Merge pull request #22382 from hrydgard/gpu-lifecycle-fixes-2

Claude code review: GPU fixes 2

Common/GPU/OpenGL/GLMemory.h

@@ -30,6 +30,8 @@ static inline int operator &(const GLBufferStrategy &lhs, const GLBufferStrategy
 
 class GLRBuffer {
 public:
+	GLRBuffer(const GLRBuffer &) = delete;
+	GLRBuffer &operator=(const GLRBuffer &) = delete;
 	GLRBuffer(GLuint target, size_t size) : target_(target), size_((int)size) {}
 	~GLRBuffer() {
 		if (buffer_) {

Common/GPU/OpenGL/GLQueueRunner.cpp

@@ -147,6 +147,7 @@ void GLQueueRunner::RunInitSteps(const FastVec<GLRInitStep> &steps, bool skipGLC
 			case GLRInitStepType::CREATE_SHADER:
 			{
 				WARN_LOG(Log::G3D, "CREATE_SHADER found with skipGLCalls, not good");
+				delete[] step.create_shader.code;
 				break;
 			}
 			default:
@@ -678,6 +679,9 @@ void GLQueueRunner::RunSteps(const std::vector<GLRStep *> &steps, GLFrameData &f
 							}
 						}
 						break;
+					case GLRRenderCommand::UNIFORMSTEREOMATRIX:
+						delete[] c.uniformStereoMatrix4.mData;
+						break;
 					default:
 						break;
 					}

Common/GPU/OpenGL/GLRenderManager.cpp

@@ -110,11 +110,11 @@ void GLRenderManager::ThreadEnd() {
 		frameData_[i].deleter_prev.Perform(this, skipGLCalls_);
 	}
 	deleter_.Perform(this, skipGLCalls_);
-	for (int i = 0; i < (int)steps_.size(); i++) {
+	// Steps that never got submitted. A dry run frees the data they own (texture uploads etc), and the steps.
-		delete steps_[i];
+	queueRunner_.RunInitSteps(initSteps_, true);
-	}
-	steps_.clear();
 	initSteps_.clear();
+	queueRunner_.RunSteps(steps_, frameData_[0], true, false, false);
+	steps_.clear();
 	INFO_LOG(Log::G3D, "GLRenderManager::ThreadEnd end");
 }
 

Common/GPU/OpenGL/GLRenderManager.h

@@ -30,6 +30,8 @@ constexpr int MAX_GL_TEXTURE_SLOTS = 8;
 
 class GLRTexture {
 public:
+	GLRTexture(const GLRTexture &) = delete;
+	GLRTexture &operator=(const GLRTexture &) = delete;
 	GLRTexture(const Draw::DeviceCaps &caps, int width, int height, int depth, int numMips);
 	~GLRTexture();
 
@@ -53,6 +55,8 @@ public:
 
 class GLRFramebuffer {
 public:
+	GLRFramebuffer(const GLRFramebuffer &) = delete;
+	GLRFramebuffer &operator=(const GLRFramebuffer &) = delete;
 	GLRFramebuffer(const Draw::DeviceCaps &caps, int _width, int _height, bool z_stencil, const char *tag)
 		: color_texture(caps, _width, _height, 1, 1), z_stencil_texture(caps, _width, _height, 1, 1),
 		width(_width), height(_height), z_stencil_(z_stencil) {
@@ -83,6 +87,8 @@ private:
 
 class GLRShader {
 public:
+	GLRShader(const GLRShader &) = delete;
+	GLRShader &operator=(const GLRShader &) = delete;
 	explicit GLRShader(std::string_view _desc) : desc(_desc) {}
 	~GLRShader() {
 		if (shader) {
@@ -116,6 +122,9 @@ public:
 
 class GLRProgram {
 public:
+	GLRProgram() = default;
+	GLRProgram(const GLRProgram &) = delete;
+	GLRProgram &operator=(const GLRProgram &) = delete;
 	~GLRProgram() {
 		if (deleteCallback_) {
 			deleteCallback_(deleteParam_);

Common/GPU/Vulkan/VulkanBarrier.h

@@ -12,6 +12,8 @@ struct VKRImage;
 
 class VulkanBarrierBatch {
 public:
+	VulkanBarrierBatch(const VulkanBarrierBatch &) = delete;
+	VulkanBarrierBatch &operator=(const VulkanBarrierBatch &) = delete;
 	VulkanBarrierBatch() : imageBarriers_(4) {}
 	~VulkanBarrierBatch();
 

Common/GPU/Vulkan/VulkanDescSet.h

@@ -18,6 +18,8 @@ enum class BindingType {
 // Only appropriate for use in a per-frame pool.
 class VulkanDescSetPool {
 public:
+	VulkanDescSetPool(const VulkanDescSetPool &) = delete;
+	VulkanDescSetPool &operator=(const VulkanDescSetPool &) = delete;
 	VulkanDescSetPool(const char *tag, bool grow = true) : tag_(tag), grow_(grow) {}
 	~VulkanDescSetPool();
 

Common/GPU/Vulkan/VulkanFramebuffer.h

@@ -60,6 +60,8 @@ struct VKRImage {
 
 class VKRFramebuffer {
 public:
+	VKRFramebuffer(const VKRFramebuffer &) = delete;
+	VKRFramebuffer &operator=(const VKRFramebuffer &) = delete;
 	VKRFramebuffer(VulkanContext *vk, VulkanBarrierBatch *barriers, int _width, int _height, int _numLayers, int _multiSampleLevel, bool createDepthStencilBuffer, const char *tag);
 	~VKRFramebuffer();
 

Common/GPU/Vulkan/VulkanImage.h

@@ -22,6 +22,8 @@ struct TextureCopyBatch {
 // ALWAYS use an allocator when calling CreateDirect.
 class VulkanTexture {
 public:
+	VulkanTexture(const VulkanTexture &) = delete;
+	VulkanTexture &operator=(const VulkanTexture &) = delete;
 	VulkanTexture(VulkanContext *vulkan, const char *tag);
 	~VulkanTexture() {
 		Destroy();

