Replace the "GetCurrentStepId"-based state invalidation with callbacks

CMakeLists.txt

@@ -606,6 +606,7 @@ add_library(Common STATIC
 	Common/File/FileDescriptor.cpp
 	Common/File/FileDescriptor.h
 	Common/GPU/DataFormat.h
+	Common/GPU/MiscTypes.h
 	Common/GPU/thin3d.cpp
 	Common/GPU/thin3d.h
 	Common/GPU/thin3d_create.h

Common/Common.vcxproj

@@ -432,6 +432,7 @@
     <ClInclude Include="GPU\D3D9\D3D9ShaderCompiler.h" />
     <ClInclude Include="GPU\D3D9\D3D9StateCache.h" />
     <ClInclude Include="GPU\DataFormat.h" />
+    <ClInclude Include="GPU\MiscTypes.h" />
     <ClInclude Include="GPU\OpenGL\DataFormatGL.h" />
     <ClInclude Include="GPU\OpenGL\gl3stub.h" />
     <ClInclude Include="GPU\OpenGL\GLFeatures.h" />

Common/Common.vcxproj.filters

@@ -452,6 +452,9 @@
     <ClInclude Include="Render\ManagedTexture.h">
       <Filter>Render</Filter>
     </ClInclude>
+    <ClInclude Include="GPU\MiscTypes.h">
+      <Filter>GPU</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="ABI.cpp" />

Common/GPU/D3D11/thin3d_d3d11.cpp

@@ -163,8 +163,8 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override;
 
-	int GetCurrentStepId() const override {
-		return stepId_;
+	void SetInvalidationCallback(InvalidationCallback callback) override {
+		invalidationCallback_ = callback;
 	}
 
 private:
@@ -177,7 +177,6 @@ private:
 	ID3D11DeviceContext *context_;
 	ID3D11Device1 *device1_;
 	ID3D11DeviceContext1 *context1_;
-	int stepId_ = -1;
 
 	ID3D11Texture2D *bbRenderTargetTex_ = nullptr; // NOT OWNED
 	ID3D11RenderTargetView *bbRenderTargetView_ = nullptr;
@@ -216,6 +215,8 @@ private:
 	ID3D11Buffer *nextIndexBuffer_ = nullptr;
 	int nextIndexBufferOffset_ = 0;
 
+	InvalidationCallback invalidationCallback_;
+
 	// Dynamic state
 	float blendFactor_[4]{};
 	bool blendFactorDirty_ = false;
@@ -399,7 +400,6 @@ 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;
 	}
 }
@@ -1500,13 +1500,11 @@ 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;
 }
 
@@ -1652,7 +1650,6 @@ bool D3D11DrawContext::CopyFramebufferToMemorySync(Framebuffer *src, int channel
 	if (!useGlobalPacktex) {
 		packTex->Release();
 	}
-	stepId_++;
 	return true;
 }
 
@@ -1699,7 +1696,9 @@ void D3D11DrawContext::BindFramebufferAsRenderTarget(Framebuffer *fbo, const Ren
 		context_->ClearDepthStencilView(curDepthStencilView_, mask, rp.clearDepth, rp.clearStencil);
 	}
 
-	stepId_++;
+	if (invalidationCallback_) {
+		invalidationCallback_(InvalidationFlags::RENDER_PASS_STATE);
+	}
 }
 
 void D3D11DrawContext::BindFramebufferAsTexture(Framebuffer *fbo, int binding, FBChannel channelBit, int layer) {

Common/GPU/D3D9/thin3d_d3d9.cpp

@@ -610,18 +610,17 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override;
 
-	int GetCurrentStepId() const override {
-		return stepId_;
-	}
-
 	void InvalidateCachedState() override;
 
+	void SetInvalidationCallback(InvalidationCallback callback) override {
+		invalidationCallback_ = callback;
+	}
+
 private:
 	LPDIRECT3D9 d3d_;
 	LPDIRECT3D9EX d3dEx_;
 	LPDIRECT3DDEVICE9 device_;
 	LPDIRECT3DDEVICE9EX deviceEx_;
-	int stepId_ = -1;
 	int adapterId_ = -1;
 	D3DADAPTER_IDENTIFIER9 identifier_{};
 	D3DCAPS9 d3dCaps_;
@@ -647,6 +646,8 @@ private:
 
