Move common EFB copy code into VideoCommon

Addded a few duplicated depth copy texture formats to the enum
in TextureDecoder.h. These texture formats were already implemented
in TextureCacheBase and the ogl/dx11 texture cache implementations.

14 files changed, 151 insertions, 249 deletions

diff --git a/Source/Core/VideoBackends/D3D/PSTextureEncoder.cpp b/Source/Core/VideoBackends/D3D/PSTextureEncoder.cpp
index 3285ff8602..a5fa6de7a5 100644
--- a/Source/Core/VideoBackends/D3D/PSTextureEncoder.cpp
+++ b/Source/Core/VideoBackends/D3D/PSTextureEncoder.cpp
@@ -87,55 +87,21 @@ void PSTextureEncoder::Shutdown()
 	SAFE_RELEASE(m_out);
 }
 
-size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat, size_t dstStride,
+void PSTextureEncoder::Encode(u8* dst, const TextureCache::TCacheEntryBase *texture_entry,
 	PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 	bool isIntensity, bool scaleByHalf)
 {
 	if (!m_ready) // Make sure we initialized OK
-		return 0;
-
-	// Clamp srcRect to 640x528. BPS: The Strike tries to encode an 800x600
-	// texture, which is invalid.
-	EFBRectangle correctSrc = srcRect;
-	correctSrc.ClampUL(0, 0, EFB_WIDTH, EFB_HEIGHT);
-
-	// Validate source rect size
-	if (correctSrc.GetWidth() <= 0 || correctSrc.GetHeight() <= 0)
-		return 0;
+		return;
 
 	HRESULT hr;
 
-	unsigned int blockW = BLOCK_WIDTHS[dstFormat];
-	unsigned int blockH = BLOCK_HEIGHTS[dstFormat];
-
-	// Round up source dims to multiple of block size
-	unsigned int actualWidth = correctSrc.GetWidth() / (scaleByHalf ? 2 : 1);
-	actualWidth = (actualWidth + blockW-1) & ~(blockW-1);
-	unsigned int actualHeight = correctSrc.GetHeight() / (scaleByHalf ? 2 : 1);
-	actualHeight = (actualHeight + blockH-1) & ~(blockH-1);
-
-	unsigned int numBlocksX = actualWidth/blockW;
-	unsigned int numBlocksY = actualHeight/blockH;
-
-	unsigned int cacheLinesPerRow;
-	if (dstFormat == 0x6) // RGBA takes two cache lines per block; all others take one
-		cacheLinesPerRow = numBlocksX*2;
-	else
-		cacheLinesPerRow = numBlocksX;
-	_assert_msg_(VIDEO, cacheLinesPerRow*32 <= MAX_BYTES_PER_BLOCK_ROW, "cache lines per row sanity check");
-
-	unsigned int totalCacheLines = cacheLinesPerRow * numBlocksY;
-	_assert_msg_(VIDEO, totalCacheLines*32 <= MAX_BYTES_PER_ENCODE, "total encode size sanity check");
-
-	size_t encodeSize = 0;
-
 	// Reset API
 	g_renderer->ResetAPIState();
 
 	// Set up all the state for EFB encoding
-
 	{
-		D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, FLOAT(cacheLinesPerRow * 8), FLOAT(numBlocksY));
+		D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, FLOAT(texture_entry->CacheLinesPerRow() * 8), FLOAT(texture_entry->NumBlocksY()));
 		D3D::context->RSSetViewports(1, &vp);
 
 		EFBRectangle fullSrcRect;
@@ -155,9 +121,9 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat, size_t dstStrid
 			FramebufferManager::GetResolvedEFBColorTexture()->GetSRV();
 
 		EFBEncodeParams params;
-		params.SrcLeft = correctSrc.left;
-		params.SrcTop = correctSrc.top;
-		params.DestWidth = actualWidth;
+		params.SrcLeft = srcRect.left;
+		params.SrcTop = srcRect.top;
+		params.DestWidth = texture_entry->native_width;
 		params.ScaleFactor = scaleByHalf ? 2 : 1;
 		D3D::context->UpdateSubresource(m_encodeParams, 0, nullptr, &params, 0, 0);
 		D3D::stateman->SetPixelConstants(m_encodeParams);
@@ -172,12 +138,12 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat, size_t dstStrid
 			targetRect.AsRECT(),
 			Renderer::GetTargetWidth(),
 			Renderer::GetTargetHeight(),
-			SetStaticShader(dstFormat, srcFormat, isIntensity, scaleByHalf),
+			SetStaticShader(texture_entry->format, srcFormat, isIntensity, scaleByHalf),
 			VertexShaderCache::GetSimpleVertexShader(),
 			VertexShaderCache::GetSimpleInputLayout());
 
