Revert "- removed the 8x8 block drawing code from softpoly"

This reverts commit d086c3d4dc.

# Conflicts:
#	src/polyrenderer/drawers/poly_triangle.cpp
#	src/polyrenderer/drawers/poly_triangle.h

The softpoly renderer crashed at high resolutions without them.

src/polyrenderer/drawers/poly_buffer.cpp

@@ -48,7 +48,8 @@ void PolyZBuffer::Resize(int newwidth, int newheight)
 {
 	width = newwidth;
 	height = newheight;
-	values.resize(width * height);
+	int count = BlockWidth() * BlockHeight();
+	values.resize(count * 64);
 }
 
 /////////////////////////////////////////////////////////////////////////////
@@ -63,6 +64,14 @@ void PolyStencilBuffer::Clear(int newwidth, int newheight, uint8_t stencil_value
 {
 	width = newwidth;
 	height = newheight;
-	values.resize(width * height);
+	int count = BlockWidth() * BlockHeight();
-	memset(Values(), stencil_value, width * height);
+	values.resize(count * 64);
+	masks.resize(count);
+
+	uint8_t *v = Values();
+	uint32_t *m = Masks();
+	for (int i = 0; i < count; i++)
+	{
+		m[i] = 0xffffff00 | stencil_value;
+	}
 }

src/polyrenderer/drawers/poly_buffer.h

@@ -33,6 +33,8 @@ public:
 	void Resize(int newwidth, int newheight);
 	int Width() const { return width; }
 	int Height() const { return height; }
+	int BlockWidth() const { return (width + 7) / 8; }
+	int BlockHeight() const { return (height + 7) / 8; }
 	float *Values() { return values.data(); }
 
 private:
@@ -48,10 +50,16 @@ public:
 	void Clear(int newwidth, int newheight, uint8_t stencil_value = 0);
 	int Width() const { return width; }
 	int Height() const { return height; }
+	int BlockWidth() const { return (width + 7) / 8; }
+	int BlockHeight() const { return (height + 7) / 8; }
 	uint8_t *Values() { return values.data(); }
+	uint32_t *Masks() { return masks.data(); }
 
 private:
 	int width;
 	int height;
+
+	// 8x8 blocks of stencil values, plus a mask for each block indicating if values are the same for early out stencil testing
 	std::vector<uint8_t> values;
+	std::vector<uint32_t> masks;
 };

src/polyrenderer/drawers/poly_triangle.cpp

@@ -53,7 +53,7 @@ bool PolyTriangleDrawer::IsBgra()
 	return isBgraRenderTarget;
 }
 
-void PolyTriangleDrawer::SetViewport(const DrawerCommandQueuePtr &queue, int x, int y, int width, int height, DCanvas *canvas)
+void PolyTriangleDrawer::SetViewport(const DrawerCommandQueuePtr &queue, int x, int y, int width, int height, DCanvas *canvas, bool span_drawers)
 {
 	uint8_t *dest = (uint8_t*)canvas->GetBuffer();
 	int dest_width = canvas->GetWidth();
@@ -75,7 +75,7 @@ void PolyTriangleDrawer::SetViewport(const DrawerCommandQueuePtr &queue, int x,
 	dest_width = clamp(viewport_x + viewport_width, 0, dest_width - offsetx);
 	dest_height = clamp(viewport_y + viewport_height, 0, dest_height - offsety);
 
-	queue->Push<PolySetViewportCommand>(viewport_x, viewport_y, viewport_width, viewport_height, dest, dest_width, dest_height, dest_pitch, dest_bgra);
+	queue->Push<PolySetViewportCommand>(viewport_x, viewport_y, viewport_width, viewport_height, dest, dest_width, dest_height, dest_pitch, dest_bgra, span_drawers);
 }
 
 void PolyTriangleDrawer::SetTransform(const DrawerCommandQueuePtr &queue, const Mat4f *objectToClip, const Mat4f *objectToWorld)
@@ -115,7 +115,7 @@ void PolyTriangleDrawer::DrawElements(const DrawerCommandQueuePtr &queue, const
 
 /////////////////////////////////////////////////////////////////////////////
 
-void PolyTriangleThreadData::SetViewport(int x, int y, int width, int height, uint8_t *new_dest, int new_dest_width, int new_dest_height, int new_dest_pitch, bool new_dest_bgra)
+void PolyTriangleThreadData::SetViewport(int x, int y, int width, int height, uint8_t *new_dest, int new_dest_width, int new_dest_height, int new_dest_pitch, bool new_dest_bgra, bool new_span_drawers)
 {
 	viewport_x = x;
 	viewport_y = y;
@@ -126,6 +126,7 @@ void PolyTriangleThreadData::SetViewport(int x, int y, int width, int height, ui
 	dest_height = new_dest_height;
 	dest_pitch = new_dest_pitch;
 	dest_bgra = new_dest_bgra;
+	span_drawers = new_span_drawers;
 	ccw = true;
 	weaponScene = false;
 }
@@ -148,7 +149,9 @@ void PolyTriangleThreadData::DrawElements(const PolyDrawArgs &drawargs, const vo
 	args.clipbottom = dest_height;
 	args.uniforms = &drawargs;
 	args.destBgra = dest_bgra;
-	args.stencilbuffer = PolyStencilBuffer::Instance()->Values();
+	args.stencilPitch = PolyStencilBuffer::Instance()->BlockWidth();
+	args.stencilValues = PolyStencilBuffer::Instance()->Values();
+	args.stencilMasks = PolyStencilBuffer::Instance()->Masks();
 	args.zbuffer = PolyZBuffer::Instance()->Values();
 	args.depthOffset = weaponScene ? 1.0f : 0.0f;
 
