u64 TextureCache::TCacheEntryBase::CalculateHash() const
{
u8* ptr = Memory::GetPointer(addr);
if (memory_stride == CacheLinesPerRow() * 32)
{
return GetHash64(ptr, size_in_bytes, g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
else
{
u32 blocks = NumBlocksY();
u64 temp_hash = size_in_bytes;
u32 samples_per_row = 0;
if (g_ActiveConfig.iSafeTextureCache_ColorSamples != 0)
{
// Hash at least 4 samples per row to avoid hashing in a bad pattern, like just on the left side of the efb copy
samples_per_row = std::max(g_ActiveConfig.iSafeTextureCache_ColorSamples / blocks, 4u);
}
for (u32 i = 0; i < blocks; i++)
{
// Multiply by a prime number to mix the hash up a bit. This prevents identical blocks from canceling each other out
temp_hash = (temp_hash * 397) ^ GetHash64(ptr, CacheLinesPerRow() * 32, samples_per_row);
ptr += memory_stride;
}
return temp_hash;
}
}
ID3D11SamplerState* GetSamplerState( D3D11_SAMPLER_DESC const& desc, char const* debugNameOnCreation ) {
auto crc = GetHash64( (u8 const*)&desc, sizeof( desc ), 0 );
auto it = sstates_.find( crc );
if ( it != sstates_.end() ) {
return it->second;
}
ID3D11SamplerState* state;
auto hr = D3D::device->CreateSamplerState( &desc, &state );
if ( FAILED( hr ) )
PanicAlert( "Failed to create sampler state at %s %d\n", __FILE__, __LINE__ );
D3D::SetDebugObjectName( state, debugNameOnCreation );
sstates_.emplace( crc, state );
return state;
}
ID3D11DepthStencilState* GetDepthStencilState( D3D11_DEPTH_STENCIL_DESC const& desc, char const* debugNameOnCreation ) {
auto crc = GetHash64( (u8 const*)&desc, sizeof( desc ), 0 );
auto it = dstates_.find( crc );
if ( it != dstates_.end() ) {
return it->second;
}
ID3D11DepthStencilState* state;
auto hr = D3D::device->CreateDepthStencilState( &desc, &state );
if ( FAILED( hr ) )
PanicAlert( "Failed to create depth stencil state at %s %d\n", __FILE__, __LINE__ );
D3D::SetDebugObjectName( state, debugNameOnCreation );
dstates_.emplace( crc, state );
return state;
}
TextureCache::TCacheEntryBase* TextureCache::Load(unsigned int const stage,
u32 const address, unsigned int width, unsigned int height, int const texformat,
unsigned int const tlutaddr, int const tlutfmt, bool const use_mipmaps, unsigned int maxlevel, bool const from_tmem)
{
if (0 == address)
return nullptr;
// 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);
const unsigned int nativeW = width;
const unsigned int nativeH = height;
u32 texID = address;
// Hash assigned to texcache entry (also used to generate filenames used for texture dumping and custom texture lookup)
u64 tex_hash = TEXHASH_INVALID;
u64 tlut_hash = TEXHASH_INVALID;
u32 full_format = texformat;
PC_TexFormat pcfmt = PC_TEX_FMT_NONE;
const bool isPaletteTexture = (texformat == GX_TF_C4 || texformat == GX_TF_C8 || texformat == GX_TF_C14X2);
if (isPaletteTexture)
full_format = texformat | (tlutfmt << 16);
const u32 texture_size = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, texformat);
const u8* src_data;
if (from_tmem)
src_data = &texMem[bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE];
else
src_data = Memory::GetPointer(address);
// TODO: This doesn't hash GB tiles for preloaded RGBA8 textures (instead, it's hashing more data from the low tmem bank than it should)
tex_hash = GetHash64(src_data, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
if (isPaletteTexture)
{
const u32 palette_size = TexDecoder_GetPaletteSize(texformat);
tlut_hash = GetHash64(&texMem[tlutaddr], palette_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
// NOTE: For non-paletted textures, texID is equal to the texture address.
// A paletted texture, however, may have multiple texIDs assigned though depending on the currently used tlut.
// This (changing texID depending on the tlut_hash) is a trick to get around
// an issue with Metroid Prime's fonts (it has multiple sets of fonts on 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 a paletted texture gets assigned multiple IDs for each tlut used.
