Fix lens flare in L.A. Rush

GPU/Common/DrawEngineCommon.cpp

@@ -35,15 +35,8 @@
 #include "GPU/ge_constants.h"
 #include "GPU/GPUState.h"
 
-#define QUAD_INDICES_MAX 65536
-
 enum {
 	TRANSFORMED_VERTEX_BUFFER_SIZE = VERTEX_BUFFER_MAX * sizeof(TransformedVertex),
-	DEPTH_TRANSFORMED_SIZE = VERTEX_BUFFER_MAX * 4 * sizeof(float),
-	DEPTH_SCREENVERTS_COMPONENT_COUNT = VERTEX_BUFFER_MAX,
-	DEPTH_SCREENVERTS_COMPONENT_SIZE = DEPTH_SCREENVERTS_COMPONENT_COUNT * sizeof(int) + 384,
-	DEPTH_SCREENVERTS_SIZE = DEPTH_SCREENVERTS_COMPONENT_SIZE * 3,
-	DEPTH_INDEXBUFFER_SIZE = VERTEX_BUFFER_MAX * 3 * sizeof(uint16_t),
 };
 
 DrawEngineCommon::DrawEngineCommon() : decoderMap_(32) {
@@ -56,33 +49,15 @@ DrawEngineCommon::DrawEngineCommon() : decoderMap_(32) {
 	decIndex_ = (u16 *)AllocateMemoryPages(DECODED_INDEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
 	indexGen.Setup(decIndex_);
 
-	switch ((DepthRasterMode)g_Config.iDepthRasterMode) {
-	case DepthRasterMode::DEFAULT:
-	case DepthRasterMode::LOW_QUALITY:
-		useDepthRaster_ = PSP_CoreParameter().compat.flags().SoftwareRasterDepth;
-		break;
-	case DepthRasterMode::FORCE_ON:
-		useDepthRaster_ = true;
-		break;
-	case DepthRasterMode::OFF:
-		useDepthRaster_ = false;
-	}
-
-	if (useDepthRaster_) {
-		depthDraws_.reserve(256);
-	}
+	InitDepthRaster();
 }
 
 DrawEngineCommon::~DrawEngineCommon() {
 	FreeMemoryPages(decoded_, DECODED_VERTEX_BUFFER_SIZE);
 	FreeMemoryPages(decIndex_, DECODED_INDEX_BUFFER_SIZE);
 	FreeMemoryPages(transformed_, TRANSFORMED_VERTEX_BUFFER_SIZE);
 	FreeMemoryPages(transformedExpanded_, 3 * TRANSFORMED_VERTEX_BUFFER_SIZE);
-	if (depthTransformed_) {
-		FreeMemoryPages(depthTransformed_, DEPTH_TRANSFORMED_SIZE);
-		FreeMemoryPages(depthScreenVerts_, DEPTH_SCREENVERTS_SIZE);
-		FreeMemoryPages(depthIndices_, DEPTH_INDEXBUFFER_SIZE);
-	}
+	ShutdownDepthRaster();
 	delete decJitCache_;
 	decoderMap_.Iterate([&](const uint32_t vtype, VertexDecoder *decoder) {
 		delete decoder;
@@ -795,11 +770,6 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 
 void DrawEngineCommon::BeginFrame() {
 	applySkinInDecode_ = g_Config.bSoftwareSkinning;
-	if (!depthTransformed_ && useDepthRaster_) {
-		depthTransformed_ = (float *)AllocateMemoryPages(DEPTH_TRANSFORMED_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
-		depthScreenVerts_ = (int *)AllocateMemoryPages(DEPTH_SCREENVERTS_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
-		depthIndices_ = (uint16_t *)AllocateMemoryPages(DEPTH_INDEXBUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
-	}
 }
 
 void DrawEngineCommon::DecodeVerts(VertexDecoder *dec, u8 *dest) {
@@ -925,6 +895,54 @@ bool DrawEngineCommon::DescribeCodePtr(const u8 *ptr, std::string &name) const {
 	}
 }
 