 	// Dynamic state
 	uint8_t stencilRef_ = 0;
+
+	InvalidationCallback invalidationCallback_;
 };
 
 void D3D9Context::InvalidateCachedState() {
@@ -1319,7 +1320,10 @@ void D3D9Context::BindFramebufferAsRenderTarget(Framebuffer *fbo, const RenderPa
 	dxstate.scissorRect.restore();
 	dxstate.scissorTest.restore();
 	dxstate.viewport.restore();
-	stepId_++;
+
+	if (invalidationCallback_) {
+		invalidationCallback_(InvalidationFlags::RENDER_PASS_STATE);
+	}
 }
 
 uintptr_t D3D9Context::GetFramebufferAPITexture(Framebuffer *fbo, int channelBits, int attachment) {
@@ -1398,7 +1402,6 @@ bool D3D9Context::BlitFramebuffer(Framebuffer *srcfb, int srcX1, int srcY1, int
 	} else {
 		return false;
 	}
-	stepId_++;
 	return SUCCEEDED(device_->StretchRect(srcSurf, &srcRect, dstSurf, &dstRect, (filter == FB_BLIT_LINEAR && channelBits == FB_COLOR_BIT) ? D3DTEXF_LINEAR : D3DTEXF_POINT));
 }
 
@@ -1519,7 +1522,6 @@ void D3D9Context::HandleEvent(Event ev, int width, int height, void *param1, voi
 		device_->GetDepthStencilSurface(&deviceDSsurf);
 		break;
 	case Event::PRESENTED:
-		stepId_ = 0;
 		break;
 	}
 }

Common/GPU/MiscTypes.h

@@ -0,0 +1,11 @@
+#pragma once
+
+#include "Common/Common.h"
+
+enum class InvalidationFlags {
+	RENDER_PASS_STATE = 1,
+	COMMAND_BUFFER_STATE = 2,
+};
+ENUM_CLASS_BITOPS(InvalidationFlags);
+
+typedef std::function<void(InvalidationFlags)> InvalidationCallback;

Common/GPU/OpenGL/GLRenderManager.cpp

@@ -354,6 +354,10 @@ void GLRenderManager::BindFramebufferAsRenderTarget(GLRFramebuffer *fb, GLRRende
 			step->dependencies.insert(fb);
 		}
 	}
+
+	if (invalidationCallback_) {
+		invalidationCallback_(InvalidationFlags::RENDER_PASS_STATE);
+	}
 }
 
 void GLRenderManager::BindFramebufferAsTexture(GLRFramebuffer *fb, int binding, int aspectBit) {
@@ -480,7 +484,6 @@ void GLRenderManager::BeginFrame() {
 	// In GL, we have to do deletes on the submission thread.
 
 	insideFrame_ = true;
-	renderStepOffset_ = 0;
 }
 
 void GLRenderManager::Finish() {
@@ -612,9 +615,6 @@ void GLRenderManager::Run(int frame) {
 }
 
 void GLRenderManager::FlushSync() {
-	// TODO: Reset curRenderStep_?
-	renderStepOffset_ += (int)steps_.size();
-
 	int curFrame = curFrame_;
 	FrameData &frameData = frameData_[curFrame];
 	{

Common/GPU/OpenGL/GLRenderManager.h

@@ -10,6 +10,7 @@
 #include <condition_variable>
 
 #include "Common/GPU/OpenGL/GLCommon.h"
+#include "Common/GPU/MiscTypes.h"
 #include "Common/Data/Convert/SmallDataConvert.h"
 #include "Common/Log.h"
 #include "GLQueueRunner.h"
@@ -368,6 +369,9 @@ public:
 	GLRenderManager() {}
 	~GLRenderManager();
 
+	void SetInvalidationCallback(InvalidationCallback callback) {
+		invalidationCallback_ = callback;
+	}
 	void SetErrorCallback(ErrorCallbackFn callback, void *userdata) {
 		queueRunner_.SetErrorCallback(callback, userdata);
 	}
@@ -980,12 +984,6 @@ public:
 		skipGLCalls_ = true;
 	}
 
-	// Gets a frame-unique ID of the current step being recorded. Can be used to figure out
-	// when the current step has changed, which means the caller will need to re-record its state.
-	int GetCurrentStepId() const {
-		return renderStepOffset_ + (int)steps_.size();
-	}
-
 private:
 	void BeginSubmitFrame(int frame);
 	void EndSubmitFrame(int frame);
@@ -1033,8 +1031,7 @@ private:
 
