- implement depth bias

src/rendering/polyrenderer/backend/poly_renderstate.cpp

@@ -108,7 +108,7 @@ void PolyRenderState::Clear(int targets)
 	//if (targets & CT_Color)
 	//	PolyTriangleDrawer::ClearColor(GetPolyFrameBuffer()->GetDrawCommands());
 	if (targets & CT_Depth)
-		PolyTriangleDrawer::ClearDepth(GetPolyFrameBuffer()->GetDrawCommands(), 0.0f);
+		PolyTriangleDrawer::ClearDepth(GetPolyFrameBuffer()->GetDrawCommands(), 65535.0f);
 	if (targets & CT_Stencil)
 		PolyTriangleDrawer::ClearStencil(GetPolyFrameBuffer()->GetDrawCommands(), 0);
 }

src/rendering/polyrenderer/drawers/poly_triangle.cpp

@@ -338,10 +338,18 @@ void PolyTriangleThreadData::SetDepthFunc(int func)
 
 void PolyTriangleThreadData::SetDepthRange(float min, float max)
 {
+	// The only two variants used by hwrenderer layer
+	if (min == 0.0f && max == 1.0f)
+	{
+	}
+	else if (min == 1.0f && max == 1.0f)
+	{
+	}
 }
 
 void PolyTriangleThreadData::SetDepthBias(float depthBiasConstantFactor, float depthBiasSlopeFactor)
 {
+	depthbias = (float)(depthBiasConstantFactor / 65536.0);
 }
 
 void PolyTriangleThreadData::SetColorMask(bool r, bool g, bool b, bool a)

src/rendering/polyrenderer/drawers/poly_triangle.h

@@ -121,7 +121,7 @@ public:
 	void SetTransform(const Mat4f *objectToClip, const Mat4f *objectToWorld);
 	void SetCullCCW(bool value) { ccw = value; }
 	void SetTwoSided(bool value) { twosided = value; }
-	void SetWeaponScene(bool value) { weaponScene = value; }
+	void SetWeaponScene(bool value) { depthbias = value ? -1.0f : 0.0f; }
 	void SetModelVertexShader(int frame1, int frame2, float interpolationFactor) { modelFrame1 = frame1; modelFrame2 = frame2; swVertexShader.modelInterpolationFactor = interpolationFactor; }
 
 	void SetInputAssembly(PolyInputAssembly *input) { inputAssembly = input; }
@@ -187,6 +187,7 @@ public:
 	int32_t texelV[MAXWIDTH];
 	uint16_t lightarray[MAXWIDTH];
 	uint32_t dynlights[MAXWIDTH];
+	float depthvalues[MAXWIDTH];
 
 	static PolyTriangleThreadData *Get(DrawerThread *thread);
 
@@ -195,7 +196,7 @@ public:
 	int dest_height = 0;
 	bool dest_bgra = false;
 	uint8_t *dest = nullptr;
-	bool weaponScene = false;
+	float depthbias = 0.0f;
 
 	int viewport_y = 0;
 

src/rendering/polyrenderer/drawers/screen_triangle.cpp

@@ -151,13 +151,37 @@ void ScreenTriangle::Draw(const TriDrawTriangleArgs *args, PolyTriangleThreadDat
 	TriangleDrawers[opt](args, thread, edges, topY, bottomY);
 }
 
+static void FillDepthValues(float *depthvalues, float posXW, float stepXW, int count, float depthbias)
+{
+#ifndef NO_SSE
+	__m128 mstepXW = _mm_set1_ps(stepXW * 4.0f);
+	__m128 mfirstStepXW = _mm_setr_ps(0.0f, stepXW, stepXW + stepXW, stepXW + stepXW + stepXW);
+	__m128 mposXW = _mm_add_ps(_mm_set1_ps(posXW), mfirstStepXW);
+	__m128 mdepthbias = _mm_set1_ps(depthbias);
+	while (count > 0)
+	{
+		_mm_storeu_ps(depthvalues, _mm_add_ps(_mm_rcp_ps(mposXW), mdepthbias));
+		mposXW = _mm_add_ps(mposXW, mstepXW);
+		depthvalues += 4;
+		count -= 4;
+	}
+#else
+	while (count > 0)
+	{
+		*(depthvalues++) = 1.0f / posXW + depthbias;
+		posXW += stepXW;
+		count--;
+	}
+#endif
+}
+
 template<typename OptT>
 void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *thread, int16_t *edges, int topY, int bottomY)
 {
 	using namespace TriScreenDrawerModes;
 
