SoftRasterizer: Do some multithreading improvements, and also clean up and refactor RasterizerUnit.

- Completely encapsulate all stray global variables into the SoftRasterizer class where they belong.
- Framebuffer clears are now fully multithreaded, significantly improving clearing performance.
- Doing multithreaded texture loads and vertex calculations now requires a minimum of 2 threads, down from 4 threads.
- The maximum amount of SoftRasterizer threads has been increased from 16 to 32.

desmume/src/OGLRender.cpp

@@ -1,7 +1,7 @@
 /*
 	Copyright (C) 2006 yopyop
 	Copyright (C) 2006-2007 shash
-	Copyright (C) 2008-2017 DeSmuME team
+	Copyright (C) 2008-2018 DeSmuME team
 
 	This file is free software: you can redistribute it and/or modify
 	it under the terms of the GNU General Public License as published by
@@ -1171,14 +1171,12 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 	
 	if (!doFramebufferFlip)
 	{
-		const size_t pixCount = this->_framebufferWidth * this->_framebufferHeight;
-		
 		if (!doFramebufferConvert)
 		{
 			if ( (dstFramebufferMain != NULL) && (dstFramebuffer16 != NULL) )
 			{
 #ifdef ENABLE_SSE2
-				const size_t ssePixCount = pixCount - (pixCount % 8);
+				const size_t ssePixCount = this->_framebufferPixCount - (this->_framebufferPixCount % 8);
 				for (; i < ssePixCount; i += 8)
 				{
 					const __m128i srcColorLo = _mm_load_si128((__m128i *)(srcFramebuffer + i + 0));
@@ -1191,7 +1189,7 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 				
 #pragma LOOPVECTORIZE_DISABLE
 #endif
-				for (; i < pixCount; i++)
+				for (; i < this->_framebufferPixCount; i++)
 				{
 					dstFramebufferMain[i].color = ColorspaceCopy32<false>(srcFramebuffer[i]);
 					dstFramebuffer16[i]         = ColorspaceConvert8888To5551<false>(srcFramebuffer[i]);
@@ -1202,12 +1200,12 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 			}
 			else if (dstFramebufferMain != NULL)
 			{
-				ColorspaceCopyBuffer32<false, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, pixCount);
+				ColorspaceCopyBuffer32<false, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, this->_framebufferPixCount);
 				this->_renderNeedsFlushMain = false;
 			}
 			else
 			{
-				ColorspaceConvertBuffer8888To5551<false, false>((u32 *)srcFramebuffer, dstFramebuffer16, pixCount);
+				ColorspaceConvertBuffer8888To5551<false, false>((u32 *)srcFramebuffer, dstFramebuffer16, this->_framebufferPixCount);
 				this->_renderNeedsFlush16 = false;
 			}
 		}
@@ -1218,7 +1216,7 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 				if ( (dstFramebufferMain != NULL) && (dstFramebuffer16 != NULL) )
 				{
 #ifdef ENABLE_SSE2
-					const size_t ssePixCount = pixCount - (pixCount % 8);
+					const size_t ssePixCount = this->_framebufferPixCount - (this->_framebufferPixCount % 8);
 					for (; i < ssePixCount; i += 8)
 					{
 						const __m128i srcColorLo = _mm_load_si128((__m128i *)(srcFramebuffer + i + 0));
@@ -1231,7 +1229,7 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 					
 #pragma LOOPVECTORIZE_DISABLE
 #endif
-					for (; i < pixCount; i++)
+					for (; i < this->_framebufferPixCount; i++)
 					{
 						dstFramebufferMain[i].color = ColorspaceConvert8888To6665<true>(srcFramebuffer[i]);
 						dstFramebuffer16[i]         = ColorspaceConvert8888To5551<true>(srcFramebuffer[i]);
@@ -1242,12 +1240,12 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 				}
 				else if (dstFramebufferMain != NULL)
 				{
-					ColorspaceConvertBuffer8888To6665<true, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, pixCount);
+					ColorspaceConvertBuffer8888To6665<true, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, this->_framebufferPixCount);
 					this->_renderNeedsFlushMain = false;
 				}
 				else
 				{
-					ColorspaceConvertBuffer8888To5551<true, false>((u32 *)srcFramebuffer, dstFramebuffer16, pixCount);
+					ColorspaceConvertBuffer8888To5551<true, false>((u32 *)srcFramebuffer, dstFramebuffer16, this->_framebufferPixCount);
 					this->_renderNeedsFlush16 = false;
 				}
 			}
@@ -1256,7 +1254,7 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 				if ( (dstFramebufferMain != NULL) && (dstFramebuffer16 != NULL) )
 				{
 #ifdef ENABLE_SSE2
-					const size_t ssePixCount = pixCount - (pixCount % 8);
+					const size_t ssePixCount = this->_framebufferPixCount - (this->_framebufferPixCount % 8);
 					for (; i < ssePixCount; i += 8)
 					{
 						const __m128i srcColorLo = _mm_load_si128((__m128i *)(srcFramebuffer + i + 0));
@@ -1269,7 +1267,7 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 					
 #pragma LOOPVECTORIZE_DISABLE
 #endif
-					for (; i < pixCount; i++)
+					for (; i < this->_framebufferPixCount; i++)
 					{
 						dstFramebufferMain[i].color = ColorspaceCopy32<true>(srcFramebuffer[i]);
 						dstFramebuffer16[i]         = ColorspaceConvert8888To5551<true>(srcFramebuffer[i]);
@@ -1280,12 +1278,12 @@ Render3DError OpenGLRenderer::_FlushFramebufferFlipAndConvertOnCPU(const Fragmen
 				}
 				else if (dstFramebufferMain != NULL)
 				{
-					ColorspaceCopyBuffer32<true, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, pixCount);
+					ColorspaceCopyBuffer32<true, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, this->_framebufferPixCount);
 					this->_renderNeedsFlushMain = false;
 				}
 				else
 				{
-					ColorspaceConvertBuffer8888To5551<true, false>((u32 *)srcFramebuffer, dstFramebuffer16, pixCount);
+					ColorspaceConvertBuffer8888To5551<true, false>((u32 *)srcFramebuffer, dstFramebuffer16, this->_framebufferPixCount);
 					this->_renderNeedsFlush16 = false;
 				}
 			}
@@ -4891,6 +4889,7 @@ Render3DError OpenGLRenderer_1_2::SetFramebufferSize(size_t w, size_t h)
 	
