Merge pull request #2679 from Tinob/master

Implement scaled partial texture updates

6 files changed, 210 insertions, 40 deletions

diff --git a/Source/Core/VideoBackends/D3D/TextureCache.cpp b/Source/Core/VideoBackends/D3D/TextureCache.cpp
index 2c820849d5..de204fa2af 100644
--- a/Source/Core/VideoBackends/D3D/TextureCache.cpp
+++ b/Source/Core/VideoBackends/D3D/TextureCache.cpp
@@ -77,11 +77,67 @@ bool TextureCache::TCacheEntry::Save(const std::string& filename, unsigned int l
 	return saved_png;
 }
 
-void TextureCache::TCacheEntry::DoPartialTextureUpdate(TCacheEntryBase* entry_, u32 x, u32 y)
+void TextureCache::TCacheEntry::CopyRectangleFromTexture(
+	const TCacheEntryBase* source,
+	const MathUtil::Rectangle<int> &srcrect,
+	const MathUtil::Rectangle<int> &dstrect)
 {
-	TCacheEntry* entry = (TCacheEntry*)entry_;
+	TCacheEntry* srcentry = (TCacheEntry*)source;
+	if (srcrect.GetWidth() == dstrect.GetWidth()
+		&& srcrect.GetHeight() == dstrect.GetHeight())
+	{
+		const D3D11_BOX *psrcbox = nullptr;
+		D3D11_BOX srcbox;
+		if (srcrect.left != 0 || srcrect.top != 0)
+		{
+			srcbox.left = srcrect.left;
+			srcbox.top = srcrect.top;
+			srcbox.right = srcrect.right;
+			srcbox.bottom = srcrect.bottom;
+			psrcbox = &srcbox;
+		}
+		D3D::context->CopySubresourceRegion(
+			texture->GetTex(),
+			0,
+			dstrect.left,
+			dstrect.top,
+			0,
+			srcentry->texture->GetTex(),
+			0,
+			psrcbox);
+		return;
+	}
+	else if (!config.rendertarget)
+	{
+		return;
+	}
+	g_renderer->ResetAPIState(); // reset any game specific settings
 
-	D3D::context->CopySubresourceRegion(texture->GetTex(), 0, x , y , 0, entry->texture->GetTex(), 0, NULL);
+	const D3D11_VIEWPORT vp = CD3D11_VIEWPORT(
+		float(dstrect.left),
+		float(dstrect.top),
+		float(dstrect.GetWidth()),
+		float(dstrect.GetHeight()));
+
+	D3D::context->OMSetRenderTargets(1, &texture->GetRTV(), nullptr);
+	D3D::context->RSSetViewports(1, &vp);
+	D3D::SetLinearCopySampler();
+	D3D11_RECT srcRC;
+	srcRC.left = srcrect.left;
+	srcRC.right = srcrect.right;
+	srcRC.top = srcrect.top;
+	srcRC.bottom = srcrect.bottom;
+	D3D::drawShadedTexQuad(srcentry->texture->GetSRV(), &srcRC,
+		srcentry->config.width, srcentry->config.height,
+		PixelShaderCache::GetColorCopyProgram(false),
+		VertexShaderCache::GetSimpleVertexShader(),
+		VertexShaderCache::GetSimpleInputLayout(), nullptr, 1.0, 0);
+
+	D3D::context->OMSetRenderTargets(1,
+		&FramebufferManager::GetEFBColorTexture()->GetRTV(),
+		FramebufferManager::GetEFBDepthTexture()->GetDSV());
+
+	g_renderer->RestoreAPIState();
 }
 
 void TextureCache::TCacheEntry::Load(unsigned int width, unsigned int height,
diff --git a/Source/Core/VideoBackends/D3D/TextureCache.h b/Source/Core/VideoBackends/D3D/TextureCache.h
index d6cb7eeef0..608fa47044 100644
--- a/Source/Core/VideoBackends/D3D/TextureCache.h
+++ b/Source/Core/VideoBackends/D3D/TextureCache.h
@@ -26,7 +26,10 @@ private:
 		TCacheEntry(const TCacheEntryConfig& config, D3DTexture2D *_tex) : TCacheEntryBase(config), texture(_tex) {}
 		~TCacheEntry();
 
