scaleByHalf, rc);
} }
} }

4
Source/Plugins/Plugin_VideoOGL/Src/Render.cpp
View File

@ -1082,7 +1082,7 @@ void Renderer::Swap(u32 xfbAddr, FieldType field, u32 fbWidth, u32 fbHeight,cons
// Disable all other stages // Disable all other stages
for (int i = 1; i < 8; ++i) for (int i = 1; i < 8; ++i)
TextureCache::DisableStage(i); OGL::TextureCache::DisableStage(i);
// Update GLViewPort // Update GLViewPort
glViewport(dst_rect.left, dst_rect.bottom, dst_rect.GetWidth(), dst_rect.GetHeight()); glViewport(dst_rect.left, dst_rect.bottom, dst_rect.GetWidth(), dst_rect.GetHeight());
@ -1246,7 +1246,7 @@ void Renderer::Swap(u32 xfbAddr, FieldType field, u32 fbWidth, u32 fbHeight,cons
} }
glBindTexture(GL_TEXTURE_RECTANGLE_ARB, 0); glBindTexture(GL_TEXTURE_RECTANGLE_ARB, 0);
TextureCache::DisableStage(0); OGL::TextureCache::DisableStage(0);
// Wireframe // Wireframe
if (g_ActiveConfig.bWireFrame) if (g_ActiveConfig.bWireFrame)

831
Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp
View File

@ -53,15 +53,11 @@
#include "VertexShaderManager.h" #include "VertexShaderManager.h"
#include "VideoConfig.h" #include "VideoConfig.h"
u8 *TextureCache::temp = NULL; namespace OGL
TextureCache::TexCache TextureCache::textures; {
extern int frameCount;
static u32 s_TempFramebuffer = 0; static u32 s_TempFramebuffer = 0;
#define TEMP_SIZE (2048*2048*4)
#define TEXTURE_KILL_THRESHOLD 200
static const GLint c_MinLinearFilter[8] = { static const GLint c_MinLinearFilter[8] = {
GL_NEAREST, GL_NEAREST,
GL_NEAREST_MIPMAP_NEAREST, GL_NEAREST_MIPMAP_NEAREST,
@ -83,352 +79,63 @@ static const GLint c_WrapSettings[4] = {
bool SaveTexture(const char* filename, u32 textarget, u32 tex, int width, int height) bool SaveTexture(const char* filename, u32 textarget, u32 tex, int width, int height)
{ {
std::vector<u32> data(width * height); std::vector<u32> data(width * height);
glBindTexture(textarget, tex); glBindTexture(textarget, tex);
glGetTexImage(textarget, 0, GL_BGRA, GL_UNSIGNED_BYTE, &data[0]); glGetTexImage(textarget, 0, GL_BGRA, GL_UNSIGNED_BYTE, &data[0]);
GLenum err = GL_REPORT_ERROR();
if (err != GL_NO_ERROR) const GLenum err = GL_REPORT_ERROR();
if (GL_NO_ERROR != err)
{ {
PanicAlert("Can't save texture, GL Error: %s", gluErrorString(err)); PanicAlert("Can't save texture, GL Error: %s", gluErrorString(err));
return false; return false;
} }
return SaveTGA(filename, width, height, &data[0]); return SaveTGA(filename, width, height, &data[0]);
} }
void TextureCache::TCacheEntry::SetTextureParameters(TexMode0 &newmode,TexMode1 &newmode1) TextureCache::TCacheEntry::~TCacheEntry()
{ {
mode = newmode; if (texture)
mode1 = newmode1;
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER,
(newmode.mag_filter || g_Config.bForceFiltering) ? GL_LINEAR : GL_NEAREST);
if (bHaveMipMaps)
{ {
if (g_ActiveConfig.bForceFiltering && newmode.min_filter < 4)
mode.min_filter += 4; // take equivalent forced linear
int filt = newmode.min_filter;
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, c_MinLinearFilter[filt & 7]);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, newmode1.min_lod >> 4);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, newmode1.max_lod >> 4);
glTexEnvf(GL_TEXTURE_FILTER_CONTROL, GL_TEXTURE_LOD_BIAS, (newmode.lod_bias/32.0f));
}
else
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER,
(g_ActiveConfig.bForceFiltering || newmode.min_filter >= 4) ? GL_LINEAR : GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, c_WrapSettings[newmode.wrap_s]);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, c_WrapSettings[newmode.wrap_t]);
if (g_Config.iMaxAnisotropy >= 1)
glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAX_ANISOTROPY_EXT, (float)(1 << g_ActiveConfig.iMaxAnisotropy));
}
void TextureCache::TCacheEntry::Destroy(bool shutdown)
{
if (texture)
glDeleteTextures(1, &texture); glDeleteTextures(1, &texture);
texture = 0; texture = 0;
if (!isRenderTarget && !shutdown && !g_ActiveConfig.bSafeTextureCache)
{
u32 *ptr = (u32*)g_VideoInitialize.pGetMemoryPointer(addr);
if (ptr && *ptr == hash)
*ptr = oldpixel;
} }
} }
void TextureCache::Init() TextureCache::TCacheEntry::TCacheEntry()
{ {
temp = (u8*)AllocateMemoryPages(TEMP_SIZE); glGenTextures(1, &texture);
TexDecoder_SetTexFmtOverlayOptions(g_ActiveConfig.bTexFmtOverlayEnable, g_ActiveConfig.bTexFmtOverlayCenter);
HiresTextures::Init(globals->unique_id);
}
void TextureCache::Invalidate(bool shutdown)
{
for (TexCache::iterator iter = textures.begin(); iter != textures.end(); ++iter)
iter->second.Destroy(shutdown);
textures.clear();
HiresTextures::Shutdown();
}
void TextureCache::Shutdown()
{
Invalidate(true);
if (s_TempFramebuffer)
{
glDeleteFramebuffersEXT(1, (GLuint *)&s_TempFramebuffer);
s_TempFramebuffer = 0;
}
FreeMemoryPages(temp, TEMP_SIZE);
temp = NULL;
}
void TextureCache::Cleanup()
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
if (frameCount > TEXTURE_KILL_THRESHOLD + iter->second.frameCount)
{
iter->second.Destroy(false);
textures.erase(iter++);
}
else
{
++iter;
}
}
}
void TextureCache::InvalidateRange(u32 start_address, u32 size)
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
int rangePosition = iter->second.IntersectsMemoryRange(start_address, size);
if (rangePosition == 0)
{
iter->second.Destroy(false);
textures.erase(iter++);
}
else
{
++iter;
}
}
}
void TextureCache::MakeRangeDynamic(u32 start_address, u32 size)
{
TexCache::iterator iter = textures.begin();
while (iter != textures.end())
{
int rangePosition = iter->second.IntersectsMemoryRange(start_address, size);
if ( rangePosition == 0)
{
iter->second.hash = 0;
}
++iter;
}
}
int TextureCache::TCacheEntry::IntersectsMemoryRange(u32 range_address, u32 range_size)
{
if (addr + size_in_bytes < range_address)
return -1;
if (addr >= range_address + range_size)
return 1;
return 0;
}
TextureCache::TCacheEntry* TextureCache::Load(int texstage, u32 address, int width, int height, u32 tex_format, int tlutaddr, int tlutfmt)
{
// notes (about "UNsafe texture cache"):
// Have to be removed soon.