 	this->_framebufferWidth = w;
 	this->_framebufferHeight = h;
+	this->_framebufferPixCount = w * h;
 	this->_framebufferColorSizeBytes = newFramebufferColorSizeBytes;
 	
 	if (this->isPBOSupported)

desmume/src/OGLRender_3_2.cpp

@@ -1,7 +1,7 @@
 /*
 	Copyright (C) 2006 yopyop
 	Copyright (C) 2006-2007 shash
-	Copyright (C) 2008-2017 DeSmuME team
+	Copyright (C) 2008-2018 DeSmuME team
 
 	This file is free software: you can redistribute it and/or modify
 	it under the terms of the GNU General Public License as published by
@@ -2064,6 +2064,7 @@ Render3DError OpenGLRenderer_3_2::SetFramebufferSize(size_t w, size_t h)
 	
 	this->_framebufferWidth = w;
 	this->_framebufferHeight = h;
+	this->_framebufferPixCount = w * h;
 	this->_framebufferColorSizeBytes = newFramebufferColorSizeBytes;
 	this->_framebufferColor = NULL; // Don't need to make a client-side buffer since we will be reading directly from the PBO.
 	

desmume/src/rasterize.h

@@ -1,5 +1,5 @@
 /*
-	Copyright (C) 2009-2017 DeSmuME team
+	Copyright (C) 2009-2018 DeSmuME team
 
 	This file is free software: you can redistribute it and/or modify
 	it under the terms of the GNU General Public License as published by
@@ -21,10 +21,20 @@
 #include "render3D.h"
 #include "gfx3d.h"
 
+#define SOFTRASTERIZER_MAX_THREADS 32
 
 extern GPU3DInterface gpu3DRasterize;
 
+class Task;
 class SoftRasterizerRenderer;
+struct edge_fx_fl;
+
+struct SoftRasterizerClearParam
+{
+	SoftRasterizerRenderer *renderer;
+	size_t startPixel;
+	size_t endPixel;
+};
 
 struct SoftRasterizerPostProcessParams
 {
@@ -82,6 +92,39 @@ public:
 	void SetScalingFactor(size_t scalingFactor);
 };
 