 		// Copy to staging buffer
-		D3D11_BOX srcBox = CD3D11_BOX(0, 0, 0, cacheLinesPerRow * 8, numBlocksY, 1);
+		D3D11_BOX srcBox = CD3D11_BOX(0, 0, 0, texture_entry->CacheLinesPerRow() * 8, texture_entry->NumBlocksY(), 1);
 		D3D::context->CopySubresourceRegion(m_outStage, 0, 0, 0, 0, m_out, 0, &srcBox);
 
 		// Transfer staging buffer to GameCube/Wii RAM
@@ -186,16 +152,14 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat, size_t dstStrid
 		CHECK(SUCCEEDED(hr), "map staging buffer (0x%x)", hr);
 
 		u8* src = (u8*)map.pData;
-		for (unsigned int y = 0; y < numBlocksY; ++y)
+		for (unsigned int y = 0; y < texture_entry->NumBlocksY(); ++y)
 		{
-			memcpy(dst, src, cacheLinesPerRow*32);
-			dst += dstStride;
+			memcpy(dst, src, texture_entry->CacheLinesPerRow() * 32);
+			dst += texture_entry->memory_stride;
 			src += map.RowPitch;
 		}
 
 		D3D::context->Unmap(m_outStage, 0);
-
-		encodeSize = dstStride * numBlocksY;
 	}
 
 	// Restore API
@@ -203,8 +167,6 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat, size_t dstStrid
 	D3D::context->OMSetRenderTargets(1,
 		&FramebufferManager::GetEFBColorTexture()->GetRTV(),
 		FramebufferManager::GetEFBDepthTexture()->GetDSV());
-
-	return encodeSize;
 }
 
 ID3D11PixelShader* PSTextureEncoder::SetStaticShader(unsigned int dstFormat, PEControl::PixelFormat srcFormat,
diff --git a/Source/Core/VideoBackends/D3D/PSTextureEncoder.h b/Source/Core/VideoBackends/D3D/PSTextureEncoder.h
index 7034a17bb9..2857e9de7e 100644
--- a/Source/Core/VideoBackends/D3D/PSTextureEncoder.h
+++ b/Source/Core/VideoBackends/D3D/PSTextureEncoder.h
@@ -6,6 +6,8 @@
 
 #include "VideoBackends/D3D/TextureEncoder.h"
 
+#include "VideoCommon/TextureCacheBase.h"
+
 struct ID3D11Texture2D;
 struct ID3D11RenderTargetView;
 struct ID3D11Buffer;
@@ -29,7 +31,7 @@ public:
 
 	void Init();
 	void Shutdown();
-	size_t Encode(u8* dst, unsigned int dstFormat, size_t dstStride,
+	void Encode(u8* dst, const TextureCache::TCacheEntryBase *texture_entry,
 	              PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 	              bool isIntensity, bool scaleByHalf);
 
diff --git a/Source/Core/VideoBackends/D3D/TextureCache.cpp b/Source/Core/VideoBackends/D3D/TextureCache.cpp
index 5616a833e2..60a5922f29 100644
--- a/Source/Core/VideoBackends/D3D/TextureCache.cpp
+++ b/Source/Core/VideoBackends/D3D/TextureCache.cpp
@@ -2,7 +2,6 @@
 // Licensed under GPLv2+
 // Refer to the license.txt file included.
 
-#include "Core/HW/Memmap.h"
 #include "VideoBackends/D3D/D3DBase.h"
 #include "VideoBackends/D3D/D3DShader.h"
 #include "VideoBackends/D3D/D3DState.h"
@@ -185,7 +184,7 @@ TextureCache::TCacheEntryBase* TextureCache::CreateTexture(const TCacheEntryConf
 	}
 }
 
-void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFormat, u32 dstStride,
+void TextureCache::TCacheEntry::FromRenderTarget(u8* dst, unsigned int dstFormat, u32 dstStride,
 	PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 	bool isIntensity, bool scaleByHalf, unsigned int cbufid,
 	const float *colmat)
@@ -226,10 +225,13 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 
 	// 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());
+		(srcFormat == PEControl::Z24 ? FramebufferManager::GetEFBDepthTexture() : 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());
 
@@ -237,14 +239,7 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 
 	if (!g_ActiveConfig.bSkipEFBCopyToRam)
 	{
-		u8* dst = Memory::GetPointer(dstAddr);
-		size_t encoded_size = g_encoder->Encode(dst, dstFormat, dstStride, 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);
+		g_encoder->Encode(dst, this, srcFormat, srcRect, isIntensity, scaleByHalf);
 	}
 }
 
diff --git a/Source/Core/VideoBackends/D3D/TextureCache.h b/Source/Core/VideoBackends/D3D/TextureCache.h
index 9cd52afcf6..4fed08ff65 100644
--- a/Source/Core/VideoBackends/D3D/TextureCache.h
+++ b/Source/Core/VideoBackends/D3D/TextureCache.h
@@ -34,7 +34,7 @@ private:
 		void Load(unsigned int width, unsigned int height,
 			unsigned int expanded_width, unsigned int levels) override;
 