 	// Submission time state
 	bool insideFrame_ = false;
-	// This is the offset within this frame, in case of a mid-frame sync.
-	int renderStepOffset_ = 0;
+
 	GLRStep *curRenderStep_ = nullptr;
 	std::vector<GLRStep *> steps_;
 	std::vector<GLRInitStep> initSteps_;
@@ -1074,4 +1071,6 @@ private:
 	GLRProgram *curProgram_ = nullptr;
 #endif
 	Draw::DeviceCaps caps_{};
+
+	InvalidationCallback invalidationCallback_;
 };

Common/GPU/OpenGL/thin3d_gl.cpp

@@ -466,12 +466,12 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override {}
 
-	int GetCurrentStepId() const override {
-		return renderManager_.GetCurrentStepId();
-	}
-
 	void InvalidateCachedState() override;
 
+	void SetInvalidationCallback(InvalidationCallback callback) {
+		renderManager_.SetInvalidationCallback(callback);
+	}
+
 private:
 	void ApplySamplers();
 

Common/GPU/Vulkan/VulkanRenderManager.cpp

@@ -668,8 +668,6 @@ void VulkanRenderManager::BeginFrame(bool enableProfiling, bool enableLogProfile
 	insideFrame_ = true;
 	vulkan_->BeginFrame(enableLogProfiler ? GetInitCmd() : VK_NULL_HANDLE);
 
-	renderStepOffset_ = 0;
-
 	frameData.profile.timestampDescriptions.clear();
 	if (frameData.profilingEnabled_) {
 		// For various reasons, we need to always use an init cmd buffer in this case to perform the vkCmdResetQueryPool,
@@ -956,6 +954,10 @@ void VulkanRenderManager::BindFramebufferAsRenderTarget(VKRFramebuffer *fb, VKRR
 		data.clear.clearMask = lateClearMask;
 		curRenderStep_->commands.push_back(data);
 	}
+
+	if (invalidationCallback_) {
+		invalidationCallback_(InvalidationFlags::RENDER_PASS_STATE);
+	}
 }
 
 bool VulkanRenderManager::CopyFramebufferToMemorySync(VKRFramebuffer *src, VkImageAspectFlags aspectBits, int x, int y, int w, int h, Draw::DataFormat destFormat, uint8_t *pixels, int pixelStride, const char *tag) {
@@ -1414,7 +1416,9 @@ void VulkanRenderManager::Run(VKRRenderThreadTask &task) {
 
 // Called from main thread.
 void VulkanRenderManager::FlushSync() {
-	renderStepOffset_ += (int)steps_.size();
+	if (invalidationCallback_) {
+		invalidationCallback_(InvalidationFlags::COMMAND_BUFFER_STATE);
+	}
 
 	int curFrame = vulkan_->GetCurFrame();
 	FrameData &frameData = frameData_[curFrame];

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -19,6 +19,7 @@
 #include "Common/Data/Convert/SmallDataConvert.h"
 #include "Common/Math/math_util.h"
 #include "Common/GPU/DataFormat.h"
+#include "Common/GPU/MiscTypes.h"
 #include "Common/GPU/Vulkan/VulkanQueueRunner.h"
 
 // Forward declaration
@@ -235,6 +236,10 @@ public:
 	// Zaps queued up commands. Use if you know there's a risk you've queued up stuff that has already been deleted. Can happen during in-game shutdown.
 	void Wipe();
 
+	void SetInvalidationCallback(InvalidationCallback callback) {
+		invalidationCallback_ = callback;
+	}
+
 	// This starts a new step containing a render pass (unless it can be trivially merged into the previous one, which is pretty common).
 	//
 	// After a "CopyFramebuffer" or the other functions that start "steps", you need to call this beforce
@@ -475,12 +480,6 @@ public:
 
 	VkCommandBuffer GetInitCmd();
 
-	// Gets a frame-unique ID of the current step being recorded. Can be used to figure out
-	// when the current step has changed, which means the caller will need to re-record its state.
-	int GetCurrentStepId() const {
-		return renderStepOffset_ + (int)steps_.size();
-	}
-
 	bool CreateBackbuffers();
 	void DestroyBackbuffers();
 
@@ -520,7 +519,6 @@ private:
 	void DrainCompileQueue();
 
 	void Run(VKRRenderThreadTask &task);
-	void BeginSubmitFrame(int frame);
 
