Merge pull request #18646 from hrydgard/through-mode-smart-filtering

Through-mode smart texture filtering

Common/GPU/Vulkan/VulkanFrameData.h

@@ -29,6 +29,7 @@ struct QueueProfileContext {
 	double cpuEndTime;
 	double descWriteTime;
 	int descriptorsWritten;
+	int descriptorsDeduped;
 #ifdef _DEBUG
 	int commandCounts[11];
 #endif

Common/GPU/Vulkan/VulkanRenderManager.cpp

@@ -687,7 +687,7 @@ void VulkanRenderManager::BeginFrame(bool enableProfiling, bool enableLogProfile
 				descUpdateTimeMs_.Update(frameData.profile.descWriteTime * 1000.0);
 				descUpdateTimeMs_.Format(line, sizeof(line));
 				str << line;
-				snprintf(line, sizeof(line), "Descriptors written: %d\n", frameData.profile.descriptorsWritten);
+				snprintf(line, sizeof(line), "Descriptors written: %d (dedup: %d)\n", frameData.profile.descriptorsWritten, frameData.profile.descriptorsDeduped);
 				str << line;
 				snprintf(line, sizeof(line), "Resource deletions: %d\n", vulkan_->GetLastDeleteCount());
 				str << line;
@@ -737,6 +737,7 @@ void VulkanRenderManager::BeginFrame(bool enableProfiling, bool enableLogProfile
 	}
 
 	frameData.profile.descriptorsWritten = 0;
+	frameData.profile.descriptorsDeduped = 0;
 
 	// Must be after the fence - this performs deletes.
 	VLOG("PUSH: BeginFrame %d", curFrame);
@@ -1747,7 +1748,7 @@ void VKRPipelineLayout::FlushDescSets(VulkanContext *vulkan, int frame, QueuePro
 	VkDescriptorBufferInfo bufferInfo[MAX_DESC_SET_BINDINGS];
 
 	size_t start = data.flushedDescriptors_;
-	int writeCount = 0;
+	int writeCount = 0, dedupCount = 0;
 
 	for (size_t index = start; index < descSets.size(); index++) {
 		auto &d = descSets[index];
@@ -1759,6 +1760,7 @@ void VKRPipelineLayout::FlushDescSets(VulkanContext *vulkan, int frame, QueuePro
 			if (descSets[index - 1].count == d.count) {
 				if (!memcmp(descData.data() + d.offset, descData.data() + descSets[index - 1].offset, d.count * sizeof(PackedDescriptor))) {
 					d.set = descSets[index - 1].set;
+					dedupCount++;
 					continue;
 				}
 			}
@@ -1841,4 +1843,5 @@ void VKRPipelineLayout::FlushDescSets(VulkanContext *vulkan, int frame, QueuePro
 
 	data.flushedDescriptors_ = (int)descSets.size();
 	profile->descriptorsWritten += writeCount;
+	profile->descriptorsDeduped += dedupCount;
 }

Core/Config.cpp

@@ -611,6 +611,7 @@ static const ConfigSetting graphicsSettings[] = {
 	ConfigSetting("HardwareTransform", &g_Config.bHardwareTransform, true, CfgFlag::PER_GAME | CfgFlag::REPORT),
 	ConfigSetting("SoftwareSkinning", &g_Config.bSoftwareSkinning, true, CfgFlag::PER_GAME | CfgFlag::REPORT),
 	ConfigSetting("TextureFiltering", &g_Config.iTexFiltering, 1, CfgFlag::PER_GAME | CfgFlag::REPORT),
+	ConfigSetting("Smart2DTexFiltering", &g_Config.bSmart2DTexFiltering, false, CfgFlag::PER_GAME | CfgFlag::REPORT),
 	ConfigSetting("InternalResolution", &g_Config.iInternalResolution, &DefaultInternalResolution, CfgFlag::PER_GAME | CfgFlag::REPORT),
 	ConfigSetting("AndroidHwScale", &g_Config.iAndroidHwScale, &DefaultAndroidHwScale, CfgFlag::DEFAULT),
 	ConfigSetting("HighQualityDepth", &g_Config.bHighQualityDepth, true, CfgFlag::PER_GAME | CfgFlag::REPORT),

Core/Config.h

@@ -172,6 +172,7 @@ struct Config {
 	bool bDisableRangeCulling;
 
 	int iTexFiltering; // 1 = auto , 2 = nearest , 3 = linear , 4 = auto max quality
+	bool bSmart2DTexFiltering;
 
 	bool bDisplayStretch;  // Automatically matches the aspect ratio of the window.
 	int iDisplayFilter;    // 1 = linear, 2 = nearest

