From 9b411af1f56af8b6854bc06e2208039dd9738286 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 00:50:20 +0200
Subject: [PATCH 1/9] It's running.

---
 GPU/Common/DrawEngineCommon.cpp      | 325 +++++++++++----------------
 GPU/Common/DrawEngineCommon.h        |  36 ++-
 GPU/D3D11/DrawEngineD3D11.cpp        |  10 +-
 GPU/D3D11/DrawEngineD3D11.h          |   6 +-
 GPU/Directx9/DrawEngineDX9.cpp       |  12 +-
 GPU/Directx9/DrawEngineDX9.h         |   6 +-
 GPU/GLES/DrawEngineGLES.cpp          |  14 +-
 GPU/GLES/DrawEngineGLES.h            |   6 +-
 GPU/Vulkan/DrawEngineVulkan.cpp      |  17 +-
 GPU/Vulkan/DrawEngineVulkan.h        |   6 +-
 GPU/Vulkan/PipelineManagerVulkan.cpp |   2 +
 11 files changed, 209 insertions(+), 231 deletions(-)

diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
index 663c0495fe..7aabfa6364 100644
--- a/GPU/Common/DrawEngineCommon.cpp
+++ b/GPU/Common/DrawEngineCommon.cpp
@@ -72,43 +72,20 @@ VertexDecoder *DrawEngineCommon::GetVertexDecoder(u32 vtype) {
 	return dec;
 }
 
-int DrawEngineCommon::ComputeNumVertsToDecode() const {
-	int vertsToDecode = 0;
-	int numDrawCalls = numDrawCalls_;
-	if (drawCalls_[0].indexType == GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT) {
-		for (int i = 0; i < numDrawCalls; i++) {
-			const DeferredDrawCall &dc = drawCalls_[i];
-			vertsToDecode += dc.vertexCount;
-		}
-	} else {
-		// TODO: Share this computation with DecodeVertsStep?
-		for (int i = 0; i < numDrawCalls; i++) {
-			const DeferredDrawCall &dc = drawCalls_[i];
-			int lastMatch = i;
-			const int total = numDrawCalls;
-			int indexLowerBound = dc.indexLowerBound;
-			int indexUpperBound = dc.indexUpperBound;
-			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;
-			}
-			vertsToDecode += indexUpperBound - indexLowerBound + 1;
-			i = lastMatch;
-		}
+void DrawEngineCommon::DecodeVerts(u8 *dest) {
+	int decodeVertsCounter = decodeVertsCounter_;
+	for (; decodeVertsCounter < numDrawVerts_; decodeVertsCounter++) {
+		DecodeVertsStep(dest, decodeVertsCounter, decodedVerts_, &drawVerts_[decodeVertsCounter].uvScale);
 	}
-	return vertsToDecode;
+	decodeVertsCounter_ = decodeVertsCounter;
 }
 
-void DrawEngineCommon::DecodeVerts(u8 *dest) {
-	int decodeCounter = decodeCounter_;
-	for (; decodeCounter < numDrawCalls_; decodeCounter++) {
-		DecodeVertsStep(dest, decodeCounter, decodedVerts_, &drawCalls_[decodeCounter].uvScale);  // NOTE! DecodeVertsStep can modify the decodeCounter parameter!
+void DrawEngineCommon::DecodeInds() {
+	int decodeIndsCounter = decodeIndsCounter_;
+	for (; decodeIndsCounter < numDrawInds_; decodeIndsCounter++) {
+		DecodeIndsStep(decodeIndsCounter);
 	}
-	decodeCounter_ = decodeCounter;
+	decodeIndsCounter_ = decodeIndsCounter;
 
 	// Sanity check
 	if (indexGen.Prim() < 0) {
@@ -619,92 +596,43 @@ void DrawEngineCommon::ApplyFramebufferRead(FBOTexState *fboTexState) {
 	gstate_c.Dirty(DIRTY_SHADERBLEND);
 }
 
-void DrawEngineCommon::DecodeVertsStep(u8 *dest, int &i, int &decodedVerts, const UVScale *uvScale) {
+void DrawEngineCommon::DecodeVertsStep(u8 *dest, int i, int &decodedVerts, const UVScale *uvScale) {
 	PROFILE_THIS_SCOPE("vertdec");
 
-	const DeferredDrawCall &dc = drawCalls_[i];
+	const DeferredVerts &dv = drawVerts_[i];
 
-	indexGen.SetIndex(decodedVerts);
-	int indexLowerBound = dc.indexLowerBound;
-	int indexUpperBound = dc.indexUpperBound;
+	int indexLowerBound = dv.indexLowerBound;
+	int indexUpperBound = dv.indexUpperBound;
 
-	if (dc.indexType == GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT) {
-		// Decode the verts (and at the same time apply morphing/skinning). Simple.
-		dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride,
-			dc.verts, uvScale, indexLowerBound, indexUpperBound);
-		decodedVerts += indexUpperBound - indexLowerBound + 1;
-		
-		bool clockwise = true;
-		if (gstate.isCullEnabled() && gstate.getCullMode() != dc.cullMode) {
-			clockwise = false;
-		}
-		indexGen.AddPrim(dc.prim, dc.vertexCount, clockwise);
-	} 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.
+	// Decode the verts (and at the same time apply morphing/skinning). Simple.
+	dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride, dv.verts, uvScale, dv.indexLowerBound, dv.indexUpperBound);
+	decodedVerts += indexUpperBound - indexLowerBound + 1;
+}
 
-		// 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;
-			// TODO: What if UV scale/offset changes between drawcalls here?
-			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++) {
-				bool clockwise = true;
-				if (gstate.isCullEnabled() && gstate.getCullMode() != drawCalls_[j].cullMode) {
-					clockwise = false;
-				}
-				indexGen.TranslatePrim(drawCalls_[j].prim, drawCalls_[j].vertexCount, (const u8 *)drawCalls_[j].inds, indexLowerBound, clockwise);
-			}
-			break;
-		case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
-			for (int j = i; j <= lastMatch; j++) {
-				bool clockwise = true;
-				if (gstate.isCullEnabled() && gstate.getCullMode() != drawCalls_[j].cullMode) {
-					clockwise = false;
-				}
-				indexGen.TranslatePrim(drawCalls_[j].prim, drawCalls_[j].vertexCount, (const u16_le *)drawCalls_[j].inds, indexLowerBound, clockwise);
-			}
-			break;
-		case GE_VTYPE_IDX_32BIT >> GE_VTYPE_IDX_SHIFT:
-			for (int j = i; j <= lastMatch; j++) {
-				bool clockwise = true;
-				if (gstate.isCullEnabled() && gstate.getCullMode() != drawCalls_[j].cullMode) {
-					clockwise = false;
-				}
-				indexGen.TranslatePrim(drawCalls_[j].prim, drawCalls_[j].vertexCount, (const u32_le *)drawCalls_[j].inds, indexLowerBound, clockwise);
-			}
-			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, uvScale, indexLowerBound, indexUpperBound);
-		decodedVerts += vertexCount;
-
-		// 4. Advance indexgen vertex counter.
-		indexGen.Advance(vertexCount);
-		i = lastMatch;
+void DrawEngineCommon::DecodeIndsStep(int i) {
+	const DeferredInds &di = drawInds_[i];
+	bool clockwise = true;
+	if (gstate.isCullEnabled() && gstate.getCullMode() != di.cullMode) {
+		clockwise = false;
 	}
+	// We've already collapsed subsequent draws with the same vertex pointer, so no tricky logic here anymore.	
+	// 2. Loop through the drawcalls, translating indices as we go.
+	switch (di.indexType) {
+	case GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT:
+		indexGen.AddPrim(di.prim, di.vertexCount, clockwise);
+		break;
+	case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
+		indexGen.TranslatePrim(di.prim, di.vertexCount, (const u8 *)di.inds, di.indexOffset, clockwise);
+		break;
+	case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
+		indexGen.TranslatePrim(di.prim, di.vertexCount, (const u16_le *)di.inds, di.indexOffset, clockwise);
+		break;
+	case GE_VTYPE_IDX_32BIT >> GE_VTYPE_IDX_SHIFT:
+		indexGen.TranslatePrim(di.prim, di.vertexCount, (const u32_le *)di.inds, di.indexOffset, clockwise);
+		break;
+	}
+	// 4. Advance indexgen vertex counter.
+	indexGen.Advance(di.vertexCount);
 }
 
 inline u32 ComputeMiniHashRange(const void *ptr, size_t sz) {
@@ -731,27 +659,59 @@ u32 DrawEngineCommon::ComputeMiniHash() {
 	const int indexSize = IndexSize(dec_->VertexType());
 
 	int step;
-	if (numDrawCalls_ < 3) {
+	if (numDrawVerts_ < 3) {
 		step = 1;
-	} else if (numDrawCalls_ < 8) {
+	} else if (numDrawVerts_ < 8) {
 		step = 4;
 	} else {
-		step = numDrawCalls_ / 8;
+		step = numDrawVerts_ / 8;
 	}
-	for (int i = 0; i < numDrawCalls_; i += step) {
-		const DeferredDrawCall &dc = drawCalls_[i];
-		if (!dc.inds) {
-			fullhash += ComputeMiniHashRange(dc.verts, vertexSize * dc.vertexCount);
-		} else {
-			int indexLowerBound = dc.indexLowerBound, indexUpperBound = dc.indexUpperBound;
-			fullhash += ComputeMiniHashRange((const u8 *)dc.verts + vertexSize * indexLowerBound, vertexSize * (indexUpperBound - indexLowerBound));
-			fullhash += ComputeMiniHashRange(dc.inds, indexSize * dc.vertexCount);
-		}
+	for (int i = 0; i < numDrawVerts_; i += step) {
+		const DeferredVerts &dc = drawVerts_[i];
+		fullhash += ComputeMiniHashRange((const u8 *)dc.verts + vertexSize * dc.indexLowerBound, vertexSize * (dc.indexUpperBound - dc.indexLowerBound));
+	}
+	for (int i = 0; i < numDrawInds_; i += step) {
+		const DeferredInds &di = drawInds_[i];
+		fullhash += ComputeMiniHashRange(di.inds, indexSize * di.vertexCount);
 	}
 
 	return fullhash;
 }
 
+// Cheap bit scrambler from https://nullprogram.com/blog/2018/07/31/
+inline uint32_t lowbias32_r(uint32_t x) {
+	x ^= x >> 16;
+	x *= 0x43021123U;
+	x ^= x >> 15 ^ x >> 30;
+	x *= 0x1d69e2a5U;
+	x ^= x >> 16;
+	return x;
+}
+
+uint32_t DrawEngineCommon::ComputeDrawcallsHash() const {
+	uint32_t dcid = 0;
+	for (int i = 0; i < numDrawVerts_; i++) {
+		u32 dhash = dcid;
+		dhash = __rotl(dhash ^ (u32)(uintptr_t)drawVerts_[i].verts, 13);
+		dhash = __rotl(dhash ^ (u32)drawInds_[i].vertexCount, 11);
+		dcid = lowbias32_r(dhash ^ (u32)drawInds_[i].prim);
+	}
+	for (int j = 0; j < numDrawInds_; j++) {
+		u32 dhash = dcid;
+		dhash = __rotl(dhash ^ (u32)(uintptr_t)drawInds_[j].inds, 19);
+		dcid = lowbias32_r(__rotl(dhash ^ (u32)drawInds_[j].indexType, 7));
+	}
+	return dcid;
+}
+
+int DrawEngineCommon::ComputeNumVertsToDecode() const {
+	int sum = 0;
+	for (int i = 0; i < numDrawVerts_; i++) {
+		sum += drawVerts_[i].indexUpperBound + 1 - drawVerts_[i].indexLowerBound;
+	}
+	return sum;
+}
+
 uint64_t DrawEngineCommon::ComputeHash() {
 	uint64_t fullhash = 0;
 	const int vertexSize = dec_->GetDecVtxFmt().stride;
@@ -759,39 +719,26 @@ uint64_t DrawEngineCommon::ComputeHash() {
 
 	// TODO: Add some caps both for numDrawCalls_ and num verts to check?
 	// It is really very expensive to check all the vertex data so often.
-	for (int i = 0; i < numDrawCalls_; i++) {
-		const DeferredDrawCall &dc = drawCalls_[i];
-		if (!dc.inds) {
-			fullhash += XXH3_64bits((const char *)dc.verts, vertexSize * dc.vertexCount);
-		} else {
-			int indexLowerBound = dc.indexLowerBound, indexUpperBound = dc.indexUpperBound;
-			int j = i + 1;
-			int lastMatch = i;
-			while (j < numDrawCalls_) {
-				if (drawCalls_[j].verts != dc.verts)
-					break;
-				indexLowerBound = std::min(indexLowerBound, (int)dc.indexLowerBound);
-				indexUpperBound = std::max(indexUpperBound, (int)dc.indexUpperBound);
-				lastMatch = j;
-				j++;
-			}
-			// This could get seriously expensive with sparse indices. Need to combine hashing ranges the same way
-			// we do when drawing.
-			fullhash += XXH3_64bits((const char *)dc.verts + vertexSize * indexLowerBound,
-				vertexSize * (indexUpperBound - indexLowerBound));
-			// Hm, we will miss some indices when combining above, but meh, it should be fine.
-			fullhash += XXH3_64bits((const char *)dc.inds, indexSize * dc.vertexCount);
-			i = lastMatch;
-		}
+	for (int i = 0; i < numDrawVerts_; i++) {
+		const DeferredVerts &dv = drawVerts_[i];
+		int indexLowerBound = dv.indexLowerBound, indexUpperBound = dv.indexUpperBound;
+		fullhash += XXH3_64bits((const char *)dv.verts + vertexSize * indexLowerBound, vertexSize * (indexUpperBound - indexLowerBound));
 	}
 