-		void FromRenderTarget(u32 dstAddr, unsigned int dstFormat, u32 dstStride,
+		void FromRenderTarget(u8* dst, unsigned int dstFormat, u32 dstStride,
 			PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 			bool isIntensity, bool scaleByHalf, unsigned int cbufid,
 			const float *colmat) override;
diff --git a/Source/Core/VideoBackends/D3D/TextureEncoder.h b/Source/Core/VideoBackends/D3D/TextureEncoder.h
index a2a2a34fdb..445fffd218 100644
--- a/Source/Core/VideoBackends/D3D/TextureEncoder.h
+++ b/Source/Core/VideoBackends/D3D/TextureEncoder.h
@@ -5,96 +5,11 @@
 #pragma once
 
 #include "VideoCommon/BPMemory.h"
+#include "VideoCommon/TextureCacheBase.h"
 #include "VideoCommon/VideoCommon.h"
 
 namespace DX11
 {
-
-// 4-bit format: 8x8 texels / cache line
-// 8-bit format: 8x4 texels / cache line
-// 16-bit format: 4x4 texels / cache line
-// 32-bit format: 4x4 texels / 2 cache lines
-// Compressed format: 8x8 texels / cache line
-
-// Document EFB encoding formats here with examples of where they are used.
-
-// Format: 0 - R4
-// Used in The Legend of Zelda: The Wind Waker for character shadows (srcFormat 1,
-// isIntensity 1, scaleByHalf 1).
-
-// Format: 1 - R8
-// FIXME: Unseen. May or may not be a duplicate of format 8.
-
-// Format: 2 - A4 R4
-// FIXME: Unseen.
-
-// Format: 3 - A8 R8
-// FIXME: Unseen.
-
-// Format: 4 - R5 G6 B5
-// Used in Wind Waker for most render-to-texture effects like heat shimmer and
-// depth-of-field.
-
-// Format: 5 - 1 R5 G5 B5 or 0 A3 R4 G4 B4
-// Used in Twilight Princess for character shadows.
-
-// Format: 6 - A8 R8 A8 R8 | G8 B8 G8 B8
-// Used in Twilight Princess for bloom effect.
-
-// Format: 7 - A8
-// Used in Metroid Prime 2 for the scan visor.
-
-// Format: 8 - R8
-// Used in Twilight Princess for the map.
-
-// Format: 9 - G8
-// FIXME: Unseen.
-
-// Format: A - B8
-// Used in Metroid Prime 2 for the scan visor.
-
-// Format: B - G8 R8
-// Used in Wind Waker for depth-of-field. Usually used with srcFormat 3 to
-// render depth textures. The bytes are swapped, so games have to correct it
-// in RAM before using it as a texture.
-
-// Format: C - B8 G8
-// FIXME: Unseen.
-
-const unsigned int BLOCK_WIDTHS[16] = {
-	8, // R4
-	8, // R8 (FIXME: duplicate of R8 below?)
-	8, // A4 R4
-	4, // A8 R8
-	4, // R5 G6 B5
-	4, // 1 R5 G5 B5 or 0 A3 R4 G4 B4
-	4, // A8 R8 A8 R8 | G8 B8 G8 B8 (two cache lines)
-	8, // A8
-	8, // R8 (FIXME: duplicate of R8 above?)
-	8, // G8
-	8, // B8
-	4, // G8 R8
-	4, // B8 G8
-	0, 0, 0 // Unknown formats
-};
-
-const unsigned int BLOCK_HEIGHTS[16] = {
-	8, // R4
-	4, // R8 (FIXME: duplicate of R8 below?)
-	4, // A4 R4
-	4, // A8 R8
-	4, // R5 G6 B5
-	4, // 1 R5 G5 B5 or 0 A3 R4 G4 B4
-	4, // A8 R8 A8 R8 | G8 B8 G8 B8 (two cache lines)
-	4, // A8
-	4, // R8 (FIXME: duplicate of R8 above?)
-	4, // G8
-	4, // B8
-	4, // G8 R8
-	4, // B8 G8
-	0, 0, 0 // Unknown formats
-};
-
 // Maximum number of bytes that can occur in a texture block-row generated by
 // the encoder
 static const UINT MAX_BYTES_PER_BLOCK_ROW = (EFB_WIDTH/4)*64;
@@ -111,7 +26,7 @@ public:
 	virtual void Init() = 0;
 	virtual void Shutdown() = 0;
 	// Returns size in bytes of encoded block of memory
-	virtual size_t Encode(u8* dst, unsigned int dstFormat, size_t dstStride,
+	virtual void Encode(u8* dst, const TextureCache::TCacheEntryBase *texture_entry,
 		PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 		bool isIntensity, bool scaleByHalf) = 0;
 
diff --git a/Source/Core/VideoBackends/OGL/TextureCache.cpp b/Source/Core/VideoBackends/OGL/TextureCache.cpp
index d388da2084..c8c71ccc29 100644
--- a/Source/Core/VideoBackends/OGL/TextureCache.cpp
+++ b/Source/Core/VideoBackends/OGL/TextureCache.cpp
@@ -213,7 +213,7 @@ void TextureCache::TCacheEntry::Load(unsigned int width, unsigned int height,
 	TextureCache::SetStage();
 }
 