GPU/Common/SoftwareTransformCommon.cpp

@@ -179,7 +179,8 @@ void SoftwareTransform::Transform(int prim, u32 vertType, const DecVtxFormat &de
 
 	float uscale = 1.0f;
 	float vscale = 1.0f;
-	if (throughmode) {
+	if (throughmode && prim != GE_PRIM_RECTANGLES) {
+		// For through rectangles, we do this scaling in Expand.
 		uscale /= gstate_c.curTextureWidth;
 		vscale /= gstate_c.curTextureHeight;
 	}
@@ -496,7 +497,7 @@ void SoftwareTransform::BuildDrawingParams(int prim, int vertexCount, u32 vertTy
 	bool useBufferedRendering = fbman->UseBufferedRendering();
 
 	if (prim == GE_PRIM_RECTANGLES) {
-		ExpandRectangles(vertexCount, maxIndex, inds, transformed, transformedExpanded, numTrans, throughmode);
+		ExpandRectangles(vertexCount, maxIndex, inds, transformed, transformedExpanded, numTrans, throughmode, &result->pixelMapped);
 		result->drawBuffer = transformedExpanded;
 		result->drawIndexed = true;
 
@@ -517,14 +518,17 @@ void SoftwareTransform::BuildDrawingParams(int prim, int vertexCount, u32 vertTy
 		ExpandPoints(vertexCount, maxIndex, inds, transformed, transformedExpanded, numTrans, throughmode);
 		result->drawBuffer = transformedExpanded;
 		result->drawIndexed = true;
+		result->pixelMapped = false;
 	} else if (prim == GE_PRIM_LINES) {
 		ExpandLines(vertexCount, maxIndex, inds, transformed, transformedExpanded, numTrans, throughmode);
 		result->drawBuffer = transformedExpanded;
 		result->drawIndexed = true;
+		result->pixelMapped = false;
 	} else {
 		// We can simply draw the unexpanded buffer.
 		numTrans = vertexCount;
 		result->drawIndexed = true;
+		result->pixelMapped = false;
 
 		// If we don't support custom cull in the shader, process it here.
 		if (!gstate_c.Use(GPU_USE_CULL_DISTANCE) && vertexCount > 0 && !throughmode) {
@@ -581,6 +585,37 @@ void SoftwareTransform::BuildDrawingParams(int prim, int vertexCount, u32 vertTy
 
 				inds = newInds;
 			}
+		} else if (throughmode && g_Config.bSmart2DTexFiltering) {
+			// We check some common cases for pixel mapping.
+			// TODO: It's not really optimal that some previous step has removed the triangle strip.
+			if (vertexCount <= 6 && prim == GE_PRIM_TRIANGLES) {
+				// It's enough to check UV deltas vs pos deltas between vertex pairs:
+				// 0-1 1-3 3-2 2-0. Maybe can even skip the last one. Probably some simple math can get us that sequence.
+				// Unfortunately we need to reverse the previous UV scaling operation. Fortunately these are powers of two
+				// so the operations are exact.
+				bool pixelMapped = true;
+				const u16 *indsIn = (const u16 *)inds;
+				for (int t = 0; t < vertexCount; t += 3) {
+					float uscale = gstate_c.curTextureWidth;
+					float vscale = gstate_c.curTextureHeight;
+					struct { int a; int b; } pairs[] = { {0, 1}, {1, 2}, {2, 0} };
+					for (int i = 0; i < ARRAY_SIZE(pairs); i++) {
+						int a = indsIn[t + pairs[i].a];
+						int b = indsIn[t + pairs[i].b];
+						float du = fabsf((transformed[a].u - transformed[b].u) * uscale);
+						float dv = fabsf((transformed[a].v - transformed[b].v) * vscale);
+						float dx = fabsf(transformed[a].x - transformed[b].x);
+						float dy = fabsf(transformed[a].y - transformed[b].y);
+						if (du != dx || dv != dy) {
+							pixelMapped = false;
+						}
+					}
+					if (!pixelMapped) {
+						break;
+					}
+				}
+				result->pixelMapped = pixelMapped;
+			}
 		}
 	}
 
@@ -610,7 +645,7 @@ void SoftwareTransform::CalcCullParams(float &minZValue, float &maxZValue) {
 		std::swap(minZValue, maxZValue);
 }
 