@@ -201,7 +204,9 @@ void PolyTriangleThreadData::DrawArray(const PolyDrawArgs &drawargs, const void
 	args.clipbottom = dest_height;
 	args.uniforms = &drawargs;
 	args.destBgra = dest_bgra;
-	args.stencilbuffer = PolyStencilBuffer::Instance()->Values();
+	args.stencilPitch = PolyStencilBuffer::Instance()->BlockWidth();
+	args.stencilValues = PolyStencilBuffer::Instance()->Values();
+	args.stencilMasks = PolyStencilBuffer::Instance()->Masks();
 	args.zbuffer = PolyZBuffer::Instance()->Values();
 	args.depthOffset = weaponScene ? 1.0f : 0.0f;
 
@@ -424,7 +429,10 @@ void PolyTriangleThreadData::DrawShadedTriangle(const ShadedTriVertex *vert, boo
 			args->v3 = &clippedvert[i - 2];
 			if (IsFrontfacing(args) == ccw && args->CalculateGradients())
 			{
-				ScreenTriangle::Draw(args, this);
+				if (!span_drawers)
+					ScreenTriangle::Draw(args, this);
+				else
+					ScreenTriangle::DrawSWRender(args, this);
 			}
 		}
 	}
@@ -437,7 +445,10 @@ void PolyTriangleThreadData::DrawShadedTriangle(const ShadedTriVertex *vert, boo
 			args->v3 = &clippedvert[i];
 			if (IsFrontfacing(args) != ccw && args->CalculateGradients())
 			{
-				ScreenTriangle::Draw(args, this);
+				if (!span_drawers)
+					ScreenTriangle::Draw(args, this);
+				else
+					ScreenTriangle::DrawSWRender(args, this);
 			}
 		}
 	}
@@ -681,14 +692,14 @@ void PolySetModelVertexShaderCommand::Execute(DrawerThread *thread)
 
 /////////////////////////////////////////////////////////////////////////////
 
-PolySetViewportCommand::PolySetViewportCommand(int x, int y, int width, int height, uint8_t *dest, int dest_width, int dest_height, int dest_pitch, bool dest_bgra)
+PolySetViewportCommand::PolySetViewportCommand(int x, int y, int width, int height, uint8_t *dest, int dest_width, int dest_height, int dest_pitch, bool dest_bgra, bool span_drawers)
-	: x(x), y(y), width(width), height(height), dest(dest), dest_width(dest_width), dest_height(dest_height), dest_pitch(dest_pitch), dest_bgra(dest_bgra)
+	: x(x), y(y), width(width), height(height), dest(dest), dest_width(dest_width), dest_height(dest_height), dest_pitch(dest_pitch), dest_bgra(dest_bgra), span_drawers(span_drawers)
 {
 }
 
 void PolySetViewportCommand::Execute(DrawerThread *thread)
 {
-	PolyTriangleThreadData::Get(thread)->SetViewport(x, y, width, height, dest, dest_width, dest_height, dest_pitch, dest_bgra);
+	PolyTriangleThreadData::Get(thread)->SetViewport(x, y, width, height, dest, dest_width, dest_height, dest_pitch, dest_bgra, span_drawers);
 }
 
 /////////////////////////////////////////////////////////////////////////////

src/polyrenderer/drawers/poly_triangle.h

@@ -33,7 +33,7 @@ class PolyTriangleDrawer
 {
 public:
 	static void ClearBuffers(DCanvas *canvas);
-	static void SetViewport(const DrawerCommandQueuePtr &queue, int x, int y, int width, int height, DCanvas *canvas);
+	static void SetViewport(const DrawerCommandQueuePtr &queue, int x, int y, int width, int height, DCanvas *canvas, bool span_drawers);
 	static void SetCullCCW(const DrawerCommandQueuePtr &queue, bool ccw);
 	static void SetTwoSided(const DrawerCommandQueuePtr &queue, bool twosided);
 	static void SetWeaponScene(const DrawerCommandQueuePtr &queue, bool enable);
@@ -49,7 +49,7 @@ class PolyTriangleThreadData
 public:
 	PolyTriangleThreadData(int32_t core, int32_t num_cores) : core(core), num_cores(num_cores) { }
 
