Do a similar thing for D3D (let the backend handle the dirtying).

GPU/Common/FramebufferCommon.cpp

@@ -547,8 +547,6 @@ void FramebufferManagerCommon::NotifyRenderFramebufferSwitched(VirtualFramebuffe
 			// This should only happen very briefly when toggling useBufferedRendering_.
 			ResizeFramebufFBO(vfb, vfb->width, vfb->height, true);
 		}
-		// Since we switched target, we need to redo viewportscissorstate.
-		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
 	} else {
 		if (vfb->fbo) {
 			// This should only happen very briefly when toggling useBufferedRendering_.
@@ -566,7 +564,7 @@ void FramebufferManagerCommon::NotifyRenderFramebufferSwitched(VirtualFramebuffe
 	}
 	textureCache_->NotifyFramebuffer(vfb->fb_address, vfb, NOTIFY_FB_UPDATED);
 
-	// ugly...
+	// ugly... is all this needed?
 	if (gstate_c.curRTWidth != vfb->width || gstate_c.curRTHeight != vfb->height) {
 		gstate_c.Dirty(DIRTY_PROJTHROUGHMATRIX | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_CULLRANGE);
 	}
@@ -1064,7 +1062,6 @@ void FramebufferManagerCommon::ResizeFramebufFBO(VirtualFramebuffer *vfb, int w,
 		INFO_LOG(FRAMEBUF, "Resizing FBO for %08x : %d x %d x %d", vfb->fb_address, w, h, vfb->format);
 		if (vfb->fbo) {
 			draw_->BindFramebufferAsRenderTarget(vfb->fbo, { Draw::RPAction::CLEAR, Draw::RPAction::CLEAR, Draw::RPAction::CLEAR }, "ResizeFramebufFBO");
-			gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
 			if (!skipCopy) {
 				BlitFramebuffer(vfb, 0, 0, &old, 0, 0, std::min((u16)oldWidth, std::min(vfb->bufferWidth, vfb->width)), std::min((u16)oldHeight, std::min(vfb->height, vfb->bufferHeight)), 0);
 			}
@@ -1075,7 +1072,6 @@ void FramebufferManagerCommon::ResizeFramebufFBO(VirtualFramebuffer *vfb, int w,
 		}
 	} else {
 		draw_->BindFramebufferAsRenderTarget(vfb->fbo, { Draw::RPAction::CLEAR, Draw::RPAction::CLEAR, Draw::RPAction::CLEAR }, "ResizeFramebufFBO");
-		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
 	}
 
 	if (!vfb->fbo) {
@@ -1845,7 +1841,7 @@ bool FramebufferManagerCommon::GetFramebuffer(u32 fb_address, int fb_stride, GEB
 	bool retval = draw_->CopyFramebufferToMemorySync(bound, Draw::FB_COLOR_BIT, 0, 0, w, h, Draw::DataFormat::R8G8B8A8_UNORM, buffer.GetData(), w, "GetFramebuffer");
 	gpuStats.numReadbacks++;
 	// After a readback we'll have flushed and started over, need to dirty a bunch of things to be safe.
-	gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_BLEND_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_RASTER_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
+	gstate_c.Dirty(DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
 	// We may have blitted to a temp FBO.
 	RebindFramebuffer();
 	return retval;
@@ -1880,7 +1876,7 @@ bool FramebufferManagerCommon::GetDepthbuffer(u32 fb_address, int fb_stride, u32
 	// No need to free on failure, that's the caller's job (it likely will reuse a buffer.)
 	bool retval = draw_->CopyFramebufferToMemorySync(vfb->fbo, Draw::FB_DEPTH_BIT, 0, 0, w, h, Draw::DataFormat::D32F, buffer.GetData(), w, "GetDepthBuffer");
 	// After a readback we'll have flushed and started over, need to dirty a bunch of things to be safe.
-	gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_BLEND_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_RASTER_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
+	gstate_c.Dirty(DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
 	// That may have unbound the framebuffer, rebind to avoid crashes when debugging.
 	RebindFramebuffer();
 	return retval;
@@ -1971,9 +1967,6 @@ void FramebufferManagerCommon::PackFramebufferSync_(VirtualFramebuffer *vfb, int
 		ERROR_LOG(G3D, "PackFramebufferSync_: Tried to readback to bad address %08x (stride = %d)", fb_address + dstByteOffset, vfb->fb_stride);
 	}
 