-void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFormat, u32 dstStride,
+void TextureCache::TCacheEntry::FromRenderTarget(u8* dstPointer, unsigned int dstFormat, u32 dstStride,
 	PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 	bool isIntensity, bool scaleByHalf, unsigned int cbufid,
 	const float *colmat)
@@ -264,24 +264,14 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 
 	if (!g_ActiveConfig.bSkipEFBCopyToRam)
 	{
-		int encoded_size = TextureConverter::EncodeToRamFromTexture(
-			dstAddr,
+		TextureConverter::EncodeToRamFromTexture(
+			dstPointer,
+			this,
 			read_texture,
 			srcFormat == PEControl::Z24,
 			isIntensity,
-			dstFormat,
 			scaleByHalf,
-			srcRect,
-			dstStride);
-
-		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(dstAddr, encoded_size);
-
-		hash = new_hash;
+			srcRect);
 	}
 
 	FramebufferManager::SetFramebuffer(0);
diff --git a/Source/Core/VideoBackends/OGL/TextureCache.h b/Source/Core/VideoBackends/OGL/TextureCache.h
index b8843c4edf..3f6f54b876 100644
--- a/Source/Core/VideoBackends/OGL/TextureCache.h
+++ b/Source/Core/VideoBackends/OGL/TextureCache.h
@@ -41,7 +41,7 @@ private:
 		void Load(unsigned int width, unsigned int height,
 			unsigned int expanded_width, unsigned int level) override;
 
-		void FromRenderTarget(u32 dstAddr, unsigned int dstFormat, u32 dstStride,
+		void FromRenderTarget(u8 *dst, unsigned int dstFormat, u32 dstStride,
 			PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 			bool isIntensity, bool scaleByHalf, unsigned int cbufid,
 			const float *colmat) override;
diff --git a/Source/Core/VideoBackends/OGL/TextureConverter.cpp b/Source/Core/VideoBackends/OGL/TextureConverter.cpp
index c85b617294..5020058da0 100644
--- a/Source/Core/VideoBackends/OGL/TextureConverter.cpp
+++ b/Source/Core/VideoBackends/OGL/TextureConverter.cpp
@@ -268,59 +268,18 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
 	}
 }
 
-int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride)
+void EncodeToRamFromTexture(u8 *dest_ptr, const TextureCache::TCacheEntryBase *texture_entry,
+                            GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, int bScaleByHalf, const EFBRectangle& source)
 {
-	u32 format = copyfmt;
-
-	if (bFromZBuffer)
-	{
-		format |= _GX_TF_ZTF;
-		if (copyfmt == 11)
-			format = GX_TF_Z16;
-		else if (format < GX_TF_Z8 || format > GX_TF_Z24X8)
-			format |= _GX_TF_CTF;
-	}
-	else
-	{
-		if (copyfmt > GX_TF_RGBA8 || (copyfmt < GX_TF_RGB565 && !bIsIntensityFmt))
-			format |= _GX_TF_CTF;
-	}
-
-	SHADER& texconv_shader = GetOrCreateEncodingShader(format);
-
-	u8 *dest_ptr = Memory::GetPointer(address);
-
-	int width = (source.right - source.left) >> bScaleByHalf;
-	int height = (source.bottom - source.top) >> bScaleByHalf;
-
-	int size_in_bytes = TexDecoder_GetTextureSizeInBytes(width, height, format);
-
-	u16 blkW = TexDecoder_GetBlockWidthInTexels(format) - 1;
-	u16 blkH = TexDecoder_GetBlockHeightInTexels(format) - 1;
-
-	// only copy on cache line boundaries
-	// extra pixels are copied but not displayed in the resulting texture
-	s32 expandedWidth = (width + blkW) & (~blkW);
-	s32 expandedHeight = (height + blkH) & (~blkH);
+	SHADER& texconv_shader = GetOrCreateEncodingShader(texture_entry->format);
 
 	texconv_shader.Bind();
-	glUniform4i(s_encodingUniforms[format],
-		source.left, source.top,
-		expandedWidth, bScaleByHalf ? 2 : 1);
-
-	unsigned int numBlocksX = expandedWidth / TexDecoder_GetBlockWidthInTexels(format);
-	unsigned int numBlocksY = expandedHeight / TexDecoder_GetBlockHeightInTexels(format);
-	unsigned int cacheLinesPerRow;
-	if ((format & 0x0f) == 6)
-		cacheLinesPerRow = numBlocksX * 2;
-	else
-		cacheLinesPerRow = numBlocksX;
+	glUniform4i(s_encodingUniforms[texture_entry->format],
+		source.left, source.top, texture_entry->native_width, bScaleByHalf ? 2 : 1);
 