//
// TODO: Because texID isn't always the same as the address now, CopyRenderTargetToTexture might be broken now
texID ^= ((u32)tlut_hash) ^(u32)(tlut_hash >> 32);
tex_hash ^= tlut_hash;
}
// D3D doesn't like when the specified mipmap count would require more than one 1x1-sized LOD in the mipmap chain
// e.g. 64x64 with 7 LODs would have the mipmap chain 64x64,32x32,16x16,8x8,4x4,2x2,1x1,1x1, so we limit the mipmap count to 6 there
while (g_ActiveConfig.backend_info.bUseMinimalMipCount && std::max(expandedWidth, expandedHeight) >> maxlevel == 0)
--maxlevel;
TCacheEntryBase *entry = textures[texID];
if (entry)
{
// 1. Calculate reference hash:
// calculated from RAM texture data for normal textures. Hashes for paletted textures are modified by tlut_hash. 0 for virtual EFB copies.
if (g_ActiveConfig.bCopyEFBToTexture && entry->IsEfbCopy())
tex_hash = TEXHASH_INVALID;
// 2. a) For EFB copies, only the hash and the texture address need to match
if (entry->IsEfbCopy() && tex_hash == entry->hash && address == entry->addr)
{
entry->type = TCET_EC_VRAM;
// TODO: Print a warning if the format changes! In this case,
// we could reinterpret the internal texture object data to the new pixel format
// (similar to what is already being done in Renderer::ReinterpretPixelFormat())
return ReturnEntry(stage, entry);
}
// 2. b) For normal textures, all texture parameters need to match
if (address == entry->addr && tex_hash == entry->hash && full_format == entry->format &&
entry->num_mipmaps > maxlevel && entry->native_width == nativeW && entry->native_height == nativeH)
{
return ReturnEntry(stage, entry);
}
// 3. If we reach this line, we'll have to upload the new texture data to VRAM.
// If we're lucky, the texture parameters didn't change and we can reuse the internal texture object instead of destroying and recreating it.
//
// TODO: Don't we need to force texture decoding to RGBA8 for dynamic EFB copies?
// TODO: Actually, it should be enough if the internal texture format matches...
if ((entry->type == TCET_NORMAL &&
width == entry->virtual_width &&
height == entry->virtual_height &&
full_format == entry->format &&
entry->num_mipmaps > maxlevel) ||
(entry->type == TCET_EC_DYNAMIC &&
entry->native_width == width &&
entry->native_height == height))
{
// reuse the texture
}
else
{
// delete the texture and make a new one
delete entry;
entry = nullptr;
}
}
bool using_custom_texture = false;
if (g_ActiveConfig.bHiresTextures)
{
// This function may modify width/height.
pcfmt = LoadCustomTexture(tex_hash, texformat, 0, width, height);
if (pcfmt != PC_TEX_FMT_NONE)
{
if (expandedWidth != width || expandedHeight != height)
{
expandedWidth = width;
expandedHeight = height;
// If we thought we could reuse the texture before, make sure to pool it now!
if (entry)
{
delete entry;
entry = nullptr;
}
}
using_custom_texture = true;
}
}
if (!using_custom_texture)
{
if (!(texformat == GX_TF_RGBA8 && from_tmem))
{
pcfmt = TexDecoder_Decode(temp, src_data, expandedWidth,
expandedHeight, texformat, tlutaddr, tlutfmt, g_ActiveConfig.backend_info.bUseRGBATextures);
}
else
{
u8* src_data_gb = &texMem[bpmem.tex[stage/4].texImage2[stage%4].tmem_odd * TMEM_LINE_SIZE];
pcfmt = TexDecoder_DecodeRGBA8FromTmem(temp, src_data, src_data_gb, expandedWidth, expandedHeight);
}
}
u32 texLevels = use_mipmaps ? (maxlevel + 1) : 1;
const bool using_custom_lods = using_custom_texture && CheckForCustomTextureLODs(tex_hash, texformat, texLevels);
// Only load native mips if their dimensions fit to our virtual texture dimensions
const bool use_native_mips = use_mipmaps && !using_custom_lods && (width == nativeW && height == nativeH);
texLevels = (use_native_mips || using_custom_lods) ? texLevels : 1; // TODO: Should be forced to 1 for non-pow2 textures (e.g. efb copies with automatically adjusted IR)
// create the entry/texture
if (nullptr == entry)
{
textures[texID] = entry = g_texture_cache->CreateTexture(width, height, expandedWidth, texLevels, pcfmt);
// Sometimes, we can get around recreating a texture if only the number of mip levels changes
// e.g. if our texture cache entry got too many mipmap levels we can limit the number of used levels by setting the appropriate render states
// Thus, we don't update this member for every Load, but just whenever the texture gets recreated
// TODO: This is the wrong value. We should be storing the number of levels our actual texture has.
// But that will currently make the above "existing entry" tests fail as "texLevels" is not calculated until after.