-	fullhash += XXH3_64bits(&drawCalls_[0].uvScale, sizeof(drawCalls_[0].uvScale) * numDrawCalls_);
+	for (int i = 0; i < numDrawInds_; i++) {
+		const DeferredInds &di = drawInds_[i];
+		// Hm, we will miss some indices when combining above, but meh, it should be fine.
+		fullhash += XXH3_64bits((const char *)di.inds, indexSize * di.vertexCount);
+	}
+
+	// this looks utterly broken??
+	// fullhash += XXH3_64bits(&drawCalls_[0].uvScale, sizeof(drawCalls_[0].uvScale) * numDrawCalls_);
 	return fullhash;
 }
 
 // vertTypeID is the vertex type but with the UVGen mode smashed into the top bits.
 void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead) {
-	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawCalls_ >= MAX_DEFERRED_DRAW_CALLS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
+	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawVerts_ >= MAX_DEFERRED_DRAW_CALLS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
 		DispatchFlush();
 	}
 
@@ -818,27 +765,48 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	if (vertexCount < 3 && ((vertexCount < 2 && prim > 0) || (prim > GE_PRIM_LINE_STRIP && prim != GE_PRIM_RECTANGLES)))
 		return;
 
-	DeferredDrawCall &dc = drawCalls_[numDrawCalls_];
-	dc.verts = verts;
-	dc.inds = inds;
-	dc.vertexCount = vertexCount;
-	dc.indexType = (vertTypeID & GE_VTYPE_IDX_MASK) >> GE_VTYPE_IDX_SHIFT;
-	dc.prim = prim;
-	dc.cullMode = cullMode;
-	dc.uvScale = gstate_c.uv;
-	if (inds) {
-		GetIndexBounds(inds, vertexCount, vertTypeID, &dc.indexLowerBound, &dc.indexUpperBound);
+	bool applySkin = (vertTypeID & GE_VTYPE_WEIGHT_MASK) && decOptions_.applySkinInDecode;
+
+	DeferredInds &di = drawInds_[numDrawInds_++];
+	di.inds = inds;
+	di.indexType = (vertTypeID & GE_VTYPE_IDX_MASK) >> GE_VTYPE_IDX_SHIFT;
+	di.prim = prim;
+	di.cullMode = cullMode;
+	di.indexOffset = 0;
+	di.vertexCount = vertexCount;
+
+	if (inds && numDrawVerts_ > decodeVertsCounter_ && drawVerts_[numDrawVerts_ - 1].verts == verts && !applySkin) {
+		// Same vertex pointer as a previous un-decoded draw call - let's just extend the decode!
+		DeferredVerts &dv = drawVerts_[numDrawVerts_ - 1];
+		u16 lb;
+		u16 ub;
+		GetIndexBounds(inds, vertexCount, vertTypeID, &lb, &ub);
+		if (lb < dv.indexLowerBound)
+			dv.indexLowerBound = lb;
+		if (ub > dv.indexUpperBound)
+			dv.indexUpperBound = ub;
+		di.indexOffset = indexOffset_;
+		// indexOffset_ += vertexCount;
 	} else {
-		dc.indexLowerBound = 0;
-		dc.indexUpperBound = vertexCount - 1;
+		// Record a new draw, and a new index gen.
+		DeferredVerts &dv = drawVerts_[numDrawVerts_++];
+		dv.verts = verts;
+		dv.vertexCount = vertexCount;
+		dv.uvScale = gstate_c.uv;
+		if (inds) {
+			GetIndexBounds(inds, vertexCount, vertTypeID, &dv.indexLowerBound, &dv.indexUpperBound);
+		} else {
+			dv.indexLowerBound = 0;
+			dv.indexUpperBound = vertexCount - 1;
+		}
+		indexOffset_ = 0;  //  vertexCount;
 	}
 
-	numDrawCalls_++;
 	vertexCountInDrawCalls_ += vertexCount;
 
-	if ((vertTypeID & GE_VTYPE_WEIGHT_MASK) && decOptions_.applySkinInDecode) {
-		DecodeVertsStep(decoded_, decodeCounter_, decodedVerts_, &dc.uvScale);
-		decodeCounter_++;
+	if (applySkin) {
+		DecodeVertsStep(decoded_, decodeVertsCounter_, decodedVerts_, &drawVerts_[numDrawVerts_ - 1].uvScale);
+		DecodeIndsStep(decodeIndsCounter_);
 	}
 
 	if (prim == GE_PRIM_RECTANGLES && (gstate.getTextureAddress(0) & 0x3FFFFFFF) == (gstate.getFrameBufAddress() & 0x3FFFFFFF)) {
@@ -861,29 +829,6 @@ bool DrawEngineCommon::CanUseHardwareTessellation(GEPatchPrimType prim) {
 	return false;
 }
 
-// Cheap bit scrambler from https://nullprogram.com/blog/2018/07/31/
-inline uint32_t lowbias32_r(uint32_t x) {
-	x ^= x >> 16;
-	x *= 0x43021123U;
-	x ^= x >> 15 ^ x >> 30;
-	x *= 0x1d69e2a5U;
-	x ^= x >> 16;
-	return x;
-}
-
-uint32_t DrawEngineCommon::ComputeDrawcallsHash() const {
-	uint32_t dcid = 0;
-	for (int i = 0; i < numDrawCalls_; i++) {
-		u32 dhash = dcid;
-		dhash = __rotl(dhash ^ (u32)(uintptr_t)drawCalls_[i].verts, 13);
-		dhash = __rotl(dhash ^ (u32)(uintptr_t)drawCalls_[i].inds, 19);
-		dhash = __rotl(dhash ^ (u32)drawCalls_[i].indexType, 7);
-		dhash = __rotl(dhash ^ (u32)drawCalls_[i].vertexCount, 11);
-		dcid = lowbias32_r(dhash ^ (u32)drawCalls_[i].prim);
-	}
-	return dcid;
-}
-
 void TessellationDataTransfer::CopyControlPoints(float *pos, float *tex, float *col, int posStride, int texStride, int colStride, const SimpleVertex *const *points, int size, u32 vertType) {
 	bool hasColor = (vertType & GE_VTYPE_COL_MASK) != 0;
 	bool hasTexCoord = (vertType & GE_VTYPE_TC_MASK) != 0;
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
index 21c0476a44..b67dbefd75 100644
--- a/GPU/Common/DrawEngineCommon.h
+++ b/GPU/Common/DrawEngineCommon.h
@@ -130,7 +130,7 @@ public:
 		return false;
 	}
 	int GetNumDrawCalls() const {
-		return numDrawCalls_;
+		return numDrawVerts_;
 	}
 
 	VertexDecoder *GetVertexDecoder(u32 vtype);
@@ -141,8 +141,8 @@ protected:
 	virtual bool UpdateUseHWTessellation(bool enabled) const { return enabled; }
 	void UpdatePlanes();
 
-	int ComputeNumVertsToDecode() const;
 	void DecodeVerts(u8 *dest);
+	void DecodeInds();
 
 	// Preprocessing for spline/bezier
 	u32 NormalizeVertices(u8 *outPtr, u8 *bufPtr, const u8 *inPtr, int lowerBound, int upperBound, u32 vertType, int *vertexSize = nullptr);
@@ -152,7 +152,10 @@ protected:
 	uint64_t ComputeHash();
 
 	// Vertex decoding
-	void DecodeVertsStep(u8 *dest, int &i, int &decodedVerts, const UVScale *uvScale);
+	void DecodeVertsStep(u8 *dest, int i, int &decodedVerts, const UVScale *uvScale);
+	void DecodeIndsStep(int i);
+
+	int ComputeNumVertsToDecode() const;
 
 	void ApplyFramebufferRead(FBOTexState *fboTexState);
 
@@ -210,25 +213,36 @@ protected:
 	TransformedVertex *transformedExpanded_ = nullptr;
 
 	// Defer all vertex decoding to a "Flush" (except when software skinning)
-	struct DeferredDrawCall {
+	struct DeferredVerts {
 		const void *verts;
-		const void *inds;
 		u32 vertexCount;
-		u8 indexType;
-		s8 prim;
-		u8 cullMode;
 		u16 indexLowerBound;
 		u16 indexUpperBound;
 		UVScale uvScale;
 	};
 
+	struct DeferredInds {
+		const void *inds;
+		u32 vertexCount;
+		u8 indexType;
+		s8 prim;
+		u8 cullMode;
+		u16 indexOffset;
+	};
+
 	enum { MAX_DEFERRED_DRAW_CALLS = 128 };
-	DeferredDrawCall drawCalls_[MAX_DEFERRED_DRAW_CALLS];
-	int numDrawCalls_ = 0;
+	DeferredVerts drawVerts_[MAX_DEFERRED_DRAW_CALLS];
+	DeferredInds drawInds_[MAX_DEFERRED_DRAW_CALLS];
+
+	int numDrawVerts_ = 0;
+	int numDrawInds_ = 0;
 	int vertexCountInDrawCalls_ = 0;
 
 	int decimationCounter_ = 0;
-	int decodeCounter_ = 0;
+	int decodeVertsCounter_ = 0;
+	int decodeIndsCounter_ = 0;
+
+	int indexOffset_ = 0;
 
 	// Vertex collector state
 	IndexGenerator indexGen;
diff --git a/GPU/D3D11/DrawEngineD3D11.cpp b/GPU/D3D11/DrawEngineD3D11.cpp
index 7780bfa28a..3163349ffb 100644
--- a/GPU/D3D11/DrawEngineD3D11.cpp
+++ b/GPU/D3D11/DrawEngineD3D11.cpp
@@ -366,7 +366,7 @@ void DrawEngineD3D11::DoFlush() {
 
 		if (useCache) {
 			// getUVGenMode can have an effect on which UV decoder we need to use! And hence what the decoded data will look like. See #9263
-			u32 dcid = (u32)XXH3_64bits(&drawCalls_, sizeof(DeferredDrawCall) * numDrawCalls_) ^ gstate.getUVGenMode();
+			u32 dcid = ComputeDrawcallsHash() ^ gstate.getUVGenMode();
 
 			VertexArrayInfoD3D11 *vai;
 			if (!vai_.Get(dcid, &vai)) {
@@ -719,14 +719,16 @@ rotateVBO:
 	}
 
 	gpuStats.numFlushes++;
-	gpuStats.numDrawCalls += numDrawCalls_;
+	gpuStats.numDrawCalls += numDrawInds_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	indexGen.Reset();
 	decodedVerts_ = 0;
-	numDrawCalls_ = 0;
+	numDrawVerts_ = 0;
+	numDrawInds_ = 0;
 	vertexCountInDrawCalls_ = 0;
-	decodeCounter_ = 0;
+	decodeVertsCounter_ = 0;
+	decodeIndsCounter_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);
 
diff --git a/GPU/D3D11/DrawEngineD3D11.h b/GPU/D3D11/DrawEngineD3D11.h
index 763588e310..2b1a9fbf79 100644
--- a/GPU/D3D11/DrawEngineD3D11.h
+++ b/GPU/D3D11/DrawEngineD3D11.h
@@ -138,19 +138,19 @@ public:
 