 	// Bad for performance but sometimes necessary for synchronous CPU readbacks (screenshots and whatnot).
 	void FlushSync();
@@ -548,7 +546,6 @@ private:
 	bool run_ = false;
 
 	// This is the offset within this frame, in case of a mid-frame sync.
-	int renderStepOffset_ = 0;
 	VKRStep *curRenderStep_ = nullptr;
 	bool curStepHasViewport_ = false;
 	bool curStepHasScissor_ = false;
@@ -587,4 +584,6 @@ private:
 	SimpleStat initTimeMs_;
 	SimpleStat totalGPUTimeMs_;
 	SimpleStat renderCPUTimeMs_;
+
+	std::function<void(InvalidationFlags)> invalidationCallback_;
 };

Common/GPU/Vulkan/thin3d_vulkan.cpp

@@ -498,14 +498,14 @@ public:
 
 	void HandleEvent(Event ev, int width, int height, void *param1, void *param2) override;
 
-	int GetCurrentStepId() const override {
-		return renderManager_.GetCurrentStepId();
-	}
-
 	void InvalidateCachedState() override;
 
 	void InvalidateFramebuffer(FBInvalidationStage stage, uint32_t channels) override;
 
+	void SetInvalidationCallback(InvalidationCallback callback) {
+		renderManager_.SetInvalidationCallback(callback);
+	}
+
 private:
 	VulkanTexture *GetNullTexture();
 	VulkanContext *vulkan_ = nullptr;

Common/GPU/thin3d.h

@@ -16,6 +16,7 @@
 #include "Common/Common.h"
 #include "Common/GPU/DataFormat.h"
 #include "Common/GPU/Shader.h"
+#include "Common/GPU/MiscTypes.h"
 #include "Common/Data/Collections/Slice.h"
 
 namespace Lin {
@@ -769,7 +770,9 @@ public:
 	// This is called when we launch a new game, so any collected internal stats in the backends don't carry over.
 	virtual void ResetStats() {}
 
-	virtual int GetCurrentStepId() const = 0;
+	// Used by the DrawEngines to know when they have to re-apply some state.
+	// Not very elegant, but more elegant than the old passId hack.
+	virtual void SetInvalidationCallback(InvalidationCallback callback) = 0;
 
 protected:
 	ShaderModule *vsPresets_[VS_MAX_PRESET];

GPU/D3D11/DrawEngineD3D11.cpp

@@ -113,6 +113,8 @@ void DrawEngineD3D11::InitDeviceObjects() {
 
 	tessDataTransferD3D11 = new TessellationDataTransferD3D11(context_, device_);
 	tessDataTransfer = tessDataTransferD3D11;
+
+	draw_->SetInvalidationCallback(std::bind(&DrawEngineD3D11::Invalidate, this, std::placeholders::_1));
 }
 
 void DrawEngineD3D11::ClearTrackedVertexArrays() {
@@ -136,6 +138,8 @@ void DrawEngineD3D11::NotifyConfigChanged() {
 }
 
 void DrawEngineD3D11::DestroyDeviceObjects() {
+	draw_->SetInvalidationCallback(InvalidationCallback());
+
 	ClearTrackedVertexArrays();
 	ClearInputLayoutMap();
 	delete tessDataTransferD3D11;
@@ -322,19 +326,18 @@ VertexArrayInfoD3D11::~VertexArrayInfoD3D11() {
 		ebo->Release();
 }
 
+// In D3D, we're synchronous and state carries over so all we reset here on a new step is the viewport/scissor.
+void DrawEngineD3D11::Invalidate(InvalidationFlags flags) {
+	if (flags & InvalidationFlags::RENDER_PASS_STATE) {
+		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
+		textureCache_->ForgetLastTexture();
+	}
+}
+
 // The inline wrapper in the header checks for numDrawCalls == 0
 void DrawEngineD3D11::DoFlush() {
 	gpuStats.numFlushes++;
 
-	// 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 | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
-		textureCache_->ForgetLastTexture();
-		lastRenderStepId_ = curRenderStepId;
-	}
-
 	bool textureNeedsApply = false;
 	if (gstate_c.IsDirty(DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS) && !gstate.isModeClear() && gstate.isTextureMapEnabled()) {
 		textureCache_->SetTexture();