// Currently, we might try to reuse a texture which appears to have more levels than actual, maybe..
entry->num_mipmaps = maxlevel + 1;
entry->type = TCET_NORMAL;
GFX_DEBUGGER_PAUSE_AT(NEXT_NEW_TEXTURE, true);
}
else
{
// load texture (CreateTexture also loads level 0)
entry->Load(width, height, expandedWidth, 0);
}
entry->SetGeneralParameters(address, texture_size, full_format, entry->num_mipmaps);
entry->SetDimensions(nativeW, nativeH, width, height);
entry->hash = tex_hash;
if (entry->IsEfbCopy() && !g_ActiveConfig.bCopyEFBToTexture)
entry->type = TCET_EC_DYNAMIC;
else
entry->type = TCET_NORMAL;
if (g_ActiveConfig.bDumpTextures && !using_custom_texture)
DumpTexture(entry, 0);
u32 level = 1;
// load mips - TODO: Loading mipmaps from tmem is untested!
if (pcfmt != PC_TEX_FMT_NONE)
{
if (use_native_mips)
{
src_data += texture_size;
const u8* ptr_even = nullptr;
const u8* ptr_odd = nullptr;
if (from_tmem)
{
ptr_even = &texMem[bpmem.tex[stage/4].texImage1[stage%4].tmem_even * TMEM_LINE_SIZE + texture_size];
ptr_odd = &texMem[bpmem.tex[stage/4].texImage2[stage%4].tmem_odd * TMEM_LINE_SIZE];
}
for (; level != texLevels; ++level)
{
const u32 mip_width = CalculateLevelSize(width, level);
const u32 mip_height = CalculateLevelSize(height, level);
const u32 expanded_mip_width = (mip_width + bsw) & (~bsw);
const u32 expanded_mip_height = (mip_height + bsh) & (~bsh);
const u8*& mip_src_data = from_tmem
? ((level % 2) ? ptr_odd : ptr_even)
: src_data;
TexDecoder_Decode(temp, mip_src_data, expanded_mip_width, expanded_mip_height, texformat, tlutaddr, tlutfmt, g_ActiveConfig.backend_info.bUseRGBATextures);
mip_src_data += TexDecoder_GetTextureSizeInBytes(expanded_mip_width, expanded_mip_height, texformat);
entry->Load(mip_width, mip_height, expanded_mip_width, level);
if (g_ActiveConfig.bDumpTextures)
DumpTexture(entry, level);
}
}
else if (using_custom_lods)
{
for (; level != texLevels; ++level)
{
unsigned int mip_width = CalculateLevelSize(width, level);
unsigned int mip_height = CalculateLevelSize(height, level);
LoadCustomTexture(tex_hash, texformat, level, mip_width, mip_height);
entry->Load(mip_width, mip_height, mip_width, level);
}
}
}
INCSTAT(stats.numTexturesCreated);
SETSTAT(stats.numTexturesAlive, textures.size());
return ReturnEntry(stage, entry);
}
void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFormat,
PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
bool isIntensity, bool scaleByHalf, unsigned int cbufid,
const float *colmat)
{
g_renderer->ResetAPIState(); // reset any game specific settings
// Make sure to resolve anything we need to read from.
const GLuint read_texture = (srcFormat == PEControl::Z24) ?
FramebufferManager::ResolveAndGetDepthTarget(srcRect) :
FramebufferManager::ResolveAndGetRenderTarget(srcRect);
GL_REPORT_ERRORD();
if (type != TCET_EC_DYNAMIC || g_ActiveConfig.bCopyEFBToTexture)
{
FramebufferManager::SetFramebuffer(framebuffer);
GL_REPORT_ERRORD();
glActiveTexture(GL_TEXTURE0+9);
glBindTexture(GL_TEXTURE_2D, read_texture);
glViewport(0, 0, virtual_width, virtual_height);
if (srcFormat == PEControl::Z24)
{
s_DepthMatrixProgram.Bind();
if (s_DepthCbufid != cbufid)
glUniform4fv(s_DepthMatrixUniform, 5, colmat);
s_DepthCbufid = cbufid;
}
else
{
s_ColorMatrixProgram.Bind();
if (s_ColorCbufid != cbufid)
glUniform4fv(s_ColorMatrixUniform, 7, colmat);
s_ColorCbufid = cbufid;
}
TargetRectangle R = g_renderer->ConvertEFBRectangle(srcRect);
glUniform4f(srcFormat == PEControl::Z24 ? s_DepthCopyPositionUniform : s_ColorCopyPositionUniform,
R.left, R.top, R.right, R.bottom);
GL_REPORT_ERRORD();
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
GL_REPORT_ERRORD();
}
if (false == g_ActiveConfig.bCopyEFBToTexture)
{
int encoded_size = TextureConverter::EncodeToRamFromTexture(
addr,
read_texture,
srcFormat == PEControl::Z24,
isIntensity,
dstFormat,
scaleByHalf,
srcRect);
u8* dst = Memory::GetPointer(addr);
u64 const new_hash = GetHash64(dst,encoded_size,g_ActiveConfig.iSafeTextureCache_ColorSamples);
// Mark texture entries in destination address range dynamic unless caching is enabled and the texture entry is up to date
if (!g_ActiveConfig.bEFBCopyCacheEnable)
TextureCache::MakeRangeDynamic(addr,encoded_size);
else if (!TextureCache::Find(addr, new_hash))
TextureCache::MakeRangeDynamic(addr,encoded_size);
hash = new_hash;
}