// But we keep it until the "safe" way became rock solid
// pros: it has an unique ID held by the texture data itself (@address) once cached.
// cons: it writes this unique ID in the gc RAM <- very dangerous (break MP1) and ugly
// notes (about "safe texture cache"):
// Metroids text issue (character table):
// Same addr, same GX_TF_C4 texture data but different TLUT (hence different outputs).
// That's why we have to hash the TLUT too for TLUT tex_format dependent textures (ie. GX_TF_C4, GX_TF_C8, GX_TF_C14X2).
// And since the address and tex data don't change, the key index in the cacheEntry map can't be the address but
// have to be a real unique ID.
// DONE but not satifiying yet -> may break copyEFBToTexture sometimes.
// Pokemon Colosseum text issue (plain text):
// Use a GX_TF_I4 512x512 text-flush-texture at a const address.
// The problem here was just the sparse hash on the texture. This texture is partly overwrited (what is needed only)
// so lot's of remaning old text. Thin white chars on black bg too.
// TODO: - clean this up when ready to kill old "unsafe texture cache"
// - fix the key index situation with CopyRenderTargetToTexture.
// Could happen only for GX_TF_C4, GX_TF_C8 and GX_TF_C14X2 fmt.
// Wonder if we can't use tex width&height to know if EFB might be copied to it...
// raw idea: TOCHECK if addresses are aligned we have few bits left...
if (address == 0)
return NULL;
TexMode0 &tm0 = bpmem.tex[texstage >> 2].texMode0[texstage & 3];
TexMode1 &tm1 = bpmem.tex[texstage >> 2].texMode1[texstage & 3];
int maxlevel = (tm1.max_lod >> 4);
bool UseNativeMips = (tm0.min_filter & 3) && (tm0.min_filter != 8) && g_ActiveConfig.bUseNativeMips;
u8 *ptr = g_VideoInitialize.pGetMemoryPointer(address);
int bsw = TexDecoder_GetBlockWidthInTexels(tex_format) - 1; // TexelSizeInNibbles(format)*width*height/16;
int bsh = TexDecoder_GetBlockHeightInTexels(tex_format) - 1; // TexelSizeInNibbles(format)*width*height/16;
int bsdepth = TexDecoder_GetTexelSizeInNibbles(tex_format);
int expandedWidth = (width + bsw) & (~bsw);
int expandedHeight = (height + bsh) & (~bsh);
u64 hash_value = 0;
u32 texID = address;
u64 texHash = 0;
u32 FullFormat = tex_format;
bool TextureisDynamic = false;
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
FullFormat = (tex_format | (tlutfmt << 16));
if (g_ActiveConfig.bSafeTextureCache || g_ActiveConfig.bHiresTextures || g_ActiveConfig.bDumpTextures)
{
texHash = GetHash64(ptr,TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
{
// WARNING! texID != address now => may break CopyRenderTargetToTexture (cf. TODO up)
// tlut size can be up to 32768B (GX_TF_C14X2) but Safer == Slower.
// This trick (to change the texID depending on the TLUT addr) is a trick to get around
// an issue with metroid prime's fonts, where it has multiple sets of fonts on top of
// each other stored in a single texture, and uses the palette to make different characters
// visible or invisible. Thus, unless we want to recreate the textures for every drawn character,
// we must make sure that texture with different tluts get different IDs.
u64 tlutHash = GetHash64(&texMem[tlutaddr], TexDecoder_GetPaletteSize(tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
texHash ^= tlutHash;
if (g_ActiveConfig.bSafeTextureCache)
{
texID = texID ^ ((u32)(tlutHash & 0xFFFFFFFF)) ^ ((u32)((tlutHash >> 32) & 0xFFFFFFFF));
}
}
if (g_ActiveConfig.bSafeTextureCache)
hash_value = texHash;
}
bool skip_texture_create = false;
TexCache::iterator iter = textures.find(texID);
if (iter != textures.end())
{
TCacheEntry &entry = iter->second;
if (!g_ActiveConfig.bSafeTextureCache)
{
if(entry.isRenderTarget || entry.isDynamic)
{
if(!g_ActiveConfig.bCopyEFBToTexture)
{
hash_value = GetHash64(ptr,TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
if ((tex_format == GX_TF_C4) || (tex_format == GX_TF_C8) || (tex_format == GX_TF_C14X2))
{
hash_value ^= GetHash64(&texMem[tlutaddr], TexDecoder_GetPaletteSize(tex_format),g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
}
else
{
hash_value = 0;
}
}
else
{
hash_value = ((u32 *)ptr)[0];
}
}
else
{
if(entry.isRenderTarget || entry.isDynamic)
{
if(g_ActiveConfig.bCopyEFBToTexture)
{
hash_value = 0;
}
}
}
if (((entry.isRenderTarget || entry.isDynamic) && hash_value == entry.hash && address == entry.addr)
|| ((address == entry.addr) && (hash_value == entry.hash) && ((int) FullFormat == entry.fmt) && entry.MipLevels >= maxlevel))
{
entry.frameCount = frameCount;
glEnable(GL_TEXTURE_2D);
glBindTexture(GL_TEXTURE_2D, entry.texture);
GL_REPORT_ERRORD();
entry.SetTextureParameters(tm0,tm1);
entry.isDynamic = false;
return &entry;
}
else
{
// Let's reload the new texture data into the same texture,
// instead of destroying it and having to create a new one.