+template <bool RENDERER>
+class RasterizerUnit
+{
+protected:
+	bool _debug_thisPoly;
+	u32 _SLI_Mask;
+	u32 _SLI_Value;
+	
+	SoftRasterizerRenderer *_softRender;
+	SoftRasterizerTexture *_currentTexture;
+	VERT *_verts[MAX_CLIPPED_VERTS];
+	size_t _polynum;
+	u8 _textureWrapMode;
+	
+	Render3DError _SetupTexture(const POLY &thePoly, size_t polyRenderIndex);
+	FORCEINLINE FragmentColor _sample(const float u, const float v);
+	FORCEINLINE float _round_s(double val);
+	
+	template<bool ISSHADOWPOLYGON> FORCEINLINE void _shade(const PolygonMode polygonMode, const FragmentColor src, FragmentColor &dst, const float texCoordU, const float texCoordV);
+	template<bool ISSHADOWPOLYGON> FORCEINLINE void _pixel(const POLYGON_ATTR polyAttr, const bool isTranslucent, const size_t fragmentIndex, FragmentColor &dstColor, float r, float g, float b, float invu, float invv, float w, float z);
+	template<bool ISSHADOWPOLYGON, bool USELINEHACK> FORCEINLINE void _drawscanline(const POLYGON_ATTR polyAttr, const bool isTranslucent, FragmentColor *dstColor, const size_t framebufferWidth, const size_t framebufferHeight, edge_fx_fl *pLeft, edge_fx_fl *pRight);
+	template<bool SLI, bool ISSHADOWPOLYGON, bool USELINEHACK, bool ISHORIZONTAL> void _runscanlines(const POLYGON_ATTR polyAttr, const bool isTranslucent, FragmentColor *dstColor, const size_t framebufferWidth, const size_t framebufferHeight, edge_fx_fl *left, edge_fx_fl *right);
+	
+	template<int TYPE> FORCEINLINE void _rot_verts();
+	template<bool ISBACKWARDS, int TYPE> void _sort_verts();
+	template<bool SLI, bool ISBACKWARDS, bool ISSHADOWPOLYGON, bool USELINEHACK> void _shape_engine(const POLYGON_ATTR polyAttr, const bool isTranslucent, FragmentColor *dstColor, const size_t framebufferWidth, const size_t framebufferHeight, int type);
+	
+public:
+	void SetSLI(u32 value, u32 mask, bool debug);
+	void SetRenderer(SoftRasterizerRenderer *theRenderer);
+	template<bool SLI, bool USELINEHACK> FORCEINLINE void Render();
+};
+
 #if defined(ENABLE_SSE2)
 class SoftRasterizerRenderer : public Render3D_SSE2
 #else
@@ -89,12 +132,27 @@ class SoftRasterizerRenderer : public Render3D
 #endif
 {
 protected:
+	Task *_task;
+	SoftRasterizerClearParam _threadClearParam[SOFTRASTERIZER_MAX_THREADS];
+	SoftRasterizerPostProcessParams _threadPostprocessParam[SOFTRASTERIZER_MAX_THREADS];
+	
+	RasterizerUnit<true> _rasterizerUnit[SOFTRASTERIZER_MAX_THREADS];
+	RasterizerUnit<false> _HACK_viewer_rasterizerUnit;
+	
+	size_t _threadCount;
+	size_t _nativeLinesPerThread;
+	size_t _nativePixelsPerThread;
+	size_t _customLinesPerThread;
+	size_t _customPixelsPerThread;
+	
+	FragmentColor _clearColor6665;
+	FragmentAttributes _clearAttributes;
+	
 	GFX3D_Clipper clipper;
 	u8 fogTable[32768];
 	FragmentColor edgeMarkTable[8];
 	bool edgeMarkDisabled[8];
 	
-	bool _stateSetupNeedsFinish;
 	bool _renderGeometryNeedsFinish;
 	
 	bool _enableHighPrecisionColorInterpolation;
@@ -124,7 +182,6 @@ public:
 	bool polyVisible[POLYLIST_SIZE];
 	bool polyBackfacing[POLYLIST_SIZE];
 	GFX3D_State *currentRenderState;
-	SoftRasterizerPostProcessParams *postprocessParam;
 	
 	bool _enableFragmentSamplingHack;
 	
