Video: respect stride of efb copies to xfb

12 files changed, 42 insertions, 46 deletions

diff --git a/Source/Core/VideoBackends/D3D/FramebufferManager.cpp b/Source/Core/VideoBackends/D3D/FramebufferManager.cpp
index 03cea928f0..85fa9a122c 100644
--- a/Source/Core/VideoBackends/D3D/FramebufferManager.cpp
+++ b/Source/Core/VideoBackends/D3D/FramebufferManager.cpp
@@ -159,10 +159,10 @@ FramebufferManager::~FramebufferManager()
 	SAFE_RELEASE(m_efb.resolved_depth_tex);
 }
 
-void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
+void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
 {
 	u8* dst = Memory::GetPointer(xfbAddr);
-	s_xfbEncoder.Encode(dst, fbWidth, fbHeight, sourceRc, Gamma);
+	s_xfbEncoder.Encode(dst, fbStride, fbHeight, sourceRc, Gamma);
 }
 
 XFBSourceBase* FramebufferManager::CreateXFBSource(unsigned int target_width, unsigned int target_height, unsigned int layers)
diff --git a/Source/Core/VideoBackends/D3D/FramebufferManager.h b/Source/Core/VideoBackends/D3D/FramebufferManager.h
index 11ee646b4c..9151c7ef56 100644
--- a/Source/Core/VideoBackends/D3D/FramebufferManager.h
+++ b/Source/Core/VideoBackends/D3D/FramebufferManager.h
@@ -83,7 +83,7 @@ private:
 	XFBSourceBase* CreateXFBSource(unsigned int target_width, unsigned int target_height, unsigned int layers) override;
 	void GetTargetSize(unsigned int *width, unsigned int *height) override;
 
-	void CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
+	void CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
 
 	static struct Efb
 	{
diff --git a/Source/Core/VideoBackends/OGL/FramebufferManager.cpp b/Source/Core/VideoBackends/OGL/FramebufferManager.cpp
index 3fee033c57..a4489a67ff 100644
--- a/Source/Core/VideoBackends/OGL/FramebufferManager.cpp
+++ b/Source/Core/VideoBackends/OGL/FramebufferManager.cpp
@@ -481,7 +481,7 @@ GLuint FramebufferManager::GetEFBDepthTexture(const EFBRectangle& sourceRc)
 	}
 }
 
-void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
+void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
 {
 	u8* xfb_in_ram = Memory::GetPointer(xfbAddr);
 	if (!xfb_in_ram)
@@ -491,7 +491,7 @@ void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, c
 	}
 
 	TargetRectangle targetRc = g_renderer->ConvertEFBRectangle(sourceRc);
-	TextureConverter::EncodeToRamYUYV(ResolveAndGetRenderTarget(sourceRc), targetRc, xfb_in_ram, fbWidth, fbHeight);
+	TextureConverter::EncodeToRamYUYV(ResolveAndGetRenderTarget(sourceRc), targetRc, xfb_in_ram, sourceRc.GetWidth(), fbStride, fbHeight);
 }
 
 void FramebufferManager::SetFramebuffer(GLuint fb)
diff --git a/Source/Core/VideoBackends/OGL/FramebufferManager.h b/Source/Core/VideoBackends/OGL/FramebufferManager.h
index bf04b41536..d0aa6a2c18 100644
--- a/Source/Core/VideoBackends/OGL/FramebufferManager.h
+++ b/Source/Core/VideoBackends/OGL/FramebufferManager.h
@@ -97,7 +97,7 @@ private:
 	XFBSourceBase* CreateXFBSource(unsigned int target_width, unsigned int target_height, unsigned int layers) override;
 	void GetTargetSize(unsigned int *width, unsigned int *height) override;
 