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -115,6 +115,8 @@ public:
 
 // Wrapped pipeline. Does own desc!
 struct VKRGraphicsPipeline {
+	VKRGraphicsPipeline(const VKRGraphicsPipeline &) = delete;
+	VKRGraphicsPipeline &operator=(const VKRGraphicsPipeline &) = delete;
 	VKRGraphicsPipeline(PipelineFlags flags, const char *tag) : flags_(flags), tag_(tag) {}
 	~VKRGraphicsPipeline();
 
@@ -195,6 +197,9 @@ static_assert(sizeof(PackedDescriptor::buffer) == 16, "PackedDescriptor should b
 // Note that we only support a single descriptor set due to compatibility with some ancient devices.
 // We should probably eventually give that up eventually.
 struct VKRPipelineLayout {
+	VKRPipelineLayout() = default;
+	VKRPipelineLayout(const VKRPipelineLayout &) = delete;
+	VKRPipelineLayout &operator=(const VKRPipelineLayout &) = delete;
 	~VKRPipelineLayout();
 
 	enum { MAX_DESC_SET_BINDINGS = 5 };

Common/Render/DrawBuffer.h

@@ -51,6 +51,8 @@ class TextDrawer;
 
 class DrawBuffer {
 public:
+	DrawBuffer(const DrawBuffer &) = delete;
+	DrawBuffer &operator=(const DrawBuffer &) = delete;
 	DrawBuffer();
 	~DrawBuffer();
 

Common/Render/TextureAtlas.h

@@ -96,6 +96,9 @@ struct AtlasFontHeader {
 };
 
 struct AtlasFont {
+	AtlasFont() = default;
+	AtlasFont(const AtlasFont &) = delete;
+	AtlasFont &operator=(const AtlasFont &) = delete;
 	~AtlasFont();
 
 	float padding;
@@ -126,6 +129,9 @@ struct AtlasHeader {
 };
 
 struct Atlas {
+	Atlas() = default;
+	Atlas(const Atlas &) = delete;
+	Atlas &operator=(const Atlas &) = delete;
 	~Atlas();
 	bool LoadMeta(const uint8_t *data, size_t data_size);
 	bool IsMetadataLoaded() const {

GPU/Common/DepthBufferCommon.cpp

@@ -199,7 +199,9 @@ bool FramebufferManagerCommon::ReadbackDepthbuffer(Draw::Framebuffer *fbo, int x
 
 		auto *blitFBO = GetTempFBO(TempFBO::Z_COPY, fbo->Width() * scaleX, fbo->Height() * scaleY);
 		draw_->BindFramebufferAsRenderTarget(blitFBO, { RPAction::DONT_CARE, RPAction::DONT_CARE, RPAction::DONT_CARE }, "ReadbackDepthbufferSync");
-		Draw::Viewport viewport = { 0.0f, 0.0f, (float)destW, (float)destH, 0.0f, 1.0f };
+		// The whole fbo is drawn, so the viewport has to cover all of it at the destination scale. Not just
+		// destW x destH, which would squeeze it whenever the read rectangle is smaller than the fbo.
+		Draw::Viewport viewport = { 0.0f, 0.0f, fbo->Width() * scaleX, fbo->Height() * scaleY, 0.0f, 1.0f };
 		draw_->SetViewport(viewport);
 		draw_->SetScissorRect(0, 0, fbo->Width() * scaleX, fbo->Height() * scaleY);
 

GPU/Common/DrawEngineCommon.cpp

@@ -51,11 +51,13 @@ DrawEngineCommon::DrawEngineCommon() : decoderMap_(32) {
 	transformedExpanded_ = (TransformedVertex *)AllocateMemoryPages(3 * TRANSFORMED_VERTEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
 	decoded_ = (u8 *)AllocateMemoryPages(DECODED_VERTEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
 	decIndex_ = (u16 *)AllocateMemoryPages(DECODED_INDEX_BUFFER_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
+	bboxScratch_ = (u8 *)AllocateMemoryPages(BBOX_SCRATCH_SIZE, MEM_PROT_READ | MEM_PROT_WRITE);
 
 	_dbg_assert_(transformed_);
 	_dbg_assert_(transformedExpanded_);
 	_dbg_assert_(decoded_);
 	_dbg_assert_(decIndex_);
+	_dbg_assert_(bboxScratch_);
 
 	indexGen.Setup(decIndex_);
 
@@ -65,6 +67,7 @@ DrawEngineCommon::DrawEngineCommon() : decoderMap_(32) {
 DrawEngineCommon::~DrawEngineCommon() {
 	FreeMemoryPages(decoded_, DECODED_VERTEX_BUFFER_SIZE);
 	FreeMemoryPages(decIndex_, DECODED_INDEX_BUFFER_SIZE);
+	FreeMemoryPages(bboxScratch_, BBOX_SCRATCH_SIZE);
 	FreeMemoryPages(transformed_, TRANSFORMED_VERTEX_BUFFER_SIZE);
 	FreeMemoryPages(transformedExpanded_, 3 * TRANSFORMED_VERTEX_BUFFER_SIZE);
 	ShutdownDepthRaster();
@@ -183,15 +186,14 @@ void DrawEngineCommon::DispatchSubmitImm(GEPrimitiveType prim, TransformedVertex
 //   - Less accurate, but..
 //   - Only requires six plane evaluations then.
 bool DrawEngineCommon::TestBoundingBox(const void *vdata, const void *inds, int vertexCount, const VertexDecoder *dec, u32 vertType) {
-	// Grab temp buffer space from large offsets in decoded_. Not exactly safe for large draws.
+	// The scratch buffer is sized for 1024 vertices. Although this may lead to drawing that shouldn't happen,
-	// Although this may lead to drawing that shouldn't happen, the viewport is more complex on VR.
+	// the viewport is more complex on VR. Let's always say objects are within bounds.
-	// Let's always say objects are within bounds.
 	if (vertexCount > 1024 || gstate_c.Use(GPU_USE_VIRTUAL_REALITY)) {
 		return true;
 	}
 