+enum {
+	DEPTH_TRANSFORMED_MAX_VERTS = VERTEX_BUFFER_MAX,
+	DEPTH_TRANSFORMED_BYTES = DEPTH_TRANSFORMED_MAX_VERTS * 4 * sizeof(float),
+	DEPTH_SCREENVERTS_COMPONENT_COUNT = VERTEX_BUFFER_MAX,
+	DEPTH_SCREENVERTS_COMPONENT_BYTES = DEPTH_SCREENVERTS_COMPONENT_COUNT * sizeof(int) + 384,
+	DEPTH_SCREENVERTS_TOTAL_BYTES = DEPTH_SCREENVERTS_COMPONENT_BYTES * 3,
+	DEPTH_INDEXBUFFER_BYTES = DEPTH_TRANSFORMED_MAX_VERTS * 3 * sizeof(uint16_t),  // hmmm
+};
+
+// We process vertices for depth rendering in several stages:
+// First, we transform and collect vertices into depthTransformed_ (4-vectors, xyzw).
+// Then, we group and cull the vertices into four-triangle groups, which are placed in
+// depthScreenVerts_, with x, y and z separated into different part of the array.
+// (Alternatively, if drawing rectangles, they're just added linearly).
+// After that, we send these groups out for SIMD setup and rasterization.
+void DrawEngineCommon::InitDepthRaster() {
+	switch ((DepthRasterMode)g_Config.iDepthRasterMode) {
+	case DepthRasterMode::DEFAULT:
+	case DepthRasterMode::LOW_QUALITY:
+		useDepthRaster_ = PSP_CoreParameter().compat.flags().SoftwareRasterDepth;
+		break;
+	case DepthRasterMode::FORCE_ON:
+		useDepthRaster_ = true;
+		break;
+	case DepthRasterMode::OFF:
+		useDepthRaster_ = false;
+	}
+
+	if (useDepthRaster_) {
+		depthDraws_.reserve(256);
+		depthTransformed_ = (float *)AllocateMemoryPages(DEPTH_TRANSFORMED_BYTES, MEM_PROT_READ | MEM_PROT_WRITE);
+		depthScreenVerts_ = (int *)AllocateMemoryPages(DEPTH_SCREENVERTS_TOTAL_BYTES, MEM_PROT_READ | MEM_PROT_WRITE);
+		depthIndices_ = (uint16_t *)AllocateMemoryPages(DEPTH_INDEXBUFFER_BYTES, MEM_PROT_READ | MEM_PROT_WRITE);
+	}
+}
+
+void DrawEngineCommon::ShutdownDepthRaster() {
+	if (depthTransformed_) {
+		FreeMemoryPages(depthTransformed_, DEPTH_TRANSFORMED_BYTES);
+	}
+	if (depthScreenVerts_) {
+		FreeMemoryPages(depthScreenVerts_, DEPTH_SCREENVERTS_TOTAL_BYTES);
+	}
+	if (depthIndices_) {
+		FreeMemoryPages(depthIndices_, DEPTH_INDEXBUFFER_BYTES);
+	}
+}
+
 Mat4F32 ComputeFinalProjMatrix() {
 	const float viewportTranslate[4] = {
 		gstate.getViewportXCenter() - gstate.getOffsetX(),
@@ -988,8 +1006,8 @@ bool DrawEngineCommon::CalculateDepthDraw(DepthDraw *draw, GEPrimitiveType prim,
 		_dbg_assert_(gstate.isDepthWriteEnabled());
 	}
 
-	if (depthVertexCount_ + vertexCount >= DEPTH_INDEXBUFFER_SIZE) {
-		// Can't add more.
+	if (depthVertexCount_ + vertexCount >= DEPTH_TRANSFORMED_MAX_VERTS) {
+		// Can't add more. We need to flush.
 		return false;
 	}
 
@@ -1008,7 +1026,7 @@ bool DrawEngineCommon::CalculateDepthDraw(DepthDraw *draw, GEPrimitiveType prim,
 	return true;
 }
 