FramebufferManager::SetFramebuffer(0);
GL_REPORT_ERRORD();
if (g_ActiveConfig.bDumpEFBTarget)
{
static int count = 0;
SaveTexture(StringFromFormat("%sefb_frame_%i.png", File::GetUserPath(D_DUMPTEXTURES_IDX).c_str(),
count++), GL_TEXTURE_2D, texture, virtual_width, virtual_height, 0);
}
g_renderer->RestoreAPIState();
}
TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
{
const FourTexUnits &tex = bpmem.tex[stage >> 2];
const u32 id = stage & 3;
const u32 address = (tex.texImage3[id].image_base/* & 0x1FFFFF*/) << 5;
u32 width = tex.texImage0[id].width + 1;
u32 height = tex.texImage0[id].height + 1;
const int texformat = tex.texImage0[id].format;
const u32 tlutaddr = tex.texTlut[id].tmem_offset << 9;
const u32 tlutfmt = tex.texTlut[id].tlut_format;
const bool use_mipmaps = (tex.texMode0[id].min_filter & 3) != 0;
u32 tex_levels = use_mipmaps ? ((tex.texMode1[id].max_lod + 0xf) / 0x10 + 1) : 1;
const bool from_tmem = tex.texImage1[id].image_type != 0;
if (0 == address)
return nullptr;
// TexelSizeInNibbles(format) * width * height / 16;
const unsigned int bsw = TexDecoder_GetBlockWidthInTexels(texformat);
const unsigned int bsh = TexDecoder_GetBlockHeightInTexels(texformat);
unsigned int expandedWidth = ROUND_UP(width, bsw);
unsigned int expandedHeight = ROUND_UP(height, bsh);
const unsigned int nativeW = width;
const unsigned int nativeH = height;
// Hash assigned to texcache entry (also used to generate filenames used for texture dumping and custom texture lookup)
u64 base_hash = TEXHASH_INVALID;
u64 full_hash = TEXHASH_INVALID;
u32 full_format = texformat;
const bool isPaletteTexture = (texformat == GX_TF_C4 || texformat == GX_TF_C8 || texformat == GX_TF_C14X2);
// Reject invalid tlut format.
if (isPaletteTexture && tlutfmt > GX_TL_RGB5A3)
return nullptr;
if (isPaletteTexture)
full_format = texformat | (tlutfmt << 16);
const u32 texture_size = TexDecoder_GetTextureSizeInBytes(expandedWidth, expandedHeight, texformat);
u32 additional_mips_size = 0; // not including level 0, which is texture_size
// GPUs don't like when the specified mipmap count would require more than one 1x1-sized LOD in the mipmap chain
// e.g. 64x64 with 7 LODs would have the mipmap chain 64x64,32x32,16x16,8x8,4x4,2x2,1x1,0x0, so we limit the mipmap count to 6 there
tex_levels = std::min<u32>(IntLog2(std::max(width, height)) + 1, tex_levels);
for (u32 level = 1; level != tex_levels; ++level)
{
// We still need to calculate the original size of the mips
const u32 expanded_mip_width = ROUND_UP(CalculateLevelSize(width, level), bsw);
const u32 expanded_mip_height = ROUND_UP(CalculateLevelSize(height, level), bsh);
additional_mips_size += TexDecoder_GetTextureSizeInBytes(expanded_mip_width, expanded_mip_height, texformat);
}
// If we are recording a FifoLog, keep track of what memory we read.
// FifiRecorder does it's own memory modification tracking independant of the texture hashing below.
if (g_bRecordFifoData && !from_tmem)
FifoRecorder::GetInstance().UseMemory(address, texture_size + additional_mips_size, MemoryUpdate::TEXTURE_MAP);
const u8* src_data;
if (from_tmem)
src_data = &texMem[bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE];
else
src_data = Memory::GetPointer(address);
// TODO: This doesn't hash GB tiles for preloaded RGBA8 textures (instead, it's hashing more data from the low tmem bank than it should)
base_hash = GetHash64(src_data, texture_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
u32 palette_size = 0;
if (isPaletteTexture)
{
palette_size = TexDecoder_GetPaletteSize(texformat);
full_hash = base_hash ^ GetHash64(&texMem[tlutaddr], palette_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
else
{
full_hash = base_hash;
}
// Search the texture cache for textures by address
//
// Find all texture cache entries for the current texture address, and decide whether to use one of
// them, or to create a new one
//
// In most cases, the fastest way is to use only one texture cache entry for the same address. Usually,
// when a texture changes, the old version of the texture is unlikely to be used again. If there were
// new cache entries created for normal texture updates, there would be a slowdown due to a huge amount
// of unused cache entries. Also thanks to texture pooling, overwriting an existing cache entry is
// faster than creating a new one from scratch.
