Implement fast-path for merging non-indexed draws quickly.

hrydgard · Oct 3, 2023 · 0260aeb · 0260aeb
1 parent e63bb04
commit 0260aeb
Show file tree

Hide file tree

Showing 3 changed files with 57 additions and 5 deletions.
diff --git 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
@@ -104,6 +104,7 @@ class DrawEngineCommon {
 
 	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
@@ -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;
@@ -1017,19 +1020,33 @@ 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;
 			break;
 		}
 		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;