-void SoftwareTransform::ExpandRectangles(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode) {
+void SoftwareTransform::ExpandRectangles(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode, bool *pixelMappedExactly) {
 	// Rectangles always need 2 vertices, disregard the last one if there's an odd number.
 	vertexCount = vertexCount & ~1;
 	numTrans = 0;
@@ -621,32 +656,59 @@ void SoftwareTransform::ExpandRectangles(int vertexCount, int &maxIndex, u16 *&i
 	u16 *indsOut = newInds;
 
 	maxIndex = 4 * (vertexCount / 2);
+
+	float uscale = 1.0f;
+	float vscale = 1.0f;
+	if (throughmode) {
+		uscale /= gstate_c.curTextureWidth;
+		vscale /= gstate_c.curTextureHeight;
+	}
+
+	bool pixelMapped = g_Config.bSmart2DTexFiltering;
+
 	for (int i = 0; i < vertexCount; i += 2) {
 		const TransformedVertex &transVtxTL = transformed[indsIn[i + 0]];
 		const TransformedVertex &transVtxBR = transformed[indsIn[i + 1]];
 
+		if (pixelMapped) {
+			float dx = transVtxBR.x - transVtxTL.x;
+			float dy = transVtxBR.y - transVtxTL.y;
+			float du = transVtxBR.u - transVtxTL.u;
+			float dv = transVtxBR.v - transVtxTL.v;
+
+			// NOTE: We will accept it as pixel mapped if only one dimension is stretched. This fixes dialog frames in FFI.
+			// Though, there could be false positives in other games due to this. Let's see if it is a problem...
+			if (dx <= 0 || dy <= 0 || (dx != du && dy != dv)) {
+				pixelMapped = false;
+			}
+		}
+
 		// We have to turn the rectangle into two triangles, so 6 points.
 		// This is 4 verts + 6 indices.
 
 		// bottom right
 		trans[0] = transVtxBR;
+		trans[0].u = transVtxBR.u * uscale;
+		trans[0].v = transVtxBR.v * vscale;
 
 		// top right
 		trans[1] = transVtxBR;
 		trans[1].y = transVtxTL.y;
-		trans[1].v = transVtxTL.v;
+		trans[1].u = transVtxBR.u * uscale;
+		trans[1].v = transVtxTL.v * vscale;
 
 		// top left
 		trans[2] = transVtxBR;
 		trans[2].x = transVtxTL.x;
 		trans[2].y = transVtxTL.y;
-		trans[2].u = transVtxTL.u;
-		trans[2].v = transVtxTL.v;
+		trans[2].u = transVtxTL.u * uscale;
+		trans[2].v = transVtxTL.v * vscale;
 
 		// bottom left
 		trans[3] = transVtxBR;
 		trans[3].x = transVtxTL.x;
-		trans[3].u = transVtxTL.u;
+		trans[3].u = transVtxTL.u * uscale;
+		trans[3].v = transVtxBR.v * vscale;
 
 		// That's the four corners. Now process UV rotation.
 		if (throughmode) {
@@ -663,12 +725,14 @@ void SoftwareTransform::ExpandRectangles(int vertexCount, int &maxIndex, u16 *&i
 		indsOut[3] = i * 2 + 3;
 		indsOut[4] = i * 2 + 0;
 		indsOut[5] = i * 2 + 2;
+
 		trans += 4;
 		indsOut += 6;
 
 		numTrans += 6;
 	}
 	inds = newInds;
+	*pixelMappedExactly = pixelMapped;
 }
 
 void SoftwareTransform::ExpandLines(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode) {

GPU/Common/SoftwareTransformCommon.h

@@ -45,6 +45,8 @@ struct SoftwareTransformResult {
 	TransformedVertex *drawBuffer;
 	int drawNumTrans;
 	bool drawIndexed;
+
+	bool pixelMapped;
 };
 
 struct SoftwareTransformParams {
@@ -70,7 +72,7 @@ class SoftwareTransform {
 
 protected:
 	void CalcCullParams(float &minZValue, float &maxZValue);
-	void ExpandRectangles(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode);
+	void ExpandRectangles(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode, bool *pixelMappedExactly);
 	void ExpandLines(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode);
 	void ExpandPoints(int vertexCount, int &maxIndex, u16 *&inds, const TransformedVertex *transformed, TransformedVertex *transformedExpanded, int &numTrans, bool throughmode);
 

GPU/Common/TextureCacheCommon.cpp

@@ -261,6 +261,9 @@ SamplerCacheKey TextureCacheCommon::GetSamplingParams(int maxLevel, const TexCac
 				if (uglyColorTest)
 					forceFiltering = TEX_FILTER_FORCE_NEAREST;
 			}
+			if (gstate_c.pixelMapped) {
+				forceFiltering = TEX_FILTER_FORCE_NEAREST;
+			}
 			break;
 		case TEX_FILTER_FORCE_LINEAR:
 			// Override to linear filtering if there's no alpha or color testing going on.
@@ -281,6 +284,9 @@ SamplerCacheKey TextureCacheCommon::GetSamplingParams(int maxLevel, const TexCac
 				if (uglyColorTest)
 					forceFiltering = TEX_FILTER_FORCE_NEAREST;
 			}
+			if (gstate_c.pixelMapped) {
+				forceFiltering = TEX_FILTER_FORCE_NEAREST;
+			}
 			break;
 		}
 	}

