- change the software renderer to render to a DSimpleCanvas like the old D3D9 target did. Then use the drawer threads to memcpy the result to the PBO

src/gl/textures/gl_hwtexture.cpp

@@ -296,10 +296,7 @@ void FHardwareTexture::AllocateBuffer(int w, int h, int texelsize)
 	{
 		glGenBuffers(1, &glBufferID);
 		glBindBuffer(GL_PIXEL_UNPACK_BUFFER, glBufferID);
-		if (screen->hwcaps & RFL_BUFFER_STORAGE)
-			glBufferStorage(GL_PIXEL_UNPACK_BUFFER, w*h*texelsize, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_CLIENT_STORAGE_BIT);
-		else
-			glBufferData(GL_PIXEL_UNPACK_BUFFER, w*h*texelsize, nullptr, GL_STREAM_DRAW);
+		glBufferData(GL_PIXEL_UNPACK_BUFFER, w*h*texelsize, nullptr, GL_STREAM_DRAW);
 		glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 	}
 }
@@ -308,16 +305,7 @@ void FHardwareTexture::AllocateBuffer(int w, int h, int texelsize)
 uint8_t *FHardwareTexture::MapBuffer()
 {
 	glBindBuffer(GL_PIXEL_UNPACK_BUFFER, glBufferID);
-	if (screen->hwcaps & RFL_BUFFER_STORAGE)
-	{
-		GLint size = -1;
-		glGetBufferParameteriv(GL_PIXEL_UNPACK_BUFFER, GL_BUFFER_SIZE, &size);
-		return (uint8_t*)glMapBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, size, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT);
-	}
-	else
-	{
-		return (uint8_t*)glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_WRITE);
-	}
+	return (uint8_t*)glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_WRITE_ONLY);
 }
 
 //===========================================================================

src/polyrenderer/drawers/poly_triangle.cpp

@@ -169,6 +169,7 @@ void PolyTriangleThreadData::DrawElements(const PolyDrawArgs &drawargs, const vo
 	args.uniforms = &drawargs;
 	args.destBgra = dest_bgra;
 	args.stencilbuffer = PolyStencilBuffer::Instance()->Values();
+	args.stencilpitch = PolyStencilBuffer::Instance()->Width();
 	args.zbuffer = PolyZBuffer::Instance()->Values();
 	args.depthOffset = weaponScene ? 1.0f : 0.0f;
 
@@ -222,6 +223,7 @@ void PolyTriangleThreadData::DrawArray(const PolyDrawArgs &drawargs, const void
 	args.uniforms = &drawargs;
 	args.destBgra = dest_bgra;
 	args.stencilbuffer = PolyStencilBuffer::Instance()->Values();
+	args.stencilpitch = PolyStencilBuffer::Instance()->Width();
 	args.zbuffer = PolyZBuffer::Instance()->Values();
 	args.depthOffset = weaponScene ? 1.0f : 0.0f;
 

src/polyrenderer/drawers/screen_triangle.cpp

@@ -153,8 +153,8 @@ void ScreenTriangle::Draw(const TriDrawTriangleArgs *args, PolyTriangleThreadDat
 		int x = leftEdge[y];
 		int xend = rightEdge[y];
 
-		float *zbufferLine = args->zbuffer + args->pitch * y;
-		uint8_t *stencilLine = args->stencilbuffer + args->pitch * y;
+		float *zbufferLine = args->zbuffer + args->stencilpitch * y;
+		uint8_t *stencilLine = args->stencilbuffer + args->stencilpitch * y;
 
 		float startX = x + (0.5f - v1X);
 		float startY = y + (0.5f - v1Y);

src/polyrenderer/drawers/screen_triangle.h

@@ -53,6 +53,7 @@ struct TriDrawTriangleArgs
 	int32_t clipright;
 	int32_t clipbottom;
 	uint8_t *stencilbuffer;
+	int stencilpitch;
 	float *zbuffer;
 	const PolyDrawArgs *uniforms;
 	bool destBgra;

src/polyrenderer/poly_renderer.cpp

@@ -58,7 +58,7 @@ PolyRenderer::PolyRenderer()
 {
 }
 
-void PolyRenderer::RenderView(player_t *player, DCanvas *target)
+void PolyRenderer::RenderView(player_t *player, DCanvas *target, void *videobuffer)
 {
 	using namespace swrenderer;
 	
@@ -70,6 +70,11 @@ void PolyRenderer::RenderView(player_t *player, DCanvas *target)
 	RenderActorView(player->mo, false);
 