 	void(*drawfunc)(int y, int x0, int x1, const TriDrawTriangleArgs *args, PolyTriangleThreadData *thread);
-	float stepXW, v1X, v1Y, v1W, posXW;
+	float stepXW, v1X, v1Y, v1W;
 	uint8_t stencilTestValue, stencilWriteValue;
 	float *zbuffer;
 	float *zbufferLine;
@@ -194,12 +218,14 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 	if (OptT::Flags & SWTRI_WriteStencil)
 		stencilWriteValue = args->uniforms->StencilWriteValue();
 
-	float weaponWOffset;
+	float depthbias;
 	if ((OptT::Flags & SWTRI_DepthTest) || (OptT::Flags & SWTRI_WriteDepth))
 	{
-		weaponWOffset = thread->weaponScene ? 1.0f : 0.0f;
+		depthbias = thread->depthbias;
 	}
 
+	float *depthvalues = thread->depthvalues;
+
 	int num_cores = thread->num_cores;
 	for (int y = topY; y < bottomY; y += num_cores)
 	{
@@ -215,16 +241,11 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 
 			float startX = x + (0.5f - v1X);
 			float startY = y + (0.5f - v1Y);
-			posXW = v1W + stepXW * startX + args->gradientY.W * startY + weaponWOffset;
+			float posXW = v1W + stepXW * startX + args->gradientY.W * startY;
+			FillDepthValues(depthvalues + x, posXW, stepXW, xend - x, depthbias);
 		}
 
 #ifndef NO_SSE
-		__m128 mstepXW, mfirstStepXW;
-		if ((OptT::Flags & SWTRI_DepthTest) || (OptT::Flags & SWTRI_WriteDepth))
-		{
-			mstepXW = _mm_set1_ps(stepXW * 4.0f);
-			mfirstStepXW = _mm_setr_ps(0.0f, stepXW, stepXW + stepXW, stepXW + stepXW + stepXW);
-		}
 		while (x < xend)
 		{
 			int xstart = x;
@@ -232,9 +253,9 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 			if ((OptT::Flags & SWTRI_DepthTest) && (OptT::Flags & SWTRI_StencilTest))
 			{
 				int xendsse = x + ((xend - x) / 4);
-				__m128 mposXW = _mm_add_ps(_mm_set1_ps(posXW), mfirstStepXW);
+				__m128 mposXW;
 				while (x < xendsse &&
-					_mm_movemask_ps(_mm_cmple_ps(_mm_loadu_ps(zbufferLine + x), mposXW)) == 15 &&
+					_mm_movemask_ps(_mm_cmpge_ps(_mm_loadu_ps(zbufferLine + x), mposXW = _mm_loadu_ps(depthvalues + x))) == 15 &&
 					stencilLine[x] == stencilTestValue &&
 					stencilLine[x + 1] == stencilTestValue &&
 					stencilLine[x + 2] == stencilTestValue &&
@@ -242,37 +263,31 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 				{
 					if (OptT::Flags & SWTRI_WriteDepth)
 						_mm_storeu_ps(zbufferLine + x, mposXW);
-					mposXW = _mm_add_ps(mposXW, mstepXW);
 					x += 4;
 				}
-				posXW = _mm_cvtss_f32(mposXW);
 
