Vulkan: Fix Prescale UV

GPU/GLES/TransformPipeline.cpp

@@ -121,14 +121,14 @@ TransformDrawEngine::TransformDrawEngine()
 	: decodedVerts_(0),
 		prevPrim_(GE_PRIM_INVALID),
 		lastVType_(-1),
-		shaderManager_(0),
+		shaderManager_(nullptr),
-		textureCache_(0),
+		textureCache_(nullptr),
-		framebufferManager_(0),
+		framebufferManager_(nullptr),
 		numDrawCalls(0),
 		vertexCountInDrawCalls(0),
 		decodeCounter_(0),
 		dcid_(0),
-		uvScale(0),
+		uvScale(nullptr),
 		fboTexNeedBind_(false),
 		fboTexBound_(false) {
 	decimationCounter_ = VERTEXCACHE_DECIMATION_INTERVAL;

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -71,6 +71,7 @@ DrawEngineVulkan::DrawEngineVulkan(VulkanContext *vulkan)
 	framebufferManager_(nullptr),
 	numDrawCalls(0),
 	vertexCountInDrawCalls(0),
+	uvScale(nullptr),
 	fboTexNeedBind_(false),
 	fboTexBound_(false),
 	curFrame_(0),
@@ -92,6 +93,10 @@ DrawEngineVulkan::DrawEngineVulkan(VulkanContext *vulkan)
 
 	indexGen.Setup(decIndex);
 
+	if (g_Config.bPrescaleUV) {
+		uvScale = new UVScale[MAX_DEFERRED_DRAW_CALLS];
+	}
+
 	// All resources we need for PSP drawing. Usually only bindings 0 and 2-4 are populated.
 	VkDescriptorSetLayoutBinding bindings[5];
 	bindings[0].descriptorCount = 1;
@@ -202,6 +207,7 @@ DrawEngineVulkan::~DrawEngineVulkan() {
 		nullTexture_->Destroy();
 		delete nullTexture_;
 	}
+	delete[] uvScale;
 }
 
 void DrawEngineVulkan::BeginFrame() {
@@ -312,6 +318,10 @@ void DrawEngineVulkan::SubmitPrim(void *verts, void *inds, GEPrimitiveType prim,
 		dc.indexUpperBound = vertexCount - 1;
 	}
 
+	if (uvScale) {
+		uvScale[numDrawCalls] = gstate_c.uv;
+	}
+
 	numDrawCalls++;
 	vertexCountInDrawCalls += vertexCount;
 
@@ -324,6 +334,70 @@ void DrawEngineVulkan::SubmitPrim(void *verts, void *inds, GEPrimitiveType prim,
 	}
 }
 
+void DrawEngineVulkan::DecodeVertsStep(u8 *dest, int &i, int &decodedVerts) {
+	const DeferredDrawCall &dc = drawCalls[i];
+
+	indexGen.SetIndex(decodedVerts);
+	int indexLowerBound = dc.indexLowerBound, indexUpperBound = dc.indexUpperBound;
+
+	void *inds = dc.inds;
+	if (dc.indexType == GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT) {
+		// Decode the verts and apply morphing. Simple.
+		dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride,
+			dc.verts, indexLowerBound, indexUpperBound);
+		decodedVerts += indexUpperBound - indexLowerBound + 1;
+		indexGen.AddPrim(dc.prim, dc.vertexCount);
+	} else {
+		// It's fairly common that games issue long sequences of PRIM calls, with differing
+		// inds pointer but the same base vertex pointer. We'd like to reuse vertices between
+		// these as much as possible, so we make sure here to combine as many as possible
+		// into one nice big drawcall, sharing data.
+
+		// 1. Look ahead to find the max index, only looking as "matching" drawcalls.
+		//    Expand the lower and upper bounds as we go.
+		int lastMatch = i;
+		const int total = numDrawCalls;
+		for (int j = i + 1; j < total; ++j) {
+			if (drawCalls[j].verts != dc.verts)
+				break;
+
+			indexLowerBound = std::min(indexLowerBound, (int)drawCalls[j].indexLowerBound);
+			indexUpperBound = std::max(indexUpperBound, (int)drawCalls[j].indexUpperBound);
+			lastMatch = j;
+		}
+
+		// 2. Loop through the drawcalls, translating indices as we go.
+		switch (dc.indexType) {
+		case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
+			for (int j = i; j <= lastMatch; j++) {
+				indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u8 *)drawCalls[j].inds, indexLowerBound);
+			}
+			break;
+		case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
+			for (int j = i; j <= lastMatch; j++) {
+				indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u16 *)drawCalls[j].inds, indexLowerBound);
+			}
+			break;
+		}
+
+		const int vertexCount = indexUpperBound - indexLowerBound + 1;
+
+		// This check is a workaround for Pangya Fantasy Golf, which sends bogus index data when switching items in "My Room" sometimes.
+		if (decodedVerts + vertexCount > VERTEX_BUFFER_MAX) {
+			return;
+		}
+
+		// 3. Decode that range of vertex data.
+		dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride,
+			dc.verts, indexLowerBound, indexUpperBound);
+		decodedVerts += vertexCount;
+
+		// 4. Advance indexgen vertex counter.
+		indexGen.Advance(vertexCount);
+		i = lastMatch;
+	}
+}
+
 void DrawEngineVulkan::DecodeVerts(VulkanPushBuffer *push, uint32_t *bindOffset, VkBuffer *vkbuf) {
 	int decodedVerts = 0;
 