-	void CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
+	void CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
 
 	static int m_targetWidth;
 	static int m_targetHeight;
diff --git a/Source/Core/VideoBackends/OGL/TextureCache.cpp b/Source/Core/VideoBackends/OGL/TextureCache.cpp
index ebc3a672f1..92f2b8a788 100644
--- a/Source/Core/VideoBackends/OGL/TextureCache.cpp
+++ b/Source/Core/VideoBackends/OGL/TextureCache.cpp
@@ -269,7 +269,8 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 			isIntensity,
 			dstFormat,
 			scaleByHalf,
-			srcRect);
+			srcRect,
+			copyMipMapStrideChannels * 32);
 
 		u8* dst = Memory::GetPointer(dstAddr);
 		u64 const new_hash = GetHash64(dst,encoded_size,g_ActiveConfig.iSafeTextureCache_ColorSamples);
diff --git a/Source/Core/VideoBackends/OGL/TextureConverter.cpp b/Source/Core/VideoBackends/OGL/TextureConverter.cpp
index c696ccf687..c25d30d088 100644
--- a/Source/Core/VideoBackends/OGL/TextureConverter.cpp
+++ b/Source/Core/VideoBackends/OGL/TextureConverter.cpp
@@ -213,12 +213,12 @@ void Shutdown()
 	s_texConvFrameBuffer[1] = 0;
 }
 
+// dst_line_size, writeStride in bytes
+
 static void EncodeToRamUsingShader(GLuint srcTexture,
-						u8* destAddr, int dstWidth, int dstHeight, int readStride,
-						bool linearFilter)
+						u8* destAddr, u32 dst_line_size, u32 dstHeight,
+						u32 writeStride, bool linearFilter)
 {
-
-
 	// switch to texture converter frame buffer
 	// attach render buffer as color destination
 	FramebufferManager::SetFramebuffer(s_texConvFrameBuffer[0]);
@@ -234,19 +234,13 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
 	else
 		g_sampler_cache->BindNearestSampler(9);
 
-	glViewport(0, 0, (GLsizei)dstWidth, (GLsizei)dstHeight);
+	glViewport(0, 0, (GLsizei)(dst_line_size / 4), (GLsizei)dstHeight);
 
 	glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 
-	// .. and then read back the results.
-	// TODO: make this less slow.
-
-	int writeStride = bpmem.copyMipMapStrideChannels * 32;
-	int dstSize = dstWidth*dstHeight*4;
-	int readHeight = readStride / dstWidth / 4; // 4 bytes per pixel
-	int readLoops = dstHeight / readHeight;
+	int dstSize = dst_line_size * dstHeight;
 
-	if (writeStride != readStride && readLoops > 1)
+	if ((writeStride != dst_line_size) && (dstHeight > 1))
 	{
 		// writing to a texture of a different size
 		// also copy more then one block line, so the different strides matters
@@ -255,13 +249,13 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
 		// CPU overhead because of the pbo
 		glBindBuffer(GL_PIXEL_PACK_BUFFER, s_PBO);
 		glBufferData(GL_PIXEL_PACK_BUFFER, dstSize, nullptr, GL_STREAM_READ);
-		glReadPixels(0, 0, (GLsizei)dstWidth, (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, nullptr);
+		glReadPixels(0, 0, (GLsizei)(dst_line_size / 4), (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, nullptr);
 		u8* pbo = (u8*)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, dstSize, GL_MAP_READ_BIT);
 
-		for (int i = 0; i < readLoops; i++)
+		for (size_t i = 0; i < dstHeight; ++i)
 		{
-			memcpy(destAddr, pbo, readStride);
-			pbo += readStride;
+			memcpy(destAddr, pbo, dst_line_size);
+			pbo += dst_line_size;
 			destAddr += writeStride;
 		}
 
@@ -270,11 +264,11 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
 	}
 	else
 	{
-		glReadPixels(0, 0, (GLsizei)dstWidth, (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, destAddr);
+		glReadPixels(0, 0, (GLsizei)(dst_line_size / 4), (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, destAddr);
 	}
 }
 
-int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source)
+int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride)
 {
 	u32 format = copyfmt;
 
@@ -323,13 +317,13 @@ int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer,
 		cacheLinesPerRow = numBlocksX;
 
 	EncodeToRamUsingShader(source_texture,
-		dest_ptr, cacheLinesPerRow * 8, numBlocksY, cacheLinesPerRow * 32,
-		bScaleByHalf > 0 && !bFromZBuffer);
+		dest_ptr, cacheLinesPerRow * 32, numBlocksY,
+		writeStride, bScaleByHalf > 0 && !bFromZBuffer);
 	return size_in_bytes; // TODO: D3D11 is calculating this value differently!
 