 	// So that this can be inlined
 	void Flush() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DoFlush();
 	}
 
 	void FinishDeferred() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DecodeVerts(decoded_);
 	}
 
 	void DispatchFlush() override {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		Flush();
 	}
diff --git a/GPU/Directx9/DrawEngineDX9.cpp b/GPU/Directx9/DrawEngineDX9.cpp
index 9efa233dd0..d5f4e4f299 100644
--- a/GPU/Directx9/DrawEngineDX9.cpp
+++ b/GPU/Directx9/DrawEngineDX9.cpp
@@ -345,7 +345,7 @@ void DrawEngineDX9::DoFlush() {
 
 		if (useCache) {
 			// getUVGenMode can have an effect on which UV decoder we need to use! And hence what the decoded data will look like. See #9263
-			u32 dcid = (u32)XXH3_64bits(&drawCalls_, sizeof(DeferredDrawCall) * numDrawCalls_) ^ gstate.getUVGenMode();
+			u32 dcid = ComputeDrawcallsHash() ^ gstate.getUVGenMode();
 			VertexArrayInfoDX9 *vai;
 			if (!vai_.Get(dcid, &vai)) {
 				vai = new VertexArrayInfoDX9();
@@ -658,14 +658,18 @@ rotateVBO:
 	}
 
 	gpuStats.numFlushes++;
-	gpuStats.numDrawCalls += numDrawCalls_;
+	gpuStats.numDrawCalls += numDrawInds_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
+	// TODO: The below should be shared.
+
 	indexGen.Reset();
 	decodedVerts_ = 0;
-	numDrawCalls_ = 0;
+	numDrawVerts_ = 0;
+	numDrawInds_ = 0;
 	vertexCountInDrawCalls_ = 0;
-	decodeCounter_ = 0;
+	decodeVertsCounter_ = 0;
+	decodeIndsCounter_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);
 
diff --git a/GPU/Directx9/DrawEngineDX9.h b/GPU/Directx9/DrawEngineDX9.h
index 9ce7b11ec8..a527523b71 100644
--- a/GPU/Directx9/DrawEngineDX9.h
+++ b/GPU/Directx9/DrawEngineDX9.h
@@ -128,19 +128,19 @@ public:
 
 	// So that this can be inlined
 	void Flush() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DoFlush();
 	}
 
 	void FinishDeferred() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DecodeVerts(decoded_);
 	}
 
 	void DispatchFlush() override {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		Flush();
 	}
diff --git a/GPU/GLES/DrawEngineGLES.cpp b/GPU/GLES/DrawEngineGLES.cpp
index d1b957ac54..f8fc3cbf34 100644
--- a/GPU/GLES/DrawEngineGLES.cpp
+++ b/GPU/GLES/DrawEngineGLES.cpp
@@ -245,9 +245,11 @@ void DrawEngineGLES::DoFlush() {
 		// can't goto bail here, skips too many variable initializations. So let's wipe the most important stuff.
 		indexGen.Reset();
 		decodedVerts_ = 0;
-		numDrawCalls_ = 0;
+		numDrawVerts_ = 0;
+		numDrawInds_ = 0;
 		vertexCountInDrawCalls_ = 0;
-		decodeCounter_ = 0;
+		decodeVertsCounter_ = 0;
+		decodeIndsCounter_ = 0;
 		return;
 	}
 
@@ -471,7 +473,7 @@ void DrawEngineGLES::DoFlush() {
 
 bail:
 	gpuStats.numFlushes++;
-	gpuStats.numDrawCalls += numDrawCalls_;
+	gpuStats.numDrawCalls += numDrawInds_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	// TODO: When the next flush has the same vertex format, we can continue with the same offset in the vertex buffer,
@@ -479,9 +481,11 @@ bail:
 	// wanted to avoid rebinding the vertex input every time).
 	indexGen.Reset();
 	decodedVerts_ = 0;
-	numDrawCalls_ = 0;
+	numDrawVerts_ = 0;
+	numDrawInds_ = 0;
 	vertexCountInDrawCalls_ = 0;
-	decodeCounter_ = 0;
+	decodeVertsCounter_ = 0;
+	decodeIndsCounter_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);
 
diff --git a/GPU/GLES/DrawEngineGLES.h b/GPU/GLES/DrawEngineGLES.h
index 8a11cbb5ea..e61951c1f1 100644
--- a/GPU/GLES/DrawEngineGLES.h
+++ b/GPU/GLES/DrawEngineGLES.h
@@ -86,19 +86,19 @@ public:
 
 	// So that this can be inlined
 	void Flush() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DoFlush();
 	}
 
 	void FinishDeferred() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DoFlush();
 	}
 
 	void DispatchFlush() override {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		Flush();
 	}
diff --git a/GPU/Vulkan/DrawEngineVulkan.cpp b/GPU/Vulkan/DrawEngineVulkan.cpp
index bacff2f4d5..70891211b9 100644
--- a/GPU/Vulkan/DrawEngineVulkan.cpp
+++ b/GPU/Vulkan/DrawEngineVulkan.cpp
@@ -748,6 +748,7 @@ void DrawEngineVulkan::DoFlush() {
 				// Decode directly into the pushbuffer
 				DecodeVertsToPushPool(pushVertex_, &vbOffset, &vbuf);
 			}
+			DecodeInds();
 			gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 		}
 
@@ -845,6 +846,7 @@ void DrawEngineVulkan::DoFlush() {
 			dec_ = GetVertexDecoder(lastVType_);
 		}
 		DecodeVerts(decoded_);
+		DecodeInds();
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
 		if (gstate.isModeThrough()) {
 			gstate_c.vertexFullAlpha = gstate_c.vertexFullAlpha && (hasColor || gstate.getMaterialAmbientA() == 255);
@@ -857,6 +859,7 @@ void DrawEngineVulkan::DoFlush() {
 		// Undo the strip optimization, not supported by the SW code yet.
 		if (prim == GE_PRIM_TRIANGLE_STRIP)
 			prim = GE_PRIM_TRIANGLES;
+		_dbg_assert_(prim != GE_PRIM_INVALID);
 
 		u16 *inds = decIndex_;
 		SoftwareTransformResult result{};
@@ -1007,14 +1010,16 @@ void DrawEngineVulkan::DoFlush() {
 	}
 
 	gpuStats.numFlushes++;
-	gpuStats.numDrawCalls += numDrawCalls_;
+	gpuStats.numDrawCalls += numDrawInds_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	indexGen.Reset();
 	decodedVerts_ = 0;
-	numDrawCalls_ = 0;
+	numDrawVerts_ = 0;
+	numDrawInds_ = 0;
 	vertexCountInDrawCalls_ = 0;
-	decodeCounter_ = 0;
+	decodeIndsCounter_ = 0;
+	decodeVertsCounter_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);
 
@@ -1030,8 +1035,10 @@ void DrawEngineVulkan::DoFlush() {
 void DrawEngineVulkan::ResetAfterDraw() {
 	indexGen.Reset();
 	decodedVerts_ = 0;
-	numDrawCalls_ = 0;
-	decodeCounter_ = 0;
+	numDrawVerts_ = 0;
+	numDrawInds_ = 0;
+	decodeIndsCounter_ = 0;
+	decodeVertsCounter_ = 0;
 	decOptions_.applySkinInDecode = g_Config.bSoftwareSkinning;
 	gstate_c.vertexFullAlpha = true;
 }
diff --git a/GPU/Vulkan/DrawEngineVulkan.h b/GPU/Vulkan/DrawEngineVulkan.h
index b599c35051..cfc03a7b6e 100644
--- a/GPU/Vulkan/DrawEngineVulkan.h
+++ b/GPU/Vulkan/DrawEngineVulkan.h
@@ -170,13 +170,13 @@ public:
 
 	// So that this can be inlined
 	void Flush() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		DoFlush();
 	}
 
 	void FinishDeferred() {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		// Decode any pending vertices. And also flush while we're at it, for simplicity.
 		// It might be possible to only decode like in the other backends, but meh, it can't matter.
@@ -185,7 +185,7 @@ public:
 	}
 
 	void DispatchFlush() override {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		Flush();
 	}
diff --git a/GPU/Vulkan/PipelineManagerVulkan.cpp b/GPU/Vulkan/PipelineManagerVulkan.cpp
index b2ad9d0739..05803e3028 100644
--- a/GPU/Vulkan/PipelineManagerVulkan.cpp
+++ b/GPU/Vulkan/PipelineManagerVulkan.cpp
@@ -291,6 +291,8 @@ static VulkanPipeline *CreateVulkanPipeline(VulkanRenderManager *renderManager,
 		desc->geometryShaderSource = gs->GetShaderString(SHADER_STRING_SOURCE_CODE);
 	}
 
+	_dbg_assert_(key.topology != VK_PRIMITIVE_TOPOLOGY_POINT_LIST);
+	_dbg_assert_(key.topology != VK_PRIMITIVE_TOPOLOGY_LINE_LIST);
 	desc->topology = (VkPrimitiveTopology)key.topology;
 
 	int vertexStride = 0;

From 92ffef262695c8407a5ff1a70f89e048c0ed8e07 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 11:25:27 +0200
Subject: [PATCH 2/9] Remove some state from IndexGenerator, fix bugs. Mostly
 works except vertex cache.

---
 GPU/Common/DrawEngineCommon.cpp    | 144 ++++++++++++++---------------
 GPU/Common/DrawEngineCommon.h      |  20 ++--
 GPU/Common/IndexGenerator.cpp      |  83 +++++++----------
 GPU/Common/IndexGenerator.h        |  25 ++---
 GPU/Common/VertexDecoderCommon.cpp |   3 +
 GPU/D3D11/DrawEngineD3D11.cpp      |  17 +++-
 GPU/Directx9/DrawEngineDX9.cpp     |  19 +++-
 GPU/GLES/DrawEngineGLES.cpp        |   5 +-
 GPU/Vulkan/DrawEngineVulkan.cpp    |  14 ++-
 GPU/Vulkan/DrawEngineVulkan.h      |   8 +-
 10 files changed, 167 insertions(+), 171 deletions(-)

diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
index 7aabfa6364..d8220720bc 100644
--- a/GPU/Common/DrawEngineCommon.cpp
+++ b/GPU/Common/DrawEngineCommon.cpp
@@ -72,29 +72,6 @@ VertexDecoder *DrawEngineCommon::GetVertexDecoder(u32 vtype) {
 	return dec;
 }
 
-void DrawEngineCommon::DecodeVerts(u8 *dest) {
-	int decodeVertsCounter = decodeVertsCounter_;
-	for (; decodeVertsCounter < numDrawVerts_; decodeVertsCounter++) {
-		DecodeVertsStep(dest, decodeVertsCounter, decodedVerts_, &drawVerts_[decodeVertsCounter].uvScale);
-	}
-	decodeVertsCounter_ = decodeVertsCounter;
-}
-
-void DrawEngineCommon::DecodeInds() {
-	int decodeIndsCounter = decodeIndsCounter_;
-	for (; decodeIndsCounter < numDrawInds_; decodeIndsCounter++) {
-		DecodeIndsStep(decodeIndsCounter);
-	}
-	decodeIndsCounter_ = decodeIndsCounter;
-
-	// 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, true);
-	}
-}
-
 std::vector<std::string> DrawEngineCommon::DebugGetVertexLoaderIDs() {
 	std::vector<std::string> ids;
 	decoderMap_.Iterate([&](const uint32_t vtype, VertexDecoder *decoder) {
@@ -596,45 +573,6 @@ void DrawEngineCommon::ApplyFramebufferRead(FBOTexState *fboTexState) {
 	gstate_c.Dirty(DIRTY_SHADERBLEND);
 }
 
-void DrawEngineCommon::DecodeVertsStep(u8 *dest, int i, int &decodedVerts, const UVScale *uvScale) {
-	PROFILE_THIS_SCOPE("vertdec");
-
-	const DeferredVerts &dv = drawVerts_[i];
-
-	int indexLowerBound = dv.indexLowerBound;
-	int indexUpperBound = dv.indexUpperBound;
-
-	// Decode the verts (and at the same time apply morphing/skinning). Simple.
-	dec_->DecodeVerts(dest + decodedVerts * (int)dec_->GetDecVtxFmt().stride, dv.verts, uvScale, dv.indexLowerBound, dv.indexUpperBound);
-	decodedVerts += indexUpperBound - indexLowerBound + 1;
-}
-
-void DrawEngineCommon::DecodeIndsStep(int i) {
-	const DeferredInds &di = drawInds_[i];
-	bool clockwise = true;
-	if (gstate.isCullEnabled() && gstate.getCullMode() != di.cullMode) {
-		clockwise = false;
-	}
-	// We've already collapsed subsequent draws with the same vertex pointer, so no tricky logic here anymore.	
-	// 2. Loop through the drawcalls, translating indices as we go.
-	switch (di.indexType) {
-	case GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT:
-		indexGen.AddPrim(di.prim, di.vertexCount, clockwise);
-		break;
-	case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
-		indexGen.TranslatePrim(di.prim, di.vertexCount, (const u8 *)di.inds, di.indexOffset, clockwise);
-		break;
-	case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
-		indexGen.TranslatePrim(di.prim, di.vertexCount, (const u16_le *)di.inds, di.indexOffset, clockwise);
-		break;
-	case GE_VTYPE_IDX_32BIT >> GE_VTYPE_IDX_SHIFT:
-		indexGen.TranslatePrim(di.prim, di.vertexCount, (const u32_le *)di.inds, di.indexOffset, clockwise);
-		break;
-	}
-	// 4. Advance indexgen vertex counter.
-	indexGen.Advance(di.vertexCount);
-}
-
 inline u32 ComputeMiniHashRange(const void *ptr, size_t sz) {
 	// Switch to u32 units, and round up to avoid unaligned accesses.
 	// Probably doesn't matter if we skip the first few bytes in some cases.
@@ -672,7 +610,9 @@ u32 DrawEngineCommon::ComputeMiniHash() {
 	}
 	for (int i = 0; i < numDrawInds_; i += step) {
 		const DeferredInds &di = drawInds_[i];
-		fullhash += ComputeMiniHashRange(di.inds, indexSize * di.vertexCount);
+		if (di.inds) {
+			fullhash += ComputeMiniHashRange(di.inds, indexSize * di.vertexCount);
+		}
 	}
 
 	return fullhash;
@@ -715,7 +655,6 @@ int DrawEngineCommon::ComputeNumVertsToDecode() const {
 uint64_t DrawEngineCommon::ComputeHash() {
 	uint64_t fullhash = 0;
 	const int vertexSize = dec_->GetDecVtxFmt().stride;
-	const int indexSize = IndexSize(dec_->VertexType());
 