-	SimpleVertex *corners = (SimpleVertex *)(decoded_ + 65536 * 12);
+	SimpleVertex *corners = (SimpleVertex *)(bboxScratch_ + BBOX_SCRATCH_CORNERS_OFFSET);
-	float *verts = (float *)(decoded_ + 65536 * 18);
+	float *verts = (float *)(bboxScratch_ + BBOX_SCRATCH_VERTS_OFFSET);
 
 	// Try to skip NormalizeVertices if it's pure positions. No need to bother with a vertex decoder
 	// and a large vertex format.
@@ -218,7 +220,7 @@ bool DrawEngineCommon::TestBoundingBox(const void *vdata, const void *inds, int
 		}
 	} else {
 		// Simplify away indices, bones, and morph before proceeding.
-		u8 *temp_buffer = decoded_ + 65536 * 24;
+		u8 *temp_buffer = bboxScratch_ + BBOX_SCRATCH_TEMP_OFFSET;
 
 		if ((inds || (vertType & (GE_VTYPE_WEIGHT_MASK | GE_VTYPE_MORPHCOUNT_MASK)))) {
 			// Need for Speed Carbon ends up on this path! With a single bone weight.
@@ -392,7 +394,8 @@ static bool TestBoundingBoxFast(const float *cullMatrix, const void *vdata, cons
 		}
 		case GE_VTYPE_IDX_32BIT:
 		{
-			u32 idx = ((u32 *)idata)[i];
+			// The PSP ignores the upper 16 bits.
+			u16 idx = (u16)((u32 *)idata)[i];
 			data = (const s8 *)srcdata + idx * stride;
 			break;
 		}
@@ -759,7 +762,7 @@ int DrawEngineCommon::ExtendNonIndexedPrim(const uint32_t *cmd, const uint32_t *
 		if (IsTrianglePrim(newPrim) != isTriangle)
 			break;
 		int vertexCount = data & 0xFFFF;
-		if (numDrawInds >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + offset + vertexCount > VERTEX_BUFFER_MAX) {
+		if (numDrawInds >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + offset + vertexCount > VERTEX_BUFFER_MAX || numVertsToDecode_ + (offset - dv.vertexCount) + vertexCount > VERTEX_BUFFER_MAX) {
 			break;
 		}
 		DeferredInds &di = drawInds_[numDrawInds++];
@@ -780,6 +783,7 @@ int DrawEngineCommon::ExtendNonIndexedPrim(const uint32_t *cmd, const uint32_t *
 	dv.vertexCount = offset;
 	dv.indexUpperBound = dv.vertexCount - 1;
 	vertexCountInDrawCalls_ += totalCount;
+	numVertsToDecode_ += totalCount;
 	*bytesRead = totalCount * dec->VertexSize();
 	return cmd - start;
 }
@@ -805,7 +809,21 @@ void DrawEngineCommon::SkipPrim(GEPrimitiveType prim, int vertexCount, const Ver
 
 // vertTypeID is the vertex type but with the UVGen mode smashed into the top bits.
 bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, const VertexDecoder *dec, u32 vertTypeID, bool clockwise, int *bytesRead, ClipInfoFlags clipInfoFlags) {
-	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawVerts_ >= MAX_DEFERRED_DRAW_VERTS || numDrawInds_ >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
+	// The index count doesn't bound how many vertices DecodeVerts will produce (the index range can be
+	// sparse), so track that separately, as the growth of the range to decode.
+	u16 lowerBound = 0;
+	u16 upperBound = 0;
+	int decodeGrowth = 0;
+	if (vertexCount > 0) {
+		GetIndexBounds(inds, vertexCount, vertTypeID, &lowerBound, &upperBound);
+		decodeGrowth = upperBound - lowerBound + 1;
+		if (CanExtendDecode(verts, inds, dec)) {
+			const DeferredVerts &last = drawVerts_[numDrawVerts_ - 1];
+			decodeGrowth = std::max(upperBound, last.indexUpperBound) - std::min(lowerBound, last.indexLowerBound) - (last.indexUpperBound - last.indexLowerBound);
+		}
+	}
+
+	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawVerts_ >= MAX_DEFERRED_DRAW_VERTS || numDrawInds_ >= MAX_DEFERRED_DRAW_INDS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX || numVertsToDecode_ + decodeGrowth > VERTEX_BUFFER_MAX) {
 		Flush();
 	}
 
@@ -855,6 +873,7 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 		const int rem = vertexCount % 3;
 		if (rem != 0) {
 			vertexCount -= rem;
+			GetIndexBounds(inds, vertexCount, vertTypeID, &lowerBound, &upperBound);
 		}
 	}
 
@@ -881,17 +900,16 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 
 	_dbg_assert_(numDrawVerts <= MAX_DEFERRED_DRAW_VERTS);
 
-	if (inds && numDrawVerts > decodeVertsCounter_ && drawVerts_[numDrawVerts - 1].verts == verts && !applySkin) {
+	if (CanExtendDecode(verts, inds, dec_)) {
 		// Same vertex pointer as a previous un-decoded draw call - let's just extend the decode!
 		di.vertDecodeIndex = numDrawVerts - 1;
-		u16 lb;
-		u16 ub;
-		GetIndexBounds(inds, vertexCount, vertTypeID, &lb, &ub);
 		DeferredVerts &dv = drawVerts_[numDrawVerts - 1];
-		if (lb < dv.indexLowerBound)
+		const int oldCount = dv.indexUpperBound - dv.indexLowerBound + 1;
-			dv.indexLowerBound = lb;
+		if (lowerBound < dv.indexLowerBound)
-		if (ub > dv.indexUpperBound)
+			dv.indexLowerBound = lowerBound;
-			dv.indexUpperBound = ub;
+		if (upperBound > dv.indexUpperBound)
+			dv.indexUpperBound = upperBound;
+		numVertsToDecode_ += dv.indexUpperBound - dv.indexLowerBound + 1 - oldCount;
 	} else {
 		// Record a new draw, and a new index gen.
 		DeferredVerts &dv = drawVerts_[numDrawVerts];
@@ -899,8 +917,9 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 		dv.verts = verts;
 		dv.vertexCount = vertexCount;
 		dv.uvScale = LoadUVScaleOffset(gstate);
-		// Does handle the unindexed case.
+		dv.indexLowerBound = lowerBound;
-		GetIndexBounds(inds, vertexCount, vertTypeID, &dv.indexLowerBound, &dv.indexUpperBound);
+		dv.indexUpperBound = upperBound;
+		numVertsToDecode_ += upperBound - lowerBound + 1;
 	}
 
