Don't forget to update the vertex addr even when culling 2D draws

Fixes #20454

GPU/Common/DrawEngineCommon.cpp

@@ -527,25 +527,21 @@ bool DrawEngineCommon::TestBoundingBoxFast(const void *vdata, int vertexCount, c
 
 // 2D bounding box test against scissor. No indexing yet.
 // Only supports non-indexed draws with float positions.
-bool DrawEngineCommon::TestBoundingBoxThrough(const void *vdata, int vertexCount, const VertexDecoder *dec, u32 vertType) {
+bool DrawEngineCommon::TestBoundingBoxThrough(const void *vdata, int vertexCount, const VertexDecoder *dec, u32 vertType, int *bytesRead) {
 	// Grab temp buffer space from large offsets in decoded_. Not exactly safe for large draws.
 	if (vertexCount > 16) {
 		return true;
 	}
 
-	float *verts = (float *)(decoded_ + 65536 * 18);
-
 	// Although this may lead to drawing that shouldn't happen, the viewport is more complex on VR.
 	// Let's always say objects are within bounds.
 	if (gstate_c.Use(GPU_USE_VIRTUAL_REALITY))
 		return true;
 
-	// Try to skip NormalizeVertices if it's pure positions. No need to bother with a vertex decoder
-	// and a large vertex format.
-	u8 *temp_buffer = decoded_ + 65536 * 24;
-	// Simple, most common case.
-	int stride = dec->VertexSize();
-	int offset = dec->posoff;
+	const int stride = dec->VertexSize();
+	const int posOffset = dec->posoff;
+
+	*bytesRead = stride * vertexCount;
 
 	bool allOutsideLeft = true;
 	bool allOutsideTop = true;
@@ -559,10 +555,11 @@ bool DrawEngineCommon::TestBoundingBoxThrough(const void *vdata, int vertexCount
 	switch (vertType & GE_VTYPE_POS_MASK) {
 	case GE_VTYPE_POS_FLOAT:
 	{
+		// TODO: This can be SIMD'd, with some trickery.
 		for (int i = 0; i < vertexCount; i++) {
-			float *pos = (float*)((const u8 *)vdata + stride * i + offset);
-			float x = pos[0];
-			float y = pos[1];
+			const float *pos = (const float*)((const u8 *)vdata + stride * i + posOffset);
+			const float x = pos[0];
+			const float y = pos[1];
 			if (x >= left) {
 				allOutsideLeft = false;
 			}
@@ -582,8 +579,9 @@ bool DrawEngineCommon::TestBoundingBoxThrough(const void *vdata, int vertexCount
 		return true;
 	}
 	default:
+		// Shouldn't end up here with the checks outside this function.
 		_dbg_assert_(false);
-		return false;
+		return true;
 	}
 }
 

GPU/Common/DrawEngineCommon.h

@@ -104,7 +104,7 @@ public:
 	// This is a less accurate version of TestBoundingBox, but faster. Can have more false positives.
 	// Doesn't support indexing.
 	bool TestBoundingBoxFast(const void *control_points, int vertexCount, const VertexDecoder *dec, u32 vertType);
-	bool TestBoundingBoxThrough(const void *vdata, int vertexCount, const VertexDecoder *dec, u32 vertType);
+	bool TestBoundingBoxThrough(const void *vdata, int vertexCount, const VertexDecoder *dec, u32 vertType, int *bytesRead);
 
 	void FlushPartialDecode() {
 		DecodeVerts(dec_, decoded_);

GPU/GPUCommonHW.cpp

@@ -1000,7 +1000,6 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		canExtend = false;
 	}
 
-	int bytesRead = 0;
 	gstate_c.UpdateUVScaleOffset();
 
 	// cull mode
@@ -1011,11 +1010,14 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 
 	// Through mode early-out for simple float 2D draws, like in Fate Extra CCC (very beneficial there due to avoiding texture loads)
 	if ((vertexType & (GE_VTYPE_THROUGH_MASK | GE_VTYPE_POS_MASK | GE_VTYPE_IDX_MASK)) == (GE_VTYPE_THROUGH_MASK | GE_VTYPE_POS_FLOAT | GE_VTYPE_IDX_NONE)) {
-		if (!drawEngineCommon_->TestBoundingBoxThrough(verts, count, decoder, vertexType)) {
+		int bytesRead = 0;
+		if (!drawEngineCommon_->TestBoundingBoxThrough(verts, count, decoder, vertexType, &bytesRead)) {
 			gpuStats.numCulledDraws++;
 			int cycles = vertexCost_ * count;
 			gpuStats.vertexGPUCycles += cycles;
 			cyclesExecuted += cycles;
+			// NOTE! We still have to advance vertex pointers!
+			gstate_c.vertexAddr += bytesRead;   // We know from the above check that it's not an indexed draw.
 			return;
 		}
 	}
@@ -1044,6 +1046,8 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		}
 	}
 
+	int bytesRead = 0;
+
 	// If the first one in a batch passes, let's assume the whole batch passes.
 	// Cuts down on checking, while not losing that much efficiency.
 	bool onePassed = false;
@@ -1357,7 +1361,7 @@ void GPUCommonHW::Execute_Bezier(u32 op, u32 diff) {
 	gstate_c.submitType = SubmitType::DRAW;
 
 	// After drawing, we advance pointers - see SubmitPrim which does the same.
-	int count = surface.num_points_u * surface.num_points_v;
+	const int count = surface.num_points_u * surface.num_points_v;
 	AdvanceVerts(gstate.vertType, count, bytesRead);
 }