// Might speed up movie playback very, very slightly.
TextureisDynamic = (entry.isRenderTarget || entry.isDynamic) && !g_ActiveConfig.bCopyEFBToTexture;
if (((!(entry.isRenderTarget || entry.isDynamic) && width == entry.w && height == entry.h && (int)FullFormat == entry.fmt) ||
((entry.isRenderTarget || entry.isDynamic) && entry.w == width && entry.h == height && entry.Scaledw == width && entry.Scaledh == height)))
{
glBindTexture(GL_TEXTURE_2D, entry.texture);
GL_REPORT_ERRORD();
entry.SetTextureParameters(tm0,tm1);
skip_texture_create = true;
}
else
{
entry.Destroy(false);
textures.erase(iter);
}
}
}
// Make an entry in the table
TCacheEntry& entry = textures[texID];
entry.isDynamic = TextureisDynamic;
entry.isRenderTarget = false;
entry.Realw = width;
entry.Realh = height;
PC_TexFormat pcfmt = PC_TEX_FMT_NONE;
if (g_ActiveConfig.bHiresTextures)
{
// Load Custom textures
char texPathTemp[MAX_PATH];
int newWidth = width;
int newHeight = height;
sprintf(texPathTemp, "%s_%08x_%i", globals->unique_id, (unsigned int) texHash, tex_format);
pcfmt = HiresTextures::GetHiresTex(texPathTemp, &newWidth, &newHeight, tex_format, temp);
if (pcfmt != PC_TEX_FMT_NONE)
{
expandedWidth = width = newWidth;
expandedHeight = height = newHeight;
}
}
if (pcfmt == PC_TEX_FMT_NONE)
pcfmt = TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt);
entry.oldpixel = ((u32 *)ptr)[0];
if (g_ActiveConfig.bSafeTextureCache || entry.isDynamic)
entry.hash = hash_value;
else
{
entry.hash = (u32)(((double)rand() / RAND_MAX) * 0xFFFFFFFF);
((u32 *)ptr)[0] = entry.hash;
}
entry.addr = address;
entry.size_in_bytes = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, tex_format);
GLenum target = GL_TEXTURE_2D;
if (!skip_texture_create) {
glGenTextures(1, (GLuint *)&entry.texture);
glBindTexture(target, entry.texture);
}
bool isPow2 = !((width & (width - 1)) || (height & (height - 1)));
int TexLevels = (width > height)?width:height;
TexLevels = (isPow2 && UseNativeMips && (maxlevel > 0)) ? (int)(log((double)TexLevels)/log((double)2)) + 1 : (isPow2? 0 : 1);
if(TexLevels > (maxlevel + 1) && maxlevel > 0)
TexLevels = (maxlevel + 1);
entry.MipLevels = maxlevel;
bool GenerateMipmaps = TexLevels > 1 || TexLevels == 0;
entry.bHaveMipMaps = GenerateMipmaps;
int gl_format = 0;
int gl_iformat = 0;
int gl_type = 0;
GL_REPORT_ERRORD(); GL_REPORT_ERRORD();
}
void TextureCache::TCacheEntry::Bind(unsigned int stage)
{
glEnable(GL_TEXTURE_2D);
glBindTexture(GL_TEXTURE_2D, texture);
GL_REPORT_ERRORD();
// TODO: is this already done somewhere else?