//
// Some games use the same address for different textures though. If the same cache entry was used in
// this case, it would be constantly overwritten, and effectively there wouldn't be any caching for
// those textures. Examples for this are Metroid Prime and Castlevania 3. Metroid Prime has multiple
// sets of fonts on each other stored in a single texture and uses the palette to make different
// characters visible or invisible. In Castlevania 3 some textures are used for 2 different things or
// at least in 2 different ways(size 1024x1024 vs 1024x256).
//
// To determine whether to use multiple cache entries or a single entry, use the following heuristic:
// If the same texture address is used several times during the same frame, assume the address is used
// for different purposes and allow creating an additional cache entry. If there's at least one entry
// that hasn't been used for the same frame, then overwrite it, in order to keep the cache as small as
// possible. If the current texture is found in the cache, use that entry.
//
// For efb copies, the entry created in CopyRenderTargetToTexture always has to be used, or else it was
// done in vain.
std::pair<TexCache::iterator, TexCache::iterator> iter_range = textures_by_address.equal_range((u64)address);
TexCache::iterator iter = iter_range.first;
TexCache::iterator oldest_entry = iter;
int temp_frameCount = 0x7fffffff;
TexCache::iterator unconverted_copy = textures_by_address.end();
while (iter != iter_range.second)
{
TCacheEntryBase* entry = iter->second;
// Do not load strided EFB copies, they are not meant to be used directly
if (entry->IsEfbCopy() && entry->native_width == nativeW && entry->native_height == nativeH &&
entry->memory_stride == entry->CacheLinesPerRow() * 32)
{
// EFB copies have slightly different rules as EFB copy formats have different
// meanings from texture formats.
if ((base_hash == entry->hash && (!isPaletteTexture || g_Config.backend_info.bSupportsPaletteConversion)) ||
IsPlayingBackFifologWithBrokenEFBCopies)
{
// TODO: We should check format/width/height/levels for EFB copies. Checking
// format is complicated because EFB copy formats don't exactly match
// texture formats. I'm not sure what effect checking width/height/levels
// would have.
if (!isPaletteTexture || !g_Config.backend_info.bSupportsPaletteConversion)
return ReturnEntry(stage, entry);
// Note that we found an unconverted EFB copy, then continue. We'll
// perform the conversion later. Currently, we only convert EFB copies to
// palette textures; we could do other conversions if it proved to be
// beneficial.
unconverted_copy = iter;
}
else
{
// Aggressively prune EFB copies: if it isn't useful here, it will probably
// never be useful again. It's theoretically possible for a game to do
// something weird where the copy could become useful in the future, but in
// practice it doesn't happen.
iter = FreeTexture(iter);
continue;
}
}
else
{
// For normal textures, all texture parameters need to match
if (entry->hash == full_hash && entry->format == full_format && entry->native_levels >= tex_levels &&
entry->native_width == nativeW && entry->native_height == nativeH)
{
entry = DoPartialTextureUpdates(iter);
return ReturnEntry(stage, entry);
}
}
// Find the texture which hasn't been used for the longest time. Count paletted
// textures as the same texture here, when the texture itself is the same. This
// improves the performance a lot in some games that use paletted textures.
// Example: Sonic the Fighters (inside Sonic Gems Collection)
// Skip EFB copies here, so they can be used for partial texture updates
if (entry->frameCount != FRAMECOUNT_INVALID && entry->frameCount < temp_frameCount &&
!entry->IsEfbCopy() && !(isPaletteTexture && entry->base_hash == base_hash))
{
temp_frameCount = entry->frameCount;
oldest_entry = iter;
}
++iter;
}
if (unconverted_copy != textures_by_address.end())
{
// Perform palette decoding.