 	vertexCountInDrawCalls_ += vertexCount;
@@ -930,8 +949,9 @@ void DrawEngineCommon::DecodeVerts(const VertexDecoder *dec, u8 *dest) {
 		drawVertexOffsets_[i] = numDecodedVerts - indexLowerBound;
 		const int indexUpperBound = dv.indexUpperBound;
 		const int count = indexUpperBound - indexLowerBound + 1;
-		if (count + numDecodedVerts >= VERTEX_BUFFER_MAX) {
+		if (count + numDecodedVerts > VERTEX_BUFFER_MAX) {
-			// Hit our limit! Stop decoding in this draw.
+			// SubmitPrim flushes before this can happen.
+			_dbg_assert_(false);
 			break;
 		}
 

GPU/Common/DrawEngineCommon.h

@@ -38,6 +38,11 @@ enum {
 	VERTEX_BUFFER_MAX = 65536,
 	DECODED_VERTEX_BUFFER_SIZE = VERTEX_BUFFER_MAX * 2 * 36,  // 36 == sizeof(SimpleVertex)
 	DECODED_INDEX_BUFFER_SIZE = VERTEX_BUFFER_MAX * 6 * 6 * 2,   // * 6 for spline tessellation, then * 6 again for converting into points/lines, and * 2 for 2 bytes per index
+	// TestBoundingBox handles up to 1025 vertices: corners (SimpleVertex), then positions, then decoded vertices.
+	BBOX_SCRATCH_CORNERS_OFFSET = 0,
+	BBOX_SCRATCH_VERTS_OFFSET = 64 * 1024,
+	BBOX_SCRATCH_TEMP_OFFSET = 128 * 1024,
+	BBOX_SCRATCH_SIZE = 256 * 1024,
 };
 
 enum {
@@ -68,6 +73,8 @@ struct alignas(16) Plane8 {
 
 class DrawEngineCommon {
 public:
+	DrawEngineCommon(const DrawEngineCommon &) = delete;
+	DrawEngineCommon &operator=(const DrawEngineCommon &) = delete;
 	DrawEngineCommon();
 	virtual ~DrawEngineCommon();
 
@@ -161,6 +168,11 @@ protected:
 	void DecodeVerts(const VertexDecoder *dec, u8 *dest);
 	int DecodeInds();
 
+	// Whether an indexed draw can share the previous draw's vertex decode, by widening its index range.
+	bool CanExtendDecode(const void *verts, const void *inds, const VertexDecoder *dec) const {
+		return inds && numDrawVerts_ > decodeVertsCounter_ && drawVerts_[numDrawVerts_ - 1].verts == verts && !dec->skinInDecode;
+	}
+
 	int ComputeNumVertsToDecode() const;
 
 	void ApplyFramebufferRead(FBOTexState *fboTexState);
@@ -211,6 +223,7 @@ protected:
 		numDrawVerts_ = 0;
 		numDrawInds_ = 0;
 		vertexCountInDrawCalls_ = 0;
+		numVertsToDecode_ = 0;
 		decodeIndsCounter_ = 0;
 		decodeVertsCounter_ = 0;
 		seenPrims_ = 0;
@@ -273,6 +286,8 @@ protected:
 	// Vertex collector buffers
 	u8 *decoded_ = nullptr;
 	u16 *decIndex_ = nullptr;
+	// Separate from decoded_, which can hold decoded vertices that haven't been flushed yet.
+	u8 *bboxScratch_ = nullptr;
 
 	// Cached vertex decoders
 	DenseHashMap<u32, VertexDecoder *> decoderMap_;
@@ -312,6 +327,8 @@ protected:
 	int numDrawVerts_ = 0;
 	int numDrawInds_ = 0;
 	int vertexCountInDrawCalls_ = 0;
+	// How many vertices DecodeVerts will produce for the queued draws. Must stay <= VERTEX_BUFFER_MAX.
+	int numVertsToDecode_ = 0;
 
 	int decodeVertsCounter_ = 0;
 	int decodeIndsCounter_ = 0;

GPU/Common/FramebufferManagerCommon.cpp

@@ -3354,7 +3354,8 @@ void FramebufferManagerCommon::FlushBeforeCopy() {
 // TODO: Replace with with depal, reading the palette from the texture on the GPU directly.
 void FramebufferManagerCommon::DownloadFramebufferForClut(u32 fb_address, u32 loadBytes) {
 	VirtualFramebuffer *vfb = GetVFBAt(fb_address);
-	if (vfb && vfb->fb_stride != 0) {
+	// Without an fbo there's nothing to read back (ReadbackFramebuffer would read the backbuffer instead).
+	if (vfb && vfb->fb_stride != 0 && vfb->fbo) {
 		const u32 bpp = BufferFormatBytesPerPixel(vfb->fb_format);
 		int x = 0;
 		int y = 0;

GPU/Common/FramebufferManagerCommon.h

@@ -291,6 +291,8 @@ struct DisplayLayoutConfig;
 
 class FramebufferManagerCommon {
 public:
+	FramebufferManagerCommon(const FramebufferManagerCommon &) = delete;
+	FramebufferManagerCommon &operator=(const FramebufferManagerCommon &) = delete;
 	FramebufferManagerCommon(Draw::DrawContext *draw);
 	virtual ~FramebufferManagerCommon();
 

GPU/Common/IndexGenerator.cpp

@@ -342,6 +342,7 @@ void IndexGenerator::TranslatePrim(int prim, int numInds, const u16_le *inds, in
 	}
 }
 