-	// A new command buffer will begin after CopyFrameBufferToMemorySync, so we need to trigger
-	// updates of any dynamic command buffer state by dirtying some stuff.
-	gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_BLEND_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_RASTER_STATE);
 	gpuStats.numReadbacks++;
 }
 
@@ -2056,7 +2049,6 @@ void FramebufferManagerCommon::RebindFramebuffer() {
 		// Should this even happen?  It could while debugging, but maybe we can just skip binding at all.
 		draw_->BindFramebufferAsRenderTarget(nullptr, { Draw::RPAction::KEEP, Draw::RPAction::KEEP, Draw::RPAction::KEEP }, "RebindFramebuffer_Bad");
 	}
-	gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
 }
 
 std::vector<FramebufferInfo> FramebufferManagerCommon::GetFramebufferList() {

GPU/D3D11/DrawEngineD3D11.cpp

@@ -307,6 +307,8 @@ void DrawEngineD3D11::BeginFrame() {
 		NOTICE_LOG(G3D, buffer);
 	}
 #endif
+
+	lastRenderStepId_ = -1;
 }
 
 VertexArrayInfoD3D11::~VertexArrayInfoD3D11() {
@@ -325,6 +327,14 @@ void DrawEngineD3D11::DoFlush() {
 	gpuStats.numFlushes++;
 	gpuStats.numTrackedVertexArrays = (int)vai_.size();
 
+	// In D3D, we're synchronous and state carries over so all we reset here on a new step is the viewport/scissor.
+	int curRenderStepId = draw_->GetCurrentStepId();
+	if (lastRenderStepId_ != curRenderStepId) {
+		// Dirty everything that has dynamic state that will need re-recording.
+		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
+		lastRenderStepId_ = curRenderStepId;
+	}
+
 	// This is not done on every drawcall, we collect vertex data
 	// until critical state changes. That's when we draw (flush).
 

GPU/D3D11/DrawEngineD3D11.h

@@ -216,4 +216,6 @@ private:
 
 	// Hardware tessellation
 	TessellationDataTransferD3D11 *tessDataTransferD3D11;
+
+	int lastRenderStepId_ = -1;
 };

GPU/Directx9/DrawEngineDX9.cpp

@@ -83,7 +83,7 @@ static const D3DVERTEXELEMENT9 TransformedVertexElements[] = {
 	D3DDECL_END()
 };
 
-DrawEngineDX9::DrawEngineDX9(Draw::DrawContext *draw) : vai_(256), vertexDeclMap_(64) {
+DrawEngineDX9::DrawEngineDX9(Draw::DrawContext *draw) : draw_(draw), vai_(256), vertexDeclMap_(64) {
 	device_ = (LPDIRECT3DDEVICE9)draw->GetNativeObject(Draw::NativeObject::DEVICE);
 	decOptions_.expandAllWeightsToFloat = true;
 	decOptions_.expand8BitNormalsToFloat = true;
@@ -299,14 +299,27 @@ static uint32_t SwapRB(uint32_t c) {
 	return (c & 0xFF00FF00) | ((c >> 16) & 0xFF) | ((c << 16) & 0xFF0000);
 }
 
+void DrawEngineDX9::BeginFrame() {
+	DecimateTrackedVertexArrays();
+
+	lastRenderStepId_ = -1;
+}
+
 // The inline wrapper in the header checks for numDrawCalls == 0
 void DrawEngineDX9::DoFlush() {
 	gpuStats.numFlushes++;
 	gpuStats.numTrackedVertexArrays = (int)vai_.size();
 
+	// In D3D, we're synchronous and state carries over so all we reset here on a new step is the viewport/scissor.
+	int curRenderStepId = draw_->GetCurrentStepId();
+	if (lastRenderStepId_ != curRenderStepId) {
+		// Dirty everything that has dynamic state that will need re-recording.
+		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE);
+		lastRenderStepId_ = curRenderStepId;
+	}
+
 	// This is not done on every drawcall, we should collect vertex data
 	// until critical state changes. That's when we draw (flush).
-
 	GEPrimitiveType prim = prevPrim_;
 	ApplyDrawState(prim);
 

GPU/Directx9/DrawEngineDX9.h

@@ -123,7 +123,8 @@ public:
 	void DestroyDeviceObjects();
 
 	void ClearTrackedVertexArrays() override;
-	void DecimateTrackedVertexArrays();
+
+	void BeginFrame();
 
 	// So that this can be inlined
 	void Flush() {
@@ -143,6 +144,7 @@ public:
 protected:
 	// Not currently supported.
 	bool UpdateUseHWTessellation(bool enable) override { return false; }
+	void DecimateTrackedVertexArrays();
 
 private:
 	void DoFlush();
@@ -156,6 +158,7 @@ private:
 	void MarkUnreliable(VertexArrayInfoDX9 *vai);
 
