Avoid decoding indices when we don't need them.

GPU/Common/DrawEngineCommon.cpp

@@ -816,6 +816,10 @@ int DrawEngineCommon::ExtendNonIndexedPrim(const uint32_t *cmd, const uint32_t *
 	_dbg_assert_(numDrawInds_ <= MAX_DEFERRED_DRAW_INDS);  // if it's equal, the check below will take care of it before any action is taken.
 	_dbg_assert_(numDrawVerts_ > 0);
 
+	if (!clockwise) {
+		anyCCWOrIndexed_ = true;
+	}
+	int seenPrims = 0;
 	while (cmd != stall) {
 		uint32_t data = *cmd;
 		if ((data & 0xFFF80000) != 0x04000000) {
@@ -831,6 +835,7 @@ int DrawEngineCommon::ExtendNonIndexedPrim(const uint32_t *cmd, const uint32_t *
 		DeferredInds &di = drawInds_[numDrawInds_++];
 		di.indexType = 0;
 		di.prim = newPrim;
+		seenPrims |= (1 << newPrim);
 		di.clockwise = clockwise;
 		di.vertexCount = vertexCount;
 		di.vertDecodeIndex = prevDrawVerts;
@@ -839,6 +844,10 @@ int DrawEngineCommon::ExtendNonIndexedPrim(const uint32_t *cmd, const uint32_t *
 		cmd++;
 	}
 
+	seenPrims_ |= seenPrims;
+
+	_dbg_assert_(cmd != start);
+
 	int totalCount = offset - dv.vertexCount;
 	dv.vertexCount = offset;
 	dv.indexUpperBound = dv.vertexCount - 1;
@@ -910,9 +919,16 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 
 	DeferredInds &di = drawInds_[numDrawInds_++];
 	di.inds = inds;
-	di.indexType = (vertTypeID & GE_VTYPE_IDX_MASK) >> GE_VTYPE_IDX_SHIFT;
+	int indexType = (vertTypeID & GE_VTYPE_IDX_MASK) >> GE_VTYPE_IDX_SHIFT;
+	if (indexType) {
+		anyCCWOrIndexed_ = true;
+	}
+	di.indexType = indexType;
 	di.prim = prim;
 	di.clockwise = clockwise;
+	if (!clockwise) {
+		anyCCWOrIndexed_ = true;
+	}
 	di.vertexCount = vertexCount;
 	di.vertDecodeIndex = numDrawVerts_;
 	di.offset = 0;
@@ -942,6 +958,7 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	}
 
 	vertexCountInDrawCalls_ += vertexCount;
+	seenPrims_ |= (1 << prim);
 
 	if (prim == GE_PRIM_RECTANGLES && (gstate.getTextureAddress(0) & 0x3FFFFFFF) == (gstate.getFrameBufAddress() & 0x3FFFFFFF)) {
 		// This prevents issues with consecutive self-renders in Ridge Racer.
@@ -952,6 +969,8 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 }
 
 void DrawEngineCommon::DecodeVerts(u8 *dest) {
+	// Note that this should be able to continue a partial decode - we don't necessarily start from zero here (although we do most of the time).
+
 	int i = decodeVertsCounter_;
 	int stride = (int)dec_->GetDecVtxFmt().stride;
 	for (; i < numDrawVerts_; i++) {
@@ -968,7 +987,9 @@ void DrawEngineCommon::DecodeVerts(u8 *dest) {
 	decodeVertsCounter_ = i;
 }
 
-void DrawEngineCommon::DecodeInds() {
+int DrawEngineCommon::DecodeInds() {
+	// Note that this should be able to continue a partial decode - we don't necessarily start from zero here (although we do most of the time).
+
 	int i = decodeIndsCounter_;
 	for (; i < numDrawInds_; i++) {
 		const DeferredInds &di = drawInds_[i];
@@ -994,12 +1015,7 @@ void DrawEngineCommon::DecodeInds() {
 	}
 	decodeIndsCounter_ = i;
 
-	// Sanity check
-	if (indexGen.Prim() < 0) {
-		ERROR_LOG_REPORT(G3D, "DecodeVerts: Failed to deduce prim: %i", indexGen.Prim());
-		// Force to points (0)
-		indexGen.AddPrim(GE_PRIM_POINTS, 0, 0, true);
-	}
+	return indexGen.VertexCount();
 }
 
 bool DrawEngineCommon::CanUseHardwareTransform(int prim) {

GPU/Common/DrawEngineCommon.h

@@ -152,7 +152,7 @@ class DrawEngineCommon {
 	void UpdatePlanes();
 
 	void DecodeVerts(u8 *dest);
-	void DecodeInds();
+	int DecodeInds();
 