 	EncodeToRamUsingShader(source_texture,
-		dest_ptr, cacheLinesPerRow * 32, numBlocksY,
-		writeStride, bScaleByHalf > 0 && !bFromZBuffer);
-	return size_in_bytes; // TODO: D3D11 is calculating this value differently!
-
+		dest_ptr, texture_entry->CacheLinesPerRow() * 32, texture_entry->NumBlocksY(),
+		texture_entry->memory_stride, bScaleByHalf > 0 && !bFromZBuffer);
 }
 
 void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* destAddr, u32 dstWidth, u32 dstStride, u32 dstHeight)
diff --git a/Source/Core/VideoBackends/OGL/TextureConverter.h b/Source/Core/VideoBackends/OGL/TextureConverter.h
index c41dfe10ac..186c9fe6d0 100644
--- a/Source/Core/VideoBackends/OGL/TextureConverter.h
+++ b/Source/Core/VideoBackends/OGL/TextureConverter.h
@@ -5,6 +5,7 @@
 #pragma once
 
 #include "VideoBackends/OGL/GLUtil.h"
+#include "VideoCommon/TextureCacheBase.h"
 #include "VideoCommon/VideoCommon.h"
 
 namespace OGL
@@ -24,7 +25,8 @@ void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc,
 void DecodeToTexture(u32 xfbAddr, int srcWidth, int srcHeight, GLuint destTexture);
 
 // returns size of the encoded data (in bytes)
-int EncodeToRamFromTexture(u32 address, GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride);
+void EncodeToRamFromTexture(u8 *dest_ptr, const TextureCache::TCacheEntryBase *texture_entry,
+	GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, int bScaleByHalf, const EFBRectangle& source);
 
 }
 
diff --git a/Source/Core/VideoCommon/TextureCacheBase.cpp b/Source/Core/VideoCommon/TextureCacheBase.cpp
index 5ab1b47bc4..62bf511863 100644
--- a/Source/Core/VideoCommon/TextureCacheBase.cpp
+++ b/Source/Core/VideoCommon/TextureCacheBase.cpp
@@ -182,24 +182,6 @@ void TextureCache::Cleanup(int _frameCount)
 	}
 }
 
-void TextureCache::MakeRangeDynamic(u32 start_address, u32 size)
-{
-	TexCache::iterator
-		iter = textures_by_address.begin();
-
-	while (iter != textures_by_address.end())
-	{
-		if (iter->second->OverlapsMemoryRange(start_address, size))
-		{
-			iter = FreeTexture(iter);
-		}
-		else
-		{
-			++iter;
-		}
-	}
-}
-
 bool TextureCache::TCacheEntryBase::OverlapsMemoryRange(u32 range_address, u32 range_size) const
 {
 	if (addr + size_in_bytes <= range_address)
@@ -243,7 +225,7 @@ TextureCache::TCacheEntryBase* TextureCache::DoPartialTextureUpdates(TexCache::i
 			&& entry_to_update->addr <= entry->addr
 			&& entry->addr + entry->size_in_bytes <= entry_to_update->addr + entry_to_update->size_in_bytes
 			&& entry->frameCount == FRAMECOUNT_INVALID
-			&& entry->copyStride == numBlocksX * block_size)
+			&& entry->memory_stride == numBlocksX * block_size)
 		{
 			u32 block_offset = (entry->addr - entry_to_update->addr) / block_size;
 			u32 block_x = block_offset % numBlocksX;
@@ -753,9 +735,11 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 		case 0: // Z4
 			colmat[3] = colmat[7] = colmat[11] = colmat[15] = 1.0f;
 			cbufid = 0;
+			dstFormat |= _GX_TF_CTF;
 			break;
+		case 8: // Z8H
+			dstFormat |= _GX_TF_CTF;
 		case 1: // Z8
-		case 8: // Z8
 			colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1.0f;
 			cbufid = 1;
 			break;
@@ -768,6 +752,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 		case 11: // Z16 (reverse order)
 			colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
 			cbufid = 3;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 6: // Z24X8
@@ -778,11 +763,13 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 		case 9: // Z8M
 			colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
 			cbufid = 5;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 10: // Z8L
 			colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
 			cbufid = 6;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 12: // Z16L - copy lower 16 depth bits
@@ -790,6 +777,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 			// Used e.g. in Zelda: Skyward Sword
 			colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1.0f;
 			cbufid = 7;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		default:
@@ -798,6 +786,8 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 			cbufid = 8;
 			break;
 		}
+
+		dstFormat |= _GX_TF_ZTF;
 	}
 	else if (isIntensity)
 	{
@@ -862,11 +852,13 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 			ColorMask[0] = 15.0f;
 			ColorMask[4] = 1.0f / 15.0f;
 			cbufid = 14;
+			dstFormat |= _GX_TF_CTF;
 			break;
 		case 1: // R8
 		case 8: // R8
 			colmat[0] = colmat[4] = colmat[8] = colmat[12] = 1;
 			cbufid = 15;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 2: // RA4
@@ -881,6 +873,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 				fConstAdd[3] = 1.0f;
 				cbufid = 17;
 			}
+			dstFormat |= _GX_TF_CTF;
 			break;
 		case 3: // RA8
 			colmat[0] = colmat[4] = colmat[8] = colmat[15] = 1.0f;
@@ -892,6 +885,7 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 				fConstAdd[3] = 1.0f;
 				cbufid = 19;
 			}
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 7: // A8
@@ -907,25 +901,30 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 				fConstAdd[3] = 1.0f;
 				cbufid = 21;
 			}
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 9: // G8
 			colmat[1] = colmat[5] = colmat[9] = colmat[13] = 1.0f;
 			cbufid = 22;
+			dstFormat |= _GX_TF_CTF;
 			break;
 		case 10: // B8
 			colmat[2] = colmat[6] = colmat[10] = colmat[14] = 1.0f;
 			cbufid = 23;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 11: // RG8
 			colmat[0] = colmat[4] = colmat[8] = colmat[13] = 1.0f;
 			cbufid = 24;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 12: // GB8
 			colmat[1] = colmat[5] = colmat[9] = colmat[14] = 1.0f;
 			cbufid = 25;
+			dstFormat |= _GX_TF_CTF;
 			break;
 