TexMode0 &tm0 = bpmem.tex[stage >> 2].texMode0[stage & 3];
TexMode1 &tm1 = bpmem.tex[stage >> 2].texMode1[stage & 3];
SetTextureParameters(tm0, tm1);
}
bool TextureCache::TCacheEntry::Save(const char filename[])
{
// TODO: make ogl dump PNGs
std::string tga_filename(filename);
tga_filename.replace(tga_filename.size() - 3, 3, "tga");
return SaveTexture(tga_filename.c_str(), GL_TEXTURE_2D, texture, w, h);
}
TextureCache::TCacheEntryBase* TextureCache::CreateTexture(unsigned int width,
unsigned int height, unsigned int expanded_width,
unsigned int tex_levels, PC_TexFormat pcfmt)
{
int gl_format = 0,
gl_iformat = 0,
gl_type = 0;
if (pcfmt != PC_TEX_FMT_DXT1) if (pcfmt != PC_TEX_FMT_DXT1)
{ {
switch (pcfmt) switch (pcfmt)
@ -441,360 +148,148 @@ TextureCache::TCacheEntry* TextureCache::Load(int texstage, u32 address, int wid
gl_iformat = 4; gl_iformat = 4;
gl_type = GL_UNSIGNED_BYTE; gl_type = GL_UNSIGNED_BYTE;
break; break;
case PC_TEX_FMT_RGBA32: case PC_TEX_FMT_RGBA32:
gl_format = GL_RGBA; gl_format = GL_RGBA;
gl_iformat = 4; gl_iformat = 4;
gl_type = GL_UNSIGNED_BYTE; gl_type = GL_UNSIGNED_BYTE;
break; break;
case PC_TEX_FMT_I4_AS_I8: case PC_TEX_FMT_I4_AS_I8:
gl_format = GL_LUMINANCE; gl_format = GL_LUMINANCE;
gl_iformat = GL_INTENSITY4; gl_iformat = GL_INTENSITY4;
gl_type = GL_UNSIGNED_BYTE; gl_type = GL_UNSIGNED_BYTE;
break; break;
case PC_TEX_FMT_IA4_AS_IA8: case PC_TEX_FMT_IA4_AS_IA8:
gl_format = GL_LUMINANCE_ALPHA; gl_format = GL_LUMINANCE_ALPHA;
gl_iformat = GL_LUMINANCE4_ALPHA4; gl_iformat = GL_LUMINANCE4_ALPHA4;
gl_type = GL_UNSIGNED_BYTE; gl_type = GL_UNSIGNED_BYTE;
break; break;
case PC_TEX_FMT_I8: case PC_TEX_FMT_I8:
gl_format = GL_LUMINANCE; gl_format = GL_LUMINANCE;
gl_iformat = GL_INTENSITY8; gl_iformat = GL_INTENSITY8;
gl_type = GL_UNSIGNED_BYTE; gl_type = GL_UNSIGNED_BYTE;
break; break;
case PC_TEX_FMT_IA8: case PC_TEX_FMT_IA8:
gl_format = GL_LUMINANCE_ALPHA; gl_format = GL_LUMINANCE_ALPHA;
gl_iformat = GL_LUMINANCE8_ALPHA8; gl_iformat = GL_LUMINANCE8_ALPHA8;
gl_type = GL_UNSIGNED_BYTE; gl_type = GL_UNSIGNED_BYTE;
break; break;
case PC_TEX_FMT_RGB565: case PC_TEX_FMT_RGB565:
gl_format = GL_RGB; gl_format = GL_RGB;
gl_iformat = GL_RGB; gl_iformat = GL_RGB;
gl_type = GL_UNSIGNED_SHORT_5_6_5; gl_type = GL_UNSIGNED_SHORT_5_6_5;
break; break;
} }
if (expandedWidth != width)
glPixelStorei(GL_UNPACK_ROW_LENGTH, expandedWidth);
//generate mipmaps even if we use native mips to suport textures with less levels
if(skip_texture_create)
{
glTexSubImage2D(target, 0,0,0,width, height, gl_format, gl_type, temp);
}
else
{
if (GenerateMipmaps)
{
if(UseNativeMips)
{
glTexImage2D(target, 0, gl_iformat, width, height, 0, gl_format, gl_type, temp);
}
else
{
glTexParameteri(GL_TEXTURE_2D, GL_GENERATE_MIPMAP, GL_TRUE);
glTexImage2D(target, 0, gl_iformat, width, height, 0, gl_format, gl_type, temp);
glTexParameteri(GL_TEXTURE_2D, GL_GENERATE_MIPMAP, GL_FALSE);
}
}
else
{
glTexImage2D(target, 0, gl_iformat, width, height, 0, gl_format, gl_type, temp);
}
}
if (expandedWidth != width) // reset
glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
}
else
{
if (skip_texture_create)
{
glCompressedTexSubImage2D(target, 0,0,0,width, height,
GL_COMPRESSED_RGBA_S3TC_DXT1_EXT,expandedWidth*expandedHeight/2, temp);
}
else
{
glCompressedTexImage2D(target, 0, GL_COMPRESSED_RGBA_S3TC_DXT1_EXT,
width, height, 0, expandedWidth*expandedHeight/2, temp);
}
}
GL_REPORT_ERRORD();
if(TexLevels > 1 && pcfmt != PC_TEX_FMT_NONE)
{
int level = 1;
int mipWidth = width >> 1;
int mipHeight = height >> 1;
ptr += entry.size_in_bytes;
while((mipHeight || mipWidth) && (level < TexLevels))
{
u32 currentWidth = (mipWidth > 0)? mipWidth : 1;
u32 currentHeight = (mipHeight > 0)? mipHeight : 1;
expandedWidth = (currentWidth + bsw) & (~bsw);
expandedHeight = (currentHeight + bsh) & (~bsh);
TexDecoder_Decode(temp, ptr, expandedWidth, expandedHeight, tex_format, tlutaddr, tlutfmt);
if (pcfmt != PC_TEX_FMT_DXT1)
{
if (expandedWidth != (int)currentWidth)
glPixelStorei(GL_UNPACK_ROW_LENGTH, expandedWidth);
glTexImage2D(target, level, gl_iformat, currentWidth, currentHeight, 0, gl_format, gl_type, temp);
if (expandedWidth != (int)currentWidth)
glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
}
else
{
glCompressedTexImage2D(target, level, GL_COMPRESSED_RGBA_S3TC_DXT1_EXT, currentWidth, currentHeight, 0, expandedWidth*expandedHeight/2, temp);
}
GL_REPORT_ERRORD();
u32 size = (max(mipWidth, bsw) * max(mipHeight, bsh) * bsdepth) >> 1;
ptr += size;
mipWidth >>= 1;
mipHeight >>= 1;
level++;
}
}
entry.frameCount = frameCount;
entry.w = width;
entry.h = height;
entry.Scaledw = width;
entry.Scaledh = height;
entry.fmt = FullFormat;
entry.SetTextureParameters(tm0,tm1);
if (g_ActiveConfig.bDumpTextures)
{
// dump texture to file
char szTemp[MAX_PATH];
char szDir[MAX_PATH];
const char* uniqueId = globals->unique_id;
static bool bCheckedDumpDir = false;
sprintf(szDir, "%s%s", File::GetUserPath(D_DUMPTEXTURES_IDX), uniqueId);
if (!bCheckedDumpDir)
{
if (!File::Exists(szDir) || !File::IsDirectory(szDir))
File::CreateDir(szDir);
bCheckedDumpDir = true;
}
sprintf(szTemp, "%s/%s_%08x_%i.tga", szDir, uniqueId, (unsigned int) texHash, tex_format);
if (!File::Exists(szTemp))
SaveTexture(szTemp, target, entry.texture, entry.w, entry.h);
} }