 	// Preprocessing for spline/bezier
 	u32 NormalizeVertices(u8 *outPtr, u8 *bufPtr, const u8 *inPtr, int lowerBound, int upperBound, u32 vertType, int *vertexSize = nullptr);
@@ -202,6 +202,8 @@ class DrawEngineCommon {
 		vertexCountInDrawCalls_ = 0;
 		decodeIndsCounter_ = 0;
 		decodeVertsCounter_ = 0;
+		seenPrims_ = 0;
+		anyCCWOrIndexed_ = false;
 		gstate_c.vertexFullAlpha = true;
 
 		// Now seems as good a time as any to reset the min/max coords, which we may examine later.
@@ -211,6 +213,34 @@ class DrawEngineCommon {
 		gstate_c.vertBounds.maxV = 0;
 	}
 
+	inline bool CollectedPureDraw() const {
+		switch (seenPrims_) {
+		case 1 << GE_PRIM_TRIANGLE_STRIP:
+			return !anyCCWOrIndexed_ && numDrawInds_ == 1;
+		case 1 << GE_PRIM_LINES:
+		case 1 << GE_PRIM_POINTS:
+		case 1 << GE_PRIM_TRIANGLES:
+			return !anyCCWOrIndexed_;
+		default:
+			return false;
+		}
+	}
+
+	inline void DecodeIndsAndGetData(GEPrimitiveType *prim, int *numVerts, int *maxIndex, bool *useElements, bool forceIndexed) {
+		if (!forceIndexed && CollectedPureDraw()) {
+			*prim = drawInds_[0].prim;
+			*numVerts = numDecodedVerts_;
+			*maxIndex = numDecodedVerts_;
+			*useElements = false;
+		} else {
+			int vertexCount = DecodeInds();
+			*numVerts = vertexCount;
+			*maxIndex = numDecodedVerts_;
+			*prim = IndexGenerator::GeneralPrim((GEPrimitiveType)drawInds_[0].prim);
+			*useElements = true;
+		}
+	}
+
 	uint32_t ComputeDrawcallsHash() const;
 
 	bool useHWTransform_ = false;
@@ -227,9 +257,7 @@ class DrawEngineCommon {
 	u16 *decIndex_ = nullptr;
 
 	// Cached vertex decoders
-	u32 lastVType_ = -1;  // corresponds to dec_.  Could really just pick it out of dec_...
 	DenseHashMap<u32, VertexDecoder *> decoderMap_;
-	VertexDecoder *dec_ = nullptr;
 	VertexDecoderJitCache *decJitCache_ = nullptr;
 	VertexDecoderOptions decOptions_{};
 
@@ -239,18 +267,18 @@ class DrawEngineCommon {
 	// Defer all vertex decoding to a "Flush" (except when software skinning)
 	struct DeferredVerts {
 		const void *verts;
+		UVScale uvScale;
 		u32 vertexCount;
 		u16 indexLowerBound;
 		u16 indexUpperBound;
-		UVScale uvScale;
 	};
 
 	struct DeferredInds {
 		const void *inds;
 		u32 vertexCount;
 		u8 vertDecodeIndex;  // index into the drawVerts_ array to look up the vertexOffset.
 		u8 indexType;
-		s8 prim;
+		GEPrimitiveType prim;
 		bool clockwise;
 		u16 offset;
 	};
@@ -261,13 +289,19 @@ class DrawEngineCommon {
 	uint32_t drawVertexOffsets_[MAX_DEFERRED_DRAW_VERTS];
 	DeferredInds drawInds_[MAX_DEFERRED_DRAW_INDS];
 
+	VertexDecoder *dec_ = nullptr;
+	u32 lastVType_ = -1;  // corresponds to dec_.  Could really just pick it out of dec_...
 	int numDrawVerts_ = 0;
 	int numDrawInds_ = 0;
 	int vertexCountInDrawCalls_ = 0;
 
 	int decodeVertsCounter_ = 0;
 	int decodeIndsCounter_ = 0;
 
+	int seenPrims_ = 0;
+	bool anyCCWOrIndexed_ = 0;
+	bool anyIndexed_ = 0;
+
 	// Vertex collector state
 	IndexGenerator indexGen;
 	int numDecodedVerts_ = 0;