Decode EFB copies used as paletted textures.

A number of games make an EFB copy in I4/I8 format, then use it as a
texture in C4/C8 format.  Detect when this happens, and decode the copy on
the GPU using the specified palette.

This has a few advantages: it allows using EFB2Tex for a few more games,
it, it preserves the resolution of scaled EFB copies, and it's probably a
bit faster.

D3D only at the moment, but porting to OpenGL should be straightforward..

Source/Core/VideoBackends/D3D/TextureCache.cpp

@@ -4,6 +4,7 @@
 
 #include "Core/HW/Memmap.h"
 #include "VideoBackends/D3D/D3DBase.h"
+#include "VideoBackends/D3D/D3DShader.h"
 #include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/D3DUtil.h"
 #include "VideoBackends/D3D/FramebufferManager.h"
@@ -14,6 +15,7 @@
 #include "VideoBackends/D3D/TextureEncoder.h"
 #include "VideoBackends/D3D/VertexShaderCache.h"
 #include "VideoCommon/ImageWrite.h"
+#include "VideoCommon/LookUpTables.h"
 #include "VideoCommon/RenderBase.h"
 #include "VideoCommon/VideoConfig.h"
 
@@ -179,17 +181,167 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 
 		size_in_bytes = (u32)encoded_size;
 
-		TextureCache::MakeRangeDynamic(addr, (u32)encoded_size);
+		TextureCache::MakeRangeDynamic(dstAddr, (u32)encoded_size);
 