-				while (zbufferLine[x] <= posXW && stencilLine[x] == stencilTestValue && x < xend)
+				while (zbufferLine[x] >= depthvalues[x] && stencilLine[x] == stencilTestValue && x < xend)
 				{
 					if (OptT::Flags & SWTRI_WriteDepth)
-						zbufferLine[x] = posXW;
+						zbufferLine[x] = depthvalues[x];
-					posXW += stepXW;
 					x++;
 				}
 			}
 			else if (OptT::Flags & SWTRI_DepthTest)
 			{
 				int xendsse = x + ((xend - x) / 4);
-				__m128 mposXW = _mm_add_ps(_mm_set1_ps(posXW), mfirstStepXW);
+				__m128 mposXW;
-				while (x < xendsse && _mm_movemask_ps(_mm_cmple_ps(_mm_loadu_ps(zbufferLine + x), mposXW)) == 15)
+				while (x < xendsse && _mm_movemask_ps(_mm_cmpge_ps(_mm_loadu_ps(zbufferLine + x), mposXW = _mm_loadu_ps(depthvalues + x))) == 15)
 				{
 					if (OptT::Flags & SWTRI_WriteDepth)
 						_mm_storeu_ps(zbufferLine + x, mposXW);
-					mposXW = _mm_add_ps(mposXW, mstepXW);
 					x += 4;
 				}
-				posXW = _mm_cvtss_f32(mposXW);
 
-				while (zbufferLine[x] <= posXW && x < xend)
+				while (zbufferLine[x] >= depthvalues[x] && x < xend)
 				{
 					if (OptT::Flags & SWTRI_WriteDepth)
-						zbufferLine[x] = posXW;
+						zbufferLine[x] = depthvalues[x];
-					posXW += stepXW;
 					x++;
 				}
 			}
@@ -315,8 +330,7 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 				{
 					for (int i = xstart; i < x; i++)
 					{
-						zbufferLine[i] = posXW;
+						zbufferLine[i] = depthvalues[i];
-						posXW += stepXW;
 					}
 				}
 			}
@@ -324,39 +338,33 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 			if ((OptT::Flags & SWTRI_DepthTest) && (OptT::Flags & SWTRI_StencilTest))
 			{
 				int xendsse = x + ((xend - x) / 4);
-				__m128 mposXW = _mm_add_ps(_mm_set1_ps(posXW), mfirstStepXW);
+				__m128 mposXW;
 				while (x < xendsse &&
-					(_mm_movemask_ps(_mm_cmple_ps(_mm_loadu_ps(zbufferLine + x), mposXW)) == 0 ||
+					(_mm_movemask_ps(_mm_cmpge_ps(_mm_loadu_ps(zbufferLine + x), mposXW = _mm_loadu_ps(depthvalues + x))) == 0 ||
 					stencilLine[x] != stencilTestValue ||
 					stencilLine[x + 1] != stencilTestValue ||
 					stencilLine[x + 2] != stencilTestValue ||
 					stencilLine[x + 3] != stencilTestValue))
 				{
-					mposXW = _mm_add_ps(mposXW, mstepXW);
 					x += 4;
 				}
-				posXW = _mm_cvtss_f32(mposXW);
 
-				while ((zbufferLine[x] > posXW || stencilLine[x] != stencilTestValue) && x < xend)
+				while ((zbufferLine[x] < depthvalues[x] || stencilLine[x] != stencilTestValue) && x < xend)
 				{
-					posXW += stepXW;
 					x++;
 				}
 			}
 			else if (OptT::Flags & SWTRI_DepthTest)
 			{
 				int xendsse = x + ((xend - x) / 4);
-				__m128 mposXW = _mm_add_ps(_mm_set1_ps(posXW), mfirstStepXW);
+				__m128 mposXW;
-				while (x < xendsse && _mm_movemask_ps(_mm_cmple_ps(_mm_loadu_ps(zbufferLine + x), mposXW)) == 0)
+				while (x < xendsse && _mm_movemask_ps(_mm_cmpge_ps(_mm_loadu_ps(zbufferLine + x), mposXW = _mm_loadu_ps(depthvalues + x))) == 0)
 				{
-					mposXW = _mm_add_ps(mposXW, mstepXW);
 					x += 4;
 				}
-				posXW = _mm_cvtss_f32(mposXW);
 