INCSTAT(stats.numTexturesCreated); TCacheEntry &entry = *new TCacheEntry;
SETSTAT(stats.numTexturesAlive, textures.size()); entry.gl_format = gl_format;
entry.gl_iformat = gl_iformat;
entry.gl_type = gl_type;
entry.pcfmt = pcfmt;
entry.bHaveMipMaps = tex_levels != 1;
return &entry; return &entry;
} }
void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle &source_rect) void TextureCache::TCacheEntry::Load(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int level)
{ {
DVSTARTPROFILE(); //glEnable(GL_TEXTURE_2D);
GL_REPORT_ERRORD(); glBindTexture(GL_TEXTURE_2D, texture);
//GL_REPORT_ERRORD();
// for intensity values, use Y of YUV format! if (pcfmt != PC_TEX_FMT_DXT1)
// for all purposes, treat 4bit equivalents as 8bit (probably just used for compression)
// RGBA8 - RGBA8
// RGB565 - RGB565
// RGB5A3 - RGB5A3
// I4,R4,Z4 - I4
// IA4,RA4 - IA4
// Z8M,G8,I8,A8,Z8,R8,B8,Z8L - I8
// Z16,GB8,RG8,Z16L,IA8,RA8 - IA8
float colmat[16];
float fConstAdd[4] = {0};
memset(colmat, 0, sizeof(colmat));
if (bFromZBuffer)
{ {
switch(copyfmt) if (expanded_width != width)
glPixelStorei(GL_UNPACK_ROW_LENGTH, expanded_width);
if (bHaveMipMaps && 0 == level)
{ {
case 0: // Z4 glTexParameteri(GL_TEXTURE_2D, GL_GENERATE_MIPMAP, GL_TRUE);
case 1: // Z8 glTexImage2D(GL_TEXTURE_2D, level, gl_iformat, width, height, 0, gl_format, gl_type, temp);
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1; glTexParameteri(GL_TEXTURE_2D, GL_GENERATE_MIPMAP, GL_FALSE);
break; }
case 3: // Z16 //? else
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1;
break;
case 11: // Z16 (reverse order)
colmat[2] = colmat[6] = colmat[10] = colmat[13] = 1;
break;
case 6: // Z24X8
colmat[2] = colmat[5] = colmat[8] = colmat[15] = 1;
break;
case 9: // Z8M
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1;
break;
case 10: // Z8L
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
break;
case 12: // Z16L
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy zbuf format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
else if (bIsIntensityFmt)
{
// TODO - verify these coefficients
fConstAdd[0] = fConstAdd[1] = fConstAdd[2] = 16.0f/255.0f;
colmat[0] = 0.257f; colmat[1] = 0.504f; colmat[2] = 0.098f;
colmat[4] = 0.257f; colmat[5] = 0.504f; colmat[6] = 0.098f;
colmat[8] = 0.257f; colmat[9] = 0.504f; colmat[10] = 0.098f;
if (copyfmt < 2)
{ {
fConstAdd[3] = 16.0f / 255.0f; glTexImage2D(GL_TEXTURE_2D, level, gl_iformat, width, height, 0, gl_format, gl_type, temp);
colmat[12] = 0.257f; colmat[13] = 0.504f; colmat[14] = 0.098f; }
}
else// alpha
colmat[15] = 1;
}
else
{
switch (copyfmt)
{
case 0: // R4
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
break;
case 8: // R8
colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
break;
case 2: // RA4
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1;
break;
case 3: // RA8
colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1;
break;
case 7: // A8
colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1;
break;
case 9: // G8
colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1;
break;
case 10: // B8
colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1;
break;
case 11: // RG8
colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1;
break;
case 12: // GB8
colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1;
break;
case 4: // RGB565
colmat[0] = colmat[5] = colmat[10] = 1;
fConstAdd[3] = 1; // set alpha to 1
break;
case 5: // RGB5A3
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
case 6: // RGBA8
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
default:
ERROR_LOG(VIDEO, "Unknown copy color format: 0x%x", copyfmt);
colmat[0] = colmat[5] = colmat[10] = colmat[15] = 1;
break;
}
}
bool bIsInit = textures.find(address) != textures.end(); if (expanded_width != width)
glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
PRIM_LOG("copytarg: addr=0x%x, fromz=%d, intfmt=%d, copyfmt=%d", address, (int)bFromZBuffer, (int)bIsIntensityFmt,copyfmt);
TCacheEntry& entry = textures[address];
entry.hash = 0;
entry.frameCount = frameCount;
int w = (abs(source_rect.GetWidth()) >> bScaleByHalf);
int h = (abs(source_rect.GetHeight()) >> bScaleByHalf);
float xScale = Renderer::GetTargetScaleX();
float yScale = Renderer::GetTargetScaleY();
int Scaledtex_w = (g_ActiveConfig.bCopyEFBScaled)?((int)(xScale * w)) : w;
int Scaledtex_h = (g_ActiveConfig.bCopyEFBScaled)?((int)(yScale * h)) : h;
GLenum gl_format = GL_RGBA;
GLenum gl_iformat = 4;
GLenum gl_type = GL_UNSIGNED_BYTE;
GL_REPORT_ERRORD();
if (!bIsInit)
{
glGenTextures(1, (GLuint *)&entry.texture);
glBindTexture(GL_TEXTURE_2D, entry.texture);
GL_REPORT_ERRORD();
glTexImage2D(GL_TEXTURE_2D, 0, gl_iformat, Scaledtex_w, Scaledtex_h, 0, gl_format, gl_type, NULL);
GL_REPORT_ERRORD();
entry.isRenderTarget = true;
entry.isDynamic = false;
} }
else else
{ {
_assert_(entry.texture); PanicAlert("PC_TEX_FMT_DXT1 support disabled");
GL_REPORT_ERRORD(); //glCompressedTexImage2D(GL_TEXTURE_2D, 0, GL_COMPRESSED_RGBA_S3TC_DXT1_EXT,
if(entry.isDynamic) //width, height, 0, expanded_width * expanded_height/2, temp);
{
Scaledtex_h = h;
Scaledtex_w = w;
}
if (((entry.isRenderTarget || entry.isDynamic) && entry.Scaledw == Scaledtex_w && entry.Scaledh == Scaledtex_h))
{
glBindTexture(GL_TEXTURE_2D, entry.texture);
// for some reason mario sunshine errors here...