-		void DoPartialTextureUpdate(TCacheEntryBase* entry, u32 x, u32 y) override;
+		void CopyRectangleFromTexture(
+			const TCacheEntryBase* source,
+			const MathUtil::Rectangle<int> &srcrect,
+			const MathUtil::Rectangle<int> &dstrect) override;
 
 		void Load(unsigned int width, unsigned int height,
 			unsigned int expanded_width, unsigned int levels) override;
diff --git a/Source/Core/VideoBackends/OGL/TextureCache.cpp b/Source/Core/VideoBackends/OGL/TextureCache.cpp
index cefc8017d6..ebc3a672f1 100644
--- a/Source/Core/VideoBackends/OGL/TextureCache.cpp
+++ b/Source/Core/VideoBackends/OGL/TextureCache.cpp
@@ -33,11 +33,13 @@
 namespace OGL
 {
 
+static SHADER s_ColorCopyProgram;
 static SHADER s_ColorMatrixProgram;
 static SHADER s_DepthMatrixProgram;
 static GLuint s_ColorMatrixUniform;
 static GLuint s_DepthMatrixUniform;
 static GLuint s_ColorCopyPositionUniform;
+static GLuint s_ColorMatrixPositionUniform;
 static GLuint s_DepthCopyPositionUniform;
 static u32 s_ColorCbufid;
 static u32 s_DepthCbufid;
@@ -137,12 +139,53 @@ TextureCache::TCacheEntryBase* TextureCache::CreateTexture(const TCacheEntryConf
 	return entry;
 }
 
-void TextureCache::TCacheEntry::DoPartialTextureUpdate(TCacheEntryBase* entry_, u32 x, u32 y)
+void TextureCache::TCacheEntry::CopyRectangleFromTexture(
+	const TCacheEntryBase* source,
+	const MathUtil::Rectangle<int> &srcrect,
+	const MathUtil::Rectangle<int> &dstrect)
 {
-
-	TCacheEntry* entry = (TCacheEntry*)entry_;
-
-	glCopyImageSubData(entry->texture, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, texture, GL_TEXTURE_2D_ARRAY, 0, x, y, 0, entry->native_width, entry->native_height, 1);
+	TCacheEntry* srcentry = (TCacheEntry*)source;
+	if (srcrect.GetWidth() == dstrect.GetWidth()
+		&& srcrect.GetHeight() == dstrect.GetHeight()
+		&& g_ActiveConfig.backend_info.bSupportsCopySubImage)
+	{
+		glCopyImageSubData(
+			srcentry->texture,
+			GL_TEXTURE_2D_ARRAY,
+			0,
+			srcrect.left,
+			srcrect.top,
+			0,
+			texture,
+			GL_TEXTURE_2D_ARRAY,
+			0,
+			dstrect.left,
+			dstrect.top,
+			0,
+			dstrect.GetWidth(),
+			dstrect.GetHeight(),
+			1);
+		return;
+	}
+	else if (!config.rendertarget)
+	{
+		return;
+	}
+	g_renderer->ResetAPIState();
+	FramebufferManager::SetFramebuffer(framebuffer);
+	glActiveTexture(GL_TEXTURE9);
+	glBindTexture(GL_TEXTURE_2D_ARRAY, srcentry->texture);
+	g_sampler_cache->BindLinearSampler(9);
+	glViewport(dstrect.left, dstrect.top, dstrect.GetWidth(), dstrect.GetHeight());
+	s_ColorCopyProgram.Bind();
+	glUniform4f(s_ColorCopyPositionUniform,
+		float(srcrect.left),
+		float(srcrect.top),
+		float(srcrect.GetWidth()),
+		float(srcrect.GetHeight()));
+	glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+	FramebufferManager::SetFramebuffer(0);
+	g_renderer->RestoreAPIState();
 }
 
 void TextureCache::TCacheEntry::Load(unsigned int width, unsigned int height,
@@ -208,7 +251,7 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 		if (s_ColorCbufid != cbufid)
 			glUniform4fv(s_ColorMatrixUniform, 7, colmat);
 		s_ColorCbufid = cbufid;
-		uniform_location = s_ColorCopyPositionUniform;
+		uniform_location = s_ColorMatrixPositionUniform;
 	}
 
 	TargetRectangle R = g_renderer->ConvertEFBRectangle(srcRect);
@@ -286,6 +329,16 @@ void TextureCache::SetStage()
 
 void TextureCache::CompileShaders()
 {
+	const char *pColorCopyProg =
+		"SAMPLER_BINDING(9) uniform sampler2DArray samp9;\n"
+		"in vec3 f_uv0;\n"
+		"out vec4 ocol0;\n"
+		"\n"
+		"void main(){\n"
+		"	vec4 texcol = texture(samp9, f_uv0);\n"
+		"	ocol0 = texcol;\n"
+		"}\n";
+
 	const char *pColorMatrixProg =
 		"SAMPLER_BINDING(9) uniform sampler2DArray samp9;\n"
 		"uniform vec4 colmat[7];\n"
@@ -357,6 +410,7 @@ void TextureCache::CompileShaders()
 	const char* prefix = (GProgram == nullptr) ? "f" : "v";
 	const char* depth_layer = (g_ActiveConfig.bStereoEFBMonoDepth) ? "0.0" : "f_uv0.z";
 