GPU/D3D11/DrawEngineD3D11.cpp

@@ -408,8 +408,11 @@ void DrawEngineD3D11::DoFlush() {
 		if (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
 			result.action = SW_NOT_READY;
 
-		if (textureNeedsApply)
+		if (textureNeedsApply) {
+			gstate_c.pixelMapped = result.pixelMapped;
 			textureCache_->ApplyTexture();
+			gstate_c.pixelMapped = false;
+		}
 
 		// Need to ApplyDrawState after ApplyTexture because depal can launch a render pass and that wrecks the state.
 		ApplyDrawState(prim);

GPU/Directx9/DrawEngineDX9.cpp

@@ -365,8 +365,11 @@ void DrawEngineDX9::DoFlush() {
 		if (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
 			result.action = SW_NOT_READY;
 
-		if (textureNeedsApply)
+		if (textureNeedsApply) {
+			gstate_c.pixelMapped = result.pixelMapped;
 			textureCache_->ApplyTexture();
+			gstate_c.pixelMapped = false;
+		}
 
 		ApplyDrawState(prim);
 

GPU/GLES/DrawEngineGLES.cpp

@@ -393,8 +393,11 @@ void DrawEngineGLES::DoFlush() {
 		if (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
 			result.action = SW_NOT_READY;
 
-		if (textureNeedsApply)
+		if (textureNeedsApply) {
+			gstate_c.pixelMapped = result.pixelMapped;
 			textureCache_->ApplyTexture();
+			gstate_c.pixelMapped = false;
+		}
 
 		// Need to ApplyDrawState after ApplyTexture because depal can launch a render pass and that wrecks the state.
 		ApplyDrawState(prim);

GPU/GPUState.h

@@ -689,6 +689,9 @@ struct GPUStateCache {
 	// We detect this case and go into a special drawing mode.
 	bool blueToAlpha;
 
+	// U/V is 1:1 to pixels. Can influence texture sampling.
+	bool pixelMapped;
+
 	// TODO: These should be accessed from the current VFB object directly.
 	u32 curRTWidth;
 	u32 curRTHeight;

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -440,7 +440,9 @@ void DrawEngineVulkan::DoFlush() {
 		// to use a "pre-clear" render pass, for high efficiency on tilers.
 		if (result.action == SW_DRAW_PRIMITIVES) {
 			if (textureNeedsApply) {
+				gstate_c.pixelMapped = result.pixelMapped;
 				textureCache_->ApplyTexture();
+				gstate_c.pixelMapped = false;
 				textureCache_->GetVulkanHandles(imageView, sampler);
 				if (imageView == VK_NULL_HANDLE)
 					imageView = (VkImageView)draw_->GetNativeObject(gstate_c.textureIsArray ? Draw::NativeObject::NULL_IMAGEVIEW_ARRAY : Draw::NativeObject::NULL_IMAGEVIEW);

GPU/Vulkan/TextureCacheVulkan.cpp

@@ -136,6 +136,7 @@ VkSampler SamplerCache::GetOrCreateSampler(const SamplerCacheKey &key) {
 	samp.magFilter = key.magFilt ? VK_FILTER_LINEAR : VK_FILTER_NEAREST;
 	samp.minFilter = key.minFilt ? VK_FILTER_LINEAR : VK_FILTER_NEAREST;
 	samp.mipmapMode = key.mipFilt ? VK_SAMPLER_MIPMAP_MODE_LINEAR : VK_SAMPLER_MIPMAP_MODE_NEAREST;
+
 	if (key.aniso) {
 		// Docs say the min of this value and the supported max are used.
 		samp.maxAnisotropy = 1 << g_Config.iAnisotropyLevel;

UI/GameSettingsScreen.cpp

@@ -553,6 +553,8 @@ void GameSettingsScreen::CreateGraphicsSettings(UI::ViewGroup *graphicsSettings)
 	static const char *texFilters[] = { "Auto", "Nearest", "Linear", "Auto Max Quality"};
 	graphicsSettings->Add(new PopupMultiChoice(&g_Config.iTexFiltering, gr->T("Texture Filter"), texFilters, 1, ARRAY_SIZE(texFilters), I18NCat::GRAPHICS, screenManager()));
 
+	graphicsSettings->Add(new CheckBox(&g_Config.bSmart2DTexFiltering, gr->T("Smart 2D texture filtering")));
+
 #if PPSSPP_PLATFORM(ANDROID) || PPSSPP_PLATFORM(IOS)
 	bool showCardboardSettings = deviceType != DEVICE_TYPE_VR;
 #else