GPU/D3D11/DrawEngineD3D11.h

@@ -155,6 +155,8 @@ public:
 	void ClearInputLayoutMap();
 
 private:
+	void Invalidate(InvalidationFlags flags);
+
 	void DoFlush();
 
 	void ApplyDrawState(int prim);

GPU/Directx9/DrawEngineDX9.cpp

@@ -126,10 +126,11 @@ DrawEngineDX9::~DrawEngineDX9() {
 }
 
 void DrawEngineDX9::InitDeviceObjects() {
-
+	draw_->SetInvalidationCallback(std::bind(&DrawEngineDX9::Invalidate, this, std::placeholders::_1));
 }
 
 void DrawEngineDX9::DestroyDeviceObjects() {
+	draw_->SetInvalidationCallback(InvalidationCallback());
 	ClearTrackedVertexArrays();
 }
 
@@ -310,18 +311,17 @@ void DrawEngineDX9::BeginFrame() {
 	lastRenderStepId_ = -1;
 }
 
+// In D3D, we're synchronous and state carries over so all we reset here on a new step is the viewport/scissor.
+void DrawEngineDX9::Invalidate(InvalidationFlags flags) {
+	if (flags & InvalidationFlags::RENDER_PASS_STATE) {
+		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
+	}
+}
+
 // The inline wrapper in the header checks for numDrawCalls == 0
 void DrawEngineDX9::DoFlush() {
 	gpuStats.numFlushes++;
 
-	// 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 | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
-		lastRenderStepId_ = curRenderStepId;
-	}
-
 	bool textureNeedsApply = false;
 	if (gstate_c.IsDirty(DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS) && !gstate.isModeClear() && gstate.isTextureMapEnabled()) {
 		textureCache_->SetTexture();

GPU/Directx9/DrawEngineDX9.h

@@ -144,6 +144,7 @@ protected:
 	void DecimateTrackedVertexArrays();
 
 private:
+	void Invalidate(InvalidationFlags flags);
 	void DoFlush();
 
 	void ApplyDrawState(int prim);

GPU/GLES/DrawEngineGLES.cpp

@@ -117,9 +117,13 @@ void DrawEngineGLES::InitDeviceObjects() {
 	entries.push_back({ ATTR_COLOR1, 3, GL_UNSIGNED_BYTE, GL_TRUE, vertexSize, offsetof(TransformedVertex, color1) });
 	entries.push_back({ ATTR_NORMAL, 1, GL_FLOAT, GL_FALSE, vertexSize, offsetof(TransformedVertex, fog) });
 	softwareInputLayout_ = render_->CreateInputLayout(entries);
+
+	draw_->SetInvalidationCallback(std::bind(&DrawEngineGLES::Invalidate, this, std::placeholders::_1));
 }
 
 void DrawEngineGLES::DestroyDeviceObjects() {
+	draw_->SetInvalidationCallback(InvalidationCallback());
+
 	// Beware: this could be called twice in a row, sometimes.
 	for (int i = 0; i < GLRenderManager::MAX_INFLIGHT_FRAMES; i++) {
 		if (!frameData_[i].pushVertex && !frameData_[i].pushIndex)
@@ -238,22 +242,22 @@ void *DrawEngineGLES::DecodeVertsToPushBuffer(GLPushBuffer *push, uint32_t *bind
 	return dest;
 }
 
+// A new render step means we need to flush any dynamic state. Really, any state that is reset in
+// GLQueueRunner::PerformRenderPass.
+void DrawEngineGLES::Invalidate(InvalidationFlags flags) {
+	if (flags & InvalidationFlags::RENDER_PASS_STATE) {
+		// Dirty everything that has dynamic state that will need re-recording.
+		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_BLEND_STATE | DIRTY_RASTER_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
+		textureCache_->ForgetLastTexture();
+	}
+}
+
 void DrawEngineGLES::DoFlush() {
 	PROFILE_THIS_SCOPE("flush");
 	FrameData &frameData = frameData_[render_->GetCurFrame()];
 	
 	gpuStats.numFlushes++;
 
-	// A new render step means we need to flush any dynamic state. Really, any state that is reset in
-	// GLQueueRunner::PerformRenderPass.
-	int curRenderStepId = render_->GetCurrentStepId();
-	if (lastRenderStepId_ != curRenderStepId) {
-		// Dirty everything that has dynamic state that will need re-recording.
-		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_BLEND_STATE | DIRTY_RASTER_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
-		textureCache_->ForgetLastTexture();
-		lastRenderStepId_ = curRenderStepId;
-	}
-
 	bool textureNeedsApply = false;
 	if (gstate_c.IsDirty(DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS) && !gstate.isModeClear() && gstate.isTextureMapEnabled()) {
 		textureCache_->SetTexture();

