Merge pull request #9706 from unknownbrackets/softgpu-vdec

SoftGPU: Use vertexjit for submitted primitives

GPU/Software/SoftGpu.cpp

@@ -55,23 +55,6 @@ static Draw::SamplerState *samplerLinear = nullptr;
 static Draw::Buffer *vdata = nullptr;
 static Draw::Buffer *idata = nullptr;
 
-class SoftwareDrawEngine : public DrawEngineCommon {
-public:
-	SoftwareDrawEngine() {
-		// All this is a LOT of memory, need to see if we can cut down somehow.  Used for splines.
-		decoded = (u8 *)AllocateMemoryPages(DECODED_VERTEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
-		decIndex = (u16 *)AllocateMemoryPages(DECODED_INDEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
-		splineBuffer = (u8 *)AllocateMemoryPages(SPLINE_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
-	}
-
-	virtual void DispatchFlush() {
-	}
-
-	virtual void DispatchSubmitPrim(void *verts, void *inds, GEPrimitiveType prim, int vertexCount, u32 vertType, int *bytesRead) {
-		TransformUnit::SubmitPrimitive(verts, inds, prim, vertexCount, vertType, bytesRead);
-	}
-};
-
 SoftGPU::SoftGPU(GraphicsContext *gfxCtx, Draw::DrawContext *draw)
 	: GPUCommon(gfxCtx, draw)
 {
@@ -121,7 +104,8 @@ SoftGPU::SoftGPU(GraphicsContext *gfxCtx, Draw::DrawContext *draw)
 	displayStride_ = 512;
 	displayFormat_ = GE_FORMAT_8888;
 
-	drawEngineCommon_ = new SoftwareDrawEngine();
+	drawEngine_ = new SoftwareDrawEngine();
+	drawEngineCommon_ = drawEngine_;
 }
 
 void SoftGPU::DeviceLost() {
@@ -384,7 +368,7 @@ void SoftGPU::ExecuteOp(u32 op, u32 diff) {
 
 			cyclesExecuted += EstimatePerVertexCost() * count;
 			int bytesRead;
-			TransformUnit::SubmitPrimitive(verts, indices, type, count, gstate.vertType, &bytesRead);
+			TransformUnit::SubmitPrimitive(verts, indices, type, count, gstate.vertType, &bytesRead, drawEngine_);
 			framebufferDirty_ = true;
 
 			// After drawing, we advance the vertexAddr (when non indexed) or indexAddr (when indexed).

GPU/Software/SoftGpu.h

@@ -45,7 +45,7 @@ struct FormatBuffer {
 	}
 };
 
-class ShaderManagerGLES;
+class SoftwareDrawEngine;
 
 class SoftGPU : public GPUCommon {
 public:
@@ -103,6 +103,8 @@ private:
 	u32 displayStride_;
 	GEBufferFormat displayFormat_;
 
+	SoftwareDrawEngine *drawEngine_ = nullptr;
+
 	Draw::Texture *fbTex;
 	Draw::Pipeline *texColor;
 	std::vector<u32> fbTexBuffer;

GPU/Software/TransformUnit.cpp

@@ -30,6 +30,31 @@
 static u8 buf[65536 * 48];  // yolo
 bool TransformUnit::outside_range_flag = false;
 