 		case 4: // RGB565
@@ -973,6 +972,13 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 		}
 	}
 
+	u8* dst = Memory::GetPointer(dstAddr);
+	if (dst == nullptr)
+	{
+		ERROR_LOG(VIDEO, "Trying to copy from EFB to invalid address 0x%8x", dstAddr);
+		return;
+	}
+
 	const unsigned int tex_w = scaleByHalf ? srcRect.GetWidth() / 2 : srcRect.GetWidth();
 	const unsigned int tex_h = scaleByHalf ? srcRect.GetHeight() / 2 : srcRect.GetHeight();
 
@@ -996,17 +1002,36 @@ void TextureCache::CopyRenderTargetToTexture(u32 dstAddr, unsigned int dstFormat
 
 	TCacheEntryBase* entry = AllocateTexture(config);
 
-	// TODO: Using the wrong dstFormat, dumb...
 	entry->SetGeneralParameters(dstAddr, 0, dstFormat);
 	entry->SetDimensions(tex_w, tex_h, 1);
 	entry->SetHashes(TEXHASH_INVALID);
 
 	entry->frameCount = FRAMECOUNT_INVALID;
-	entry->is_efb_copy = true;
+	entry->SetEfbCopy(dstStride);
 	entry->is_custom_tex = false;
-	entry->copyStride = dstStride;
 
-	entry->FromRenderTarget(dstAddr, dstFormat, dstStride, srcFormat, srcRect, isIntensity, scaleByHalf, cbufid, colmat);
+	entry->FromRenderTarget(dst, dstFormat, dstStride, srcFormat, srcRect, isIntensity, scaleByHalf, cbufid, colmat);
+
+	if (!g_ActiveConfig.bSkipEFBCopyToRam)
+	{
+		entry->hash = GetHash64(dst, (int)entry->size_in_bytes, g_ActiveConfig.iSafeTextureCache_ColorSamples);
+
+		// Invalidate all textures that overlap the range of our texture
+		TexCache::iterator
+			iter = textures_by_address.begin();
+
+		while (iter != textures_by_address.end())
+		{
+			if (iter->second->OverlapsMemoryRange(dstAddr, entry->size_in_bytes))
+			{
+				iter = FreeTexture(iter);
+			}
+			else
+			{
+				++iter;
+			}
+		}
+	}
 
 	if (g_ActiveConfig.bDumpEFBTarget)
 	{
@@ -1052,3 +1077,36 @@ TextureCache::TexCache::iterator TextureCache::FreeTexture(TexCache::iterator it
 