GPU/GLES/DrawEngineGLES.h

@@ -84,7 +84,6 @@ public:
 	void BeginFrame();
 	void EndFrame();
 
-
 	// So that this can be inlined
 	void Flush() {
 		if (!numDrawCalls)
@@ -117,6 +116,8 @@ protected:
 	bool UpdateUseHWTessellation(bool enable) override;
 
 private:
+	void Invalidate(InvalidationFlags flags);
+
 	void InitDeviceObjects();
 	void DestroyDeviceObjects();
 

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -16,6 +16,7 @@
 // https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
 
 #include <algorithm>
+#include <functional>
 
 #include "Common/Data/Convert/SmallDataConvert.h"
 #include "Common/Profiler/Profiler.h"
@@ -150,7 +151,7 @@ void DrawEngineVulkan::InitDeviceObjects() {
 	dpTypes[2].type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
 	dpTypes[3].descriptorCount = DEFAULT_DESC_POOL_SIZE;  // TODO: Use a separate layout when no spline stuff is needed to reduce the need for these.
 	dpTypes[3].type = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT;
-	dpTypes[4].descriptorCount = 1;  // For the frame global uniform buffer.
+	dpTypes[4].descriptorCount = DEFAULT_DESC_POOL_SIZE;  // For the frame global uniform buffer. Might need to allocate multiple times.
 	dpTypes[4].type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
 
 	VkDescriptorPoolCreateInfo dp{ VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO };
@@ -204,6 +205,8 @@ void DrawEngineVulkan::InitDeviceObjects() {
 
 	tessDataTransferVulkan = new TessellationDataTransferVulkan(vulkan);
 	tessDataTransfer = tessDataTransferVulkan;
+
+	draw_->SetInvalidationCallback(std::bind(&DrawEngineVulkan::Invalidate, this, std::placeholders::_1));
 }
 
 DrawEngineVulkan::~DrawEngineVulkan() {
@@ -234,7 +237,14 @@ void DrawEngineVulkan::FrameData::Destroy(VulkanContext *vulkan) {
 }
 
 void DrawEngineVulkan::DestroyDeviceObjects() {
-	VulkanContext *vulkan = draw_ ? (VulkanContext *)draw_->GetNativeObject(Draw::NativeObject::CONTEXT) : nullptr;
+	if (!draw_) {
+		// We've already done this from LostDevice.
+		return;
+	}
+
+	VulkanContext *vulkan = (VulkanContext *)draw_->GetNativeObject(Draw::NativeObject::CONTEXT);
+
+	draw_->SetInvalidationCallback(InvalidationCallback());
 
 	delete tessDataTransferVulkan;
 	tessDataTransfer = nullptr;
@@ -283,8 +293,6 @@ void DrawEngineVulkan::BeginFrame() {
 
 	lastPipeline_ = nullptr;
 
-	lastRenderStepId_ = -1;
-
 	FrameData *frame = &GetCurFrame();
 
 	// First reset all buffers, then begin. This is so that Reset can free memory and Begin can allocate it,
@@ -536,6 +544,21 @@ void MarkUnreliable(VertexArrayInfoVulkan *vai) {
 	// For now we just leave it in the pushbuffer.
 }
 
+void DrawEngineVulkan::Invalidate(InvalidationFlags flags) {
+	if (flags & InvalidationFlags::COMMAND_BUFFER_STATE) {
+		GetCurFrame().frameDescSetUpdated = false;
+	}
+	if (flags & InvalidationFlags::RENDER_PASS_STATE) {
+		// If have a new render pass, dirty our dynamic state so it gets re-set.
+		// We have to do this again after the last possible place in DoFlush that can cause a renderpass switch
+		// like a shader blend blit or similar. But before we actually set the state!
+		//
+		// Dirty everything that has dynamic state that will need re-recording.
+		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_BLEND_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
+		lastPipeline_ = nullptr;
+	}
+}
+
 // The inline wrapper in the header checks for numDrawCalls == 0
 void DrawEngineVulkan::DoFlush() {
 	VulkanRenderManager *renderManager = (VulkanRenderManager *)draw_->GetNativeObject(Draw::NativeObject::RENDER_MANAGER);
@@ -545,18 +568,6 @@ void DrawEngineVulkan::DoFlush() {
 
 	gpuStats.numFlushes++;
 	
