It's running.

hrydgard · Oct 3, 2023 · 9b411af · 9b411af
1 parent 6a2e5dd
commit 9b411af
Show file tree

Hide file tree

Showing 11 changed files with 209 additions and 231 deletions.
diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
@@ -130,7 +130,7 @@ class DrawEngineCommon {
 		return false;
 	}
 	int GetNumDrawCalls() const {
-		return numDrawCalls_;
+		return numDrawVerts_;
 	}
 
 	VertexDecoder *GetVertexDecoder(u32 vtype);
@@ -141,8 +141,8 @@ class DrawEngineCommon {
 	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 @@ class DrawEngineCommon {
 	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 @@ class DrawEngineCommon {
 	TransformedVertex *transformedExpanded_ = nullptr;
 
 	// Defer all vertex decoding to a "Flush" (except when software skinning)
-	struct DeferredDrawCall {
+	struct DeferredVerts {
 		const void *verts;
+		u32 vertexCount;
+		u16 indexLowerBound;
+		u16 indexUpperBound;
+		UVScale uvScale;
+	};
+
+	struct DeferredInds {
 		const void *inds;
 		u32 vertexCount;
 		u8 indexType;
 		s8 prim;
 		u8 cullMode;
-		u16 indexLowerBound;
-		u16 indexUpperBound;
-		UVScale uvScale;
+		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
@@ -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 @@ void DrawEngineD3D11::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;
+	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
@@ -138,19 +138,19 @@ class DrawEngineD3D11 : public DrawEngineCommon {
 
 	// 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
@@ -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 @@ void DrawEngineDX9::DoFlush() {
 	}
 
 	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
@@ -128,19 +128,19 @@ class DrawEngineDX9 : public DrawEngineCommon {
 
 	// 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
@@ -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;
 	}
 
@@ -471,17 +473,19 @@ 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,
 	// and start indexing from a higher value. This is very friendly to OpenGL (where we can't rely on baseindex if we
 	// 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
@@ -86,19 +86,19 @@ class DrawEngineGLES : public DrawEngineCommon {
 
 	// 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
@@ -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
@@ -170,13 +170,13 @@ class DrawEngineVulkan : public DrawEngineCommon {
 
 	// 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 @@ class DrawEngineVulkan : public DrawEngineCommon {
 	}
 
 	void DispatchFlush() override {
-		if (!numDrawCalls_)
+		if (!numDrawVerts_)
 			return;
 		Flush();
 	}

diff --git 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;