+// The PSP ignores the upper 16 bits of 32-bit indices. The u16 output drops them the same way.
 void IndexGenerator::TranslatePrim(int prim, int numInds, const u32_le *inds, int indexOffset, bool clockwise) {
 	switch (prim) {
 	case GE_PRIM_POINTS: TranslatePoints<u32_le>(numInds, inds, indexOffset); break;

GPU/Common/PresentationCommon.cpp

@@ -555,8 +555,12 @@ static void DoRelease(T *&obj) {
 
 template <typename T>
 static void DoReleaseVector(std::vector<T *> &list) {
-	for (auto &obj : list)
+	for (auto &obj : list) {
-		obj->Release();
+		// Can be null when a creation failed partway.
+		if (obj) {
+			obj->Release();
+		}
+	}
 	list.clear();
 }
 

GPU/Common/PresentationCommon.h

@@ -80,6 +80,8 @@ ENUM_CLASS_BITOPS(OutputFlags);
 
 class PresentationCommon {
 public:
+	PresentationCommon(const PresentationCommon &) = delete;
+	PresentationCommon &operator=(const PresentationCommon &) = delete;
 	PresentationCommon(Draw::DrawContext *draw);
 	~PresentationCommon();
 

GPU/Common/ReplacedTexture.cpp

@@ -278,6 +278,14 @@ void ReplacedTexture::Prepare(VFSBackend *vfs) {
 		}
 
 		result = LoadLevelData(fileRef, desc_.filenames[i], i, &pixelFormat);
+		// A level that got loaded owns the reference. Otherwise (error, or nothing to load) it's ours to free.
+		bool kept = false;
+		for (const ReplacedTextureLevel &level : levels_) {
+			kept = kept || level.fileRef == fileRef;
+		}
+		if (!kept) {
+			vfs_->ReleaseFile(fileRef);
+		}
 		if (result == LoadLevelResult::DONE) {
 			// Loaded all the levels we're gonna get.
 			fmt = pixelFormat;
@@ -730,6 +738,7 @@ ReplacedTexture::LoadLevelResult ReplacedTexture::LoadLevelData(VFSFileReference
 		}
 		if (png.width > (uint32_t)level.w || png.height > (uint32_t)level.h) {
 			ERROR_LOG(Log::TexReplacement, "Texture replacement changed since header read: %s", filename.c_str());
+			png_image_free(&png);
 			return LoadLevelResult::LOAD_ERROR;
 		}
 

GPU/Common/ReplacedTexture.h

@@ -140,6 +140,8 @@ struct ReplacedTextureLevel {
 
 class ReplacedTexture {
 public:
+	ReplacedTexture(const ReplacedTexture &) = delete;
+	ReplacedTexture &operator=(const ReplacedTexture &) = delete;
 	ReplacedTexture(VFSBackend *vfs, const ReplacementDesc &desc);
 	~ReplacedTexture();
 

GPU/Common/ShaderCommon.h

@@ -123,6 +123,8 @@ enum : uint64_t {
 
 class ShaderManagerCommon {
 public:
+	ShaderManagerCommon(const ShaderManagerCommon &) = delete;
+	ShaderManagerCommon &operator=(const ShaderManagerCommon &) = delete;
 	ShaderManagerCommon(Draw::DrawContext *draw) : draw_(draw) {}
 	virtual ~ShaderManagerCommon() {}
 

GPU/Common/SoftwareTransformCommon.cpp

@@ -198,7 +198,7 @@ SoftwareTransformAction RunSoftwareTransform(SoftwareTransformParams &params, in
 				float depth = std::clamp(transformed[1].z, 0.0f, 65535.0f) / 65535.0f;
 				// Non-zero depth clears are unusual, but some drivers don't match drawn depth values to cleared values.
 				// Games sometimes expect exact matches (see #12626, for example) for equal comparisons.
-				if (!(params.everUsedEqualDepth && gstate.isClearModeDepthMask() && result->depth > 0.0f && result->depth < 1.0f)) {
+				if (!(params.everUsedEqualDepth && gstate.isClearModeDepthMask() && depth > 0.0f && depth < 1.0f)) {
 					result->color = transformed[1].color0_32;
 					result->depth = depth;
 					gpuStats.perFrame.numClears++;

GPU/Common/TextureCacheCommon.cpp

@@ -114,6 +114,8 @@ TextureCacheCommon::TextureCacheCommon(Draw::DrawContext *draw, Draw2D *draw2D)
 }
 
 TextureCacheCommon::~TextureCacheCommon() {
+	// Only DeviceLost cleared these, which the GLES and D3D11 backends don't go through at shutdown.
+	clutTextureCache_.Clear();
 	FreeAlignedMemory(clutBufConverted_);
 	FreeAlignedMemory(clutBufRaw_);
 	FreeAlignedMemory(expandClut_);
@@ -342,6 +344,9 @@ SamplerCacheKey TextureCacheCommon::GetSamplingParams(int maxLevel, const TexCac
 		break;
 	}
 
+	if (key.aniso) {
+		key.anisoLevel = (uint8_t)g_Config.iAnisotropyLevel;
+	}
 	return key;
 }
 

GPU/Common/TextureCacheCommon.h

@@ -86,6 +86,8 @@ struct SamplerCacheKey {
 			bool tClamp : 1;
 			bool aniso : 1;
 			bool texture3d : 1;
+			// Baked into the sampler objects, so it has to be part of the key, or a change wouldn't apply.
+			uint8_t anisoLevel;
 		};
 	};
 	bool operator < (const SamplerCacheKey &other) const {
@@ -158,6 +160,9 @@ ENUM_CLASS_BITOPS(TexStatus);
 
 // TODO: Shrink this struct. There is some fluff.
 struct TexCacheEntry {
+	TexCacheEntry() = default;
+	TexCacheEntry(const TexCacheEntry &) = delete;
+	TexCacheEntry &operator=(const TexCacheEntry &) = delete;
 	~TexCacheEntry() {
 #ifdef _DEBUG
 		if (texturePtr || textureName || vkTex)
@@ -336,6 +341,8 @@ struct TextureApplyResult {
 
 class TextureCacheCommon {
 public:
+	TextureCacheCommon(const TextureCacheCommon &) = delete;
+	TextureCacheCommon &operator=(const TextureCacheCommon &) = delete;
 	TextureCacheCommon(Draw::DrawContext *draw, Draw2D *draw2D);
 	virtual ~TextureCacheCommon();
 