 	// TODO: Add some caps both for numDrawCalls_ and num verts to check?
 	// It is really very expensive to check all the vertex data so often.
@@ -727,8 +666,11 @@ uint64_t DrawEngineCommon::ComputeHash() {
 
 	for (int i = 0; i < numDrawInds_; i++) {
 		const DeferredInds &di = drawInds_[i];
-		// Hm, we will miss some indices when combining above, but meh, it should be fine.
-		fullhash += XXH3_64bits((const char *)di.inds, indexSize * di.vertexCount);
+		if (di.indexType != 0) {
+			int indexSize = IndexSize(di.indexType << GE_VTYPE_IDX_SHIFT);
+			// Hm, we will miss some indices when combining above, but meh, it should be fine.
+			fullhash += XXH3_64bits((const char *)di.inds, indexSize * di.vertexCount);
+		}
 	}
 
 	// this looks utterly broken??
@@ -738,9 +680,11 @@ uint64_t DrawEngineCommon::ComputeHash() {
 
 // vertTypeID is the vertex type but with the UVGen mode smashed into the top bits.
 void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead) {
-	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawVerts_ >= MAX_DEFERRED_DRAW_CALLS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
+	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawVerts_ >= MAX_DEFERRED_DRAW_VERTS || numDrawInds_ >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
 		DispatchFlush();
 	}
+	_dbg_assert_(numDrawVerts_ < MAX_DEFERRED_DRAW_VERTS);
+	_dbg_assert_(numDrawInds_ < MAX_DEFERRED_DRAW_INDS);
 
 	// This isn't exactly right, if we flushed, since prims can straddle previous calls.
 	// But it generally works for common usage.
@@ -772,11 +716,15 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	di.indexType = (vertTypeID & GE_VTYPE_IDX_MASK) >> GE_VTYPE_IDX_SHIFT;
 	di.prim = prim;
 	di.cullMode = cullMode;
-	di.indexOffset = 0;
 	di.vertexCount = vertexCount;
+	di.vertDecodeIndex = numDrawVerts_;
+
+	_dbg_assert_(numDrawVerts_ <= MAX_DEFERRED_DRAW_VERTS);
+	_dbg_assert_(numDrawInds_ <= MAX_DEFERRED_DRAW_INDS);
 
 	if (inds && numDrawVerts_ > decodeVertsCounter_ && drawVerts_[numDrawVerts_ - 1].verts == verts && !applySkin) {
 		// Same vertex pointer as a previous un-decoded draw call - let's just extend the decode!
+		di.vertDecodeIndex = numDrawVerts_ - 1;
 		DeferredVerts &dv = drawVerts_[numDrawVerts_ - 1];
 		u16 lb;
 		u16 ub;
@@ -785,8 +733,6 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 			dv.indexLowerBound = lb;
 		if (ub > dv.indexUpperBound)
 			dv.indexUpperBound = ub;
-		di.indexOffset = indexOffset_;
-		// indexOffset_ += vertexCount;
 	} else {
 		// Record a new draw, and a new index gen.
 		DeferredVerts &dv = drawVerts_[numDrawVerts_++];
@@ -799,14 +745,12 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 			dv.indexLowerBound = 0;
 			dv.indexUpperBound = vertexCount - 1;
 		}
-		indexOffset_ = 0;  //  vertexCount;
 	}
 
 	vertexCountInDrawCalls_ += vertexCount;
 
 	if (applySkin) {
-		DecodeVertsStep(decoded_, decodeVertsCounter_, decodedVerts_, &drawVerts_[numDrawVerts_ - 1].uvScale);
-		DecodeIndsStep(decodeIndsCounter_);
+		DecodeVerts(decoded_);
 	}
 
 	if (prim == GE_PRIM_RECTANGLES && (gstate.getTextureAddress(0) & 0x3FFFFFFF) == (gstate.getFrameBufAddress() & 0x3FFFFFFF)) {
@@ -816,6 +760,60 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	}
 }
 
+void DrawEngineCommon::DecodeVerts(u8 *dest) {
+	int i = decodeVertsCounter_;
+	int stride = (int)dec_->GetDecVtxFmt().stride;
+	for (; i < numDrawVerts_; i++) {
+		DeferredVerts &dv = drawVerts_[i];
+
+		int indexLowerBound = dv.indexLowerBound;
+		drawVertexOffsets_[i] = decodedVerts_ - indexLowerBound;
+
+		int indexUpperBound = dv.indexUpperBound;
+		// Decode the verts (and at the same time apply morphing/skinning). Simple.
+		dec_->DecodeVerts(dest + decodedVerts_ * stride, dv.verts, &dv.uvScale, indexLowerBound, indexUpperBound);
+		decodedVerts_ += indexUpperBound - indexLowerBound + 1;
+	}
+	decodeVertsCounter_ = i;
+}
+
+void DrawEngineCommon::DecodeInds() {
+	int i = decodeIndsCounter_;
+	for (; i < numDrawInds_; i++) {
+		const DeferredInds &di = drawInds_[i];
+
+		int indexOffset = drawVertexOffsets_[di.vertDecodeIndex];
+		bool clockwise = true;
+		if (gstate.isCullEnabled() && gstate.getCullMode() != di.cullMode) {
+			clockwise = false;
+		}
+		// We've already collapsed subsequent draws with the same vertex pointer, so no tricky logic here anymore.
+		// 2. Loop through the drawcalls, translating indices as we go.
+		switch (di.indexType) {
+		case GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT:
+			indexGen.AddPrim(di.prim, di.vertexCount, indexOffset, clockwise);
+			break;
+		case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
+			indexGen.TranslatePrim(di.prim, di.vertexCount, (const u8 *)di.inds, indexOffset, clockwise);
+			break;
+		case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
+			indexGen.TranslatePrim(di.prim, di.vertexCount, (const u16_le *)di.inds, indexOffset, clockwise);
+			break;
+		case GE_VTYPE_IDX_32BIT >> GE_VTYPE_IDX_SHIFT:
+			indexGen.TranslatePrim(di.prim, di.vertexCount, (const u32_le *)di.inds, indexOffset, clockwise);
+			break;
+		}
+	}
+	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);
+	}
+}
+
 bool DrawEngineCommon::CanUseHardwareTransform(int prim) {
 	if (!useHWTransform_)
 		return false;
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
index b67dbefd75..d299c3d7de 100644
--- a/GPU/Common/DrawEngineCommon.h
+++ b/GPU/Common/DrawEngineCommon.h
@@ -144,6 +144,10 @@ protected:
 	void DecodeVerts(u8 *dest);
 	void DecodeInds();
 
+	int MaxIndex() const {
+		return decodedVerts_;
+	}
+
 	// Preprocessing for spline/bezier
 	u32 NormalizeVertices(u8 *outPtr, u8 *bufPtr, const u8 *inPtr, int lowerBound, int upperBound, u32 vertType, int *vertexSize = nullptr);
 
@@ -151,10 +155,6 @@ protected:
 	u32 ComputeMiniHash();
 	uint64_t ComputeHash();
 
-	// Vertex decoding
-	void DecodeVertsStep(u8 *dest, int i, int &decodedVerts, const UVScale *uvScale);
-	void DecodeIndsStep(int i);
-
 	int ComputeNumVertsToDecode() const;
 
 	void ApplyFramebufferRead(FBOTexState *fboTexState);
@@ -224,15 +224,17 @@ protected:
 	struct DeferredInds {
 		const void *inds;
 		u32 vertexCount;
+		u8 vertDecodeIndex;  // index into the drawVerts_ array to look up the vertexOffset.
 		u8 indexType;
 		s8 prim;
 		u8 cullMode;
-		u16 indexOffset;
 	};
 
-	enum { MAX_DEFERRED_DRAW_CALLS = 128 };
-	DeferredVerts drawVerts_[MAX_DEFERRED_DRAW_CALLS];
-	DeferredInds drawInds_[MAX_DEFERRED_DRAW_CALLS];
+	enum { MAX_DEFERRED_DRAW_VERTS = 128 };  // If you change this to more than 256, change type of DeferredInds::vertDecodeIndex.
+	enum { MAX_DEFERRED_DRAW_INDS = 512 };  // Monster Hunter spams indexed calls that we end up merging.
+	DeferredVerts drawVerts_[MAX_DEFERRED_DRAW_VERTS];
+	uint32_t drawVertexOffsets_[MAX_DEFERRED_DRAW_VERTS];
+	DeferredInds drawInds_[MAX_DEFERRED_DRAW_INDS];
 
 	int numDrawVerts_ = 0;
 	int numDrawInds_ = 0;
@@ -242,8 +244,6 @@ protected:
 	int decodeVertsCounter_ = 0;
 	int decodeIndsCounter_ = 0;
 
-	int indexOffset_ = 0;
-
 	// Vertex collector state
 	IndexGenerator indexGen;
 	int decodedVerts_ = 0;
diff --git a/GPU/Common/IndexGenerator.cpp b/GPU/Common/IndexGenerator.cpp
index d1635e807d..b361cbcbce 100644
--- a/GPU/Common/IndexGenerator.cpp
+++ b/GPU/Common/IndexGenerator.cpp
@@ -50,44 +50,40 @@ void IndexGenerator::Setup(u16 *inds) {
 	Reset();
 }
 
-void IndexGenerator::AddPrim(int prim, int vertexCount, bool clockwise) {
+void IndexGenerator::AddPrim(int prim, int vertexCount, int indexOffset, bool clockwise) {
 	switch (prim) {
-	case GE_PRIM_POINTS: AddPoints(vertexCount); break;
-	case GE_PRIM_LINES: AddLineList(vertexCount); break;
-	case GE_PRIM_LINE_STRIP: AddLineStrip(vertexCount); break;
-	case GE_PRIM_TRIANGLES: AddList(vertexCount, clockwise); break;
-	case GE_PRIM_TRIANGLE_STRIP: AddStrip(vertexCount, clockwise); break;
-	case GE_PRIM_TRIANGLE_FAN: AddFan(vertexCount, clockwise); break;
-	case GE_PRIM_RECTANGLES: AddRectangles(vertexCount); break;  // Same
+	case GE_PRIM_POINTS: AddPoints(vertexCount, indexOffset); break;
+	case GE_PRIM_LINES: AddLineList(vertexCount, indexOffset); break;
+	case GE_PRIM_LINE_STRIP: AddLineStrip(vertexCount, indexOffset); break;
+	case GE_PRIM_TRIANGLES: AddList(vertexCount, indexOffset, clockwise); break;
+	case GE_PRIM_TRIANGLE_STRIP: AddStrip(vertexCount, indexOffset, clockwise); break;
+	case GE_PRIM_TRIANGLE_FAN: AddFan(vertexCount, indexOffset, clockwise); break;
+	case GE_PRIM_RECTANGLES: AddRectangles(vertexCount, indexOffset); break;  // Same
 	}
 }
 
-void IndexGenerator::AddPoints(int numVerts) {
+void IndexGenerator::AddPoints(int numVerts, int indexOffset) {
 	u16 *outInds = inds_;
-	const int startIndex = index_;
 	for (int i = 0; i < numVerts; i++)
-		*outInds++ = startIndex + i;
+		*outInds++ = indexOffset + i;
 	inds_ = outInds;
 	// ignore overflow verts
-	index_ += numVerts;
 	count_ += numVerts;
 	prim_ = GE_PRIM_POINTS;
 	seenPrims_ |= 1 << GE_PRIM_POINTS;
 }
 
-void IndexGenerator::AddList(int numVerts, bool clockwise) {
+void IndexGenerator::AddList(int numVerts, int indexOffset, bool clockwise) {
 	u16 *outInds = inds_;
-	const int startIndex = index_;
 	const int v1 = clockwise ? 1 : 2;
 	const int v2 = clockwise ? 2 : 1;
 	for (int i = 0; i < numVerts; i += 3) {
-		*outInds++ = startIndex + i;
-		*outInds++ = startIndex + i + v1;
-		*outInds++ = startIndex + i + v2;
+		*outInds++ = indexOffset + i;
+		*outInds++ = indexOffset + i + v1;
+		*outInds++ = indexOffset + i + v2;
 	}
 	inds_ = outInds;
 	// ignore overflow verts
-	index_ += numVerts;
 	count_ += numVerts;
 	prim_ = GE_PRIM_TRIANGLES;
 	seenPrims_ |= 1 << GE_PRIM_TRIANGLES;
@@ -119,7 +115,7 @@ alignas(16) static const uint16_t offsets_counter_clockwise[24] = {
 	7, (u16)(7 + 1), (u16)(7 + 2),
 };
 