// Beyond Good and Evil does too, occasionally.
GL_REPORT_ERRORD();
} else {
// Delete existing texture.
glDeleteTextures(1,(GLuint *)&entry.texture);
glGenTextures(1, (GLuint *)&entry.texture);
glBindTexture(GL_TEXTURE_2D, entry.texture);
glTexImage2D(GL_TEXTURE_2D, 0, gl_iformat, Scaledtex_w, Scaledtex_h, 0, gl_format, gl_type, NULL);
GL_REPORT_ERRORD();
entry.isRenderTarget = !entry.isDynamic;
}
} }
GL_REPORT_ERRORD();
}
if (!bIsInit) TextureCache::TCacheEntryBase* TextureCache::CreateRenderTargetTexture(
unsigned int scaled_tex_w, unsigned int scaled_tex_h)
{
TCacheEntry *const entry = new TCacheEntry;
glBindTexture(GL_TEXTURE_2D, entry->texture);
GL_REPORT_ERRORD();
const GLenum
gl_format = GL_RGBA,
gl_iformat = 4,
gl_type = GL_UNSIGNED_BYTE;
glTexImage2D(GL_TEXTURE_2D, 0, gl_iformat, scaled_tex_w, scaled_tex_h, 0, gl_format, gl_type, NULL);
GL_REPORT_ERRORD();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
if (GL_REPORT_ERROR() != GL_NO_ERROR)
{ {
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP);
GL_REPORT_ERRORD();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
if (GL_REPORT_ERROR() != GL_NO_ERROR) {
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP);
GL_REPORT_ERRORD();
}
} }
entry.addr = address; return entry;
entry.w = entry.Realw = w; }
entry.h = entry.Realh = h;
entry.Scaledw = Scaledtex_w; void TextureCache::TCacheEntry::FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
entry.Scaledh = Scaledtex_h; unsigned int cbufid, const float colmat[], const EFBRectangle &source_rect,
entry.fmt = copyfmt; bool bIsIntensityFmt, u32 copyfmt)
entry.hash = 0; {
glBindTexture(GL_TEXTURE_2D, texture);
// Make sure to resolve anything we need to read from. // Make sure to resolve anything we need to read from.
GLuint read_texture = bFromZBuffer ? g_framebufferManager.ResolveAndGetDepthTarget(source_rect) : g_framebufferManager.ResolveAndGetRenderTarget(source_rect); const GLuint read_texture = bFromZBuffer ?
g_framebufferManager.ResolveAndGetDepthTarget(source_rect) :
g_framebufferManager.ResolveAndGetRenderTarget(source_rect);
// TODO: move
const float xScale = Renderer::GetTargetScaleX();
const float yScale = Renderer::GetTargetScaleY();
GL_REPORT_ERRORD(); GL_REPORT_ERRORD();
// We have to run a pixel shader, for color conversion. if (false == isDynamic || g_ActiveConfig.bCopyEFBToTexture)
Renderer::ResetAPIState(); // reset any game specific settings
if(!entry.isDynamic || g_ActiveConfig.bCopyEFBToTexture)
{ {
if (s_TempFramebuffer == 0) if (s_TempFramebuffer == 0)
glGenFramebuffersEXT(1, (GLuint *)&s_TempFramebuffer); glGenFramebuffersEXT(1, (GLuint*)&s_TempFramebuffer);
g_framebufferManager.SetFramebuffer(s_TempFramebuffer); g_framebufferManager.SetFramebuffer(s_TempFramebuffer);
// Bind texture to temporary framebuffer // Bind texture to temporary framebuffer
glFramebufferTexture2DEXT(GL_FRAMEBUFFER_EXT, GL_COLOR_ATTACHMENT0_EXT, GL_TEXTURE_2D, entry.texture, 0); glFramebufferTexture2DEXT(GL_FRAMEBUFFER_EXT, GL_COLOR_ATTACHMENT0_EXT, GL_TEXTURE_2D, texture, 0);
GL_REPORT_FBO_ERROR(); GL_REPORT_FBO_ERROR();
GL_REPORT_ERRORD(); GL_REPORT_ERRORD();
@ -803,9 +298,10 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
glEnable(GL_TEXTURE_RECTANGLE_ARB); glEnable(GL_TEXTURE_RECTANGLE_ARB);
glBindTexture(GL_TEXTURE_RECTANGLE_ARB, read_texture); glBindTexture(GL_TEXTURE_RECTANGLE_ARB, read_texture);
glViewport(0, 0, Scaledtex_w, Scaledtex_h); glViewport(0, 0, scaledW, scaledH);
PixelShaderCache::SetCurrentShader(bFromZBuffer ? PixelShaderCache::GetDepthMatrixProgram() : PixelShaderCache::GetColorMatrixProgram()); PixelShaderCache::SetCurrentShader(bFromZBuffer ? PixelShaderCache::GetDepthMatrixProgram() : PixelShaderCache::GetColorMatrixProgram());
const float* const fConstAdd = colmat + 16; // fConstAdd is the last 4 floats of colmat
PixelShaderManager::SetColorMatrix(colmat, fConstAdd); // set transformation PixelShaderManager::SetColorMatrix(colmat, fConstAdd); // set transformation
GL_REPORT_ERRORD(); GL_REPORT_ERRORD();
@ -824,10 +320,10 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
glFramebufferTexture2DEXT(GL_FRAMEBUFFER_EXT, GL_COLOR_ATTACHMENT0_EXT, GL_TEXTURE_2D, 0, 0); glFramebufferTexture2DEXT(GL_FRAMEBUFFER_EXT, GL_COLOR_ATTACHMENT0_EXT, GL_TEXTURE_2D, 0, 0);
} }
if(!g_ActiveConfig.bCopyEFBToTexture) if (false == g_ActiveConfig.bCopyEFBToTexture)
{ {
textures[address].hash = TextureConverter::EncodeToRamFromTexture( hash = TextureConverter::EncodeToRamFromTexture(
address, addr,
read_texture, read_texture,
xScale, xScale,
yScale, yScale,
@ -836,32 +332,69 @@ void TextureCache::CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, boo
copyfmt, copyfmt,
bScaleByHalf, bScaleByHalf,
source_rect); source_rect);
} }
// Return to the EFB.