+	ProgramShaderCache::CompileShader(s_ColorCopyProgram, StringFromFormat(VProgram, prefix, prefix).c_str(), pColorCopyProg, GProgram);
 	ProgramShaderCache::CompileShader(s_ColorMatrixProgram, StringFromFormat(VProgram, prefix, prefix).c_str(), pColorMatrixProg, GProgram);
 	ProgramShaderCache::CompileShader(s_DepthMatrixProgram, StringFromFormat(VProgram, prefix, prefix).c_str(), StringFromFormat(pDepthMatrixProg, depth_layer).c_str(), GProgram);
 
@@ -365,7 +419,8 @@ void TextureCache::CompileShaders()
 	s_ColorCbufid = -1;
 	s_DepthCbufid = -1;
 
-	s_ColorCopyPositionUniform = glGetUniformLocation(s_ColorMatrixProgram.glprogid, "copy_position");
+	s_ColorCopyPositionUniform = glGetUniformLocation(s_ColorCopyProgram.glprogid, "copy_position");
+	s_ColorMatrixPositionUniform = glGetUniformLocation(s_ColorMatrixProgram.glprogid, "copy_position");
 	s_DepthCopyPositionUniform = glGetUniformLocation(s_DepthMatrixProgram.glprogid, "copy_position");
 
 	std::string palette_shader =
diff --git a/Source/Core/VideoBackends/OGL/TextureCache.h b/Source/Core/VideoBackends/OGL/TextureCache.h
index 42c83d8dda..1ed85b0539 100644
--- a/Source/Core/VideoBackends/OGL/TextureCache.h
+++ b/Source/Core/VideoBackends/OGL/TextureCache.h
@@ -33,7 +33,10 @@ private:
 		TCacheEntry(const TCacheEntryConfig& config);
 		~TCacheEntry();
 
-		void DoPartialTextureUpdate(TCacheEntryBase* entry, u32 x, u32 y) override;
+		void CopyRectangleFromTexture(
+			const TCacheEntryBase* source,
+			const MathUtil::Rectangle<int> &srcrect,
+			const MathUtil::Rectangle<int> &dstrect) override;
 
 		void Load(unsigned int width, unsigned int height,
 			unsigned int expanded_width, unsigned int level) override;
diff --git a/Source/Core/VideoCommon/TextureCacheBase.cpp b/Source/Core/VideoCommon/TextureCacheBase.cpp
index 6baca12326..7d49e698da 100644
--- a/Source/Core/VideoCommon/TextureCacheBase.cpp
+++ b/Source/Core/VideoCommon/TextureCacheBase.cpp
@@ -211,46 +211,98 @@ bool TextureCache::TCacheEntryBase::OverlapsMemoryRange(u32 range_address, u32 r
 	return true;
 }
 
-void TextureCache::TCacheEntryBase::DoPartialTextureUpdates()
+TextureCache::TCacheEntryBase* TextureCache::DoPartialTextureUpdates(TexCache::iterator iter_t)
 {
-	const bool isPaletteTexture = (format== GX_TF_C4 || format == GX_TF_C8 || format == GX_TF_C14X2 || format >= 0x10000);
+	TCacheEntryBase* entry_to_update = iter_t->second;
+	const bool isPaletteTexture = (entry_to_update->format == GX_TF_C4
+		|| entry_to_update->format == GX_TF_C8
+		|| entry_to_update->format == GX_TF_C14X2
+		|| entry_to_update->format >= 0x10000);
 
 	// Efb copies and paletted textures are excluded from these updates, until there's an example where a game would
 	// benefit from this. Both would require more work to be done.
 	// TODO: Implement upscaling support for normal textures, and then remove the efb to ram and the scaled efb restrictions
-	if (!g_ActiveConfig.backend_info.bSupportsCopySubImage || !g_ActiveConfig.bSkipEFBCopyToRam || IsEfbCopy()
-		|| isPaletteTexture || (g_ActiveConfig.bCopyEFBScaled && g_ActiveConfig.iEFBScale != SCALE_1X))
-		return;
-
-	u32 block_width = TexDecoder_GetBlockWidthInTexels(format);
-	u32 block_height = TexDecoder_GetBlockHeightInTexels(format);
-	u32 block_size = block_width * block_height * TexDecoder_GetTexelSizeInNibbles(format) / 2;
+	if (entry_to_update->IsEfbCopy()
+		|| isPaletteTexture)
+		return entry_to_update;
 