 	Threads.MainThread()->FlushDrawQueue();
+
+	auto copyqueue = std::make_shared<DrawerCommandQueue>(Threads.MainThread()->FrameMemory.get());
+	copyqueue->Push<MemcpyCommand>(videobuffer, target->GetPixels(), target->GetWidth(), target->GetHeight(), target->GetPitch(), target->IsBgra() ? 4 : 1);
+	DrawerThreads::Execute(copyqueue);
+
 	PolyDrawerWaitCycles.Clock();
 	DrawerThreads::WaitForWorkers();
 	PolyDrawerWaitCycles.Unclock();

src/polyrenderer/poly_renderer.h

@@ -49,7 +49,7 @@ class PolyRenderer
 public:
 	PolyRenderer();
 	
-	void RenderView(player_t *player, DCanvas *target);
+	void RenderView(player_t *player, DCanvas *target, void *videobuffer);
 	void RenderViewToCanvas(AActor *actor, DCanvas *canvas, int x, int y, int width, int height, bool dontmaplines);
 	void RenderRemainingPlayerSprites();
 

src/r_renderer.h

@@ -23,7 +23,7 @@ struct FRenderer
 	virtual void Precache(uint8_t *texhitlist, TMap<PClassActor*, bool> &actorhitlist) = 0;
 
 	// render 3D view
-	virtual void RenderView(player_t *player, DCanvas *target) = 0;
+	virtual void RenderView(player_t *player, DCanvas *target, void *videobuffer) = 0;
 
 	// renders view to a savegame picture
 	virtual void WriteSavePic(player_t *player, FileWriter *file, int width, int height) = 0;

src/swrenderer/drawers/r_thread.cpp

@@ -223,3 +223,27 @@ void GroupMemoryBarrierCommand::Execute(DrawerThread *thread)
 	condition.notify_all();
 	condition.wait(lock, [&]() { return count >= (size_t)thread->num_cores; });
 }
+
+/////////////////////////////////////////////////////////////////////////////
+
+MemcpyCommand::MemcpyCommand(void *dest, const void *src, int width, int height, int srcpitch, int pixelsize)
+	: dest(dest), src(src), width(width), height(height), srcpitch(srcpitch), pixelsize(pixelsize)
+{
+}
+
+void MemcpyCommand::Execute(DrawerThread *thread)
+{
+	int start = thread->skipped_by_thread(0);
+	int count = thread->count_for_thread(0, height);
+	int sstep = thread->num_cores * srcpitch * pixelsize;
+	int dstep = thread->num_cores * width * pixelsize;
+	int size = width * pixelsize;
+	uint8_t *d = (uint8_t*)dest + start * width * pixelsize;
+	const uint8_t *s = (const uint8_t*)src + start * srcpitch * pixelsize;
+	for (int i = 0; i < count; i++)
+	{
+		memcpy(d, s, size);
+		d += dstep;
+		s += sstep;
+	}
+}

src/swrenderer/drawers/r_thread.h

@@ -109,6 +109,22 @@ private:
 	size_t count = 0;
 };
 
+// Copy finished rows to video memory
+class MemcpyCommand : public DrawerCommand
+{
+public:
+	MemcpyCommand(void *dest, const void *src, int width, int height, int srcpitch, int pixelsize);
+	void Execute(DrawerThread *thread);
+
+private:
+	void *dest;
+	const void *src;
+	int width;
+	int height;
+	int srcpitch;
+	int pixelsize;
+};
+
 class DrawerCommandQueue;
 typedef std::shared_ptr<DrawerCommandQueue> DrawerCommandQueuePtr;
 

src/swrenderer/r_swrenderer.cpp

@@ -157,13 +157,13 @@ void FSoftwareRenderer::Precache(uint8_t *texhitlist, TMap<PClassActor*, bool> &
 	}
 }
 
-void FSoftwareRenderer::RenderView(player_t *player, DCanvas *target)
+void FSoftwareRenderer::RenderView(player_t *player, DCanvas *target, void *videobuffer)
 {
 	if (V_IsPolyRenderer())
 	{
 		PolyRenderer::Instance()->Viewpoint = r_viewpoint;
 		PolyRenderer::Instance()->Viewwindow = r_viewwindow;
-		PolyRenderer::Instance()->RenderView(player, target);
+		PolyRenderer::Instance()->RenderView(player, target, videobuffer);
 		r_viewpoint = PolyRenderer::Instance()->Viewpoint;
 		r_viewwindow = PolyRenderer::Instance()->Viewwindow;
 	}
@@ -171,7 +171,7 @@ void FSoftwareRenderer::RenderView(player_t *player, DCanvas *target)
 	{
 		mScene.MainThread()->Viewport->viewpoint = r_viewpoint;
 		mScene.MainThread()->Viewport->viewwindow = r_viewwindow;
-		mScene.RenderView(player, target);
+		mScene.RenderView(player, target, videobuffer);
 		r_viewpoint = mScene.MainThread()->Viewport->viewpoint;
 		r_viewwindow = mScene.MainThread()->Viewport->viewwindow;
 	}

src/swrenderer/r_swrenderer.h

@@ -12,7 +12,7 @@ struct FSoftwareRenderer : public FRenderer
 	void Precache(uint8_t *texhitlist, TMap<PClassActor*, bool> &actorhitlist) override;
 