-				while (zbufferLine[x] > posXW && x < xend)
+				while (zbufferLine[x] < depthvalues[x] && x < xend)
 				{
-					posXW += stepXW;
 					x++;
 				}
 			}
@@ -381,21 +389,19 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 
 			if ((OptT::Flags & SWTRI_DepthTest) && (OptT::Flags & SWTRI_StencilTest))
 			{
-				while (zbufferLine[x] <= posXW && stencilLine[x] == stencilTestValue && x < xend)
+				while (zbufferLine[x] >= depthvalues[x] && stencilLine[x] == stencilTestValue && x < xend)
 				{
 					if (OptT::Flags & SWTRI_WriteDepth)
-						zbufferLine[x] = posXW;
+						zbufferLine[x] = depthvalues[x];
-					posXW += stepXW;
 					x++;
 				}
 			}
 			else if (OptT::Flags & SWTRI_DepthTest)
 			{
-				while (zbufferLine[x] <= posXW && x < xend)
+				while (zbufferLine[x] >= depthvalues[x] && x < xend)
 				{
 					if (OptT::Flags & SWTRI_WriteDepth)
-						zbufferLine[x] = posXW;
+						zbufferLine[x] = depthvalues[x];
-					posXW += stepXW;
 					x++;
 				}
 			}
@@ -424,25 +430,22 @@ void DrawTriangle(const TriDrawTriangleArgs *args, PolyTriangleThreadData *threa
 				{
 					for (int i = xstart; i < x; i++)
 					{
-						zbufferLine[i] = posXW;
+						zbufferLine[i] = depthvalues[i];
-						posXW += stepXW;
 					}
 				}
 			}
 
 			if ((OptT::Flags & SWTRI_DepthTest) && (OptT::Flags & SWTRI_StencilTest))
 			{
-				while ((zbufferLine[x] > posXW || stencilLine[x] != stencilTestValue) && x < xend)
+				while ((zbufferLine[x] < depthvalues[x] || stencilLine[x] != stencilTestValue) && x < xend)
 				{
-					posXW += stepXW;
 					x++;
 				}
 			}
 			else if (OptT::Flags & SWTRI_DepthTest)
 			{
-				while (zbufferLine[x] > posXW && x < xend)
+				while (zbufferLine[x] < depthvalues[x] && x < xend)
 				{
-					posXW += stepXW;
 					x++;
 				}
 			}

src/rendering/swrenderer/drawers/r_draw.cpp

@@ -220,7 +220,7 @@ namespace swrenderer
 	class DepthColumnCommand : public DrawerCommand
 	{
 	public:
-		DepthColumnCommand(const WallDrawerArgs &args, float idepth) : idepth(idepth)
+		DepthColumnCommand(const WallDrawerArgs &args, float idepth) : idepth(1.0f / idepth)
 		{
 			auto rendertarget = args.Viewport()->RenderTarget;
 			if (rendertarget->IsBgra())
@@ -244,7 +244,7 @@ namespace swrenderer
 			count = args.Count();
 		}
 
-		DepthColumnCommand(const SkyDrawerArgs &args, float idepth) : idepth(idepth)
+		DepthColumnCommand(const SkyDrawerArgs &args, float idepth) : idepth(1.0f / idepth)
 		{
 			auto rendertarget = args.Viewport()->RenderTarget;
 			if (rendertarget->IsBgra())
@@ -319,7 +319,7 @@ namespace swrenderer
 
 			if (idepth1 == idepth2)
 			{
-				float depth = idepth1;
+				float depth = 1.0f / idepth1;
 				#ifdef DEPTH_DEBUG
 				uint32_t gray = clamp<int32_t>((int32_t)(1.0f / depth / 4.0f), 0, 255);
 				uint32_t color = MAKEARGB(255, gray, gray, gray);
@@ -344,7 +344,7 @@ namespace swrenderer
 					dest[x] = color;
 					#endif
 
-					values[x] = depth;
+					values[x] = 1.0f / depth;
 					depth += step;
 				}
 			}