-	void SetViewport(int x, int y, int width, int height, uint8_t *dest, int dest_width, int dest_height, int dest_pitch, bool dest_bgra);
+	void SetViewport(int x, int y, int width, int height, uint8_t *dest, int dest_width, int dest_height, int dest_pitch, bool dest_bgra, bool span_drawers);
 	void SetTransform(const Mat4f *objectToClip, const Mat4f *objectToWorld);
 	void SetCullCCW(bool value) { ccw = value; }
 	void SetTwoSided(bool value) { twosided = value; }
@@ -95,6 +95,7 @@ private:
 	int modelFrame1 = -1;
 	int modelFrame2 = -1;
 	float modelInterpolationFactor = 0.0f;
+	bool span_drawers = false;
 
 	enum { max_additional_vertices = 16 };
 };
@@ -160,7 +161,7 @@ private:
 class PolySetViewportCommand : public DrawerCommand
 {
 public:
-	PolySetViewportCommand(int x, int y, int width, int height, uint8_t *dest, int dest_width, int dest_height, int dest_pitch, bool dest_bgra);
+	PolySetViewportCommand(int x, int y, int width, int height, uint8_t *dest, int dest_width, int dest_height, int dest_pitch, bool dest_bgra, bool span_drawers);
 
 	void Execute(DrawerThread *thread) override;
 
@@ -174,6 +175,7 @@ private:
 	int dest_height;
 	int dest_pitch;
 	bool dest_bgra;
+	bool span_drawers;
 };
 
 class DrawPolyTrianglesCommand : public DrawerCommand

src/polyrenderer/drawers/screen_triangle.h

@@ -52,7 +52,9 @@ struct TriDrawTriangleArgs
 	ShadedTriVertex *v3;
 	int32_t clipright;
 	int32_t clipbottom;
-	uint8_t *stencilbuffer;
+	uint8_t *stencilValues;
+	uint32_t *stencilMasks;
+	int32_t stencilPitch;
 	float *zbuffer;
 	const PolyDrawArgs *uniforms;
 	bool destBgra;
@@ -140,6 +142,7 @@ class ScreenTriangle
 {
 public:
 	static void Draw(const TriDrawTriangleArgs *args, PolyTriangleThreadData *thread);
+	static void DrawSWRender(const TriDrawTriangleArgs *args, PolyTriangleThreadData *thread);
 
 	static void(*SpanDrawers8[])(int y, int x0, int x1, const TriDrawTriangleArgs *args);
 	static void(*SpanDrawers32[])(int y, int x0, int x1, const TriDrawTriangleArgs *args);

src/polyrenderer/poly_renderer.cpp

@@ -210,11 +210,11 @@ void PolyRenderer::SetSceneViewport()
 			height = (screenblocks*SCREENHEIGHT / 10) & ~7;
 
 		int bottom = SCREENHEIGHT - (height + viewwindowy - ((height - viewheight) / 2));
-		PolyTriangleDrawer::SetViewport(Threads.MainThread()->DrawQueue, viewwindowx, SCREENHEIGHT - bottom - height, viewwidth, height, RenderTarget);
+		PolyTriangleDrawer::SetViewport(Threads.MainThread()->DrawQueue, viewwindowx, SCREENHEIGHT - bottom - height, viewwidth, height, RenderTarget, false);
 	}
 	else // Rendering to camera texture
 	{
-		PolyTriangleDrawer::SetViewport(Threads.MainThread()->DrawQueue, 0, 0, RenderTarget->GetWidth(), RenderTarget->GetHeight(), RenderTarget);
+		PolyTriangleDrawer::SetViewport(Threads.MainThread()->DrawQueue, 0, 0, RenderTarget->GetWidth(), RenderTarget->GetHeight(), RenderTarget, false);
 	}
 }
 

src/swrenderer/drawers/r_draw.cpp

@@ -270,7 +270,7 @@ namespace swrenderer
 		void Execute(DrawerThread *thread) override
 		{
 			auto zbuffer = PolyZBuffer::Instance();
-			int pitch = PolyStencilBuffer::Instance()->Width();
+			int pitch = PolyStencilBuffer::Instance()->BlockWidth() * 8;
 			float *values = zbuffer->Values() + y * pitch + x;
 			int cnt = count;
 
@@ -312,7 +312,7 @@ namespace swrenderer
 				return;
 
 			auto zbuffer = PolyZBuffer::Instance();
-			int pitch = PolyStencilBuffer::Instance()->Width();
+			int pitch = PolyStencilBuffer::Instance()->BlockWidth() * 8;
 			float *values = zbuffer->Values() + y * pitch;
 			int end = x2;
 

src/swrenderer/scene/r_scene.cpp

@@ -285,7 +285,7 @@ namespace swrenderer
 		thread->OpaquePass->ResetFakingUnderwater(); // [RH] Hack to make windows into underwater areas possible
 		thread->Portal->SetMainPortal();
 
-		PolyTriangleDrawer::SetViewport(thread->DrawQueue, viewwindowx, viewwindowy, viewwidth, viewheight, thread->Viewport->RenderTarget);
+		PolyTriangleDrawer::SetViewport(thread->DrawQueue, viewwindowx, viewwindowy, viewwidth, viewheight, thread->Viewport->RenderTarget, true);
 
 		// Cull things outside the range seen by this thread
 		VisibleSegmentRenderer visitor;