 	LPDIRECT3DDEVICE9 device_ = nullptr;
+	Draw::DrawContext *draw_;
 
 	PrehashMap<VertexArrayInfoDX9 *, nullptr> vai_;
 	DenseHashMap<u32, IDirect3DVertexDeclaration9 *, nullptr> vertexDeclMap_;
@@ -170,6 +173,8 @@ private:
 
 	// Hardware tessellation
 	TessellationDataTransferDX9 *tessDataTransferDX9;
+
+	int lastRenderStepId_ = -1;
 };
 
 }  // namespace

GPU/Directx9/GPU_DX9.cpp

@@ -282,7 +282,7 @@ void GPU_DX9::ReapplyGfxState() {
 
 void GPU_DX9::BeginFrame() {
 	textureCacheDX9_->StartFrame();
-	drawEngine_.DecimateTrackedVertexArrays();
+	drawEngine_.BeginFrame();
 	depalShaderCache_.Decimate();
 	// fragmentTestCache_.Decimate();
 

GPU/GLES/DrawEngineGLES.cpp

@@ -156,6 +156,8 @@ void DrawEngineGLES::ClearInputLayoutMap() {
 }
 
 void DrawEngineGLES::BeginFrame() {
+	DecimateTrackedVertexArrays();
+
 	FrameData &frameData = frameData_[render_->GetCurFrame()];
 	render_->BeginPushBuffer(frameData.pushIndex);
 	render_->BeginPushBuffer(frameData.pushVertex);

GPU/GLES/DrawEngineGLES.h

@@ -150,7 +150,6 @@ public:
 	void DeviceRestore(Draw::DrawContext *draw);
 
 	void ClearTrackedVertexArrays() override;
-	void DecimateTrackedVertexArrays();
 
 	void BeginFrame();
 	void EndFrame();
@@ -186,6 +185,7 @@ public:
 
 protected:
 	bool UpdateUseHWTessellation(bool enable) override;
+	void DecimateTrackedVertexArrays();
 
 private:
 	void InitDeviceObjects();

GPU/GLES/GPU_GLES.cpp

@@ -373,7 +373,6 @@ void GPU_GLES::ReapplyGfxState() {
 
 void GPU_GLES::BeginFrame() {
 	textureCacheGL_->StartFrame();
-	drawEngine_.DecimateTrackedVertexArrays();
 	depalShaderCache_.Decimate();
 	fragmentTestCache_.Decimate();
 

ext/native/thin3d/thin3d.h

@@ -673,6 +673,8 @@ public:
 	// Flush state like scissors etc so the caller can do its own custom drawing.
 	virtual void FlushState() {}
 
+	virtual int GetCurrentStepId() const = 0;
+
 protected:
 	ShaderModule *vsPresets_[VS_MAX_PRESET];
 	ShaderModule *fsPresets_[FS_MAX_PRESET];

ext/native/thin3d/thin3d_d3d11.cpp

@@ -158,6 +158,10 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override;
 
+	int GetCurrentStepId() const override {
+		return stepId_;
+	}
+
 private:
 	void ApplyCurrentState();
 
@@ -166,6 +170,7 @@ private:
 	ID3D11DeviceContext *context_;
 	ID3D11Device1 *device1_;
 	ID3D11DeviceContext1 *context1_;
+	int stepId_ = -1;
 
 	ID3D11Texture2D *bbRenderTargetTex_ = nullptr; // NOT OWNED
 	ID3D11RenderTargetView *bbRenderTargetView_ = nullptr;
@@ -353,6 +358,7 @@ void D3D11DrawContext::HandleEvent(Event ev, int width, int height, void *param1
 		// Make sure that we don't eliminate the next time the render target is set.
 		curRenderTargetView_ = nullptr;
 		curDepthStencilView_ = nullptr;
+		stepId_ = 0;
 		break;
 	}
 }
@@ -1394,11 +1400,13 @@ void D3D11DrawContext::CopyFramebufferImage(Framebuffer *srcfb, int level, int x
 		D3D11_BOX srcBox{ (UINT)x, (UINT)y, (UINT)z, (UINT)(x + width), (UINT)(y + height), (UINT)(z + depth) };
 		context_->CopySubresourceRegion(dstTex, dstLevel, dstX, dstY, dstZ, srcTex, level, &srcBox);
 	}
+	stepId_++;
 }
 
 bool D3D11DrawContext::BlitFramebuffer(Framebuffer *srcfb, int srcX1, int srcY1, int srcX2, int srcY2, Framebuffer *dstfb, int dstX1, int dstY1, int dstX2, int dstY2, int channelBits, FBBlitFilter filter, const char *tag) {
 	// Unfortunately D3D11 has no equivalent to this, gotta render a quad. Well, in some cases we can issue a copy instead.
 	Crash();
+	stepId_++;
 	return false;
 }
 
