Fix portals not working in softpoly (#1323)

src/common/rendering/polyrenderer/drawers/poly_thread.cpp

@@ -186,13 +186,8 @@ void PolyTriangleThreadData::SetDepthFunc(int func)
 
 void PolyTriangleThreadData::SetDepthRange(float min, float max)
 {
-	// The only two variants used by hwrenderer layer
+	DepthRangeStart = min;
-	if (min == 0.0f && max == 1.0f)
+	DepthRangeScale = max - min;
-	{
-	}
-	else if (min == 1.0f && max == 1.0f)
-	{
-	}
 }
 
 void PolyTriangleThreadData::SetDepthBias(float depthBiasConstantFactor, float depthBiasSlopeFactor)
@@ -210,19 +205,18 @@ void PolyTriangleThreadData::SetStencil(int stencilRef, int op)
 	StencilTestValue = stencilRef;
 	if (op == SOP_Increment)
 	{
-		WriteStencil = StencilTest;
 		StencilWriteValue = MIN(stencilRef + 1, (int)255);
 	}
 	else if (op == SOP_Decrement)
 	{
-		WriteStencil = StencilTest;
 		StencilWriteValue = MAX(stencilRef - 1, (int)0);
 	}
 	else // SOP_Keep
 	{
-		WriteStencil = false;
 		StencilWriteValue = stencilRef;
 	}
+
+	WriteStencil = StencilTest && (StencilTestValue != StencilWriteValue);
 }
 
 void PolyTriangleThreadData::SetCulling(int mode)

src/common/rendering/polyrenderer/drawers/poly_thread.h

@@ -174,6 +174,8 @@ public:
 	bool WriteDepth = true;
 	uint8_t StencilTestValue = 0;
 	uint8_t StencilWriteValue = 0;
+	float DepthRangeStart = 0.0f;
+	float DepthRangeScale = 1.0f;
 
 	void (*FragmentShader)(int x0, int x1, PolyTriangleThreadData* thread) = nullptr;
 	void (*WriteColorFunc)(int y, int x0, int x1, PolyTriangleThreadData* thread) = nullptr;

src/common/rendering/polyrenderer/drawers/screen_scanline_setup.cpp

@@ -70,11 +70,13 @@ void WriteW(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangle
 		mposW = _mm_add_ps(mposW, mstepW);
 	}
 
-	posW += ssecount * stepW;
+	mstepW = _mm_set_ss(stepW);
 	for (int x = sseend; x < x1; x++)
 	{
-		w[x] = 1.0f / posW;
+		__m128 res = _mm_rcp_ss(mposW);
-		posW += stepW;
+		__m128 muls = _mm_mul_ss(mposW, _mm_mul_ss(res, res));
+		_mm_store_ss(w + x, _mm_sub_ss(_mm_add_ss(res, res), muls));
+		mposW = _mm_add_ss(mposW, mstepW);
 	}
 }
 #endif

src/common/rendering/polyrenderer/drawers/screen_triangle.cpp

@@ -34,19 +34,28 @@
 
 static void WriteDepth(int y, int x0, int x1, PolyTriangleThreadData* thread)
 {
-	size_t pitch = thread->depthstencil->Width();
+	float* line = thread->depthstencil->DepthValues() + (size_t)thread->depthstencil->Width() * y;
-	float* line = thread->depthstencil->DepthValues() + pitch * y;
+
+	if (thread->DepthRangeScale != 0.0f)
+	{
 		float* w = thread->scanline.W;
 		for (int x = x0; x < x1; x++)
 		{
 			line[x] = w[x];
 		}
 	}
+	else // portal fills always uses DepthRangeStart = 1 and DepthRangeScale = 0
+	{
+		for (int x = x0; x < x1; x++)
+		{
+			line[x] = 65536.0f;
+		}
+	}
+}
 
 static void WriteStencil(int y, int x0, int x1, PolyTriangleThreadData* thread)
 {
-	size_t pitch = thread->depthstencil->Width();
+	uint8_t* line = thread->depthstencil->StencilValues() + (size_t)thread->depthstencil->Width() * y;
-	uint8_t* line = thread->depthstencil->StencilValues() + pitch * y;
 	uint8_t value = thread->StencilWriteValue;
 	for (int x = x0; x < x1; x++)
 	{
@@ -54,11 +63,16 @@ static void WriteStencil(int y, int x0, int x1, PolyTriangleThreadData* thread)
 	}
 }
 
-static void DrawSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
+static void RunFragmentShader(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
 {
 	WriteVaryings(y, x0, x1, args, thread);
-
 	thread->FragmentShader(x0, x1, thread);
+}
+
+static void DrawSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
+{
+	if (thread->WriteColor || thread->AlphaTest)
+		RunFragmentShader(y, x0, x1, args, thread);
 
 	if (!thread->AlphaTest)
 	{
@@ -92,36 +106,19 @@ static void DrawSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, Pol
 	}
 }
 
-template<typename OptT>
+static void NoTestSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
-static void TestSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
+{
+	WriteW(y, x0, x1, args, thread);
+	DrawSpan(y, x0, x1, args, thread);
+}
+
+static void DepthTestSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
 {
 	WriteW(y, x0, x1, args, thread);
 