g_framebufferManager.SetFramebuffer(0); g_framebufferManager.SetFramebuffer(0);
Renderer::RestoreAPIState();
VertexShaderManager::SetViewportChanged(); VertexShaderManager::SetViewportChanged();
TextureCache::DisableStage(0); DisableStage(0);
GL_REPORT_ERRORD(); GL_REPORT_ERRORD();
if (g_ActiveConfig.bDumpEFBTarget) if (g_ActiveConfig.bDumpEFBTarget)
{ {
static int count = 0; static int count = 0;
SaveTexture(StringFromFormat("%sefb_frame_%i.tga", File::GetUserPath(D_DUMPTEXTURES_IDX), count++).c_str(), GL_TEXTURE_2D, entry.texture, entry.w, entry.h); SaveTexture(StringFromFormat("%sefb_frame_%i.tga", File::GetUserPath(D_DUMPTEXTURES_IDX),
} count++).c_str(), GL_TEXTURE_2D, texture, w, h);
}
} }
void TextureCache::DisableStage(int stage) void TextureCache::TCacheEntry::SetTextureParameters(const TexMode0 &newmode, const TexMode1 &newmode1)
{
// TODO: not used anywhere
TexMode0 mode = newmode;
//mode1 = newmode1;
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER,
(newmode.mag_filter || g_Config.bForceFiltering) ? GL_LINEAR : GL_NEAREST);
if (bHaveMipMaps)
{
// TODO: not used anywhere
if (g_ActiveConfig.bForceFiltering && newmode.min_filter < 4)
mode.min_filter += 4; // take equivalent forced linear
int filt = newmode.min_filter;
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, c_MinLinearFilter[filt & 7]);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, newmode1.min_lod >> 4);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, newmode1.max_lod >> 4);
glTexEnvf(GL_TEXTURE_FILTER_CONTROL, GL_TEXTURE_LOD_BIAS, (newmode.lod_bias / 32.0f));
}
else
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER,
(g_ActiveConfig.bForceFiltering || newmode.min_filter >= 4) ? GL_LINEAR : GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, c_WrapSettings[newmode.wrap_s]);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, c_WrapSettings[newmode.wrap_t]);
if (g_Config.iMaxAnisotropy >= 1)
glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAX_ANISOTROPY_EXT,
(float)(1 << g_ActiveConfig.iMaxAnisotropy));
}
TextureCache::~TextureCache()
{
if (s_TempFramebuffer)
{
glDeleteFramebuffersEXT(1, (GLuint*)&s_TempFramebuffer);
s_TempFramebuffer = 0;
}
}
void TextureCache::DisableStage(unsigned int stage)
{ {
glActiveTexture(GL_TEXTURE0 + stage); glActiveTexture(GL_TEXTURE0 + stage);
glDisable(GL_TEXTURE_2D); glDisable(GL_TEXTURE_2D);
glDisable(GL_TEXTURE_RECTANGLE_ARB); glDisable(GL_TEXTURE_RECTANGLE_ARB);
} }
void TextureCache::ClearRenderTargets()
{
for (TexCache::iterator iter = textures.begin(); iter != textures.end(); ++iter)
iter->second.isRenderTarget = false;
} }

91
Source/Plugins/Plugin_VideoOGL/Src/TextureCache.h
View File

@ -24,61 +24,62 @@
#include "GLUtil.h" #include "GLUtil.h"
#include "BPStructs.h" #include "BPStructs.h"
class TextureCache #include "TextureCacheBase.h"
namespace OGL
{
class TextureCache : public ::TextureCache
{ {
public: public:
struct TCacheEntry static void DisableStage(unsigned int stage);
{
TCacheEntry() : texture(0), addr(0), size_in_bytes(0), hash(0),
w(0), h(0), MipLevels(0),Scaledw(0), Scaledh(0),
isRenderTarget(false), isDynamic(false), bHaveMipMaps(false)
{ mode.hex = 0xFCFCFCFC; }
GLuint texture;
u32 addr;
u32 size_in_bytes;
u64 hash;
u32 paletteHash;
u32 oldpixel; // used for simple cleanup
TexMode0 mode; // current filter and clamp modes that texture is set to
TexMode1 mode1; // current filter and clamp modes that texture is set to
int frameCount;
int w, h, fmt,MipLevels, Realw, Realh;
int Scaledw, Scaledh;
bool isRenderTarget; // if render texture, then rendertex is filled with the direct copy of the render target
// later conversions would have to convert properly from rendertexfmt to texfmt
bool isDynamic; // modified from cpu
bool bHaveMipMaps;
void SetTextureParameters(TexMode0& newmode,TexMode1 &newmode1);
void Destroy(bool shutdown);
void ConvertFromRenderTarget(u32 taddr, int twidth, int theight, int tformat, int tlutaddr, int tlutfmt);
int IntersectsMemoryRange(u32 range_address, u32 range_size);
};
private: private:
typedef std::map<u32, TCacheEntry> TexCache; struct TCacheEntry : TCacheEntryBase
{
GLuint texture;
static u8 *temp; PC_TexFormat pcfmt;
static TexCache textures;
public: int gl_format;
static void Init(); int gl_iformat;
static void Cleanup(); int gl_type;
static void Shutdown();
static void Invalidate(bool shutdown);
static void InvalidateRange(u32 start_address, u32 size);