GPU/Common/TextureReplacer.cpp

@@ -230,6 +230,7 @@ bool TextureReplacer::LoadIni(std::string *error, bool notify) {
 
 			if (filenameMap.empty()) {
 				WARN_LOG(Log::TexReplacement, "No replacement textures found.");
+				delete dir;
 				return false;
 			}
 

GPU/Common/TextureReplacer.h

@@ -76,6 +76,8 @@ enum class ReplacerDecimateMode {
 
 class TextureReplacer {
 public:
+	TextureReplacer(const TextureReplacer &) = delete;
+	TextureReplacer &operator=(const TextureReplacer &) = delete;
 	// The draw context is checked for supported texture formats.
 	TextureReplacer(Draw::DrawContext *draw);
 	~TextureReplacer();

GPU/Common/TextureShaderCommon.cpp

@@ -252,7 +252,9 @@ ClutTexture ClutTextureCache::GetClutTexture(GEPaletteFormat clutFormat, const u
 
 void ClutTextureCache::Clear() {
 	for (auto tex = texCache_.begin(); tex != texCache_.end(); ++tex) {
-		tex->second->texture->Release();
+		if (tex->second->texture) {
+			tex->second->texture->Release();
+		}
 		delete tex->second;
 	}
 	texCache_.clear();
@@ -261,7 +263,9 @@ void ClutTextureCache::Clear() {
 void ClutTextureCache::Decimate() {
 	for (auto tex = texCache_.begin(); tex != texCache_.end(); ) {
 		if (tex->second->lastFrame + DEPAL_TEXTURE_OLD_AGE < gpuStats.totals.numFlips) {
-			tex->second->texture->Release();
+			if (tex->second->texture) {
+				tex->second->texture->Release();
+			}
 			delete tex->second;
 			texCache_.erase(tex++);
 		} else {

GPU/Common/VertexDecoderCommon.cpp

@@ -170,8 +170,9 @@ void GetIndexBounds(const void *inds, int count, u32 vertType, u16 *indexLowerBo
 		bool oob = false;
 		const u32_le *ind32 = (const u32_le *)inds;
 		for (int i = 0; i < count; i++) {
+			// The PSP ignores the upper 16 bits, so only the low ones count.
 			const u16 value = (u16)ind32[i];
-			// These aren't documented and should be rare.  Let's bounds check each one.
+			// Games setting the upper bits should be rare, so report them.
 			if (ind32[i] != value) {
 				oob = true;
 			}
@@ -1373,6 +1374,12 @@ void VertexDecoder::SetVertexType(u32 fmt, const VertexDecoderOptions &options,
 	// Attempt to JIT as well. But only do that if the main CPU JIT is enabled, in order to aid
 	// debugging attempts - if the main JIT doesn't work, this one won't do any better, probably.
 	if (jitCache) {
+		// Compile doesn't check for space. We can't clear the cache here since other decoders point into it,
+		// so when it's full (only seen with garbage display lists), new decoders use the interpreter.
+		if (jitCache->GetSpaceLeft() < 4096) {
+			WARN_LOG(Log::G3D, "Vertex decoder JIT cache full, using the interpreter for %08x", fmt_);
+			return;
+		}
 		jitted_ = jitCache->Compile(*this, &jittedSize_);
 		if (!jitted_) {
 			WARN_LOG(Log::G3D, "Vertex decoder JIT failed! fmt = %08x (%s)", fmt_, GetString(SHADER_STRING_SHORT_DESC).c_str());

GPU/Common/VertexDecoderCommon.h

@@ -115,7 +115,8 @@ public:
 		case GE_VTYPE_IDX_16BIT:
 			return indices16[index];
 		case GE_VTYPE_IDX_32BIT:
-			return indices32[index];
+			// The PSP supports 32-bit indices in name only: it ignores the upper 16 bits.
+			return indices32[index] & 0xFFFF;
 		default:
 			return index;
 		}

GPU/D3D11/DrawEngineD3D11.h

@@ -63,7 +63,9 @@ public:
 	void Flush() override;
 
 	void FinishDeferred() {
-		DecodeVerts(dec_, decoded_);
+		// Decoding only the vertices isn't enough: the indices are still read from PSP memory at flush
+		// time, and the game may change them once it regains control (#10095).
+		Flush();
 	}
 
 	void NotifyConfigChanged() override;

GPU/D3D11/TextureCacheD3D11.cpp

@@ -90,7 +90,7 @@ HRESULT SamplerCacheD3D11::GetOrCreateSampler(ID3D11Device *device, const Sample
 	samp.AddressV = key.tClamp ? D3D11_TEXTURE_ADDRESS_CLAMP : D3D11_TEXTURE_ADDRESS_WRAP;
 	samp.AddressW = samp.AddressU;  // Mali benefits from all clamps being the same, and this one is irrelevant.
 	if (key.aniso) {
-		samp.MaxAnisotropy = (float)(1 << g_Config.iAnisotropyLevel);
+		samp.MaxAnisotropy = (float)(1 << key.anisoLevel);
 	} else {
 		samp.MaxAnisotropy = 1.0f;
 	}

GPU/Debugger/Record.cpp

@@ -126,13 +126,16 @@ void Recorder::DirtyDrawnVRAM() {
 }
 
 bool Recorder::BeginRecording() {
+	std::unique_lock<std::mutex> guard(callbackLock_);
+	nextFrame = false;
 	if (PSP_CoreParameter().fileType == IdentifiedFileType::PPSSPP_GE_DUMP) {
-		// Can't record a GE dump.
+		// Can't record a GE dump. RecordNextFrame refuses this too.
+		writeCallback = nullptr;
 		return false;
 	}
 
 	active = true;
-	nextFrame = false;
+	guard.unlock();
 	lastTextures.clear();
 	lastRenderTargets.clear();
 	flipLastAction = gpuStats.totals.numFlips;
@@ -180,6 +183,10 @@ Path Recorder::WriteRecording() {
 	NOTICE_LOG(Log::G3D, "Recording filename: %s", filename.c_str());
 