 		this->hash = hash;
 	}
 }
 
+const char palette_shader[] =
+R"HLSL(
+sampler samp0 : register(s0);
+Texture2DArray Tex0 : register(t0);
+Buffer<uint> Tex1 : register(t1);
+uniform float Multiply;
+
+uint Convert3To8(uint v)
+{
+	// Swizzle bits: 00000123 -> 12312312
+	return (v << 5) | (v << 2) | (v >> 1);
+}
+
+uint Convert4To8(uint v)
+{
+	// Swizzle bits: 00001234 -> 12341234
+	return (v << 4) | v;
+}
+
+uint Convert5To8(uint v)
+{
+	// Swizzle bits: 00012345 -> 12345123
+	return (v << 3) | (v >> 2);
+}
+
+uint Convert6To8(uint v)
+{
+	// Swizzle bits: 00123456 -> 12345612
+	return (v << 2) | (v >> 4);
+}
+
+float4 DecodePixel_RGB5A3(uint val)
+{
+	int r,g,b,a;
+	if ((val&0x8000))
+	{
+		r=Convert5To8((val>>10) & 0x1f);
+		g=Convert5To8((val>>5 ) & 0x1f);
+		b=Convert5To8((val    ) & 0x1f);
+		a=0xFF;
+	}
+	else
+	{
+		a=Convert3To8((val>>12) & 0x7);
+		r=Convert4To8((val>>8 ) & 0xf);
+		g=Convert4To8((val>>4 ) & 0xf);
+		b=Convert4To8((val    ) & 0xf);
+	}
+	return float4(r, g, b, a) / 255;
+}
+
+float4 DecodePixel_RGB565(uint val)
+{
+	int r, g, b, a;
+	r = Convert5To8((val >> 11) & 0x1f);
+	g = Convert6To8((val >> 5) & 0x3f);
+	b = Convert5To8((val) & 0x1f);
+	a = 0xFF;
+	return float4(r, g, b, a) / 255;
+}
+
+float4 DecodePixel_IA8(uint val)
+{
+	int i = val & 0xFF;
+	int a = val >> 8;
+	return float4(i, i, i, a) / 255;
+}
+
+void main(
+	out float4 ocol0 : SV_Target,
+	in float4 pos : SV_Position,
+	in float3 uv0 : TEXCOORD0)
+{
+	uint src = round(Tex0.Sample(samp0,uv0) * Multiply).r;
+	src = Tex1.Load(src);
+	src = ((src << 8) & 0xFF00) | (src >> 8);
+	ocol0 = DECODE(src);
+}
+)HLSL";
+
+void TextureCache::ConvertTexture(TCacheEntryBase* entry, TCacheEntryBase* unconverted, void* palette, TlutFormat format)
+{
+	g_renderer->ResetAPIState();
+
+	// stretch picture with increased internal resolution
+	const D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, (float)unconverted->config.width, (float)unconverted->config.height);
+	D3D::context->RSSetViewports(1, &vp);
+
+	D3D11_BOX box{ 0, 0, 0, 512, 1, 1 };
+	D3D::context->UpdateSubresource(palette_buf, 0, &box, palette, 0, 0);
+
+	D3D::stateman->SetTexture(1, palette_buf_srv);
+
+	// TODO: Add support for C14X2 format.  (Different multiplier, more palette entries.)
+	float params[4] = { unconverted->format == 0 ? 15.f : 255.f };
+	D3D::context->UpdateSubresource(palette_uniform, 0, nullptr, &params, 0, 0);
+	D3D::stateman->SetPixelConstants(palette_uniform);
+
+	const D3D11_RECT sourcerect = CD3D11_RECT(0, 0, unconverted->config.width, unconverted->config.height);
+
+	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(static_cast<TCacheEntry*>(entry)->texture->GetSRV());
+
+	D3D::context->OMSetRenderTargets(1, &static_cast<TCacheEntry*>(entry)->texture->GetRTV(), nullptr);
+
+	// Create texture copy
+	D3D::drawShadedTexQuad(
+		static_cast<TCacheEntry*>(unconverted)->texture->GetSRV(),
+		&sourcerect, unconverted->config.width, unconverted->config.height,
+		palette_pixel_shader[format],
+		VertexShaderCache::GetSimpleVertexShader(), VertexShaderCache::GetSimpleInputLayout(),
+		GeometryShaderCache::GetCopyGeometryShader());
+
+	D3D::context->OMSetRenderTargets(1, &FramebufferManager::GetEFBColorTexture()->GetRTV(), FramebufferManager::GetEFBDepthTexture()->GetDSV());
+
+	g_renderer->RestoreAPIState();
+}
+
+ID3D11PixelShader *GetConvertShader(const char* Type)
+{
+	std::string shader = "#define DECODE DecodePixel_";
+	shader.append(Type);
+	shader.append("\n");
+	shader.append(palette_shader);
+	return D3D::CompileAndCreatePixelShader(shader);
+}
+
 TextureCache::TextureCache()
 {
 	// FIXME: Is it safe here?
 	g_encoder = new PSTextureEncoder;
 	g_encoder->Init();
+
+	palette_buf = nullptr;
+	palette_buf_srv = nullptr;
+	palette_uniform = nullptr;
+	palette_pixel_shader[GX_TL_IA8] = GetConvertShader("IA8");
+	palette_pixel_shader[GX_TL_RGB565] = GetConvertShader("RGB565");
+	palette_pixel_shader[GX_TL_RGB5A3] = GetConvertShader("RGB5A3");
+	auto lutBd = CD3D11_BUFFER_DESC(sizeof(u16) * 256, D3D11_BIND_SHADER_RESOURCE);
+	HRESULT hr = D3D::device->CreateBuffer(&lutBd, nullptr, &palette_buf);
+	CHECK(SUCCEEDED(hr), "create palette decoder lut buffer");
+	D3D::SetDebugObjectName(palette_buf, "texture decoder lut buffer");
+	// TODO: C14X2 format.
+	auto outlutUavDesc = CD3D11_SHADER_RESOURCE_VIEW_DESC(palette_buf, DXGI_FORMAT_R16_UINT, 0, 256, 0);
+	hr = D3D::device->CreateShaderResourceView(palette_buf, &outlutUavDesc, &palette_buf_srv);
+	CHECK(SUCCEEDED(hr), "create palette decoder lut srv");
+	D3D::SetDebugObjectName(palette_buf_srv, "texture decoder lut srv");
+	const D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(16, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DEFAULT);
+	hr = D3D::device->CreateBuffer(&cbdesc, nullptr, &palette_uniform);
+	CHECK(SUCCEEDED(hr), "Create palette decoder constant buffer");
+	D3D::SetDebugObjectName((ID3D11DeviceChild*)palette_uniform, "a constant buffer used in TextureCache::CopyRenderTargetToTexture");
 }
 