-void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder *dec, uint32_t vertTypeID, int vertexCount) {
+void DrawEngineCommon::DepthRasterSubmitRaw(GEPrimitiveType prim, VertexDecoder *dec, uint32_t vertTypeID, int vertexCount) {
 	if (!gstate.isModeClear() && (!gstate.isDepthTestEnabled() || !gstate.isDepthWriteEnabled())) {
 		return;
 	}
@@ -1033,8 +1051,9 @@ void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder
 	int numDecoded = 0;
 	for (int i = 0; i < numDrawVerts_; i++) {
 		const DeferredVerts &dv = drawVerts_[i];
-		if (dv.indexUpperBound + 1 - dv.indexLowerBound + numDecoded >= VERTEX_BUFFER_MAX) {
+		if (dv.indexUpperBound + 1 - dv.indexLowerBound + numDecoded >= DEPTH_TRANSFORMED_MAX_VERTS) {
 			// Hit our limit! Stop decoding in this draw.
+			// We should have already broken out in CalculateDepthDraw.
 			break;
 		}
 		// Decode the verts (and at the same time apply morphing/skinning). Simple.
@@ -1070,6 +1089,7 @@ void DrawEngineCommon::DepthRasterPredecoded(GEPrimitiveType prim, const void *i
 
 	TimeCollector collectStat(&gpuStats.msPrepareDepth, coreCollectDebugStats);
 
+	// Make sure these have already been indexed away.
 	_dbg_assert_(prim != GE_PRIM_TRIANGLE_STRIP && prim != GE_PRIM_TRIANGLE_FAN);
 
 	if (dec->throughmode) {

GPU/Common/DrawEngineCommon.h

@@ -177,7 +177,9 @@ class DrawEngineCommon {
 
 	void ApplyFramebufferRead(FBOTexState *fboTexState);
 
-	void DepthRasterTransform(GEPrimitiveType prim, VertexDecoder *dec, uint32_t vertTypeID, int vertexCount);
+	void InitDepthRaster();
+	void ShutdownDepthRaster();
+	void DepthRasterSubmitRaw(GEPrimitiveType prim, VertexDecoder *dec, uint32_t vertTypeID, int vertexCount);
 	void DepthRasterPredecoded(GEPrimitiveType prim, const void *inVerts, int numDecoded, VertexDecoder *dec, int vertexCount);
 	bool CalculateDepthDraw(DepthDraw *draw, GEPrimitiveType prim, int vertexCount);
 

GPU/D3D11/DrawEngineD3D11.cpp

@@ -327,7 +327,7 @@ void DrawEngineD3D11::Flush() {
 			}
 		}
 		if (useDepthRaster_) {
-			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+			DepthRasterSubmitRaw(prim, dec_, dec_->VertexType(), vertexCount);
 		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");

GPU/Directx9/DrawEngineDX9.cpp

@@ -293,7 +293,7 @@ void DrawEngineDX9::Flush() {
 			}
 		}
 		if (useDepthRaster_) {
-			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+			DepthRasterSubmitRaw(prim, dec_, dec_->VertexType(), vertexCount);
 		}
 	} else {
 		VertexDecoder *swDec = dec_;

GPU/GLES/DrawEngineGLES.cpp

@@ -316,7 +316,7 @@ void DrawEngineGLES::Flush() {
 				glprim[prim], 0, vertexCount);
 		}
 		if (useDepthRaster_) {
-			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+			DepthRasterSubmitRaw(prim, dec_, dec_->VertexType(), vertexCount);
 		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -371,7 +371,7 @@ void DrawEngineVulkan::Flush() {
 			renderManager->Draw(descSetIndex, ARRAY_SIZE(dynamicUBOOffsets), dynamicUBOOffsets, vbuf, vbOffset, vertexCount);
 		}
 		if (useDepthRaster_) {
-			DepthRasterTransform(prim, dec_, dec_->VertexType(), vertexCount);
+			DepthRasterSubmitRaw(prim, dec_, dec_->VertexType(), vertexCount);
 		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");

assets/compat.ini

@@ -1712,6 +1712,10 @@ UCAS40289 = true
 NPUH10025 = true
 NPEH00047 = true
 
+# L.A. Rush (also known as just Rush) (sun lens flare)
+ULES00554 = true
+ULUS10174 = true
+
 [BlockTransferDepth]
 # Iron Man - see issue #16530
 # Note that this option also requires IntraVRAMBlockTransferAllowCreateFB.