 	return textures_by_address.erase(iter);
 }
+
+u32 TextureCache::TCacheEntryBase::CacheLinesPerRow() const
+{
+	u32 blockW = TexDecoder_GetBlockWidthInTexels(format) - 1;
+	// Round up source height to multiple of block size
+	u32 actualWidth = (native_width + blockW) & ~(blockW);
+
+	u32 numBlocksX = actualWidth / TexDecoder_GetBlockWidthInTexels(format);
+
+	// RGBA takes two cache lines per block; all others take one
+	if (format == GX_TF_RGBA8)
+		numBlocksX = numBlocksX * 2;
+	return numBlocksX;
+}
+
+u32 TextureCache::TCacheEntryBase::NumBlocksY() const
+{
+	u32 blockH = TexDecoder_GetBlockHeightInTexels(format) - 1;
+	// Round up source height to multiple of block size
+	u32 actualHeight = (native_height + blockH) & ~(blockH);
+
+	return actualHeight / TexDecoder_GetBlockHeightInTexels(format);
+}
+
+void TextureCache::TCacheEntryBase::SetEfbCopy(u32 stride)
+{
+	is_efb_copy = true;
+	memory_stride = stride;
+
+	_assert_msg_(VIDEO, memory_stride >= CacheLinesPerRow(), "Memory stride is too small");
+
+	size_in_bytes = memory_stride * NumBlocksY();
+}
diff --git a/Source/Core/VideoCommon/TextureCacheBase.h b/Source/Core/VideoCommon/TextureCacheBase.h
index 6bb80f945c..c012cad0fc 100644
--- a/Source/Core/VideoCommon/TextureCacheBase.h
+++ b/Source/Core/VideoCommon/TextureCacheBase.h
@@ -53,7 +53,7 @@ public:
 		u32 format;
 		bool is_efb_copy;
 		bool is_custom_tex;
-		u32 copyStride;
+		u32 memory_stride;
 
 		unsigned int native_width, native_height; // Texture dimensions from the GameCube's point of view
 		unsigned int native_levels;
@@ -76,6 +76,7 @@ public:
 			native_width = _native_width;
 			native_height = _native_height;
 			native_levels = _native_levels;
+			memory_stride = _native_width;
 		}
 
 		void SetHashes(u64 _hash)
@@ -83,6 +84,8 @@ public:
 			hash = _hash;
 		}
 
+		void SetEfbCopy(u32 stride);
+
 		TCacheEntryBase(const TCacheEntryConfig& c) : config(c) {}
 		virtual ~TCacheEntryBase();
 
@@ -96,7 +99,7 @@ public:
 
 		virtual void Load(unsigned int width, unsigned int height,
 			unsigned int expanded_width, unsigned int level) = 0;
-		virtual void FromRenderTarget(u32 dstAddr, unsigned int dstFormat, u32 dstStride,
+		virtual void FromRenderTarget(u8* dst, unsigned int dstFormat, u32 dstStride,
 			PEControl::PixelFormat srcFormat, const EFBRectangle& srcRect,
 			bool isIntensity, bool scaleByHalf, unsigned int cbufid,
 			const float *colmat) = 0;
@@ -104,6 +107,11 @@ public:
 		bool OverlapsMemoryRange(u32 range_address, u32 range_size) const;
 
 		bool IsEfbCopy() const { return is_efb_copy; }
+
+		u32 NumBlocksY() const;
+		u32 CacheLinesPerRow() const;
+
+		void Memset(u8* ptr, u32 tag);
 	};
 
 	virtual ~TextureCache(); // needs virtual for DX11 dtor
@@ -115,7 +123,6 @@ public:
 	static void Cleanup(int _frameCount);
 
 	static void Invalidate();
-	static void MakeRangeDynamic(u32 start_address, u32 size);
 
 	virtual TCacheEntryBase* CreateTexture(const TCacheEntryConfig& config) = 0;
 
diff --git a/Source/Core/VideoCommon/TextureConversionShader.cpp b/Source/Core/VideoCommon/TextureConversionShader.cpp
index 9375a984c8..033a56b4aa 100644
--- a/Source/Core/VideoCommon/TextureConversionShader.cpp
+++ b/Source/Core/VideoCommon/TextureConversionShader.cpp
@@ -650,9 +650,11 @@ const char *GenerateEncodingShader(u32 format,API_TYPE ApiType)
 	case GX_CTF_GB8:
 		WriteCC8Encoder(p, "gb", ApiType);
 		break;
+	case GX_CTF_Z8H:
 	case GX_TF_Z8:
 		WriteC8Encoder(p, "r", ApiType);
 		break;
+	case GX_CTF_Z16R:
 	case GX_TF_Z16:
 		WriteZ16Encoder(p, ApiType);
 		break;
diff --git a/Source/Core/VideoCommon/TextureDecoder.h b/Source/Core/VideoCommon/TextureDecoder.h
index 1abcc5705c..fd3ee43ff7 100644
--- a/Source/Core/VideoCommon/TextureDecoder.h
+++ b/Source/Core/VideoCommon/TextureDecoder.h
@@ -35,7 +35,6 @@ enum TextureFormat
 	GX_CTF_R4    = 0x0 | _GX_TF_CTF,
 	GX_CTF_RA4   = 0x2 | _GX_TF_CTF,
 	GX_CTF_RA8   = 0x3 | _GX_TF_CTF,
-	GX_CTF_YUVA8 = 0x6 | _GX_TF_CTF,
 	GX_CTF_A8    = 0x7 | _GX_TF_CTF,
 	GX_CTF_R8    = 0x8 | _GX_TF_CTF,
 	GX_CTF_G8    = 0x9 | _GX_TF_CTF,
@@ -43,13 +42,17 @@ enum TextureFormat
 	GX_CTF_RG8   = 0xB | _GX_TF_CTF,
 	GX_CTF_GB8   = 0xC | _GX_TF_CTF,
 