-void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
+void IndexGenerator::AddStrip(int numVerts, int indexOffset, bool clockwise) {
 	int numTris = numVerts - 2;
 #ifdef _M_SSE
 	// In an SSE2 register we can fit 8 16-bit integers.
@@ -130,7 +126,7 @@ void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
 	// We allow ourselves to write some extra indices to avoid the fallback loop.
 	// That's alright as we're appending to a buffer - they will get overwritten anyway.
 	int numChunks = (numTris + 7) >> 3;
-	__m128i ibase8 = _mm_set1_epi16(index_);
+	__m128i ibase8 = _mm_set1_epi16(indexOffset);
 	const __m128i *offsets = (const __m128i *)(clockwise ? offsets_clockwise : offsets_counter_clockwise);
 	__m128i *dst = (__m128i *)inds_;
 	__m128i offsets0 = _mm_add_epi16(ibase8, _mm_load_si128(offsets));
@@ -158,7 +154,7 @@ void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
 	// wind doesn't need to be updated, an even number of triangles have been drawn.
 #elif PPSSPP_ARCH(ARM_NEON)
 	int numChunks = (numTris + 7) >> 3;
-	uint16x8_t ibase8 = vdupq_n_u16(index_);
+	uint16x8_t ibase8 = vdupq_n_u16(indexOffset);
 	const u16 *offsets = clockwise ? offsets_clockwise : offsets_counter_clockwise;
 	u16 *dst = inds_;
 	uint16x8_t offsets0 = vaddq_u16(ibase8, vld1q_u16(offsets));
@@ -185,7 +181,7 @@ void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
 #else
 	// Slow fallback loop.
 	int wind = clockwise ? 1 : 2;
-	int ibase = index_;
+	int ibase = indexOffset;
 	size_t numPairs = numTris / 2;
 	u16 *outInds = inds_;
 	while (numPairs > 0) {
@@ -207,7 +203,6 @@ void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
 	inds_ = outInds;
 #endif
 
-	index_ += numVerts;
 	if (numTris > 0)
 		count_ += numTris * 3;
 	// This is so we can detect one single strip by just looking at seenPrims_.
@@ -222,19 +217,17 @@ void IndexGenerator::AddStrip(int numVerts, bool clockwise) {
 	}
 }
 
-void IndexGenerator::AddFan(int numVerts, bool clockwise) {
+void IndexGenerator::AddFan(int numVerts, int indexOffset, bool clockwise) {
 	const int numTris = numVerts - 2;
 	u16 *outInds = inds_;
-	const int startIndex = index_;
 	const int v1 = clockwise ? 1 : 2;
 	const int v2 = clockwise ? 2 : 1;
 	for (int i = 0; i < numTris; i++) {
-		*outInds++ = startIndex;
-		*outInds++ = startIndex + i + v1;
-		*outInds++ = startIndex + i + v2;
+		*outInds++ = indexOffset;
+		*outInds++ = indexOffset + i + v1;
+		*outInds++ = indexOffset + i + v2;
 	}
 	inds_ = outInds;
-	index_ += numVerts;
 	count_ += numTris * 3;
 	prim_ = GE_PRIM_TRIANGLES;
 	seenPrims_ |= 1 << GE_PRIM_TRIANGLE_FAN;
@@ -245,46 +238,40 @@ void IndexGenerator::AddFan(int numVerts, bool clockwise) {
 }
 
 //Lines
-void IndexGenerator::AddLineList(int numVerts) {
+void IndexGenerator::AddLineList(int numVerts, int indexOffset) {
 	u16 *outInds = inds_;
-	const int startIndex = index_;
 	for (int i = 0; i < numVerts; i += 2) {
-		*outInds++ = startIndex + i;
-		*outInds++ = startIndex + i + 1;
+		*outInds++ = indexOffset + i;
+		*outInds++ = indexOffset + i + 1;
 	}
 	inds_ = outInds;
-	index_ += numVerts;
 	count_ += numVerts;
 	prim_ = GE_PRIM_LINES;
 	seenPrims_ |= 1 << prim_;
 }
 
-void IndexGenerator::AddLineStrip(int numVerts) {
+void IndexGenerator::AddLineStrip(int numVerts, int indexOffset) {
 	const int numLines = numVerts - 1;
 	u16 *outInds = inds_;
-	const int startIndex = index_;
 	for (int i = 0; i < numLines; i++) {
-		*outInds++ = startIndex + i;
-		*outInds++ = startIndex + i + 1;
+		*outInds++ = indexOffset + i;
+		*outInds++ = indexOffset + i + 1;
 	}
 	inds_ = outInds;
-	index_ += numVerts;
 	count_ += numLines * 2;
 	prim_ = GE_PRIM_LINES;
 	seenPrims_ |= 1 << GE_PRIM_LINE_STRIP;
 }
 
-void IndexGenerator::AddRectangles(int numVerts) {
+void IndexGenerator::AddRectangles(int numVerts, int indexOffset) {
 	u16 *outInds = inds_;
-	const int startIndex = index_;
 	//rectangles always need 2 vertices, disregard the last one if there's an odd number
 	numVerts = numVerts & ~1;
 	for (int i = 0; i < numVerts; i += 2) {
-		*outInds++ = startIndex + i;
-		*outInds++ = startIndex + i + 1;
+		*outInds++ = indexOffset + i;
+		*outInds++ = indexOffset + i + 1;
 	}
 	inds_ = outInds;
-	index_ += numVerts;
 	count_ += numVerts;
 	prim_ = GE_PRIM_RECTANGLES;
 	seenPrims_ |= 1 << GE_PRIM_RECTANGLES;
@@ -292,7 +279,6 @@ void IndexGenerator::AddRectangles(int numVerts) {
 
 template <class ITypeLE, int flag>
 void IndexGenerator::TranslatePoints(int numInds, const ITypeLE *inds, int indexOffset) {
-	indexOffset = index_ - indexOffset;
 	u16 *outInds = inds_;
 	for (int i = 0; i < numInds; i++)
 		*outInds++ = indexOffset + inds[i];
@@ -304,7 +290,6 @@ void IndexGenerator::TranslatePoints(int numInds, const ITypeLE *inds, int index
 
 template <class ITypeLE, int flag>
 void IndexGenerator::TranslateLineList(int numInds, const ITypeLE *inds, int indexOffset) {
-	indexOffset = index_ - indexOffset;
 	u16 *outInds = inds_;
 	numInds = numInds & ~1;
 	for (int i = 0; i < numInds; i += 2) {
@@ -319,7 +304,6 @@ void IndexGenerator::TranslateLineList(int numInds, const ITypeLE *inds, int ind
 
 template <class ITypeLE, int flag>
 void IndexGenerator::TranslateLineStrip(int numInds, const ITypeLE *inds, int indexOffset) {
-	indexOffset = index_ - indexOffset;
 	int numLines = numInds - 1;
 	u16 *outInds = inds_;
 	for (int i = 0; i < numLines; i++) {
@@ -334,7 +318,6 @@ void IndexGenerator::TranslateLineStrip(int numInds, const ITypeLE *inds, int in
 
 template <class ITypeLE, int flag>
 void IndexGenerator::TranslateList(int numInds, const ITypeLE *inds, int indexOffset, bool clockwise) {
-	indexOffset = index_ - indexOffset;
 	// We only bother doing this minor optimization in triangle list, since it's by far the most
 	// common operation that can benefit.
 	if (sizeof(ITypeLE) == sizeof(inds_[0]) && indexOffset == 0 && clockwise) {
@@ -347,6 +330,7 @@ void IndexGenerator::TranslateList(int numInds, const ITypeLE *inds, int indexOf
 		numInds = numTris * 3;
 		const int v1 = clockwise ? 1 : 2;
 		const int v2 = clockwise ? 2 : 1;
+		// TODO: This can actually be SIMD-d, although will need complex shuffles if clockwise.
 		for (int i = 0; i < numInds; i += 3) {
 			*outInds++ = indexOffset + inds[i];
 			*outInds++ = indexOffset + inds[i + v1];
@@ -362,7 +346,6 @@ void IndexGenerator::TranslateList(int numInds, const ITypeLE *inds, int indexOf
 template <class ITypeLE, int flag>
 void IndexGenerator::TranslateStrip(int numInds, const ITypeLE *inds, int indexOffset, bool clockwise) {
 	int wind = clockwise ? 1 : 2;
-	indexOffset = index_ - indexOffset;
 	int numTris = numInds - 2;
 	u16 *outInds = inds_;
 	for (int i = 0; i < numTris; i++) {
@@ -380,7 +363,6 @@ void IndexGenerator::TranslateStrip(int numInds, const ITypeLE *inds, int indexO
 template <class ITypeLE, int flag>
 void IndexGenerator::TranslateFan(int numInds, const ITypeLE *inds, int indexOffset, bool clockwise) {
 	if (numInds <= 0) return;
-	indexOffset = index_ - indexOffset;
 	int numTris = numInds - 2;
 	u16 *outInds = inds_;
 	const int v1 = clockwise ? 1 : 2;
@@ -398,7 +380,6 @@ void IndexGenerator::TranslateFan(int numInds, const ITypeLE *inds, int indexOff
 
 template <class ITypeLE, int flag>
 inline void IndexGenerator::TranslateRectangles(int numInds, const ITypeLE *inds, int indexOffset) {
-	indexOffset = index_ - indexOffset;
 	u16 *outInds = inds_;
 	//rectangles always need 2 vertices, disregard the last one if there's an odd number
 	numInds = numInds & ~1;
diff --git a/GPU/Common/IndexGenerator.h b/GPU/Common/IndexGenerator.h
index e8c2578409..b5df11aab8 100644
--- a/GPU/Common/IndexGenerator.h
+++ b/GPU/Common/IndexGenerator.h
@@ -28,7 +28,6 @@ public:
 	void Reset() {
 		prim_ = GE_PRIM_INVALID;
 		count_ = 0;
-		index_ = 0;
 		seenPrims_ = 0;
 		pureCount_ = 0;
 		this->inds_ = indsBase_;
@@ -57,19 +56,12 @@ public:
 		}
 	}
 
-	void AddPrim(int prim, int vertexCount, bool clockwise);
+	void AddPrim(int prim, int vertexCount, int indexOffset, bool clockwise);
 	void TranslatePrim(int prim, int numInds, const u8 *inds, int indexOffset, bool clockwise);
 	void TranslatePrim(int prim, int numInds, const u16_le *inds, int indexOffset, bool clockwise);
 	void TranslatePrim(int prim, int numInds, const u32_le *inds, int indexOffset, bool clockwise);
 
-	void Advance(int numVerts) {
-		index_ += numVerts;
-	}
-
-	void SetIndex(int ind) { index_ = ind; }
-	int MaxIndex() const { return index_; }  // Really NextIndex rather than MaxIndex, it's one more than the highest index generated
 	int VertexCount() const { return count_; }
-	bool Empty() const { return index_ == 0; }
 	int SeenPrims() const { return seenPrims_; }
 	int PureCount() const { return pureCount_; }
 	bool SeenOnlyPurePrims() const {
@@ -81,16 +73,16 @@ public:
 
 private:
 	// Points (why index these? code simplicity)
-	void AddPoints(int numVerts);
+	void AddPoints(int numVerts, int indexOffset);
 	// Triangles
-	void AddList(int numVerts, bool clockwise);
-	void AddStrip(int numVerts, bool clockwise);
-	void AddFan(int numVerts, bool clockwise);
+	void AddList(int numVerts, int indexOffset, bool clockwise);
+	void AddStrip(int numVerts, int indexOffset, bool clockwise);
+	void AddFan(int numVerts, int indexOffset, bool clockwise);
 	// Lines
-	void AddLineList(int numVerts);
-	void AddLineStrip(int numVerts);
+	void AddLineList(int numVerts, int indexOffset);
+	void AddLineStrip(int numVerts, int indexOffset);
 	// Rectangles
-	void AddRectangles(int numVerts);
+	void AddRectangles(int numVerts, int indexOffset);
 