@@ -148,6 +205,7 @@ public:
 	virtual Render3DError Render(const GFX3D &engine);
 	virtual Render3DError RenderFinish();
 	virtual Render3DError RenderFlush(bool willFlushBuffer32, bool willFlushBuffer16);
+	virtual void ClearUsingValuesLoop(const size_t startPixel, const size_t endPixel);
 	virtual Render3DError SetFramebufferSize(size_t w, size_t h);
 };
 
@@ -155,7 +213,22 @@ public:
 
 class SoftRasterizerRenderer_SSE2 : public SoftRasterizerRenderer
 {
+protected:
+	v128u32 _clearColor_v128u32;
+	v128u32 _clearDepth_v128u32;
+	v128u8 _clearAttrOpaquePolyID_v128u8;
+	v128u8 _clearAttrTranslucentPolyID_v128u8;
+	v128u8 _clearAttrStencil_v128u8;
+	v128u8 _clearAttrIsFogged_v128u8;
+	v128u8 _clearAttrIsTranslucentPoly_v128u8;
+	
 	virtual Render3DError ClearUsingValues(const FragmentColor &clearColor6665, const FragmentAttributes &clearAttributes);
+	
+public:
+	SoftRasterizerRenderer_SSE2();
+	
+	virtual void ClearUsingValuesLoop(const size_t startPixel, const size_t endPixel);
+	virtual Render3DError SetFramebufferSize(size_t w, size_t h);
 };
 
 #endif

desmume/src/render3D.cpp

@@ -1,6 +1,6 @@
 /*
 	Copyright (C) 2006-2007 shash
-	Copyright (C) 2008-2017 DeSmuME team
+	Copyright (C) 2008-2018 DeSmuME team
 
 	This file is free software: you can redistribute it and/or modify
 	it under the terms of the GNU General Public License as published by
@@ -244,6 +244,8 @@ Render3D::Render3D()
 	
 	_framebufferWidth = GPU_FRAMEBUFFER_NATIVE_WIDTH;
 	_framebufferHeight = GPU_FRAMEBUFFER_NATIVE_HEIGHT;
+	_framebufferPixCount = _framebufferWidth * _framebufferHeight;
+	_framebufferSIMDPixCount = 0;
 	_framebufferColorSizeBytes = 0;
 	_framebufferColor = NULL;
 	
@@ -332,6 +334,7 @@ Render3DError Render3D::SetFramebufferSize(size_t w, size_t h)
 	
 	this->_framebufferWidth = w;
 	this->_framebufferHeight = h;
+	this->_framebufferPixCount = w * h;
 	this->_framebufferColorSizeBytes = w * h * sizeof(FragmentColor);
 	this->_framebufferColor = GPU->GetEngineMain()->Get3DFramebufferMain(); // Just use the buffer that is already present on the main GPU engine
 	
@@ -478,22 +481,20 @@ Render3DError Render3D::FlushFramebuffer(const FragmentColor *__restrict srcFram
 		return RENDER3DERROR_NOERR;
 	}
 	
-	const size_t pixCount = this->_framebufferWidth * this->_framebufferHeight;
-	
 	if (dstFramebufferMain != NULL)
 	{
 		if ( (this->_internalRenderingFormat == NDSColorFormat_BGR888_Rev) && (this->_outputFormat == NDSColorFormat_BGR666_Rev) )
 		{
-			ColorspaceConvertBuffer8888To6665<false, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, pixCount);
+			ColorspaceConvertBuffer8888To6665<false, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, this->_framebufferPixCount);
 		}
 		else if ( (this->_internalRenderingFormat == NDSColorFormat_BGR666_Rev) && (this->_outputFormat == NDSColorFormat_BGR888_Rev) )
 		{
-			ColorspaceConvertBuffer6665To8888<false, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, pixCount);
+			ColorspaceConvertBuffer6665To8888<false, false>((u32 *)srcFramebuffer, (u32 *)dstFramebufferMain, this->_framebufferPixCount);
 		}
 		else if ( ((this->_internalRenderingFormat == NDSColorFormat_BGR666_Rev) && (this->_outputFormat == NDSColorFormat_BGR666_Rev)) ||
 		          ((this->_internalRenderingFormat == NDSColorFormat_BGR888_Rev) && (this->_outputFormat == NDSColorFormat_BGR888_Rev)) )
 		{
-			memcpy(dstFramebufferMain, srcFramebuffer, pixCount * sizeof(FragmentColor));
+			memcpy(dstFramebufferMain, srcFramebuffer, this->_framebufferPixCount * sizeof(FragmentColor));
 		}
 		