@@ -1519,6 +1527,7 @@ bool D3D11DrawContext::CopyFramebufferToMemorySync(Framebuffer *src, int channel
 	if (!useGlobalPacktex) {
 		packTex->Release();
 	}
+	stepId_++;
 	return true;
 }
 
@@ -1562,6 +1571,8 @@ void D3D11DrawContext::BindFramebufferAsRenderTarget(Framebuffer *fbo, const Ren
 	if (mask && curDepthStencilView_) {
 		context_->ClearDepthStencilView(curDepthStencilView_, mask, rp.clearDepth, rp.clearStencil);
 	}
+
+	stepId_++;
 }
 
 // color must be 0, for now.

ext/native/thin3d/thin3d_d3d9.cpp

@@ -599,11 +599,16 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override;
 
+	int GetCurrentStepId() const override {
+		return stepId_;
+	}
+
 private:
 	LPDIRECT3D9 d3d_;
 	LPDIRECT3D9EX d3dEx_;
 	LPDIRECT3DDEVICE9 device_;
 	LPDIRECT3DDEVICE9EX deviceEx_;
+	int stepId_ = -1;
 	int adapterId_ = -1;
 	D3DADAPTER_IDENTIFIER9 identifier_{};
 	D3DCAPS9 d3dCaps_;
@@ -1160,6 +1165,7 @@ void D3D9Context::BindFramebufferAsRenderTarget(Framebuffer *fbo, const RenderPa
 
 	dxstate.scissorRect.restore();
 	dxstate.viewport.restore();
+	stepId_++;
 }
 
 uintptr_t D3D9Context::GetFramebufferAPITexture(Framebuffer *fbo, int channelBits, int attachment) {
@@ -1187,10 +1193,6 @@ uintptr_t D3D9Context::GetFramebufferAPITexture(Framebuffer *fbo, int channelBit
 	}
 }
 
-LPDIRECT3DSURFACE9 fbo_get_color_for_read(D3D9Framebuffer *fbo) {
-	return fbo->surf;
-}
-
 void D3D9Context::BindFramebufferAsTexture(Framebuffer *fbo, int binding, FBChannel channelBit, int color) {
 	D3D9Framebuffer *fb = (D3D9Framebuffer *)fbo;
 	switch (channelBit) {
@@ -1228,6 +1230,7 @@ bool D3D9Context::BlitFramebuffer(Framebuffer *srcfb, int srcX1, int srcY1, int
 	RECT dstRect{ (LONG)dstX1, (LONG)dstY1, (LONG)dstX2, (LONG)dstY2 };
 	LPDIRECT3DSURFACE9 srcSurf = src ? src->surf : deviceRTsurf;
 	LPDIRECT3DSURFACE9 dstSurf = dst ? dst->surf : deviceRTsurf;
+	stepId_++;
 	return SUCCEEDED(device_->StretchRect(srcSurf, &srcRect, dstSurf, &dstRect, filter == FB_BLIT_LINEAR ? D3DTEXF_LINEAR : D3DTEXF_POINT));
 }
 
@@ -1245,6 +1248,9 @@ void D3D9Context::HandleEvent(Event ev, int width, int height, void *param1, voi
 		device_->GetRenderTarget(0, &deviceRTsurf);
 		device_->GetDepthStencilSurface(&deviceDSsurf);
 		break;
+	case Event::PRESENTED:
+		stepId_ = 0;
+		break;
 	}
 }
 

ext/native/thin3d/thin3d_gl.cpp

@@ -478,6 +478,10 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override {}
 
+	int GetCurrentStepId() const {
+		return renderManager_.GetCurrentStepId();
+	}
+
 private:
 	void ApplySamplers();
 

ext/native/thin3d/thin3d_vulkan.cpp

@@ -482,6 +482,10 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override;
 
+	int GetCurrentStepId() const {
+		return renderManager_.GetCurrentStepId();
+	}
+
 private:
 	VulkanTexture *GetNullTexture();
 	VulkanContext *vulkan_ = nullptr;