+	// Texture formats (which match the color formats above)
 	GX_TF_Z8     = 0x1 | _GX_TF_ZTF,
 	GX_TF_Z16    = 0x3 | _GX_TF_ZTF,
 	GX_TF_Z24X8  = 0x6 | _GX_TF_ZTF,
 
+	// copy only texture formats
 	GX_CTF_Z4    = 0x0 | _GX_TF_ZTF | _GX_TF_CTF,
+	GX_CTF_Z8H   = 0x8 | _GX_TF_ZTF | _GX_TF_CTF, // This produces an identical result to to GX_TF_Z8
 	GX_CTF_Z8M   = 0x9 | _GX_TF_ZTF | _GX_TF_CTF,
 	GX_CTF_Z8L   = 0xA | _GX_TF_ZTF | _GX_TF_CTF,
+	GX_CTF_Z16R  = 0xB | _GX_TF_ZTF | _GX_TF_CTF, // Reversed version of GX_TF_Z16
 	GX_CTF_Z16L  = 0xC | _GX_TF_ZTF | _GX_TF_CTF,
 };
 
diff --git a/Source/Core/VideoCommon/TextureDecoder_Common.cpp b/Source/Core/VideoCommon/TextureDecoder_Common.cpp
index 092b8ec1e0..8fc1631b93 100644
--- a/Source/Core/VideoCommon/TextureDecoder_Common.cpp
+++ b/Source/Core/VideoCommon/TextureDecoder_Common.cpp
@@ -35,7 +35,6 @@ int TexDecoder_GetTexelSizeInNibbles(int format)
 	case GX_CTF_R4:    return 1;
 	case GX_CTF_RA4:   return 2;
 	case GX_CTF_RA8:   return 4;
-	case GX_CTF_YUVA8: return 8;
 	case GX_CTF_A8:    return 2;
 	case GX_CTF_R8:    return 2;
 	case GX_CTF_G8:    return 2;
@@ -48,10 +47,14 @@ int TexDecoder_GetTexelSizeInNibbles(int format)
 	case GX_TF_Z24X8:  return 8;
 
 	case GX_CTF_Z4:    return 1;
+	case GX_CTF_Z8H:   return 2;
 	case GX_CTF_Z8M:   return 2;
 	case GX_CTF_Z8L:   return 2;
+	case GX_CTF_Z16R:  return 4;
 	case GX_CTF_Z16L:  return 4;
-	default: return 1;
+	default:
+		PanicAlert("Unsupported Texture Format (%08x)! (GetTexelSizeInNibbles)", format);
+		return 1;
 	}
 }
 
@@ -88,11 +91,13 @@ int TexDecoder_GetBlockWidthInTexels(u32 format)
 	case GX_TF_Z16: return 4;
 	case GX_TF_Z24X8: return 4;
 	case GX_CTF_Z4: return 8;
+	case GX_CTF_Z8H: return 8;
 	case GX_CTF_Z8M: return 8;
 	case GX_CTF_Z8L: return 8;
+	case GX_CTF_Z16R: return 4;
 	case GX_CTF_Z16L: return 4;
 	default:
-		ERROR_LOG(VIDEO, "Unsupported Texture Format (%08x)! (GetBlockWidthInTexels)", format);
+		PanicAlert("Unsupported Texture Format (%08x)! (GetBlockWidthInTexels)", format);
 		return 8;
 	}
 }
@@ -125,11 +130,13 @@ int TexDecoder_GetBlockHeightInTexels(u32 format)
 	case GX_TF_Z16: return 4;
 	case GX_TF_Z24X8: return 4;
 	case GX_CTF_Z4: return 8;
+	case GX_CTF_Z8H: return 4;
 	case GX_CTF_Z8M: return 4;
 	case GX_CTF_Z8L: return 4;
+	case GX_CTF_Z16R: return 4;
 	case GX_CTF_Z16L: return 4;
 	default:
-		ERROR_LOG(VIDEO, "Unsupported Texture Format (%08x)! (GetBlockHeightInTexels)", format);
+		PanicAlert("Unsupported Texture Format (%08x)! (GetBlockHeightInTexels)", format);
 		return 4;
 	}
 }