-	// If have a new render pass, dirty our dynamic state so it gets re-set.
-	// We have to do this again after the last possible place in DoFlush that can cause a renderpass switch
-	// like a shader blend blit or similar. But before we actually set the state!
-	int curRenderStepId = renderManager->GetCurrentStepId();
-	if (lastRenderStepId_ != curRenderStepId) {
-		// Dirty everything that has dynamic state that will need re-recording.
-		gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_BLEND_STATE | DIRTY_TEXTURE_IMAGE | DIRTY_TEXTURE_PARAMS);
-		textureCache_->ForgetLastTexture();
-		lastRenderStepId_ = curRenderStepId;
-		lastPipeline_ = nullptr;
-	}
-
 	bool tess = gstate_c.submitType == SubmitType::HW_BEZIER || gstate_c.submitType == SubmitType::HW_SPLINE;
 
 	bool textureNeedsApply = false;
@@ -785,16 +796,6 @@ void DrawEngineVulkan::DoFlush() {
 			}
 			BindShaderBlendTex();  // This might cause copies so important to do before BindPipeline.
 
-			// If have a new render pass, dirty our dynamic state so it gets re-set.
-			// WARNING: We have to do this AFTER the last possible place in DoFlush that can cause a renderpass switch
-			// like a shader blend blit or similar. But before we actually set the state!
-			int curRenderStepId = renderManager->GetCurrentStepId();
-			if (lastRenderStepId_ != curRenderStepId) {
-				// Dirty everything that has dynamic state that will need re-recording.
-				gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_BLEND_STATE);
-				lastRenderStepId_ = curRenderStepId;
-			}
-
 			renderManager->BindPipeline(pipeline->pipeline, pipeline->pipelineFlags, pipelineLayout_);
 			if (pipeline != lastPipeline_) {
 				if (lastPipeline_ && !(lastPipeline_->UsesBlendConstant() && pipeline->UsesBlendConstant())) {
@@ -928,16 +929,6 @@ void DrawEngineVulkan::DoFlush() {
 				}
 				BindShaderBlendTex();  // This might cause copies so super important to do before BindPipeline.
 
-				// If have a new render pass, dirty our dynamic state so it gets re-set.
-				// WARNING: We have to do this AFTER the last possible place in DoFlush that can cause a renderpass switch
-				// like a shader blend blit or similar. But before we actually set the state!
-				int curRenderStepId = renderManager->GetCurrentStepId();
-				if (lastRenderStepId_ != curRenderStepId) {
-					// Dirty everything that has dynamic state that will need re-recording.
-					gstate_c.Dirty(DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_BLEND_STATE);
-					lastRenderStepId_ = curRenderStepId;
-				}
-
 				renderManager->BindPipeline(pipeline->pipeline, pipeline->pipelineFlags, pipelineLayout_);
 				if (pipeline != lastPipeline_) {
 					if (lastPipeline_ && !lastPipeline_->UsesBlendConstant() && pipeline->UsesBlendConstant()) {

GPU/Vulkan/DrawEngineVulkan.h

@@ -43,6 +43,7 @@
 #include "GPU/Common/DrawEngineCommon.h"
 #include "GPU/Common/GPUStateUtils.h"
 #include "GPU/Vulkan/StateMappingVulkan.h"
+#include "GPU/Vulkan/VulkanRenderManager.h"
 
 struct DecVtxFormat;
 struct UVScale;
@@ -208,6 +209,8 @@ public:
 	}
 
 private:
+	void Invalidate(InvalidationFlags flags);
+
 	struct FrameData;
 	void ApplyDrawStateLate(VulkanRenderManager *renderManager, bool applyStencilRef, uint8_t stencilRef, bool useBlendConstant);
 	void ConvertStateToVulkanKey(FramebufferManagerVulkan &fbManager, ShaderManagerVulkan *shaderManager, int prim, VulkanPipelineRasterStateKey &key, VulkanDynamicState &dynState);
@@ -308,6 +311,4 @@ private:
 
 	// Hardware tessellation
 	TessellationDataTransferVulkan *tessDataTransferVulkan;
-
-	int lastRenderStepId_ = -1;
 };