+SoftwareDrawEngine::SoftwareDrawEngine() {
+	// All this is a LOT of memory, need to see if we can cut down somehow.  Used for splines.
+	decoded = (u8 *)AllocateMemoryPages(DECODED_VERTEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
+	decIndex = (u16 *)AllocateMemoryPages(DECODED_INDEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
+	splineBuffer = (u8 *)AllocateMemoryPages(SPLINE_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
+}
+
+SoftwareDrawEngine::~SoftwareDrawEngine() {
+	FreeMemoryPages(decoded, DECODED_VERTEX_BUFFER_SIZE);
+	FreeMemoryPages(decIndex, DECODED_INDEX_BUFFER_SIZE);
+	FreeMemoryPages(splineBuffer, SPLINE_BUFFER_SIZE);
+}
+
+void SoftwareDrawEngine::DispatchFlush() {
+}
+
+void SoftwareDrawEngine::DispatchSubmitPrim(void *verts, void *inds, GEPrimitiveType prim, int vertexCount, u32 vertType, int *bytesRead) {
+	TransformUnit::SubmitPrimitive(verts, inds, prim, vertexCount, vertType, bytesRead, this);
+}
+
+VertexDecoder *SoftwareDrawEngine::FindVertexDecoder(u32 vtype) {
+	const u32 vertTypeID = (vtype & 0xFFFFFF) | (gstate.getUVGenMode() << 24);
+	return DrawEngineCommon::GetVertexDecoder(vertTypeID);
+}
+
 WorldCoords TransformUnit::ModelToWorld(const ModelCoords& coords)
 {
 	Mat3x3<float> world_matrix(gstate.worldMatrix);
@@ -216,13 +241,10 @@ struct SplinePatch {
 	int pad[3];
 };
 
-void TransformUnit::SubmitPrimitive(void* vertices, void* indices, GEPrimitiveType prim_type, int vertex_count, u32 vertex_type, int *bytesRead)
+void TransformUnit::SubmitPrimitive(void* vertices, void* indices, GEPrimitiveType prim_type, int vertex_count, u32 vertex_type, int *bytesRead, SoftwareDrawEngine *drawEngine)
 {
-	// TODO: Cache VertexDecoder objects
+	VertexDecoder &vdecoder = *drawEngine->FindVertexDecoder(vertex_type);
-	VertexDecoder vdecoder;
+	const DecVtxFormat &vtxfmt = vdecoder.GetDecVtxFmt();
-	VertexDecoderOptions options{};
-	vdecoder.SetVertexType(vertex_type, options);
-	const DecVtxFormat& vtxfmt = vdecoder.GetDecVtxFmt();
 
 	if (bytesRead)
 		*bytesRead = vertex_count * vdecoder.VertexSize();

GPU/Software/TransformUnit.h

@@ -18,6 +18,7 @@
 #pragma once
 
 #include "CommonTypes.h"
+#include "GPU/Common/DrawEngineCommon.h"
 #include "GPU/Common/GPUDebugInterface.h"
 #include "GPU/Math3D.h"
 
@@ -30,6 +31,17 @@ typedef Vec3<float> WorldCoords;
 typedef Vec3<float> ViewCoords;
 typedef Vec4<float> ClipCoords; // Range: -w <= x/y/z <= w
 
+class SoftwareDrawEngine : public DrawEngineCommon {
+public:
+	SoftwareDrawEngine();
+	~SoftwareDrawEngine();
+
+	void DispatchFlush() override;
+	void DispatchSubmitPrim(void *verts, void *inds, GEPrimitiveType prim, int vertexCount, u32 vertType, int *bytesRead) override;
+
+	VertexDecoder *FindVertexDecoder(u32 vtype);
+};
+
 struct SplinePatch;
 
 struct ScreenCoords
@@ -126,7 +138,7 @@ public:
 	static DrawingCoords ScreenToDrawing(const ScreenCoords& coords);
 	static ScreenCoords DrawingToScreen(const DrawingCoords& coords);
 
-	static void SubmitPrimitive(void* vertices, void* indices, GEPrimitiveType prim_type, int vertex_count, u32 vertex_type, int *bytesRead);
+	static void SubmitPrimitive(void* vertices, void* indices, GEPrimitiveType prim_type, int vertex_count, u32 vertex_type, int *bytesRead, SoftwareDrawEngine *drawEngine);
 
 	static bool GetCurrentSimpleVertices(int count, std::vector<GPUDebugVertex> &vertices, std::vector<u16> &indices);
 	static VertexData ReadVertex(VertexReader& vreader);