TCacheEntryBase *entry = unconverted_copy->second;
TCacheEntryConfig config;
config.rendertarget = true;
config.width = entry->config.width;
config.height = entry->config.height;
config.layers = FramebufferManagerBase::GetEFBLayers();
TCacheEntryBase *decoded_entry = AllocateTexture(config);
decoded_entry->SetGeneralParameters(address, texture_size, full_format);
decoded_entry->SetDimensions(entry->native_width, entry->native_height, 1);
decoded_entry->SetHashes(base_hash, full_hash);
decoded_entry->frameCount = FRAMECOUNT_INVALID;
decoded_entry->is_efb_copy = false;
g_texture_cache->ConvertTexture(decoded_entry, entry, &texMem[tlutaddr], (TlutFormat)tlutfmt);
textures_by_address.emplace((u64)address, decoded_entry);
return ReturnEntry(stage, decoded_entry);
}
// Search the texture cache for normal textures by hash
//
// If the texture was fully hashed, the address does not need to match. Identical duplicate textures cause unnecessary slowdowns
// Example: Tales of Symphonia (GC) uses over 500 small textures in menus, but only around 70 different ones
if (g_ActiveConfig.iSafeTextureCache_ColorSamples == 0 ||
std::max(texture_size, palette_size) <= (u32)g_ActiveConfig.iSafeTextureCache_ColorSamples * 8)
{
iter_range = textures_by_hash.equal_range(full_hash);
iter = iter_range.first;
while (iter != iter_range.second)
{
TCacheEntryBase* entry = iter->second;
// All parameters, except the address, need to match here
if (entry->format == full_format && entry->native_levels >= tex_levels &&
entry->native_width == nativeW && entry->native_height == nativeH)
{
entry = DoPartialTextureUpdates(iter);
return ReturnEntry(stage, entry);
}
++iter;
}
}
// If at least one entry was not used for the same frame, overwrite the oldest one
if (temp_frameCount != 0x7fffffff)
{
// pool this texture and make a new one later
FreeTexture(oldest_entry);
}
std::shared_ptr<HiresTexture> hires_tex;
if (g_ActiveConfig.bHiresTextures)
{
hires_tex = HiresTexture::Search(
src_data, texture_size,
&texMem[tlutaddr], palette_size,
width, height,
texformat, use_mipmaps
);
if (hires_tex)
{
auto& l = hires_tex->m_levels[0];
if (l.width != width || l.height != height)
{
width = l.width;
height = l.height;
}
expandedWidth = l.width;
expandedHeight = l.height;
CheckTempSize(l.data_size);
memcpy(temp, l.data, l.data_size);
}
}
if (!hires_tex)
{
if (!(texformat == GX_TF_RGBA8 && from_tmem))
{
const u8* tlut = &texMem[tlutaddr];
TexDecoder_Decode(temp, src_data, expandedWidth, expandedHeight, texformat, tlut, (TlutFormat)tlutfmt);
}
else
{
u8* src_data_gb = &texMem[bpmem.tex[stage / 4].texImage2[stage % 4].tmem_odd * TMEM_LINE_SIZE];
TexDecoder_DecodeRGBA8FromTmem(temp, src_data, src_data_gb, expandedWidth, expandedHeight);
}
}
// how many levels the allocated texture shall have
const u32 texLevels = hires_tex ? (u32)hires_tex->m_levels.size() : tex_levels;
// create the entry/texture
TCacheEntryConfig config;
config.width = width;
config.height = height;
config.levels = texLevels;
TCacheEntryBase* entry = AllocateTexture(config);
GFX_DEBUGGER_PAUSE_AT(NEXT_NEW_TEXTURE, true);
iter = textures_by_address.emplace((u64)address, entry);
if (g_ActiveConfig.iSafeTextureCache_ColorSamples == 0 ||
std::max(texture_size, palette_size) <= (u32)g_ActiveConfig.iSafeTextureCache_ColorSamples * 8)
{
entry->textures_by_hash_iter = textures_by_hash.emplace(full_hash, entry);
}
entry->SetGeneralParameters(address, texture_size, full_format);
entry->SetDimensions(nativeW, nativeH, tex_levels);
entry->SetHashes(base_hash, full_hash);
entry->is_efb_copy = false;
entry->is_custom_tex = hires_tex != nullptr;
// load texture
entry->Load(width, height, expandedWidth, 0);
std::string basename = "";
if (g_ActiveConfig.bDumpTextures && !hires_tex)
{
basename = HiresTexture::GenBaseName(
src_data, texture_size,
&texMem[tlutaddr], palette_size,
width, height,
texformat, use_mipmaps,
true
);
DumpTexture(entry, basename, 0);
}
if (hires_tex)
{
for (u32 level = 1; level != texLevels; ++level)
{
auto& l = hires_tex->m_levels[level];
CheckTempSize(l.data_size);
memcpy(temp, l.data, l.data_size);
entry->Load(l.width, l.height, l.width, level);
}
}
else
{
// load mips - TODO: Loading mipmaps from tmem is untested!
src_data += texture_size;
const u8* ptr_even = nullptr;
const u8* ptr_odd = nullptr;
if (from_tmem)
{
ptr_even = &texMem[bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE + texture_size];
ptr_odd = &texMem[bpmem.tex[stage / 4].texImage2[stage % 4].tmem_odd * TMEM_LINE_SIZE];
}
for (u32 level = 1; level != texLevels; ++level)
{
const u32 mip_width = CalculateLevelSize(width, level);
const u32 mip_height = CalculateLevelSize(height, level);
const u32 expanded_mip_width = ROUND_UP(mip_width, bsw);
const u32 expanded_mip_height = ROUND_UP(mip_height, bsh);
const u8*& mip_src_data = from_tmem
? ((level % 2) ? ptr_odd : ptr_even)
: src_data;
const u8* tlut = &texMem[tlutaddr];
TexDecoder_Decode(temp, mip_src_data, expanded_mip_width, expanded_mip_height, texformat, tlut, (TlutFormat)tlutfmt);
mip_src_data += TexDecoder_GetTextureSizeInBytes(expanded_mip_width, expanded_mip_height, texformat);
entry->Load(mip_width, mip_height, expanded_mip_width, level);
if (g_ActiveConfig.bDumpTextures)
DumpTexture(entry, basename, level);
}
}
INCSTAT(stats.numTexturesUploaded);
SETSTAT(stats.numTexturesAlive, textures_by_address.size());
entry = DoPartialTextureUpdates(iter);
return ReturnEntry(stage, entry);
}
void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFormat,
PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
bool isIntensity, bool scaleByHalf, unsigned int cbufid,
const float *colmat)
{
g_renderer->ResetAPIState();
// stretch picture with increased internal resolution
const D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, (float)config.width, (float)config.height);
D3D::context->RSSetViewports(1, &vp);
// set transformation
if (nullptr == efbcopycbuf[cbufid])
{
const D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(28 * 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::stateman->SetPixelConstants(efbcopycbuf[cbufid]);
const TargetRectangle targetSource = g_renderer->ConvertEFBRectangle(srcRect);
// 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 (scaleByHalf)
D3D::SetLinearCopySampler();
else
D3D::SetPointCopySampler();
// Make sure we don't draw with the texture set as both a source and target.