 TextureCache::~TextureCache()
@@ -200,6 +352,12 @@ TextureCache::~TextureCache()
 	g_encoder->Shutdown();
 	delete g_encoder;
 	g_encoder = nullptr;
+
+	SAFE_RELEASE(palette_buf);
+	SAFE_RELEASE(palette_buf_srv);
+	SAFE_RELEASE(palette_uniform);
+	for (ID3D11PixelShader*& shader : palette_pixel_shader)
+		SAFE_RELEASE(shader);
 }
 
 }

Source/Core/VideoBackends/D3D/TextureCache.h

@@ -42,8 +42,15 @@ private:
 
 	u64 EncodeToRamFromTexture(u32 address, void* source_texture, u32 SourceW, u32 SourceH, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source) {return 0;};
 
+	void ConvertTexture(TCacheEntryBase* entry, TCacheEntryBase* unconverted, void* palette, TlutFormat format) override;
+
 	void CompileShaders() override { }
 	void DeleteShaders() override { }
+
+	ID3D11Buffer* palette_buf;
+	ID3D11ShaderResourceView* palette_buf_srv;
+	ID3D11Buffer* palette_uniform;
+	ID3D11PixelShader* palette_pixel_shader[3];
 };
 
 }

Source/Core/VideoBackends/D3D/main.cpp

@@ -82,6 +82,7 @@ void InitBackendInfo()
 	g_Config.backend_info.bSupportsGeometryShaders = true;
 	g_Config.backend_info.bSupports3DVision = true;
 	g_Config.backend_info.bSupportsPostProcessing = false;
+	g_Config.backend_info.bSupportsPaletteConversion = true;
 
 	IDXGIFactory* factory;
 	IDXGIAdapter* ad;

Source/Core/VideoBackends/OGL/TextureCache.cpp

@@ -204,7 +204,7 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 	if (false == g_ActiveConfig.bCopyEFBToTexture)
 	{
 		int encoded_size = TextureConverter::EncodeToRamFromTexture(
-			addr,
+			dstAddr,
 			read_texture,
 			srcFormat == PEControl::Z24,
 			isIntensity,
@@ -212,12 +212,12 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 			scaleByHalf,
 			srcRect);
 
-		u8* dst = Memory::GetPointer(addr);
+		u8* dst = Memory::GetPointer(dstAddr);
 		u64 const new_hash = GetHash64(dst,encoded_size,g_ActiveConfig.iSafeTextureCache_ColorSamples);
 
 		size_in_bytes = (u32)encoded_size;
 
-		TextureCache::MakeRangeDynamic(addr,encoded_size);
+		TextureCache::MakeRangeDynamic(dstAddr, encoded_size);
 
 		hash = new_hash;
 	}
@@ -359,4 +359,10 @@ void TextureCache::DeleteShaders()
 	s_DepthMatrixProgram.Destroy();
 }
 
+void TextureCache::ConvertTexture(TextureCache::TCacheEntryBase* entry, TCacheEntryBase* unconverted, void* palette, TlutFormat format)
+{
+	// TODO: Implement.
+	return;
+}
+
 }

Source/Core/VideoBackends/OGL/TextureCache.h

@@ -48,6 +48,7 @@ private:
 	~TextureCache();
 