 	FILE *fp = File::OpenCFile(filename, "wb");
+	if (!fp) {
+		ERROR_LOG(Log::G3D, "Failed to open '%s' for writing the recording", filename.c_str());
+		return Path();
+	}
 	Header header{};
 	memcpy(header.magic, HEADER_MAGIC, sizeof(header.magic));
 	header.version = VERSION;
@@ -574,14 +581,19 @@ void Recorder::EmitBezierSpline(u32 op) {
 }
 
 bool Recorder::RecordNextFrame(const std::function<void(const Path &)> callback) {
-	if (!nextFrame) {
+	if (PSP_CoreParameter().fileType == IdentifiedFileType::PPSSPP_GE_DUMP) {
-		flipLastAction = gpuStats.totals.numFlips;
+		return false;
-		flipFinishAt = -1;
-		writeCallback = callback;
-		nextFrame = true;
-		return true;
 	}
-	return false;
+	std::lock_guard<std::mutex> guard(callbackLock_);
+	// Don't take over a recording in progress, it would get the wrong callback and end point.
+	if (nextFrame || active) {
+		return false;
+	}
+	flipLastAction = gpuStats.totals.numFlips;
+	flipFinishAt = -1;
+	writeCallback = callback;
+	nextFrame = true;
+	return true;
 }
 
 void Recorder::FinishRecording() {
@@ -596,15 +608,21 @@ void Recorder::FinishRecording() {
 	lastVRAM.clear();
 
 	NOTICE_LOG(Log::System, "Recording finished");
-	active = false;
 	flipLastAction = gpuStats.totals.numFlips;
 	flipFinishAt = -1;
 	lastEdramTrans = 0x400;
 
-	if (writeCallback) {
+	std::function<void(const Path &)> callback;
-		writeCallback(filename);
+	{
+		std::lock_guard<std::mutex> guard(callbackLock_);
+		callback = std::move(writeCallback);
+		writeCallback = nullptr;
+		active = false;
+	}
+
+	if (callback && !filename.empty()) {
+		callback(filename);
 	}
-	writeCallback = nullptr;
 }
 
 void Recorder::CheckEdramTrans() {

GPU/Debugger/Record.h

@@ -19,6 +19,7 @@
 
 #include <functional>
 #include <atomic>
+#include <mutex>
 #include <vector>
 #include <set>
 
@@ -50,7 +51,7 @@ public:
 	}
 	bool RecordNextFrame(const std::function<void(const Path &)> callback);
 	void ClearCallback() {
-		// Not super thread safe..
+		std::lock_guard<std::mutex> guard(callbackLock_);
 		writeCallback = nullptr;
 	}
 
@@ -95,6 +96,8 @@ private:
 	int flipFinishAt = -1;
 	uint32_t lastEdramTrans = 0x400;
 	std::function<void(const Path &)> writeCallback;
+	// RecordNextFrame is called from other threads. Guards writeCallback, nextFrame and the writes to active.
+	std::mutex callbackLock_;
 
 	std::vector<u8> pushbuf;
 	std::vector<Command> commands;

GPU/GLES/DrawEngineGLES.cpp

@@ -229,6 +229,7 @@ void DrawEngineGLES::Flush() {
 		numDrawVerts_ = 0;
 		numDrawInds_ = 0;
 		vertexCountInDrawCalls_ = 0;
+		numVertsToDecode_ = 0;
 		decodeVertsCounter_ = 0;
 		decodeIndsCounter_ = 0;
 		return;

GPU/GLES/FragmentTestCacheGLES.cpp

@@ -144,7 +144,7 @@ GLRTexture *FragmentTestCacheGLES::CreateTestTexture(const GEComparison funcs[4]
 }
 
 void FragmentTestCacheGLES::Clear(bool deleteThem) {
-	if (deleteThem) {
+	if (deleteThem && render_) {
 		for (const auto &[_, v] : cache_) {
 			render_->DeleteTexture(v.texture);
 		}

GPU/GLES/FragmentTestCacheGLES.h

@@ -68,7 +68,8 @@ public:
 	void BindTestTexture(int slot);
 
 	void DeviceLost() {
-		Clear(false);
+		// Queue the deletes anyway, the deleter frees the GLRTexture objects and skips the GL calls if needed.
+		Clear(true);
 		render_ = nullptr;
 	}
 	void DeviceRestore(Draw::DrawContext *draw);

GPU/GLES/ShaderManagerGLES.h

@@ -37,6 +37,8 @@ class IOFile;
 
 class LinkedShader {
 public:
+	LinkedShader(const LinkedShader &) = delete;
+	LinkedShader &operator=(const LinkedShader &) = delete;
 	LinkedShader(GLRenderManager *render, VShaderID VSID, Shader *vs, FShaderID FSID, Shader *fs, bool useHWTransform, bool preloading = false);
 	~LinkedShader();
 
@@ -139,6 +141,8 @@ struct ShaderDescGLES {
 
 class Shader {
 public:
+	Shader(const Shader &) = delete;
+	Shader &operator=(const Shader &) = delete;
 	Shader(GLRenderManager *render, const char *code, const std::string &desc, const ShaderDescGLES &params);
 	~Shader();
 	GLRShader *shader;

GPU/GLES/TextureCacheGLES.cpp

@@ -51,10 +51,10 @@ void TextureCacheGLES::SetFramebufferManager(FramebufferManagerGLES *fbManager)
 }
 
 void TextureCacheGLES::ReleaseTexture(TexCacheEntry *entry, bool delete_them) {
-	if (delete_them) {
+	// Delete even when !delete_them (device lost): the GLRTexture is a heap object that only the deleter
-		if (entry->textureName) {
+	// frees, and the deleter skips the GL calls itself once the context is gone.
-			render_->DeleteTexture(entry->textureName);
+	if (entry->textureName && render_) {
-		}
+		render_->DeleteTexture(entry->textureName);
 	}
 	entry->textureName = nullptr;
 }