-	u32 numBlocksX = (native_width + block_width - 1) / block_width;
+	u32 block_width = TexDecoder_GetBlockWidthInTexels(entry_to_update->format);
+	u32 block_height = TexDecoder_GetBlockHeightInTexels(entry_to_update->format);
+	u32 block_size = block_width * block_height * TexDecoder_GetTexelSizeInNibbles(entry_to_update->format) / 2;
 
-	TexCache::iterator iter = textures_by_address.lower_bound(addr);
-	TexCache::iterator iterend = textures_by_address.upper_bound(addr + size_in_bytes);
+	u32 numBlocksX = (entry_to_update->native_width + block_width - 1) / block_width;
 
+	TexCache::iterator iter = textures_by_address.lower_bound(entry_to_update->addr);
+	TexCache::iterator iterend = textures_by_address.upper_bound(entry_to_update->addr + entry_to_update->size_in_bytes);
+	bool entry_need_scaling = true;
 	while (iter != iterend)
 	{
 		TCacheEntryBase* entry = iter->second;
-		if (entry->IsEfbCopy() && addr <= entry->addr && entry->addr + entry->size_in_bytes <= addr + size_in_bytes
-			&& entry->frameCount == FRAMECOUNT_INVALID && entry->copyMipMapStrideChannels * 32 == numBlocksX * block_size)
+		if (entry != entry_to_update
+			&& entry->IsEfbCopy()
+			&& 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->copyMipMapStrideChannels * 32 == numBlocksX * block_size)
 		{
-			u32 block_offset = (entry->addr - addr) / block_size;
+			u32 block_offset = (entry->addr - entry_to_update->addr) / block_size;
 			u32 block_x = block_offset % numBlocksX;
 			u32 block_y = block_offset / numBlocksX;
 
 			u32 x = block_x * block_width;
 			u32 y = block_y * block_height;
-
-			DoPartialTextureUpdate(entry, x, y);
-
+			MathUtil::Rectangle<int> srcrect, dstrect;
+			srcrect.left = 0;
+			srcrect.top = 0;
+			dstrect.left = 0;
+			dstrect.top = 0;
+			if (entry_need_scaling)
+			{
+				entry_need_scaling = false;
+				u32 w = entry_to_update->native_width * entry->config.width / entry->native_width;
+				u32 h = entry_to_update->native_height * entry->config.height / entry->native_height;
+				u32 max = g_renderer->GetMaxTextureSize();
+				if (max < w || max < h)
+				{
+					iter++;
+					continue;
+				}
+				if (entry_to_update->config.width != w || entry_to_update->config.height != h)
+				{
+					TextureCache::TCacheEntryConfig newconfig;
+					newconfig.width = w;
+					newconfig.height = h;
+					newconfig.rendertarget = true;
+					TCacheEntryBase* newentry = AllocateTexture(newconfig);
+					newentry->SetGeneralParameters(entry_to_update->addr, entry_to_update->size_in_bytes, entry_to_update->format);
+					newentry->SetDimensions(entry_to_update->native_width, entry_to_update->native_height, 1);
+					newentry->SetHashes(entry_to_update->hash);
+					newentry->frameCount = frameCount;
+					newentry->is_efb_copy = false;
+					srcrect.right = entry_to_update->config.width;
+					srcrect.bottom = entry_to_update->config.height;
+					dstrect.right = w;
+					dstrect.bottom = h;
+					newentry->CopyRectangleFromTexture(entry_to_update, srcrect, dstrect);
+					entry_to_update = newentry;
+					u64 key = iter_t->first;
+					iter_t = FreeTexture(iter_t);
+					textures_by_address.emplace(key, entry_to_update);
+				}
+			}
+			srcrect.right = entry->config.width;
+			srcrect.bottom = entry->config.height;
+			dstrect.left = x * entry_to_update->config.width / entry_to_update->native_width;
+			dstrect.top = y * entry_to_update->config.height / entry_to_update->native_height;
+			dstrect.right = (x + entry->native_width) * entry_to_update->config.width / entry_to_update->native_width;
+			dstrect.bottom = (y + entry->native_height) * entry_to_update->config.height / entry_to_update->native_height;
+			entry_to_update->CopyRectangleFromTexture(entry, srcrect, dstrect);
 			// Mark the texture update as used, so it isn't applied more than once
 			entry->frameCount = frameCount;
 		}
 		++iter;
 	}
+	return entry_to_update;
 }
 
 void TextureCache::DumpTexture(TCacheEntryBase* entry, std::string basename, unsigned int level)
@@ -323,7 +375,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	const unsigned int bsw = TexDecoder_GetBlockWidthInTexels(texformat) - 1;
 	const unsigned int bsh = TexDecoder_GetBlockHeightInTexels(texformat) - 1;
 