 	TCacheEntryBase* CreateTexture(const TCacheEntryConfig& config) override;
+	void ConvertTexture(TCacheEntryBase* entry, TCacheEntryBase* unconverted, void* palette, TlutFormat format) override;
 
 	void CompileShaders() override;
 	void DeleteShaders() override;

Source/Core/VideoBackends/OGL/main.cpp

@@ -140,6 +140,7 @@ static void InitBackendInfo()
 	g_Config.backend_info.bSupportsGeometryShaders = true;
 	g_Config.backend_info.bSupports3DVision = false;
 	g_Config.backend_info.bSupportsPostProcessing = true;
+	g_Config.backend_info.bSupportsPaletteConversion = false;
 
 	g_Config.backend_info.Adapters.clear();
 

Source/Core/VideoCommon/TextureCacheBase.cpp

@@ -32,6 +32,7 @@ size_t TextureCache::temp_size;
 
 TextureCache::TexCache TextureCache::textures;
 TextureCache::TexPool TextureCache::texture_pool;
+TextureCache::TCacheEntryBase* TextureCache::bound_textures[8];
 
 TextureCache::BackupConfig TextureCache::backup_config;
 
@@ -74,6 +75,8 @@ void TextureCache::RequestInvalidateTextureCache()
 
 void TextureCache::Invalidate()
 {
+	UnbindTextures();
+
 	for (auto& tex : textures)
 	{
 		delete tex.second;
@@ -143,7 +146,7 @@ void TextureCache::Cleanup(int _frameCount)
 		}
 		if (_frameCount > TEXTURE_KILL_THRESHOLD + iter->second->frameCount &&
 		    // EFB copies living on the host GPU are unrecoverable and thus shouldn't be deleted
-		    !iter->second->IsEfbCopy())
+		    !iter->second->IsUnrecoverable())
 		{
 			FreeTexture(iter->second);
 			iter = textures.erase(iter);
@@ -174,17 +177,17 @@ void TextureCache::Cleanup(int _frameCount)
 	}
 }
 
-void TextureCache::InvalidateRange(u32 start_address, u32 size)
+void TextureCache::MakeRangeDynamic(u32 start_address, u32 size)
 {
 	TexCache::iterator
-		iter = textures.begin(),
+		iter = textures.begin();
-		tcend = textures.end();
+
-	while (iter != tcend)
+	while (iter != textures.end())
 	{
 		if (iter->second->OverlapsMemoryRange(start_address, size))
 		{
 			FreeTexture(iter->second);
-			textures.erase(iter++);
+			iter = textures.erase(iter);
 		}
 		else
 		{
@@ -193,55 +196,21 @@ void TextureCache::InvalidateRange(u32 start_address, u32 size)
 	}
 }
 
-void TextureCache::MakeRangeDynamic(u32 start_address, u32 size)
-{
-	TexCache::iterator
-		iter = textures.lower_bound(start_address),
-		tcend = textures.upper_bound(start_address + size);
-
-	if (iter != textures.begin())
-		--iter;
-
-	for (; iter != tcend; ++iter)
-	{
-		if (iter->second->OverlapsMemoryRange(start_address, size))
-		{
-			iter->second->SetHashes(TEXHASH_INVALID);
-		}
-	}
-}
-
 bool TextureCache::TCacheEntryBase::OverlapsMemoryRange(u32 range_address, u32 range_size) const
 {
-	if (addr + size_in_bytes <= range_address)
+	if (!addr.HasMemAddress())
 		return false;
 
-	if (addr >= range_address + range_size)
+	u32 memaddr = addr.GetMemAddress();
+	if (memaddr + size_in_bytes <= range_address)
+		return false;
+
+	if (memaddr >= range_address + range_size)
 		return false;
 
 	return true;
 }
 
-void TextureCache::ClearRenderTargets()
-{
-	TexCache::iterator
-		iter = textures.begin(),
-		tcend = textures.end();
-
-	while (iter != tcend)
-	{
-		if (iter->second->IsEfbCopy())
-		{
-			FreeTexture(iter->second);
-			textures.erase(iter++);
-		}
-		else
-		{
-			++iter;
-		}
-	}
-}
-
 void TextureCache::DumpTexture(TCacheEntryBase* entry, std::string basename, unsigned int level)
 {
 	std::string szDir = File::GetUserPath(D_DUMPTEXTURES_IDX) +
@@ -267,16 +236,30 @@ static u32 CalculateLevelSize(u32 level_0_size, u32 level)
 }
 