// (This can happen because we don't unbind textures when we free them.)
D3D::stateman->UnsetTexture(texture->GetSRV());
D3D::context->OMSetRenderTargets(1, &texture->GetRTV(), nullptr);
// Create texture copy
D3D::drawShadedTexQuad(
(srcFormat == PEControl::Z24) ? FramebufferManager::GetEFBDepthTexture()->GetSRV() : FramebufferManager::GetEFBColorTexture()->GetSRV(),
&sourcerect, Renderer::GetTargetWidth(), Renderer::GetTargetHeight(),
(srcFormat == PEControl::Z24) ? PixelShaderCache::GetDepthMatrixProgram(true) : PixelShaderCache::GetColorMatrixProgram(true),
VertexShaderCache::GetSimpleVertexShader(), VertexShaderCache::GetSimpleInputLayout(), GeometryShaderCache::GetCopyGeometryShader());
D3D::context->OMSetRenderTargets(1, &FramebufferManager::GetEFBColorTexture()->GetRTV(), FramebufferManager::GetEFBDepthTexture()->GetDSV());
g_renderer->RestoreAPIState();
if (!g_ActiveConfig.bSkipEFBCopyToRam)
{
u8* dst = Memory::GetPointer(dstAddr);
size_t encoded_size = g_encoder->Encode(dst, dstFormat, srcFormat, srcRect, isIntensity, scaleByHalf);
size_in_bytes = (u32)encoded_size;
TextureCache::MakeRangeDynamic(dstAddr, (u32)encoded_size);
this->hash = GetHash64(dst, (int)encoded_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
}
}
void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFormat,
unsigned int srcFormat, const EFBRectangle& srcRect,
bool isIntensity, bool scaleByHalf, unsigned int cbufid,
const float *colmat)
{
const LPDIRECT3DTEXTURE9 read_texture = (srcFormat == PIXELFMT_Z24) ?
FramebufferManager::GetEFBDepthTexture() :
FramebufferManager::GetEFBColorTexture();
if (type != TCET_EC_DYNAMIC || g_ActiveConfig.bCopyEFBToTexture)
{
LPDIRECT3DSURFACE9 Rendersurf = NULL;
texture->GetSurfaceLevel(0, &Rendersurf);
D3D::dev->SetDepthStencilSurface(NULL);
D3D::dev->SetRenderTarget(0, Rendersurf);
D3DVIEWPORT9 vp;
// Stretch picture with increased internal resolution
vp.X = 0;
vp.Y = 0;
vp.Width = virtual_width;
vp.Height = virtual_height;
vp.MinZ = 0.0f;
vp.MaxZ = 1.0f;
D3D::dev->SetViewport(&vp);
RECT destrect;
destrect.bottom = virtual_height;
destrect.left = 0;
destrect.right = virtual_width;
destrect.top = 0;
PixelShaderManager::SetColorMatrix(colmat); // set transformation
TargetRectangle targetSource = g_renderer->ConvertEFBRectangle(srcRect);
RECT sourcerect;
sourcerect.bottom = targetSource.bottom;
sourcerect.left = targetSource.left;
sourcerect.right = targetSource.right;
sourcerect.top = targetSource.top;
if (srcFormat == PIXELFMT_Z24)
{
if (scaleByHalf || g_ActiveConfig.iMultisampleMode)
{
D3D::ChangeSamplerState(0, D3DSAMP_MINFILTER, D3DTEXF_LINEAR);
D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR);
}
else
{
D3D::ChangeSamplerState(0, D3DSAMP_MINFILTER, D3DTEXF_POINT);
D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_POINT);
}
}
else
{
D3D::ChangeSamplerState(0, D3DSAMP_MINFILTER, D3DTEXF_LINEAR);
D3D::ChangeSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR);
}
D3DFORMAT bformat = FramebufferManager::GetEFBDepthRTSurfaceFormat();
int SSAAMode = g_ActiveConfig.iMultisampleMode;
D3D::drawShadedTexQuad(read_texture, &sourcerect,
Renderer::GetTargetWidth(), Renderer::GetTargetHeight(),
virtual_width, virtual_height,
// TODO: why is D3DFMT_D24X8 singled out here? why not D3DFMT_D24X4S4/D24S8/D24FS8/D32/D16/D15S1 too, or none of them?