 }
 
-void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* destAddr, int dstWidth, int dstHeight)
+void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* destAddr, u32 dstWidth, u32 dstStride, u32 dstHeight)
 {
 	g_renderer->ResetAPIState();
 
@@ -341,7 +335,7 @@ void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* des
 	// We enable linear filtering, because the GameCube does filtering in the vertical direction when
 	// yscale is enabled.
 	// Otherwise we get jaggies when a game uses yscaling (most PAL games)
-	EncodeToRamUsingShader(srcTexture, destAddr, dstWidth / 2, dstHeight, dstWidth*dstHeight*2, true);
+	EncodeToRamUsingShader(srcTexture, destAddr, dstWidth * 2, dstHeight, dstStride * 2, true);
 	FramebufferManager::SetFramebuffer(0);
 	TextureCache::DisableStage(0);
 	g_renderer->RestoreAPIState();
diff --git a/Source/Core/VideoBackends/OGL/TextureConverter.h b/Source/Core/VideoBackends/OGL/TextureConverter.h
index e0c6cbd5fb..c41dfe10ac 100644
--- a/Source/Core/VideoBackends/OGL/TextureConverter.h
+++ b/Source/Core/VideoBackends/OGL/TextureConverter.h
@@ -19,12 +19,12 @@ void Init();
 void Shutdown();
 
 void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc,
-					 u8* destAddr, int dstWidth, int dstHeight);
+	u8* destAddr, u32 dstWidth, u32 dstStride, u32 dstHeight);
 
 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);
+int EncodeToRamFromTexture(u32 address, GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride);
 
 }
 
diff --git a/Source/Core/VideoCommon/BPStructs.cpp b/Source/Core/VideoCommon/BPStructs.cpp
index 410d4d44cc..92ab0bf63f 100644
--- a/Source/Core/VideoCommon/BPStructs.cpp
+++ b/Source/Core/VideoCommon/BPStructs.cpp
@@ -251,9 +251,10 @@ static void BPWritten(const BPCmd& bp)
 					height = MAX_XFB_HEIGHT;
 				}
 
-				u32 width = bpmem.copyMipMapStrideChannels << 4;
-
-				Renderer::RenderToXFB(destAddr, srcRect, width, height, s_gammaLUT[PE_copy.gamma]);
+				u32 stride = bpmem.copyMipMapStrideChannels << 4;
+				WARN_LOG(VIDEO, "RenderToXFB: destAddr: %08x | srcRect {%d %d %d %d} | fbWidth: %u | fbStride: %u | fbHeight: %u",
+					destAddr, srcRect.left, srcRect.top, srcRect.right, srcRect.bottom, bpmem.copyTexSrcWH.x + 1, stride, height);
+				Renderer::RenderToXFB(destAddr, srcRect, stride, height, s_gammaLUT[PE_copy.gamma]);
 			}
 
 			// Clear the rectangular region after copying it.
diff --git a/Source/Core/VideoCommon/FramebufferManagerBase.cpp b/Source/Core/VideoCommon/FramebufferManagerBase.cpp
index 10d1cfa414..711b117ef6 100644
--- a/Source/Core/VideoCommon/FramebufferManagerBase.cpp
+++ b/Source/Core/VideoCommon/FramebufferManagerBase.cpp
@@ -115,15 +115,15 @@ const XFBSourceBase* const* FramebufferManagerBase::GetVirtualXFBSource(u32 xfbA
 	return &m_overlappingXFBArray[0];
 }
 
-void FramebufferManagerBase::CopyToXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
+void FramebufferManagerBase::CopyToXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma)
 {
 	if (g_ActiveConfig.bUseRealXFB)
-		g_framebuffer_manager->CopyToRealXFB(xfbAddr, fbWidth, fbHeight, sourceRc,Gamma);
+		g_framebuffer_manager->CopyToRealXFB(xfbAddr, fbStride, fbHeight, sourceRc, Gamma);
 	else
-		CopyToVirtualXFB(xfbAddr, fbWidth, fbHeight, sourceRc,Gamma);
+		CopyToVirtualXFB(xfbAddr, fbStride, fbHeight, sourceRc, Gamma);
 }
 