 // Used by TextureCache::Load
-static TextureCache::TCacheEntryBase* ReturnEntry(unsigned int stage, TextureCache::TCacheEntryBase* entry)
+TextureCache::TCacheEntryBase* TextureCache::ReturnEntry(unsigned int stage, TCacheEntryBase* entry)
 {
 	entry->frameCount = FRAMECOUNT_INVALID;
-	entry->Bind(stage);
+	bound_textures[stage] = entry;
 
 	GFX_DEBUGGER_PAUSE_AT(NEXT_TEXTURE_CHANGE, true);
 
 	return entry;
 }
 
+void TextureCache::BindTextures()
+{
+	for (int i = 0; i < 8; ++i)
+	{
+		if (bound_textures[i])
+			bound_textures[i]->Bind(i);
+	}
+}
+
+void TextureCache::UnbindTextures()
+{
+	std::fill(std::begin(bound_textures), std::end(bound_textures), nullptr);
+}
+
 TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 {
 	const FourTexUnits &tex = bpmem.tex[stage >> 2];
@@ -309,6 +292,11 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	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);
 
@@ -323,19 +311,43 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	// 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);
 	u32 palette_size = 0;
+	u64 tlut_hash = 0;
 	if (isPaletteTexture)
 	{
 		palette_size = TexDecoder_GetPaletteSize(texformat);
-		u64 tlut_hash = GetHash64(&texMem[tlutaddr], palette_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
+		tlut_hash = GetHash64(&texMem[tlutaddr], palette_size, g_ActiveConfig.iSafeTextureCache_ColorSamples);
-
-		// Mix the tlut hash into the texture hash. So we only have to compare it once.
-		tex_hash ^= tlut_hash;
 	}
 
 	// 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);
 