-	if ((OptT::Flags & SWTRI_DepthTest) || (OptT::Flags & SWTRI_StencilTest))
+	float* zbufferLine = thread->depthstencil->DepthValues() + (size_t)thread->depthstencil->Width() * y;
-	{
+	float* w = thread->scanline.W;
-		size_t pitch = thread->depthstencil->Width();
+	float depthbias = thread->depthbias;
-
-		uint8_t* stencilbuffer;
-		uint8_t* stencilLine;
-		uint8_t stencilTestValue;
-		if (OptT::Flags & SWTRI_StencilTest)
-		{
-			stencilbuffer = thread->depthstencil->StencilValues();
-			stencilLine = stencilbuffer + pitch * y;
-			stencilTestValue = thread->StencilTestValue;
-		}
-
-		float* zbuffer;
-		float* zbufferLine;
-		float* w;
-		float depthbias;
-		if (OptT::Flags & SWTRI_DepthTest)
-		{
-			zbuffer = thread->depthstencil->DepthValues();
-			zbufferLine = zbuffer + pitch * y;
-			w = thread->scanline.W;
-			depthbias = thread->depthbias;
-		}
 
 	int x = x0;
 	int xend = x1;
@@ -129,51 +126,63 @@ static void TestSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, Pol
 	{
 		int xstart = x;
 
-			if ((OptT::Flags & SWTRI_DepthTest) && (OptT::Flags & SWTRI_StencilTest))
-			{
-				while (zbufferLine[x] >= w[x] + depthbias && stencilLine[x] == stencilTestValue && x < xend)
-					x++;
-			}
-			else if (OptT::Flags & SWTRI_DepthTest)
-			{
 		while (zbufferLine[x] >= w[x] + depthbias && x < xend)
 			x++;
-			}
-			else if (OptT::Flags & SWTRI_StencilTest)
-			{
-				while (stencilLine[x] == stencilTestValue && x < xend)
-					x++;
-			}
-			else
-			{
-				x = xend;
-			}
 
 		if (x > xstart)
 		{
 			DrawSpan(y, xstart, x, args, thread);
 		}
 
-			if ((OptT::Flags & SWTRI_DepthTest) && (OptT::Flags & SWTRI_StencilTest))
-			{
-				while ((zbufferLine[x] < w[x] + depthbias || stencilLine[x] != stencilTestValue) && x < xend)
-					x++;
-			}
-			else if (OptT::Flags & SWTRI_DepthTest)
-			{
 		while (zbufferLine[x] < w[x] + depthbias && x < xend)
 			x++;
 	}
-			else if (OptT::Flags & SWTRI_StencilTest)
+}
+
+static void DepthStencilTestSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
 {
+	uint8_t* stencilLine = thread->depthstencil->StencilValues() + (size_t)thread->depthstencil->Width() * y;
+	uint8_t stencilTestValue = thread->StencilTestValue;
+
+	int x = x0;
+	int xend = x1;
+	while (x < xend)
+	{
+		int xstart = x;
+		while (stencilLine[x] == stencilTestValue && x < xend)
+			x++;
+
+		if (x > xstart)
+		{
+			DepthTestSpan(y, xstart, x, args, thread);
+		}
+
 		while (stencilLine[x] != stencilTestValue && x < xend)
 			x++;
 	}
 }
-	}
+
-	else
+static void StencilTestSpan(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread)
 {
-		DrawSpan(y, x0, x1, args, thread);
+	uint8_t* stencilLine = thread->depthstencil->StencilValues() + (size_t)thread->depthstencil->Width() * y;
+	uint8_t stencilTestValue = thread->StencilTestValue;
+
+	int x = x0;
+	int xend = x1;
+	while (x < xend)
+	{
+		int xstart = x;
+		while (stencilLine[x] == stencilTestValue && x < xend)
+			x++;
+
+		if (x > xstart)
+		{
+			WriteW(y, x0, x1, args, thread);
+			DrawSpan(y, xstart, x, args, thread);
+		}
+
+		while (stencilLine[x] != stencilTestValue && x < xend)
+			x++;
 	}
 }
 
@@ -287,8 +296,8 @@ void ScreenTriangle::Draw(const TriDrawTriangleArgs* args, PolyTriangleThreadDat
 
 void(*ScreenTriangle::TestSpanOpts[])(int y, int x0, int x1, const TriDrawTriangleArgs* args, PolyTriangleThreadData* thread) =
 {
-	&TestSpan<TestSpanOpt0>,
+	&NoTestSpan,
-	&TestSpan<TestSpanOpt1>,
+	&DepthTestSpan,
-	&TestSpan<TestSpanOpt2>,
+	&StencilTestSpan,
-	&TestSpan<TestSpanOpt3>
+	&DepthStencilTestSpan
 };

src/common/rendering/polyrenderer/drawers/screen_triangle.h

@@ -122,8 +122,3 @@ enum SWTestSpan
 	SWTRI_DepthTest = 1,
 	SWTRI_StencilTest = 2
 };
-
-struct TestSpanOpt0 { static const int Flags = 0; };
-struct TestSpanOpt1 { static const int Flags = 1; };
-struct TestSpanOpt2 { static const int Flags = 2; };
-struct TestSpanOpt3 { static const int Flags = 3; };