 	// render 3D view
-	void RenderView(player_t *player, DCanvas *target) override;
+	void RenderView(player_t *player, DCanvas *target, void *videobuffer) override;
 
 	// renders view to a savegame picture
 	void WriteSavePic (player_t *player, FileWriter *file, int width, int height) override;

src/swrenderer/r_swscene.cpp

@@ -99,7 +99,6 @@ sector_t *SWSceneDrawer::RenderView(player_t *player)
 	FBTextureIndex = (FBTextureIndex + 1) % 2;
 	auto &fbtex = FBTexture[FBTextureIndex];
 
-	DCanvas buffer(screen->GetWidth(), screen->GetHeight(), V_IsTrueColor());
 	if (fbtex == nullptr || fbtex->SystemTexture[0] == nullptr || 
 		fbtex->GetWidth() != screen->GetWidth() || 
 		fbtex->GetHeight() != screen->GetHeight() || 
@@ -111,12 +110,14 @@ sector_t *SWSceneDrawer::RenderView(player_t *player)
 		fbtex->SystemTexture[0]->AllocateBuffer(screen->GetWidth(), screen->GetHeight(), V_IsTrueColor() ? 4 : 1);
 		auto mat = FMaterial::ValidateTexture(fbtex.get(), false);
 		mat->AddTextureLayer(PaletteTexture.get());
+
+		Canvas.reset();
+		Canvas.reset(new DSimpleCanvas(screen->GetWidth(), screen->GetHeight(), V_IsTrueColor()));
 	}
 
 	auto buf = fbtex->SystemTexture[0]->MapBuffer();
 	if (!buf) I_FatalError("Unable to map buffer for software rendering");
-	buffer.SetBuffer(screen->GetWidth(), screen->GetHeight(), screen->GetWidth(), buf);
-	SWRenderer->RenderView(player, &buffer);
+	SWRenderer->RenderView(player, Canvas.get(), buf);
 	fbtex->SystemTexture[0]->CreateTexture(nullptr, screen->GetWidth(), screen->GetHeight(), 0, false, 0, "swbuffer");
 
 	auto map = swrenderer::CameraLight::Instance()->ShaderColormap();

src/swrenderer/r_swscene.h

@@ -16,6 +16,7 @@ class SWSceneDrawer
 	std::unique_ptr<FSWSceneTexture> FBTexture[2];
 	int FBTextureIndex = 0;
 	bool FBIsTruecolor = false;
+	std::unique_ptr<DSimpleCanvas> Canvas;
 
 public:
 	SWSceneDrawer();

src/swrenderer/scene/r_scene.cpp

@@ -90,7 +90,7 @@ namespace swrenderer
 		clearcolor = color;
 	}
 
-	void RenderScene::RenderView(player_t *player, DCanvas *target)
+	void RenderScene::RenderView(player_t *player, DCanvas *target, void *videobuffer)
 	{
 		auto viewport = MainThread()->Viewport.get();
 		viewport->RenderTarget = target;
@@ -131,6 +131,10 @@ namespace swrenderer
 
 		RenderActorView(player->mo);
 
+		auto copyqueue = std::make_shared<DrawerCommandQueue>(MainThread()->FrameMemory.get());
+		copyqueue->Push<MemcpyCommand>(videobuffer, target->GetPixels(), target->GetWidth(), target->GetHeight(), target->GetPitch(), target->IsBgra() ? 4 : 1);
+		DrawerThreads::Execute(copyqueue);
+
 		DrawerWaitCycles.Clock();
 		DrawerThreads::WaitForWorkers();
 		DrawerWaitCycles.Unclock();

src/swrenderer/scene/r_scene.h

@@ -48,7 +48,7 @@ namespace swrenderer
 
 		void SetClearColor(int color);
 		
-		void RenderView(player_t *player, DCanvas *target);
+		void RenderView(player_t *player, DCanvas *target, void *videobuffer);
 		void RenderViewToCanvas(AActor *actor, DCanvas *canvas, int x, int y, int width, int height, bool dontmaplines = false);
 	
 		bool DontMapLines() const { return dontmaplines; }