+	// Compute a texture ID; this isn't everything about a texture, rather just
+	// enough to group together textures with related memory addresses.
+	TextureAddress texID;
+	TextureAddress paletteDecodedID;
+	if (from_tmem)
+	{
+		u32 tmem_addr = bpmem.tex[stage / 4].texImage1[stage % 4].tmem_even * TMEM_LINE_SIZE;
+		if (texformat == GX_TF_RGBA8 && from_tmem)
+		{
+			u32 tmem_odd_addr = bpmem.tex[stage / 4].texImage2[stage % 4].tmem_odd * TMEM_LINE_SIZE;
+			texID = TextureAddress::TMemRGBA8(tmem_addr, tmem_odd_addr);
+		}
+		else
+		{
+			texID = TextureAddress::TMem(tmem_addr);
+			if (isPaletteTexture)
+				paletteDecodedID = TextureAddress::TMemPalette(tmem_addr, tlutaddr);
+		}
+	}
+	else
+	{
+		texID = TextureAddress::Mem(address);
+		if (isPaletteTexture)
+			paletteDecodedID = TextureAddress::MemPalette(address, tlutaddr);
+	}
+
 	// Find all texture cache entries for the current texture address, and decide whether to use one of
 	// them, or to create a new one
 	//
@@ -360,7 +372,13 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	//
 	// 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.equal_range(address);
+	std::pair<TexCache::iterator, TexCache::iterator> iter_range = textures.equal_range(texID);
+	bool palette_decoded_entry = false;
+	if (isPaletteTexture && iter_range.first == iter_range.second)
+	{
+		iter_range = textures.equal_range(paletteDecodedID);
+		palette_decoded_entry = true;
+	}
 	TexCache::iterator iter = iter_range.first;
 	TexCache::iterator oldest_entry = iter;
 	int temp_frameCount = 0x7fffffff;
@@ -370,14 +388,39 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 		TCacheEntryBase* entry = iter->second;
 		if (entry->IsEfbCopy())
 		{
-			// For EFB copies, only the hash and the texture address need to match. Ignore the hash when
+			// EFB copies have slightly different rules: the hash doesn't need to match
-			// using EFB to texture, because there's no hash in this case
+			// in EFB2Tex mode, and EFB copy formats have different meanings from texture
-			if (g_ActiveConfig.bCopyEFBToTexture || entry->hash == tex_hash)
+			// formats.
+			if (g_ActiveConfig.bCopyEFBToTexture ||
+				(tex_hash == entry->hash && (!isPaletteTexture || g_Config.backend_info.bSupportsPaletteConversion)))
 			{
-				// TODO: Print a warning if the format changes! In this case,
+				// TODO: We should check format/width/height/levels for EFB copies. Checking
-				// we could reinterpret the internal texture object data to the new pixel format
+				// format is complicated because EFB copy formats don't exactly match
-				// (similar to what is already being done in Renderer::ReinterpretPixelFormat())
+				// texture formats. I'm not sure what effect checking width/height/levels
-				// TODO: Convert paletted textures, which are efb copies, using the right palette, so they display correctly
+				// would have.
+				if (!palette_decoded_entry && isPaletteTexture && g_Config.backend_info.bSupportsPaletteConversion)
+				{
+					// Perform palette decoding.
+					// TODO: Skip decoding if we find a match.
+					std::pair<TexCache::iterator, TexCache::iterator> decoded_iter_range = textures.equal_range(paletteDecodedID);
+					while (decoded_iter_range.first != decoded_iter_range.second)
+					{
+						// Pool this texture and make a new one later.
+						FreeTexture(decoded_iter_range.first->second);
+						decoded_iter_range.first = textures.erase(decoded_iter_range.first);
+					}
+
+					TCacheEntryBase *decoded_entry = AllocateTexture(entry->config);
+
+					decoded_entry->SetGeneralParameters(paletteDecodedID, texture_size, full_format);
+					decoded_entry->SetDimensions(entry->native_width, entry->native_height, 1);
+					decoded_entry->SetHashes(TEXHASH_INVALID);
+					decoded_entry->frameCount = FRAMECOUNT_INVALID;
+
+					g_texture_cache->ConvertTexture(decoded_entry, entry, &texMem[tlutaddr], (TlutFormat)tlutfmt);
+					textures.insert(TexCache::value_type(paletteDecodedID, decoded_entry));
+					entry = decoded_entry;
+				}
 				return ReturnEntry(stage, entry);
 			}
 			else
@@ -389,12 +432,14 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 				continue;
 			}
 		}
-
+		else
-		// For normal textures, all texture parameters need to match
-		if (entry->hash == tex_hash && entry->format == full_format && entry->native_levels >= tex_levels &&
-			entry->native_width == nativeW && entry->native_height == nativeH)
 		{
-			return ReturnEntry(stage, entry);
+			// For normal textures, all texture parameters need to match
+			if (entry->hash == (tex_hash ^ tlut_hash) && entry->format == full_format && entry->native_levels >= tex_levels &&
+				entry->native_width == nativeW && entry->native_height == nativeH)
+			{
+				return ReturnEntry(stage, entry);
+			}
 		}
 
 		// Find the entry which hasn't been used for the longest time
@@ -468,11 +513,11 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	TCacheEntryBase* entry = AllocateTexture(config);
 	GFX_DEBUGGER_PAUSE_AT(NEXT_NEW_TEXTURE, true);
 
-	textures.insert(TexCache::value_type(address, entry));
+	textures.insert(TexCache::value_type(isPaletteTexture ? paletteDecodedID : texID, entry));
 