-void FramebufferManagerBase::CopyToVirtualXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
+void FramebufferManagerBase::CopyToVirtualXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma)
 {
 	VirtualXFBListType::iterator vxfb = FindVirtualXFB(xfbAddr, fbWidth, fbHeight);
 
diff --git a/Source/Core/VideoCommon/FramebufferManagerBase.h b/Source/Core/VideoCommon/FramebufferManagerBase.h
index bd6360877c..6127037268 100644
--- a/Source/Core/VideoCommon/FramebufferManagerBase.h
+++ b/Source/Core/VideoCommon/FramebufferManagerBase.h
@@ -45,7 +45,7 @@ public:
 	FramebufferManagerBase();
 	virtual ~FramebufferManagerBase();
 
-	static void CopyToXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma);
+	static void CopyToXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma);
 	static const XFBSourceBase* const* GetXFBSource(u32 xfbAddr, u32 fbWidth, u32 fbHeight, u32* xfbCount);
 
 	static void SetLastXfbWidth(unsigned int width) { s_last_xfb_width = width; }
@@ -87,7 +87,7 @@ private:
 	static void ReplaceVirtualXFB();
 
 	// TODO: merge these virtual funcs, they are nearly all the same
-	virtual void CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma = 1.0f) = 0;
+	virtual void CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma = 1.0f) = 0;
 	static void CopyToVirtualXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma = 1.0f);
 
 	static const XFBSourceBase* const* GetRealXFBSource(u32 xfbAddr, u32 fbWidth, u32 fbHeight, u32* xfbCount);
diff --git a/Source/Core/VideoCommon/RenderBase.cpp b/Source/Core/VideoCommon/RenderBase.cpp
index a688c86792..cd98f5dbf0 100644
--- a/Source/Core/VideoCommon/RenderBase.cpp
+++ b/Source/Core/VideoCommon/RenderBase.cpp
@@ -112,22 +112,22 @@ Renderer::~Renderer()
 #endif
 }
 
-void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidth, u32 fbHeight, float Gamma)
+void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbStride, u32 fbHeight, float Gamma)
 {
 	CheckFifoRecording();
 
-	if (!fbWidth || !fbHeight)
+	if (!fbStride || !fbHeight)
 		return;
 
 	XFBWrited = true;
 
 	if (g_ActiveConfig.bUseXFB)
 	{
-		FramebufferManagerBase::CopyToXFB(xfbAddr, fbWidth, fbHeight, sourceRc, Gamma);
+		FramebufferManagerBase::CopyToXFB(xfbAddr, fbStride, fbHeight, sourceRc, Gamma);
 	}
 	else
 	{
-		Swap(xfbAddr, fbWidth, fbWidth, fbHeight, sourceRc, Gamma);
+		Swap(xfbAddr, sourceRc.GetWidth(), fbStride, fbHeight, sourceRc, Gamma);
 	}
 }
 
diff --git a/Source/Core/VideoCommon/RenderBase.h b/Source/Core/VideoCommon/RenderBase.h
index a17525ca2e..c717483d1e 100644
--- a/Source/Core/VideoCommon/RenderBase.h
+++ b/Source/Core/VideoCommon/RenderBase.h
@@ -109,7 +109,7 @@ public:
 
 	virtual void ClearScreen(const EFBRectangle& rc, bool colorEnable, bool alphaEnable, bool zEnable, u32 color, u32 z) = 0;
 	virtual void ReinterpretPixelData(unsigned int convtype) = 0;
-	static void RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidth, u32 fbHeight, float Gamma = 1.0f);
+	static void RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbStride, u32 fbHeight, float Gamma = 1.0f);
 
 	virtual u32 AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data) = 0;
 	virtual void PokeEFB(EFBAccessType type, const std::vector<EfbPokeData>& data);