static void MakeRangeDynamic(u32 start_address, u32 size);
static TCacheEntry* Load(int texstage, u32 address, int width, int height, u32 format, int tlutaddr, int tlutfmt);
static void CopyRenderTargetToTexture(u32 address, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle &source);
static void DisableStage(int stage); // sets active texture bool bHaveMipMaps;
static void ClearRenderTargets(); // sets render target value of all textures to false //TexMode0 mode; // current filter and clamp modes that texture is set to
//TexMode1 mode1; // current filter and clamp modes that texture is set to
TCacheEntry();
~TCacheEntry();
void Load(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int level);
void FromRenderTarget(bool bFromZBuffer, bool bScaleByHalf,
unsigned int cbufid, const float colmat[], const EFBRectangle &source_rect,
bool bIsIntensityFmt, u32 copyfmt);
void Bind(unsigned int stage);
bool Save(const char filename[]);
private:
void SetTextureParameters(const TexMode0 &newmode, const TexMode1 &newmode1);
};
~TextureCache();
TCacheEntryBase* CreateTexture(unsigned int width, unsigned int height,
unsigned int expanded_width, unsigned int tex_levels, PC_TexFormat pcfmt);
TCacheEntryBase* CreateRenderTargetTexture(unsigned int scaled_tex_w, unsigned int scaled_tex_h);
private:
bool isOGL() { return true; }
}; };
bool SaveTexture(const char* filename, u32 textarget, u32 tex, int width, int height); bool SaveTexture(const char* filename, u32 textarget, u32 tex, int width, int height);
}
#endif // _TEXTUREMNGR_H_ #endif // _TEXTUREMNGR_H_

2
Source/Plugins/Plugin_VideoOGL/Src/TextureConverter.cpp
View File

@ -34,6 +34,8 @@
namespace TextureConverter namespace TextureConverter
{ {
using OGL::TextureCache;
static GLuint s_texConvFrameBuffer = 0; static GLuint s_texConvFrameBuffer = 0;
static GLuint s_srcTexture = 0; // for decoding from RAM static GLuint s_srcTexture = 0; // for decoding from RAM
static GLuint s_srcTextureWidth = 0; static GLuint s_srcTextureWidth = 0;

10
Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp
View File

@ -154,23 +154,25 @@ void VertexManager::vFlush()
{ {
glActiveTexture(GL_TEXTURE0 + i); glActiveTexture(GL_TEXTURE0 + i);
FourTexUnits &tex = bpmem.tex[i >> 2]; FourTexUnits &tex = bpmem.tex[i >> 2];
TextureCache::TCacheEntry* tentry = TextureCache::Load(i, TextureCache::TCacheEntryBase* tentry = TextureCache::Load(i,
(tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5, (tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,
tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1, tex.texImage0[i&3].width + 1, tex.texImage0[i&3].height + 1,
tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9, tex.texImage0[i&3].format, tex.texTlut[i&3].tmem_offset<<9,
tex.texTlut[i&3].tlut_format); tex.texTlut[i&3].tlut_format,
(tex.texMode0[i&3].min_filter & 3) && (tex.texMode0[i&3].min_filter != 8) && g_ActiveConfig.bUseNativeMips,
(tex.texMode1[i&3].max_lod >> 4));
if (tentry) if (tentry)
{ {
// 0s are probably for no manual wrapping needed. // 0s are probably for no manual wrapping needed.
PixelShaderManager::SetTexDims(i, tentry->Realw, tentry->Realh, 0, 0); PixelShaderManager::SetTexDims(i, tentry->nativeW, tentry->nativeH, 0, 0);
if (g_ActiveConfig.iLog & CONF_SAVETEXTURES) if (g_ActiveConfig.iLog & CONF_SAVETEXTURES)
{ {
// save the textures // save the textures
char strfile[255]; char strfile[255];
sprintf(strfile, "%stex%.3d_%d.tga", File::GetUserPath(D_DUMPFRAMES_IDX), g_Config.iSaveTargetId, i); sprintf(strfile, "%stex%.3d_%d.tga", File::GetUserPath(D_DUMPFRAMES_IDX), g_Config.iSaveTargetId, i);
SaveTexture(strfile, GL_TEXTURE_2D, tentry->texture, tentry->w, tentry->h); tentry->Save(strfile);
} }
} }
else else

4
Source/Plugins/Plugin_VideoOGL/Src/main.cpp
View File

@ -223,7 +223,7 @@ void Video_Prepare()
CommandProcessor::Init(); CommandProcessor::Init();
PixelEngine::Init(); PixelEngine::Init();
TextureCache::Init(); g_texture_cache = new OGL::TextureCache;
BPInit(); BPInit();
g_vertex_manager = new OGL::VertexManager; g_vertex_manager = new OGL::VertexManager;
@ -267,7 +267,7 @@ void Shutdown()
PixelShaderManager::Shutdown(); PixelShaderManager::Shutdown();
PixelShaderCache::Shutdown(); PixelShaderCache::Shutdown();
delete g_vertex_manager; delete g_vertex_manager;
TextureCache::Shutdown(); delete g_texture_cache;
OpcodeDecoder_Shutdown(); OpcodeDecoder_Shutdown();
Renderer::Shutdown(); Renderer::Shutdown();
OpenGL_Shutdown(); OpenGL_Shutdown();