-	entry->SetGeneralParameters(address, texture_size, full_format);
+	entry->SetGeneralParameters(isPaletteTexture ? paletteDecodedID : texID, texture_size, full_format);
 	entry->SetDimensions(nativeW, nativeH, tex_levels);
-	entry->hash = tex_hash;
+	entry->hash = tex_hash ^ tlut_hash;
 
 	// load texture
 	entry->Load(width, height, expandedWidth, 0);
@@ -828,7 +873,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 	unsigned int scaled_tex_h = g_ActiveConfig.bCopyEFBScaled ? Renderer::EFBToScaledY(tex_h) : tex_h;
 
 	// remove all texture cache entries at dstAddr
-	std::pair <TexCache::iterator, TexCache::iterator> iter_range = textures.equal_range(dstAddr);
+	std::pair <TexCache::iterator, TexCache::iterator> iter_range = textures.equal_range(TextureAddress::Mem(dstAddr));
 	TexCache::iterator iter = iter_range.first;
 	while (iter != iter_range.second)
 	{
@@ -846,7 +891,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 	TCacheEntryBase* entry = AllocateTexture(config);
 
 	// TODO: Using the wrong dstFormat, dumb...
-	entry->SetGeneralParameters(dstAddr, 0, dstFormat);
+	entry->SetGeneralParameters(TextureAddress::Mem(dstAddr), 0, dstFormat);
 	entry->SetDimensions(tex_w, tex_h, 1);
 	entry->SetHashes(TEXHASH_INVALID);
 
@@ -854,7 +899,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 
 	entry->FromRenderTarget(dstAddr, dstFormat, srcFormat, srcRect, isIntensity, scaleByHalf, cbufid, colmat);
 
-	textures.insert(TexCache::value_type(dstAddr, entry));
+	textures.insert(TexCache::value_type(TextureAddress::Mem(dstAddr), entry));
 }
 
 TextureCache::TCacheEntryBase* TextureCache::AllocateTexture(const TCacheEntryConfig& config)

Source/Core/VideoCommon/TextureCacheBase.h

@@ -43,13 +43,56 @@ public:
 		};
 
 	};
-
+	class TextureAddress
+	{
+		u32 address1;
+		u32 address2;
+		enum AddressKind
+		{
+			// A texture in RAM
+			RAM,
+			// A texture loaded into TMEM
+			TMEM,
+			// A texture in RAM, fully decoded using a palette.
+			RAM_PALETTE,
+			// An RGBA8 texture in TMEM.
+			TMEM_RGBA8,
+			// A palette texture in TMEM.
+			TMEM_PALETTE,
+			// Uninitialized address.
+			INVALID
+		};
+		AddressKind kind;
+		TextureAddress(u32 a, u32 b, AddressKind k) : address1(a), address2(b), kind(k) {}
+	public:
+		TextureAddress() : kind(INVALID), address1(0), address2(0) {}
+		static TextureAddress Mem(u32 a) { return TextureAddress(a, 0, RAM); }
+		static TextureAddress MemPalette(u32 a, u32 b) { return TextureAddress(a, b, RAM_PALETTE); }
+		static TextureAddress TMem(u32 a) { return TextureAddress(a, 0, TMEM); }
+		static TextureAddress TMemRGBA8(u32 a, u32 b) { return TextureAddress(a, b, TMEM_RGBA8); }
+		static TextureAddress TMemPalette(u32 a, u32 b) { return TextureAddress(a, b, TMEM_PALETTE); }
+		bool operator == (const TextureAddress& b) const
+		{
+			return kind == b.kind && address1 == b.address1 && address2 == b.address2;
+		}
+		bool operator < (const TextureAddress& b) const
+		{
+			if (kind != b.kind)
+				return kind < b.kind;
+			if (address1 != b.address1)
+				return address1 < b.address1;
+			 return address2 < b.address2;
+		}
+		bool IsMemOnlyAddress() const { return kind == RAM; }
+		bool HasMemAddress() const { return kind == RAM || kind == RAM_PALETTE; }
+		u32 GetMemAddress() const { return address1; }
+	};
 	struct TCacheEntryBase
 	{
 		const TCacheEntryConfig config;
 