@@ -343,69 +417,19 @@ void DrawEngineVulkan::DecodeVerts(VulkanPushBuffer *push, uint32_t *bindOffset,
 		dest = (u8 *)push->Push(vertsToDecode * dec_->GetDecVtxFmt().stride, bindOffset, vkbuf);
 	}
 
-	for (int i = 0; i < numDrawCalls; i++) {
+	if (uvScale) {
-		const DeferredDrawCall &dc = drawCalls[i];
+		const UVScale origUV = gstate_c.uv;
-
+		for (int i = 0; i < numDrawCalls; i++) {
-		indexGen.SetIndex(decodedVerts);
+			gstate_c.uv = uvScale[i];
-		int indexLowerBound = dc.indexLowerBound, indexUpperBound = dc.indexUpperBound;
+			DecodeVertsStep(dest, i, decodedVerts);  // Note that this can modify i
-
+		}
-		void *inds = dc.inds;
+		gstate_c.uv = origUV;
-		if (dc.indexType == GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT) {
+	} else {
-			// Decode the verts and apply morphing. Simple.
+		for (int i = 0; i < numDrawCalls; i++) {
-			dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride,
+			DecodeVertsStep(dest, i, decodedVerts);  // Note that this can modify i
-				dc.verts, indexLowerBound, indexUpperBound);
-			decodedVerts += indexUpperBound - indexLowerBound + 1;
-			indexGen.AddPrim(dc.prim, dc.vertexCount);
-		} else {
-			// It's fairly common that games issue long sequences of PRIM calls, with differing
-			// inds pointer but the same base vertex pointer. We'd like to reuse vertices between
-			// these as much as possible, so we make sure here to combine as many as possible
-			// into one nice big drawcall, sharing data.
-
-			// 1. Look ahead to find the max index, only looking as "matching" drawcalls.
-			//    Expand the lower and upper bounds as we go.
-			int lastMatch = i;
-			const int total = numDrawCalls;
-			for (int j = i + 1; j < total; ++j) {
-				if (drawCalls[j].verts != dc.verts)
-					break;
-
-				indexLowerBound = std::min(indexLowerBound, (int)drawCalls[j].indexLowerBound);
-				indexUpperBound = std::max(indexUpperBound, (int)drawCalls[j].indexUpperBound);
-				lastMatch = j;
-			}
-
-			// 2. Loop through the drawcalls, translating indices as we go.
-			switch (dc.indexType) {
-			case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
-				for (int j = i; j <= lastMatch; j++) {
-					indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u8 *)drawCalls[j].inds, indexLowerBound);
-				}
-				break;
-			case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
-				for (int j = i; j <= lastMatch; j++) {
-					indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u16 *)drawCalls[j].inds, indexLowerBound);
-				}
-				break;
-			}
-
-			const int vertexCount = indexUpperBound - indexLowerBound + 1;
-
-			// This check is a workaround for Pangya Fantasy Golf, which sends bogus index data when switching items in "My Room" sometimes.
-			if (decodedVerts + vertexCount > VERTEX_BUFFER_MAX) {
-				return;
-			}
-
-			// 3. Decode that range of vertex data.
-			dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride,
-				dc.verts, indexLowerBound, indexUpperBound);
-			decodedVerts += vertexCount;
-
-			// 4. Advance indexgen vertex counter.
-			indexGen.Advance(vertexCount);
-			i = lastMatch;
 		}
 	}
+
 	// Sanity check
 	if (indexGen.Prim() < 0) {
 		ERROR_LOG_REPORT(G3D, "DecodeVerts: Failed to deduce prim: %i", indexGen.Prim());

GPU/Vulkan/DrawEngineVulkan.h

@@ -147,6 +147,8 @@ private:
 	struct FrameData;
 
 	void DecodeVerts(VulkanPushBuffer *push, uint32_t *bindOffset, VkBuffer *vkbuf);
+	void DecodeVertsStep(u8 *dest, int &i, int &decodedVerts);
+
 	void DoFlush(VkCommandBuffer cmd);
 	void UpdateUBOs(FrameData *frame);
 
@@ -244,6 +246,7 @@ private:
 	DeferredDrawCall drawCalls[MAX_DEFERRED_DRAW_CALLS];
 	int numDrawCalls;
 	int vertexCountInDrawCalls;
+	UVScale *uvScale;
 
 	bool fboTexNeedBind_;
 	bool fboTexBound_;

UI/GameSettingsScreen.cpp

@@ -311,6 +311,9 @@ void GameSettingsScreen::CreateViews() {
 	CheckBox *depthWrite = graphicsSettings->Add(new CheckBox(&g_Config.bAlwaysDepthWrite, gr->T("Always Depth Write")));
 	depthWrite->SetDisabledPtr(&g_Config.bSoftwareRendering);
 
+	graphicsSettings->Add(new CheckBox(&g_Config.bPrescaleUV, gr->T("Texture Coord Speedhack")));
+	depthWrite->SetDisabledPtr(&g_Config.bSoftwareRendering);
+
 	static const char *bloomHackOptions[] = { "Off", "Safe", "Balanced", "Aggressive" };
 	PopupMultiChoice *bloomHack = graphicsSettings->Add(new PopupMultiChoice(&g_Config.iBloomHack, gr->T("Lower resolution for effects (reduces artifacts)"), bloomHackOptions, 0, ARRAY_SIZE(bloomHackOptions), gr->GetName(), screenManager()));
 	bloomHackEnable_ = !g_Config.bSoftwareRendering && (g_Config.iInternalResolution != 1);