-	unsigned int expandedWidth  = (width  + bsw) & (~bsw);
+	unsigned int expandedWidth = (width + bsw) & (~bsw);
 	unsigned int expandedHeight = (height + bsh) & (~bsh);
 	const unsigned int nativeW = width;
 	const unsigned int nativeH = height;
@@ -440,7 +492,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 			if (entry->hash == full_hash && entry->format == full_format && entry->native_levels >= tex_levels &&
 				entry->native_width == nativeW && entry->native_height == nativeH)
 			{
-				entry->DoPartialTextureUpdates();
+				entry = DoPartialTextureUpdates(iter);
 
 				return ReturnEntry(stage, entry);
 			}
@@ -494,7 +546,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 			if (entry->format == full_format && entry->native_levels >= tex_levels &&
 				entry->native_width == nativeW && entry->native_height == nativeH)
 			{
-				entry->DoPartialTextureUpdates();
+				entry = DoPartialTextureUpdates(iter);
 
 				return ReturnEntry(stage, entry);
 			}
@@ -539,7 +591,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 		if (!(texformat == GX_TF_RGBA8 && from_tmem))
 		{
 			const u8* tlut = &texMem[tlutaddr];
-			TexDecoder_Decode(temp, src_data, expandedWidth, expandedHeight, texformat, tlut, (TlutFormat) tlutfmt);
+			TexDecoder_Decode(temp, src_data, expandedWidth, expandedHeight, texformat, tlut, (TlutFormat)tlutfmt);
 		}
 		else
 		{
@@ -560,7 +612,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	TCacheEntryBase* entry = AllocateTexture(config);
 	GFX_DEBUGGER_PAUSE_AT(NEXT_NEW_TEXTURE, true);
 
-	textures_by_address.emplace((u64)address, entry);
+	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)
 	{
@@ -636,7 +688,7 @@ TextureCache::TCacheEntryBase* TextureCache::Load(const u32 stage)
 	INCSTAT(stats.numTexturesUploaded);
 	SETSTAT(stats.numTexturesAlive, textures_by_address.size());
 
-	entry->DoPartialTextureUpdates();
+	entry = DoPartialTextureUpdates(iter);
 
 	return ReturnEntry(stage, entry);
 }
diff --git a/Source/Core/VideoCommon/TextureCacheBase.h b/Source/Core/VideoCommon/TextureCacheBase.h
index 5d4f9204fc..0a51ef64b9 100644
--- a/Source/Core/VideoCommon/TextureCacheBase.h
+++ b/Source/Core/VideoCommon/TextureCacheBase.h
@@ -89,7 +89,10 @@ public:
 		virtual void Bind(unsigned int stage) = 0;
 		virtual bool Save(const std::string& filename, unsigned int level) = 0;
 
-		virtual void DoPartialTextureUpdate(TCacheEntryBase* entry, u32 x, u32 y) = 0;
+		virtual void CopyRectangleFromTexture(
+			const TCacheEntryBase* source,
+			const MathUtil::Rectangle<int> &srcrect,
+			const MathUtil::Rectangle<int> &dstrect) = 0;
 
 		virtual void Load(unsigned int width, unsigned int height,
 			unsigned int expanded_width, unsigned int level) = 0;
@@ -100,8 +103,6 @@ public:
 
 		bool OverlapsMemoryRange(u32 range_address, u32 range_size) const;
 
-		void DoPartialTextureUpdates();
-
 		bool IsEfbCopy() const { return is_efb_copy; }
 	};
 
@@ -140,7 +141,7 @@ protected:
 private:
 	typedef std::multimap<u64, TCacheEntryBase*> TexCache;
 	typedef std::unordered_multimap<TCacheEntryConfig, TCacheEntryBase*, TCacheEntryConfig::Hasher> TexPool;
-
+	static TCacheEntryBase* DoPartialTextureUpdates(TexCache::iterator iter);
 	static void DumpTexture(TCacheEntryBase* entry, std::string basename, unsigned int level);
 	static void CheckTempSize(size_t required_size);