 		// common members
-		u32 addr;
+		TextureAddress addr;
 		u32 size_in_bytes;
 		u64 hash;
 		u32 format;
@@ -61,7 +104,7 @@ public:
 		int frameCount;
 
 
-		void SetGeneralParameters(u32 _addr, u32 _size, u32 _format)
+		void SetGeneralParameters(TextureAddress _addr, u32 _size, u32 _format)
 		{
 			addr = _addr;
 			size_in_bytes = _size;
@@ -96,6 +139,7 @@ public:
 		bool OverlapsMemoryRange(u32 range_address, u32 range_size) const;
 
 		bool IsEfbCopy() { return config.rendertarget; }
+		bool IsUnrecoverable() { return IsEfbCopy() && addr.IsMemOnlyAddress(); }
 	};
 
 	virtual ~TextureCache(); // needs virtual for DX11 dtor
@@ -107,9 +151,7 @@ public:
 	static void Cleanup(int frameCount);
 
 	static void Invalidate();
-	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(const TCacheEntryConfig& config) = 0;
 
@@ -117,11 +159,15 @@ public:
 	virtual void DeleteShaders() = 0; // currently only implemented by OGL
 
 	static TCacheEntryBase* Load(const u32 stage);
+	static void UnbindTextures();
+	static void BindTextures();
 	static void CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat, PEControl::PixelFormat srcFormat,
 		const EFBRectangle& srcRect, bool isIntensity, bool scaleByHalf);
 
 	static void RequestInvalidateTextureCache();
 
+	virtual void ConvertTexture(TCacheEntryBase* entry, TCacheEntryBase* unconverted, void* palette, TlutFormat format) = 0;
+
 protected:
 	TextureCache();
 
@@ -135,11 +181,14 @@ private:
 	static TCacheEntryBase* AllocateTexture(const TCacheEntryConfig& config);
 	static void FreeTexture(TCacheEntryBase* entry);
 
-	typedef std::multimap<u32, TCacheEntryBase*> TexCache;
+	static TCacheEntryBase* ReturnEntry(unsigned int stage, TCacheEntryBase* entry);
+
+	typedef std::multimap<TextureAddress, TCacheEntryBase*> TexCache;
 	typedef std::unordered_multimap<TCacheEntryConfig, TCacheEntryBase*, TCacheEntryConfig::Hasher> TexPool;
 
 	static TexCache textures;
 	static TexPool texture_pool;
+	static TCacheEntryBase* bound_textures[8];
 
 	// Backup configuration values
 	static struct BackupConfig

Source/Core/VideoCommon/VertexManagerBase.cpp

@@ -209,6 +209,7 @@ void VertexManager::Flush()
 				if (bpmem.tevind[i].IsActive() && bpmem.tevind[i].bt < bpmem.genMode.numindstages)
 					usedtextures[bpmem.tevindref.getTexMap(bpmem.tevind[i].bt)] = true;
 
+		TextureCache::UnbindTextures();
 		for (unsigned int i : usedtextures)
 		{
 			g_renderer->SetSamplerState(i & 3, i >> 2);
@@ -224,6 +225,7 @@ void VertexManager::Flush()
 				ERROR_LOG(VIDEO, "error loading texture");
 			}
 		}
+		TextureCache::BindTextures();
 	}
 
 	// set global vertex constants

Source/Core/VideoCommon/VideoConfig.h

@@ -159,6 +159,7 @@ struct VideoConfig final
 		bool bSupportsBBox;
 		bool bSupportsGSInstancing; // Needed by GeometryShaderGen, so must stay in VideoCommon
 		bool bSupportsPostProcessing;
+		bool bSupportsPaletteConversion;
 	} backend_info;
 
 	// Utility