PixelShaderCache::GetDepthMatrixProgram(SSAAMode, (srcFormat == PIXELFMT_Z24) && bformat != FOURCC_RAWZ && bformat != D3DFMT_D24X8),
VertexShaderCache::GetSimpleVertexShader(SSAAMode));
Rendersurf->Release();
}
if (!g_ActiveConfig.bCopyEFBToTexture)
{
int encoded_size = TextureConverter::EncodeToRamFromTexture(
addr,
read_texture,
Renderer::GetTargetWidth(),
Renderer::GetTargetHeight(),
srcFormat == PIXELFMT_Z24,
isIntensity,
dstFormat,
scaleByHalf,
srcRect);
u8* dst = Memory::GetPointer(addr);
hash = GetHash64(dst,encoded_size,g_ActiveConfig.iSafeTextureCache_ColorSamples);
// Mark texture entries in destination address range dynamic unless caching is enabled and the texture entry is up to date
if (!g_ActiveConfig.bEFBCopyCacheEnable)
TextureCache::MakeRangeDynamic(addr,encoded_size);
else if (!TextureCache::Find(addr, hash))
TextureCache::MakeRangeDynamic(addr,encoded_size);
}
D3D::RefreshSamplerState(0, D3DSAMP_MINFILTER);
D3D::RefreshSamplerState(0, D3DSAMP_MAGFILTER);
D3D::SetTexture(0, NULL);
D3D::dev->SetRenderTarget(0, FramebufferManager::GetEFBColorRTSurface());
D3D::dev->SetDepthStencilSurface(FramebufferManager::GetEFBDepthRTSurface());
}
void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFormat,
PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
bool isIntensity, bool scaleByHalf, unsigned int cbufid,
const float *colmat)
{
if (type != TCET_EC_DYNAMIC || g_ActiveConfig.bCopyEFBToTexture)
{
g_renderer->ResetAPIState();
// stretch picture with increased internal resolution
const D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, (float)virtual_width, (float)virtual_height);
D3D::context->RSSetViewports(1, &vp);
// set transformation
if (nullptr == efbcopycbuf[cbufid])
{
const D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(28 * 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::stateman->SetPixelConstants(efbcopycbuf[cbufid]);
const TargetRectangle targetSource = g_renderer->ConvertEFBRectangle(srcRect);
// 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 (scaleByHalf)
D3D::SetLinearCopySampler();
else
D3D::SetPointCopySampler();
// if texture is currently in use, it needs to be temporarily unset
u32 textureSlotMask = D3D::stateman->UnsetTexture(texture->GetSRV());
D3D::stateman->Apply();
D3D::context->OMSetRenderTargets(1, &texture->GetRTV(), nullptr);
// Create texture copy
D3D::drawShadedTexQuad(
(srcFormat == PEControl::Z24) ? FramebufferManager::GetEFBDepthTexture()->GetSRV() : FramebufferManager::GetEFBColorTexture()->GetSRV(),
&sourcerect, Renderer::GetTargetWidth(), Renderer::GetTargetHeight(),
(srcFormat == PEControl::Z24) ? PixelShaderCache::GetDepthMatrixProgram(true) : PixelShaderCache::GetColorMatrixProgram(true),
VertexShaderCache::GetSimpleVertexShader(), VertexShaderCache::GetSimpleInputLayout(), GeometryShaderCache::GetCopyGeometryShader());
D3D::context->OMSetRenderTargets(1, &FramebufferManager::GetEFBColorTexture()->GetRTV(), FramebufferManager::GetEFBDepthTexture()->GetDSV());
g_renderer->RestoreAPIState();
// Restore old texture in all previously used slots, if any
D3D::stateman->SetTextureByMask(textureSlotMask, texture->GetSRV());
}
if (!g_ActiveConfig.bCopyEFBToTexture)
{
u8* dst = Memory::GetPointer(dstAddr);
size_t encoded_size = g_encoder->Encode(dst, dstFormat, srcFormat, srcRect, isIntensity, scaleByHalf);
u64 hash = GetHash64(dst, (int)encoded_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
// Mark texture entries in destination address range dynamic unless caching is enabled and the texture entry is up to date
if (!g_ActiveConfig.bEFBCopyCacheEnable)
TextureCache::MakeRangeDynamic(addr, (u32)encoded_size);
else if (!TextureCache::Find(addr, hash))
TextureCache::MakeRangeDynamic(addr, (u32)encoded_size);
this->hash = hash;
}
}