 	// These translate already indexed lists
 	template <class ITypeLE, int flag>
@@ -118,7 +110,6 @@ private:
 
 	u16 *indsBase_;
 	u16 *inds_;
-	int index_;
 	int count_;
 	int pureCount_;
 	GEPrimitiveType prim_;
diff --git a/GPU/Common/VertexDecoderCommon.cpp b/GPU/Common/VertexDecoderCommon.cpp
index c31a5f1d58..ec1af7c80e 100644
--- a/GPU/Common/VertexDecoderCommon.cpp
+++ b/GPU/Common/VertexDecoderCommon.cpp
@@ -1293,6 +1293,9 @@ void VertexDecoder::SetVertexType(u32 fmt, const VertexDecoderOptions &options,
 }
 
 void VertexDecoder::DecodeVerts(u8 *decodedptr, const void *verts, const UVScale *uvScaleOffset, int indexLowerBound, int indexUpperBound) const {
+	// A single 0 is acceptable for point lists.
+	_dbg_assert_(indexLowerBound <= indexUpperBound);
+
 	// Decode the vertices within the found bounds, once each
 	// decoded_ and ptr_ are used in the steps, so can't be turned into locals for speed.
 	const u8 *startPtr = (const u8*)verts + indexLowerBound * size;
diff --git a/GPU/D3D11/DrawEngineD3D11.cpp b/GPU/D3D11/DrawEngineD3D11.cpp
index 3163349ffb..3711fee648 100644
--- a/GPU/D3D11/DrawEngineD3D11.cpp
+++ b/GPU/D3D11/DrawEngineD3D11.cpp
@@ -384,9 +384,10 @@ void DrawEngineD3D11::DoFlush() {
 					vai->status = VertexArrayInfoD3D11::VAI_HASHING;
 					vai->drawsUntilNextFullHash = 0;
 					DecodeVerts(decoded_); // writes to indexGen
+					DecodeInds();
 					vai->numVerts = indexGen.VertexCount();
 					vai->prim = indexGen.Prim();
-					vai->maxIndex = indexGen.MaxIndex();
+					vai->maxIndex = MaxIndex();
 					vai->flags = gstate_c.vertexFullAlpha ? VAI11_FLAG_VERTEXFULLALPHA : 0;
 					goto rotateVBO;
 				}
@@ -409,6 +410,7 @@ void DrawEngineD3D11::DoFlush() {
 						if (newMiniHash != vai->minihash || newHash != vai->hash) {
 							MarkUnreliable(vai);
 							DecodeVerts(decoded_);
+							DecodeInds();
 							goto rotateVBO;
 						}
 						if (vai->numVerts > 64) {
@@ -428,15 +430,17 @@ void DrawEngineD3D11::DoFlush() {
 						if (newMiniHash != vai->minihash) {
 							MarkUnreliable(vai);
 							DecodeVerts(decoded_);
+							DecodeInds();
 							goto rotateVBO;
 						}
 					}
 
 					if (vai->vbo == 0) {
 						DecodeVerts(decoded_);
+						DecodeInds();
 						vai->numVerts = indexGen.VertexCount();
 						vai->prim = indexGen.Prim();
-						vai->maxIndex = indexGen.MaxIndex();
+						vai->maxIndex = MaxIndex();
 						vai->flags = gstate_c.vertexFullAlpha ? VAI11_FLAG_VERTEXFULLALPHA : 0;
 						useElements = !indexGen.SeenOnlyPurePrims() || prim == GE_PRIM_TRIANGLE_FAN;
 						if (!useElements && indexGen.PureCount()) {
@@ -446,7 +450,7 @@ void DrawEngineD3D11::DoFlush() {
 						_dbg_assert_msg_(gstate_c.vertBounds.minV >= gstate_c.vertBounds.maxV, "Should not have checked UVs when caching.");
 
 						// TODO: Combine these two into one buffer?
-						u32 size = dec_->GetDecVtxFmt().stride * indexGen.MaxIndex();
+						u32 size = dec_->GetDecVtxFmt().stride * MaxIndex();
 						D3D11_BUFFER_DESC desc{ size, D3D11_USAGE_IMMUTABLE, D3D11_BIND_VERTEX_BUFFER, 0 };
 						D3D11_SUBRESOURCE_DATA data{ decoded_ };
 						ASSERT_SUCCESS(device_->CreateBuffer(&desc, &data, &vai->vbo));
@@ -500,6 +504,7 @@ void DrawEngineD3D11::DoFlush() {
 						vai->numFrames++;
 					}
 					DecodeVerts(decoded_);
+					DecodeInds();
 					goto rotateVBO;
 				}
 			}
@@ -507,11 +512,12 @@ void DrawEngineD3D11::DoFlush() {
 			vai->lastFrame = gpuStats.numFlips;
 		} else {
 			DecodeVerts(decoded_);
+			DecodeInds();
 rotateVBO:
 			gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 			useElements = !indexGen.SeenOnlyPurePrims() || prim == GE_PRIM_TRIANGLE_FAN;
 			vertexCount = indexGen.VertexCount();
-			maxIndex = indexGen.MaxIndex();
+			maxIndex = MaxIndex();
 			if (!useElements && indexGen.PureCount()) {
 				vertexCount = indexGen.PureCount();
 			}
@@ -584,6 +590,7 @@ rotateVBO:
 			dec_ = GetVertexDecoder(lastVType_);
 		}
 		DecodeVerts(decoded_);
+		DecodeInds();
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
 		if (gstate.isModeThrough()) {
 			gstate_c.vertexFullAlpha = gstate_c.vertexFullAlpha && (hasColor || gstate.getMaterialAmbientA() == 255);
@@ -622,7 +629,7 @@ rotateVBO:
 			UpdateCachedViewportState(vpAndScissor);
 		}
 
-		int maxIndex = indexGen.MaxIndex();
+		int maxIndex = MaxIndex();
 		SoftwareTransform swTransform(params);
 
 		const Lin::Vec3 trans(gstate_c.vpXOffset, -gstate_c.vpYOffset, gstate_c.vpZOffset * 0.5f + 0.5f);
diff --git a/GPU/Directx9/DrawEngineDX9.cpp b/GPU/Directx9/DrawEngineDX9.cpp
index d5f4e4f299..890daa4cc2 100644
--- a/GPU/Directx9/DrawEngineDX9.cpp
+++ b/GPU/Directx9/DrawEngineDX9.cpp
@@ -362,9 +362,10 @@ void DrawEngineDX9::DoFlush() {
 					vai->status = VertexArrayInfoDX9::VAI_HASHING;
 					vai->drawsUntilNextFullHash = 0;
 					DecodeVerts(decoded_); // writes to indexGen
+					DecodeInds();
 					vai->numVerts = indexGen.VertexCount();
 					vai->prim = indexGen.Prim();
-					vai->maxIndex = indexGen.MaxIndex();
+					vai->maxIndex = MaxIndex();
 					vai->flags = gstate_c.vertexFullAlpha ? VAI_FLAG_VERTEXFULLALPHA : 0;
 
 					goto rotateVBO;
@@ -388,6 +389,7 @@ void DrawEngineDX9::DoFlush() {
 						if (newMiniHash != vai->minihash || newHash != vai->hash) {
 							MarkUnreliable(vai);
 							DecodeVerts(decoded_);
+							DecodeInds();
 							goto rotateVBO;
 						}
 						if (vai->numVerts > 64) {
@@ -407,15 +409,17 @@ void DrawEngineDX9::DoFlush() {
 						if (newMiniHash != vai->minihash) {
 							MarkUnreliable(vai);
 							DecodeVerts(decoded_);
+							DecodeInds();
 							goto rotateVBO;
 						}
 					}
 
 					if (vai->vbo == 0) {
 						DecodeVerts(decoded_);
+						DecodeInds();
 						vai->numVerts = indexGen.VertexCount();
 						vai->prim = indexGen.Prim();
-						vai->maxIndex = indexGen.MaxIndex();
+						vai->maxIndex = MaxIndex();
 						vai->flags = gstate_c.vertexFullAlpha ? VAI_FLAG_VERTEXFULLALPHA : 0;
 						useElements = !indexGen.SeenOnlyPurePrims();
 						if (!useElements && indexGen.PureCount()) {
@@ -425,7 +429,7 @@ void DrawEngineDX9::DoFlush() {
 						_dbg_assert_msg_(gstate_c.vertBounds.minV >= gstate_c.vertBounds.maxV, "Should not have checked UVs when caching.");
 
 						void * pVb;
-						u32 size = dec_->GetDecVtxFmt().stride * indexGen.MaxIndex();
+						u32 size = dec_->GetDecVtxFmt().stride * MaxIndex();
 						device_->CreateVertexBuffer(size, D3DUSAGE_WRITEONLY, 0, D3DPOOL_DEFAULT, &vai->vbo, NULL);
 						vai->vbo->Lock(0, size, &pVb, 0);
 						memcpy(pVb, decoded_, size);
@@ -482,6 +486,7 @@ void DrawEngineDX9::DoFlush() {
 						vai->numFrames++;
 					}
 					DecodeVerts(decoded_);
+					DecodeInds();
 					goto rotateVBO;
 				}
 			}
@@ -489,17 +494,20 @@ void DrawEngineDX9::DoFlush() {
 			vai->lastFrame = gpuStats.numFlips;
 		} else {
 			DecodeVerts(decoded_);
+			DecodeInds();
 rotateVBO:
 			gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 			useElements = !indexGen.SeenOnlyPurePrims();
 			vertexCount = indexGen.VertexCount();
-			maxIndex = indexGen.MaxIndex();
+			maxIndex = MaxIndex();
 			if (!useElements && indexGen.PureCount()) {
 				vertexCount = indexGen.PureCount();
 			}
 			prim = indexGen.Prim();
 		}
 
+		_dbg_assert_((int)prim > 0);
+
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
 		if (gstate.isModeThrough()) {
 			gstate_c.vertexFullAlpha = gstate_c.vertexFullAlpha && (hasColor || gstate.getMaterialAmbientA() == 255);
@@ -544,6 +552,7 @@ rotateVBO:
 			dec_ = GetVertexDecoder(lastVType_);
 		}
 		DecodeVerts(decoded_);
+		DecodeInds();
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
 		if (gstate.isModeThrough()) {
 			gstate_c.vertexFullAlpha = gstate_c.vertexFullAlpha && (hasColor || gstate.getMaterialAmbientA() == 255);
@@ -582,7 +591,7 @@ rotateVBO:
 			UpdateCachedViewportState(vpAndScissor);
 		}
 
-		int maxIndex = indexGen.MaxIndex();
+		int maxIndex = MaxIndex();
 		SoftwareTransform swTransform(params);
 
 		// Half pixel offset hack.
diff --git a/GPU/GLES/DrawEngineGLES.cpp b/GPU/GLES/DrawEngineGLES.cpp
index f8fc3cbf34..c202eb3634 100644
--- a/GPU/GLES/DrawEngineGLES.cpp
+++ b/GPU/GLES/DrawEngineGLES.cpp
@@ -286,9 +286,9 @@ void DrawEngineGLES::DoFlush() {
 			// Figure out how much pushbuffer space we need to allocate.
 			int vertsToDecode = ComputeNumVertsToDecode();
 			u8 *dest = (u8 *)frameData.pushVertex->Allocate(vertsToDecode * dec_->GetDecVtxFmt().stride, 4, &vertexBuffer, &vertexBufferOffset);
-			// Indices are decoded in here.
 			DecodeVerts(dest);
 		}
+		DecodeInds();
 
 		gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 
@@ -345,6 +345,7 @@ void DrawEngineGLES::DoFlush() {
 			dec_ = GetVertexDecoder(lastVType_);
 		}
 		DecodeVerts(decoded_);
+		DecodeInds();
 
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
 		if (gstate.isModeThrough()) {
@@ -383,7 +384,7 @@ void DrawEngineGLES::DoFlush() {
 			UpdateCachedViewportState(vpAndScissor_);
 		}
 
-		int maxIndex = indexGen.MaxIndex();
+		int maxIndex = MaxIndex();
 		int vertexCount = indexGen.VertexCount();
 