 		this->_renderNeedsFlushMain = false;
@@ -503,11 +504,11 @@ Render3DError Render3D::FlushFramebuffer(const FragmentColor *__restrict srcFram
 	{
 		if (this->_outputFormat == NDSColorFormat_BGR666_Rev)
 		{
-			ColorspaceConvertBuffer6665To5551<false, false>((u32 *)srcFramebuffer, dstFramebuffer16, pixCount);
+			ColorspaceConvertBuffer6665To5551<false, false>((u32 *)srcFramebuffer, dstFramebuffer16, this->_framebufferPixCount);
 		}
 		else if (this ->_outputFormat == NDSColorFormat_BGR888_Rev)
 		{
-			ColorspaceConvertBuffer8888To5551<false, false>((u32 *)srcFramebuffer, dstFramebuffer16, pixCount);
+			ColorspaceConvertBuffer8888To5551<false, false>((u32 *)srcFramebuffer, dstFramebuffer16, this->_framebufferPixCount);
 		}
 		
 		this->_renderNeedsFlush16 = false;
@@ -687,6 +688,42 @@ Render3DError Render3D::VramReconfigureSignal()
 	return RENDER3DERROR_NOERR;
 }
 
+Render3D_SIMD128::Render3D_SIMD128()
+{
+	_framebufferSIMDPixCount = _framebufferPixCount - (_framebufferPixCount % 16);
+}
+
+Render3DError Render3D_SIMD128::SetFramebufferSize(size_t w, size_t h)
+{
+	Render3DError error = this->Render3D::SetFramebufferSize(w, h);
+	if (error != RENDER3DERROR_NOERR)
+	{
+		return RENDER3DERROR_NOERR;
+	}
+	
+	this->_framebufferSIMDPixCount = this->_framebufferPixCount - (this->_framebufferPixCount % 16);
+	
+	return error;
+}
+
+Render3D_SIMD256::Render3D_SIMD256()
+{
+	_framebufferSIMDPixCount = _framebufferPixCount - (_framebufferPixCount % 32);
+}
+
+Render3DError Render3D_SIMD256::SetFramebufferSize(size_t w, size_t h)
+{
+	Render3DError error = this->Render3D::SetFramebufferSize(w, h);
+	if (error != RENDER3DERROR_NOERR)
+	{
+		return RENDER3DERROR_NOERR;
+	}
+	
+	this->_framebufferSIMDPixCount = this->_framebufferPixCount - (this->_framebufferPixCount % 32);
+	
+	return error;
+}
+
 #ifdef ENABLE_SSE2
 
 Render3DError Render3D_SSE2::ClearFramebuffer(const GFX3D_State &renderState)

desmume/src/render3D.h

@@ -1,6 +1,6 @@
 /*
 	Copyright (C) 2006-2007 shash
-	Copyright (C) 2007-2017 DeSmuME team
+	Copyright (C) 2007-2018 DeSmuME team
 
 	This file is free software: you can redistribute it and/or modify
 	it under the terms of the GNU General Public License as published by
@@ -144,6 +144,8 @@ protected:
 	
 	size_t _framebufferWidth;
 	size_t _framebufferHeight;
+	size_t _framebufferPixCount;
+	size_t _framebufferSIMDPixCount;
 	size_t _framebufferColorSizeBytes;
 	FragmentColor *_framebufferColor;
 	
@@ -246,9 +248,25 @@ public:
 	Render3DTexture* GetTextureByPolygonRenderIndex(size_t polyRenderIndex) const;
 };
 
+class Render3D_SIMD128 : public Render3D
+{
+public:
+	Render3D_SIMD128();
+	
+	virtual Render3DError SetFramebufferSize(size_t w, size_t h);
+};
+
+class Render3D_SIMD256 : public Render3D
+{
+public:
+	Render3D_SIMD256();
+	
+	virtual Render3DError SetFramebufferSize(size_t w, size_t h);
+};
+
 #ifdef ENABLE_SSE2
 
-class Render3D_SSE2 : public Render3D
+class Render3D_SSE2 : public Render3D_SIMD128
 {
 public:
 	virtual Render3DError ClearFramebuffer(const GFX3D_State &renderState);