 		// TODO: Split up into multiple draw calls for GLES 2.0 where you can't guarantee support for more than 0x10000 verts.
diff --git a/GPU/Vulkan/DrawEngineVulkan.cpp b/GPU/Vulkan/DrawEngineVulkan.cpp
index 70891211b9..2a463fb9e0 100644
--- a/GPU/Vulkan/DrawEngineVulkan.cpp
+++ b/GPU/Vulkan/DrawEngineVulkan.cpp
@@ -566,10 +566,11 @@ bool DrawEngineVulkan::VertexCacheLookup(int &vertexCount, GEPrimitiveType &prim
 		vai->minihash = ComputeMiniHash();
 		vai->status = VertexArrayInfoVulkan::VAI_HASHING;
 		vai->drawsUntilNextFullHash = 0;
-		DecodeVertsToPushPool(pushVertex_, &vbOffset, &vbuf);  // writes to indexGen
+		DecodeVertsToPushPool(pushVertex_, &vbOffset, &vbuf);
+		DecodeInds();
 		vai->numVerts = indexGen.VertexCount();
 		vai->prim = indexGen.Prim();
-		vai->maxIndex = indexGen.MaxIndex();
+		vai->maxIndex = MaxIndex();
 		vai->flags = gstate_c.vertexFullAlpha ? VAIVULKAN_FLAG_VERTEXFULLALPHA : 0;
 		return true;
 	}
@@ -593,6 +594,7 @@ bool DrawEngineVulkan::VertexCacheLookup(int &vertexCount, GEPrimitiveType &prim
 			if (newMiniHash != vai->minihash || newHash != vai->hash) {
 				MarkUnreliable(vai);
 				DecodeVertsToPushPool(pushVertex_, &vbOffset, &vbuf);
+				DecodeInds();
 				return true;
 			}
 			if (vai->numVerts > 64) {
@@ -612,6 +614,7 @@ bool DrawEngineVulkan::VertexCacheLookup(int &vertexCount, GEPrimitiveType &prim
 			if (newMiniHash != vai->minihash) {
 				MarkUnreliable(vai);
 				DecodeVertsToPushPool(pushVertex_, &vbOffset, &vbuf);
+				DecodeInds();
 				return true;
 			}
 		}
@@ -619,9 +622,10 @@ bool DrawEngineVulkan::VertexCacheLookup(int &vertexCount, GEPrimitiveType &prim
 		if (!vai->vb) {
 			// Directly push to the vertex cache.
 			DecodeVertsToPushBuffer(vertexCache_, &vai->vbOffset, &vai->vb);
+			DecodeInds();
 			_dbg_assert_msg_(gstate_c.vertBounds.minV >= gstate_c.vertBounds.maxV, "Should not have checked UVs when caching.");
 			vai->numVerts = indexGen.VertexCount();
-			vai->maxIndex = indexGen.MaxIndex();
+			vai->maxIndex = MaxIndex();
 			vai->flags = gstate_c.vertexFullAlpha ? VAIVULKAN_FLAG_VERTEXFULLALPHA : 0;
 			if (forceIndexed) {
 				vai->prim = indexGen.GeneralPrim();
@@ -684,6 +688,7 @@ bool DrawEngineVulkan::VertexCacheLookup(int &vertexCount, GEPrimitiveType &prim
 			vai->numFrames++;
 		}
 		DecodeVertsToPushPool(pushVertex_, &vbOffset, &vbuf);
+		DecodeInds();
 		return true;
 	}
 	default:
@@ -889,7 +894,7 @@ void DrawEngineVulkan::DoFlush() {
 			UpdateCachedViewportState(vpAndScissor);
 		}
 
-		int maxIndex = indexGen.MaxIndex();
+		int maxIndex = MaxIndex();
 		SoftwareTransform swTransform(params);
 
 		const Lin::Vec3 trans(gstate_c.vpXOffset, gstate_c.vpYOffset, gstate_c.vpZOffset * 0.5f + 0.5f);
@@ -1037,6 +1042,7 @@ void DrawEngineVulkan::ResetAfterDraw() {
 	decodedVerts_ = 0;
 	numDrawVerts_ = 0;
 	numDrawInds_ = 0;
+	vertexCountInDrawCalls_ = 0;
 	decodeIndsCounter_ = 0;
 	decodeVertsCounter_ = 0;
 	decOptions_.applySkinInDecode = g_Config.bSoftwareSkinning;
diff --git a/GPU/Vulkan/DrawEngineVulkan.h b/GPU/Vulkan/DrawEngineVulkan.h
index cfc03a7b6e..ae6124a04c 100644
--- a/GPU/Vulkan/DrawEngineVulkan.h
+++ b/GPU/Vulkan/DrawEngineVulkan.h
@@ -170,13 +170,13 @@ public:
 
 	// So that this can be inlined
 	void Flush() {
-		if (!numDrawVerts_)
+		if (!numDrawInds_)
 			return;
 		DoFlush();
 	}
 
 	void FinishDeferred() {
-		if (!numDrawVerts_)
+		if (!numDrawInds_)
 			return;
 		// Decode any pending vertices. And also flush while we're at it, for simplicity.
 		// It might be possible to only decode like in the other backends, but meh, it can't matter.
@@ -185,9 +185,9 @@ public:
 	}
 
 	void DispatchFlush() override {
-		if (!numDrawVerts_)
+		if (!numDrawInds_)
 			return;
-		Flush();
+		DoFlush();
 	}
 
 	VkPipelineLayout GetPipelineLayout() const {

From 1c49d5718c33acccd19933de454ef9a92962c9fa Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 16:03:49 +0200
Subject: [PATCH 3/9] Add an offset field that we'll need later

---
 GPU/Common/DrawEngineCommon.cpp | 3 ++-
 GPU/Common/DrawEngineCommon.h   | 1 +
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
index d8220720bc..dfe127fa3e 100644
--- a/GPU/Common/DrawEngineCommon.cpp
+++ b/GPU/Common/DrawEngineCommon.cpp
@@ -718,6 +718,7 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	di.cullMode = cullMode;
 	di.vertexCount = vertexCount;
 	di.vertDecodeIndex = numDrawVerts_;
+	di.offset = 0;
 
 	_dbg_assert_(numDrawVerts_ <= MAX_DEFERRED_DRAW_VERTS);
 	_dbg_assert_(numDrawInds_ <= MAX_DEFERRED_DRAW_INDS);
@@ -782,7 +783,7 @@ void DrawEngineCommon::DecodeInds() {
 	for (; i < numDrawInds_; i++) {
 		const DeferredInds &di = drawInds_[i];
 
-		int indexOffset = drawVertexOffsets_[di.vertDecodeIndex];
+		int indexOffset = drawVertexOffsets_[di.vertDecodeIndex] + di.offset;
 		bool clockwise = true;
 		if (gstate.isCullEnabled() && gstate.getCullMode() != di.cullMode) {
 			clockwise = false;
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
index d299c3d7de..f53988700f 100644
--- a/GPU/Common/DrawEngineCommon.h
+++ b/GPU/Common/DrawEngineCommon.h
@@ -228,6 +228,7 @@ protected:
 		u8 indexType;
 		s8 prim;
 		u8 cullMode;
+		u16 offset;
 	};
 
 	enum { MAX_DEFERRED_DRAW_VERTS = 128 };  // If you change this to more than 256, change type of DeferredInds::vertDecodeIndex.

From e63bb0459c817e4d0add5c8ccd197746f4d288ac Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 17:28:25 +0200
Subject: [PATCH 4/9] Add a new stat, so we can see per game if the
 optimization has an effect

---
 GPU/D3D11/DrawEngineD3D11.cpp   | 1 +
 GPU/Directx9/DrawEngineDX9.cpp  | 1 +
 GPU/GLES/DrawEngineGLES.cpp     | 1 +
 GPU/GPU.h                       | 2 ++
 GPU/GPUCommonHW.cpp             | 3 ++-
 GPU/Vulkan/DrawEngineVulkan.cpp | 1 +
 6 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/GPU/D3D11/DrawEngineD3D11.cpp b/GPU/D3D11/DrawEngineD3D11.cpp
index 3711fee648..0deb056ef4 100644
--- a/GPU/D3D11/DrawEngineD3D11.cpp
+++ b/GPU/D3D11/DrawEngineD3D11.cpp
@@ -727,6 +727,7 @@ rotateVBO:
 
 	gpuStats.numFlushes++;
 	gpuStats.numDrawCalls += numDrawInds_;
+	gpuStats.numVertexDecodes += numDrawVerts_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	indexGen.Reset();
diff --git a/GPU/Directx9/DrawEngineDX9.cpp b/GPU/Directx9/DrawEngineDX9.cpp
index 890daa4cc2..276fde10a3 100644
--- a/GPU/Directx9/DrawEngineDX9.cpp
+++ b/GPU/Directx9/DrawEngineDX9.cpp
@@ -668,6 +668,7 @@ rotateVBO:
 
 	gpuStats.numFlushes++;
 	gpuStats.numDrawCalls += numDrawInds_;
+	gpuStats.numVertexDecodes += numDrawVerts_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	// TODO: The below should be shared.
diff --git a/GPU/GLES/DrawEngineGLES.cpp b/GPU/GLES/DrawEngineGLES.cpp
index c202eb3634..9949d0c37b 100644
--- a/GPU/GLES/DrawEngineGLES.cpp
+++ b/GPU/GLES/DrawEngineGLES.cpp
@@ -475,6 +475,7 @@ void DrawEngineGLES::DoFlush() {
 bail:
 	gpuStats.numFlushes++;
 	gpuStats.numDrawCalls += numDrawInds_;
+	gpuStats.numVertexDecodes += numDrawVerts_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	// TODO: When the next flush has the same vertex format, we can continue with the same offset in the vertex buffer,
diff --git a/GPU/GPU.h b/GPU/GPU.h
index 9e9dd6049a..5db94e2d39 100644
--- a/GPU/GPU.h
+++ b/GPU/GPU.h
@@ -75,6 +75,7 @@ struct GPUStatistics {
 
 	void ResetFrame() {
 		numDrawCalls = 0;
+		numVertexDecodes = 0;
 		numDrawSyncs = 0;
 		numListSyncs = 0;
 		numCachedDrawCalls = 0;
@@ -111,6 +112,7 @@ struct GPUStatistics {
 
 	// Per frame statistics
 	int numDrawCalls;
+	int numVertexDecodes;
 	int numDrawSyncs;
 	int numListSyncs;
 	int numCachedDrawCalls;
diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index a9b5ded875..544f1c58cc 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1647,7 +1647,7 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 	float vertexAverageCycles = gpuStats.numVertsSubmitted > 0 ? (float)gpuStats.vertexGPUCycles / (float)gpuStats.numVertsSubmitted : 0.0f;
 	return snprintf(buffer, size,
 		"DL processing time: %0.2f ms, %d drawsync, %d listsync\n"
-		"Draw calls: %d, flushes %d, clears %d, bbox jumps %d (%d updates)\n"
+		"Draw: %d (%d dec), flushes %d, clears %d, bbox jumps %d (%d updates)\n"
 		"Cached draws: %d (tracked: %d)\n"
 		"Vertices: %d cached: %d uncached: %d\n"
 		"FBOs active: %d (evaluations: %d)\n"
@@ -1660,6 +1660,7 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		gpuStats.numDrawSyncs,
 		gpuStats.numListSyncs,
 		gpuStats.numDrawCalls,
+		gpuStats.numVertexDecodes,
 		gpuStats.numFlushes,
 		gpuStats.numClears,
 		gpuStats.numBBOXJumps,
diff --git a/GPU/Vulkan/DrawEngineVulkan.cpp b/GPU/Vulkan/DrawEngineVulkan.cpp
index 2a463fb9e0..ed83c7f207 100644
--- a/GPU/Vulkan/DrawEngineVulkan.cpp
+++ b/GPU/Vulkan/DrawEngineVulkan.cpp
@@ -1016,6 +1016,7 @@ void DrawEngineVulkan::DoFlush() {
 
 	gpuStats.numFlushes++;
 	gpuStats.numDrawCalls += numDrawInds_;
+	gpuStats.numVertexDecodes += numDrawVerts_;
 	gpuStats.numVertsSubmitted += vertexCountInDrawCalls_;
 
 	indexGen.Reset();

From 0260aebc265c9f2df75e732a5a6ad0aeabe21d32 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 16:26:12 +0200
Subject: [PATCH 5/9] Implement fast-path for merging non-indexed draws
 quickly.

---
 GPU/Common/DrawEngineCommon.cpp | 40 +++++++++++++++++++++++++++++----
 GPU/Common/DrawEngineCommon.h   |  1 +
 GPU/GPUCommonHW.cpp             | 21 ++++++++++++++++-
 3 files changed, 57 insertions(+), 5 deletions(-)

diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
index dfe127fa3e..76be8eb225 100644
--- a/GPU/Common/DrawEngineCommon.cpp
+++ b/GPU/Common/DrawEngineCommon.cpp
@@ -678,6 +678,38 @@ uint64_t DrawEngineCommon::ComputeHash() {
 	return fullhash;
 }
 
+bool DrawEngineCommon::ExtendNonIndexedPrim(GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead) {
+	if (numDrawInds_ >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
+		return false;
+	}
+
+	bool applySkin = (vertTypeID & GE_VTYPE_WEIGHT_MASK) && decOptions_.applySkinInDecode;
+	if (applySkin) {
+		// TODO: Support this somehow.
+		return false;
+	}
+
+	_dbg_assert_(numDrawInds_ < MAX_DEFERRED_DRAW_INDS);
+	_dbg_assert_(numDrawVerts_ > 0);
+	*bytesRead = vertexCount * dec_->VertexSize();
+
+	DeferredInds &di = drawInds_[numDrawInds_++];
+	di.inds = nullptr;
+	di.indexType = 0;
+	di.prim = prim;
+	di.cullMode = cullMode;
+	di.vertexCount = vertexCount;
+	di.vertDecodeIndex = numDrawVerts_ - 1;
+
+	DeferredVerts &dv = drawVerts_[numDrawVerts_ - 1];
+	int offset = dv.vertexCount;
+	di.offset = offset;
+	dv.vertexCount += vertexCount;
+	dv.indexUpperBound = dv.vertexCount - 1;
+	vertexCountInDrawCalls_ += vertexCount;
+	return true;
+}
+
 // vertTypeID is the vertex type but with the UVGen mode smashed into the top bits.
 void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead) {
 	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawVerts_ >= MAX_DEFERRED_DRAW_VERTS || numDrawInds_ >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
@@ -750,15 +782,15 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 
 	vertexCountInDrawCalls_ += vertexCount;
 
-	if (applySkin) {
-		DecodeVerts(decoded_);
-	}
-
 	if (prim == GE_PRIM_RECTANGLES && (gstate.getTextureAddress(0) & 0x3FFFFFFF) == (gstate.getFrameBufAddress() & 0x3FFFFFFF)) {
 		// This prevents issues with consecutive self-renders in Ridge Racer.
 		gstate_c.Dirty(DIRTY_TEXTURE_PARAMS);
 		DispatchFlush();
 	}
+
+	if (applySkin) {
+		DecodeVerts(decoded_);
+	}
 }
 
 void DrawEngineCommon::DecodeVerts(u8 *dest) {
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
index f53988700f..a6461c9cfb 100644
--- a/GPU/Common/DrawEngineCommon.h
+++ b/GPU/Common/DrawEngineCommon.h
@@ -104,6 +104,7 @@ public:
 
 	bool TestBoundingBox(const void *control_points, const void *inds, int vertexCount, u32 vertType);
 
+	bool ExtendNonIndexedPrim(GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead);
 	void SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead);
 	template<class Surface>
 	void SubmitCurve(const void *control_points, const void *indices, Surface &surface, u32 vertType, int *bytesRead, const char *scope);
diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index 544f1c58cc..5bf7e9873c 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -967,6 +967,8 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 
 	const void *verts = Memory::GetPointerUnchecked(gstate_c.vertexAddr);
 	const void *inds = nullptr;
+
+	bool canExtend = true;
 	u32 vertexType = gstate.vertType;
 	if ((vertexType & GE_VTYPE_IDX_MASK) != GE_VTYPE_IDX_NONE) {
 		u32 indexAddr = gstate_c.indexAddr;
@@ -975,6 +977,7 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			return;
 		}
 		inds = Memory::GetPointerUnchecked(indexAddr);
+		canExtend = false;
 	}
 
 	int bytesRead = 0;
@@ -1017,12 +1020,25 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			if (IsTrianglePrim(newPrim) != isTriangle)
 				goto bail;  // Can't join over this boundary. Might as well exit and get this on the next time around.
 			// TODO: more efficient updating of verts/inds
+
+			u32 count = data & 0xFFFF;
+			if (canExtend) {
+				// Non-indexed draws can be cheaply merged if vertexAddr hasn't changed, that means the vertices
+				// are consecutive in memory.
+				_dbg_assert_((vertexType & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_NONE);
+				if (drawEngineCommon_->ExtendNonIndexedPrim(newPrim, count, vertTypeID, cullMode, &bytesRead)) {
+					gstate_c.vertexAddr += bytesRead;
+					totalVertCount += count;
+					break;
+				}
+			}
+
+			// Failed, or can't extend? Do a normal submit.
 			verts = Memory::GetPointerUnchecked(gstate_c.vertexAddr);
 			inds = nullptr;
 			if ((vertexType & GE_VTYPE_IDX_MASK) != GE_VTYPE_IDX_NONE) {
 				inds = Memory::GetPointerUnchecked(gstate_c.indexAddr);
 			}
-			u32 count = data & 0xFFFF;
 			drawEngineCommon_->SubmitPrim(verts, inds, newPrim, count, vertTypeID, cullMode, &bytesRead);
 			AdvanceVerts(vertexType, count, bytesRead);
 			totalVertCount += count;
@@ -1030,6 +1046,7 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		}
 		case GE_CMD_VERTEXTYPE:
 		{
+			canExtend = false;  // TODO: Might support extending between some vertex types in the future.
 			uint32_t diff = data ^ vertexType;
 			// don't mask upper bits, vertexType is unmasked
 			if (diff & vtypeCheckMask) {
@@ -1043,6 +1060,7 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		case GE_CMD_VADDR:
 			gstate.cmdmem[GE_CMD_VADDR] = data;
 			gstate_c.vertexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
+			canExtend = false;
 			break;
 		case GE_CMD_IADDR:
 			gstate.cmdmem[GE_CMD_IADDR] = data;
@@ -1051,6 +1069,7 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		case GE_CMD_OFFSETADDR:
 			gstate.cmdmem[GE_CMD_OFFSETADDR] = data;
 			gstate_c.offsetAddr = data << 8;
+			canExtend = false;
 			break;
 		case GE_CMD_BASE:
 			gstate.cmdmem[GE_CMD_BASE] = data;

From 4d952500521c7d2d2bb4c002cad9f20df0d672d5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 16:53:54 +0200
Subject: [PATCH 6/9] Optimize further

---
 GPU/Common/DrawEngineCommon.cpp | 11 +----------
 GPU/Common/DrawEngineCommon.h   |  7 +++++++
 GPU/GPUCommonHW.cpp             | 10 ++++++++--
 3 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
index 76be8eb225..b61447a3d9 100644
--- a/GPU/Common/DrawEngineCommon.cpp
+++ b/GPU/Common/DrawEngineCommon.cpp
@@ -683,12 +683,6 @@ bool DrawEngineCommon::ExtendNonIndexedPrim(GEPrimitiveType prim, int vertexCoun
 		return false;
 	}
 
-	bool applySkin = (vertTypeID & GE_VTYPE_WEIGHT_MASK) && decOptions_.applySkinInDecode;
-	if (applySkin) {
-		// TODO: Support this somehow.
-		return false;
-	}
-
 	_dbg_assert_(numDrawInds_ < MAX_DEFERRED_DRAW_INDS);
 	_dbg_assert_(numDrawVerts_ > 0);
 	*bytesRead = vertexCount * dec_->VertexSize();
@@ -707,6 +701,7 @@ bool DrawEngineCommon::ExtendNonIndexedPrim(GEPrimitiveType prim, int vertexCoun
 	dv.vertexCount += vertexCount;
 	dv.indexUpperBound = dv.vertexCount - 1;
 	vertexCountInDrawCalls_ += vertexCount;
+
 	return true;
 }
 
@@ -787,10 +782,6 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 		gstate_c.Dirty(DIRTY_TEXTURE_PARAMS);
 		DispatchFlush();
 	}
-
-	if (applySkin) {
-		DecodeVerts(decoded_);
-	}
 }
 
 void DrawEngineCommon::DecodeVerts(u8 *dest) {
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
index a6461c9cfb..3e9228bbfd 100644
--- a/GPU/Common/DrawEngineCommon.h
+++ b/GPU/Common/DrawEngineCommon.h
@@ -104,6 +104,13 @@ public:
 
 	bool TestBoundingBox(const void *control_points, const void *inds, int vertexCount, u32 vertType);
 
+	void FlushSkin() {
+		bool applySkin = (lastVType_ & GE_VTYPE_WEIGHT_MASK) && decOptions_.applySkinInDecode;
+		if (applySkin) {
+			DecodeVerts(decoded_);
+		}
+	}
+
 	bool ExtendNonIndexedPrim(GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead);
 	void SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead);
 	template<class Surface>
diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index 5bf7e9873c..41b0a7ec57 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1038,6 +1038,9 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			inds = nullptr;
 			if ((vertexType & GE_VTYPE_IDX_MASK) != GE_VTYPE_IDX_NONE) {
 				inds = Memory::GetPointerUnchecked(gstate_c.indexAddr);
+			} else {
+				// We can extend again after submitting a normal draw.
+				canExtend = true;
 			}
 			drawEngineCommon_->SubmitPrim(verts, inds, newPrim, count, vertTypeID, cullMode, &bytesRead);
 			AdvanceVerts(vertexType, count, bytesRead);
@@ -1058,18 +1061,18 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			break;
 		}
 		case GE_CMD_VADDR:
+			canExtend = false;
 			gstate.cmdmem[GE_CMD_VADDR] = data;
 			gstate_c.vertexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
-			canExtend = false;
 			break;
 		case GE_CMD_IADDR:
 			gstate.cmdmem[GE_CMD_IADDR] = data;
 			gstate_c.indexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
 			break;
 		case GE_CMD_OFFSETADDR:
+			canExtend = false;
 			gstate.cmdmem[GE_CMD_OFFSETADDR] = data;
 			gstate_c.offsetAddr = data << 8;
-			canExtend = false;
 			break;
 		case GE_CMD_BASE:
 			gstate.cmdmem[GE_CMD_BASE] = data;
@@ -1124,6 +1127,8 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 				(Memory::ReadUnchecked_U32(target + 12 * 4) >> 24) == GE_CMD_RET &&
 				(target > currentList->stall || target + 12 * 4 < currentList->stall) &&
 				(gstate.boneMatrixNumber & 0x00FFFFFF) <= 96 - 12) {
+				drawEngineCommon_->FlushSkin();
+				canExtend = false;
 				FastLoadBoneMatrix(target);
 			} else {
 				goto bail;
@@ -1145,6 +1150,7 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 	}
 
 bail:
+	drawEngineCommon_->FlushSkin();
 	gstate.cmdmem[GE_CMD_VERTEXTYPE] = vertexType;
 	int cmdCount = src - start;
 	// Skip over the commands we just read out manually.

From 3aa0f5b543b890d21f4ba1ebbe95aede085e1bd7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 17:00:46 +0200
Subject: [PATCH 7/9] A bit more

---
 GPU/GPUCommonHW.cpp | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index 41b0a7ec57..47e1ef56a0 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1049,19 +1049,19 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		}
 		case GE_CMD_VERTEXTYPE:
 		{
-			canExtend = false;  // TODO: Might support extending between some vertex types in the future.
 			uint32_t diff = data ^ vertexType;
 			// don't mask upper bits, vertexType is unmasked
 			if (diff & vtypeCheckMask) {
 				goto bail;
-			} else {
+			} else if (data != vertexType) {
+				canExtend = false;  // TODO: Might support extending between some vertex types in the future.
 				vertexType = data;
 				vertTypeID = GetVertTypeID(vertexType, gstate.getUVGenMode(), g_Config.bSoftwareSkinning);
 			}
 			break;
 		}
 		case GE_CMD_VADDR:
-			canExtend = false;
+			canExtend = false;  // TODO: See if we can do a more lenient check.
 			gstate.cmdmem[GE_CMD_VADDR] = data;
 			gstate_c.vertexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
 			break;
@@ -1070,7 +1070,6 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			gstate_c.indexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
 			break;
 		case GE_CMD_OFFSETADDR:
-			canExtend = false;
 			gstate.cmdmem[GE_CMD_OFFSETADDR] = data;
 			gstate_c.offsetAddr = data << 8;
 			break;

From 200575b2bca91d1dfb94403f629719e54573320e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 17:07:44 +0200
Subject: [PATCH 8/9] Allow the new optimization through redundant VADDR
 instructions, very common

---
 GPU/GPUCommonHW.cpp | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index 47e1ef56a0..465adb3f74 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1051,9 +1051,9 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 		{
 			uint32_t diff = data ^ vertexType;
 			// don't mask upper bits, vertexType is unmasked
-			if (diff & vtypeCheckMask) {
-				goto bail;
-			} else if (data != vertexType) {
+			if (diff) {
+				if (diff & vtypeCheckMask)
+					goto bail;
 				canExtend = false;  // TODO: Might support extending between some vertex types in the future.
 				vertexType = data;
 				vertTypeID = GetVertTypeID(vertexType, gstate.getUVGenMode(), g_Config.bSoftwareSkinning);
@@ -1061,10 +1061,15 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			break;
 		}
 		case GE_CMD_VADDR:
-			canExtend = false;  // TODO: See if we can do a more lenient check.
+		{
 			gstate.cmdmem[GE_CMD_VADDR] = data;
-			gstate_c.vertexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
+			uint32_t newAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);
+			if (gstate_c.vertexAddr != newAddr) {
+				canExtend = false;
+				gstate_c.vertexAddr = newAddr;
+			}
 			break;
+		}
 		case GE_CMD_IADDR:
 			gstate.cmdmem[GE_CMD_IADDR] = data;
 			gstate_c.indexAddr = gstate_c.getRelativeAddress(data & 0x00FFFFFF);

From af7efe4b5d358cbbcf7f89c4b005a12907c7716c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Mon, 2 Oct 2023 19:18:27 +0200
Subject: [PATCH 9/9] Fix. Need to flush soft-skinned vertices when changing
 vertex format.

---
 GPU/GPUCommonHW.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index 465adb3f74..a28a16ff68 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1052,6 +1052,7 @@ void GPUCommonHW::Execute_Prim(u32 op, u32 diff) {
 			uint32_t diff = data ^ vertexType;
 			// don't mask upper bits, vertexType is unmasked
 			if (diff) {
+				drawEngineCommon_->FlushSkin();
 				if (diff & vtypeCheckMask)
 					goto bail;
 				canExtend = false;  // TODO: Might support extending between some vertex types in the future.