GPU/GPUCommon.cpp

@@ -957,6 +957,7 @@ DLResult GPUCommon::ProcessDLQueue() {
 				// Nothing to execute here, and leaving it at the head of the queue would block everything
 				// behind it for good.  Treat it like a list that ran into an error.
 				ERROR_LOG(Log::G3D, "Display list %d has a bad pc %08x (state %d), dropping it", listIndex, list.pc, (int)list.state);
+				CompleteFailedList(list);
 				dlQueue.erase(std::remove(dlQueue.begin(), dlQueue.end(), listIndex), dlQueue.end());
 				continue;
 			}
@@ -1045,7 +1046,8 @@ DLResult GPUCommon::ProcessDLQueue() {
 			}
 			break;
 		case GPUSTATE_ERROR:
-			// don't do anything - though dunno about error...
+			// The list can't continue. It's removed from the queue below.
+			CompleteFailedList(list);
 			break;
 		case GPUSTATE_STALL:
 			// Resume work on this same display list later. The GE is still busy with what it has
@@ -1390,6 +1392,20 @@ void GPUCommon::Execute_End(u32 op, u32 diff) {
 	}
 }
 
+// A list dropped on an error still has to complete like a finished one, or its ID is never freed and
+// sceGeListSync on it never returns.
+void GPUCommon::CompleteFailedList(DisplayList &list) {
+	if (list.started && list.context.IsValid()) {
+		gstate.Restore(list.context);
+		ReapplyGfxState();
+		list.started = false;
+	}
+	list.state = PSP_GE_DL_STATE_COMPLETED;
+	list.waitUntilTicks = startingTicks + cyclesExecuted;
+	busyTicks = std::max(busyTicks, list.waitUntilTicks);
+	__GeTriggerSync(GPU_SYNC_LIST, list.id, list.waitUntilTicks);
+}
+
 void GPUCommon::Execute_BoundingBox(u32 op, u32 diff) {
 	// Just resetting, nothing to check bounds for.
 	const u32 count = op & 0xFFFF;
@@ -1552,8 +1568,10 @@ void GPUCommon::FlushImm() {
 
 	bool changed = texturing != prevTexturing || cullEnable != prevCullEnable || dither != prevDither;
 	changed = changed || prevShading != shading || prevFog != fog;
+	// Always flush, even if the flags match: DispatchSubmitImm switches to through mode and a different
+	// vertex decoder, which would otherwise apply to the draws already queued.
+	Flush();
 	if (changed) {
-		Flush();
 		gstate.antiAliasEnable = (GE_CMD_ANTIALIASENABLE << 24) | (int)antialias;
 		gstate.shademodel = (GE_CMD_SHADEMODE << 24) | (int)shading;
 		gstate.cullfaceEnable = (GE_CMD_CULLFACEENABLE << 24) | (int)cullEnable;
@@ -2290,12 +2308,16 @@ bool GPUCommon::NeedsSlowInterpreter() const {
 void GPUCommon::ClearBreakNext() {
 	breakNext_ = GPUDebug::BreakNext::NONE;
 	breakAtCount_ = -1;
+	// A step that never reached its target leaves these behind, and they'd trip unexpectedly later.
+	breakpoints_.ClearTempBreakpoints();
 	GPUStepping::ResumeFromStepping();
 }
 
 void GPUCommon::SetBreakNext(GPUDebug::BreakNext next) {
 	breakNext_ = next;
 	breakAtCount_ = -1;
+	// Drop the ones from a previous step that didn't get there, before adding this one's.
+	breakpoints_.ClearTempBreakpoints();
 	switch (next) {
 	case GPUDebug::BreakNext::TEX:
 		breakpoints_.AddTextureChangeTempBreakpoint();

GPU/GPUCommon.h

@@ -56,6 +56,8 @@ inline bool IsTrianglePrim(GEPrimitiveType prim) {
 struct TransformStats;
 class GPUCommon {
 public:
+	GPUCommon(const GPUCommon &) = delete;
+	GPUCommon &operator=(const GPUCommon &) = delete;
 	// The constructor might run on the loader thread.
 	GPUCommon(GraphicsContext *gfxCtx, Draw::DrawContext *draw);
 	virtual ~GPUCommon() = default;
@@ -452,4 +454,5 @@ protected:
 private:
 	void DoExecuteCall(u32 target);
 	void PopDLQueue();
+	void CompleteFailedList(DisplayList &list);
 };

GPU/Software/BinManager.cpp

@@ -246,21 +246,32 @@ void BinManager::UpdateState() {
 		if (newMaxTasks > MAX_POSSIBLE_TASKS)
 			newMaxTasks = MAX_POSSIBLE_TASKS;
 		// We don't want to overlap wrong, so flush any pending.
+		bool flushed = false;
 		if (maxTasks_ != newMaxTasks) {
 			maxTasks_ = newMaxTasks;
 			Flush("selfrender");
+			flushed = true;
 		}
-		pendingOverlap_ = pendingOverlap_ || selfRender;
 
 		// Lastly, we have to check if we're newly writing depth we were texturing before.
 		// This happens in Call of Duty (depth clear after depth texture), for example.
-		if (!hadDepth && state.pixelID.depthWrite) {
+		if (!flushed && !hadDepth && state.pixelID.depthWrite) {
 			for (size_t i = 0; i < states_.Size(); ++i) {
 				if (HasTextureWrite(states_.Peek(i))) {
 					Flush("selfdepth");
+					flushed = true;
+					break;
 				}
 			}
 		}
+
+		if (flushed) {
+			// The flush forgot what this draw writes and reads, so record it again.
+			MarkPendingWrites(state);
+			MarkPendingReads(state);
+			ClearDirty(SoftDirty::BINNER_RANGE);
+		}
+		pendingOverlap_ = pendingOverlap_ || selfRender;
 		ClearDirty(SoftDirty::BINNER_OVERLAP);
 	}
 }

GPU/Software/BinManager.h

@@ -57,6 +57,8 @@ struct BinItem {
 
 template <typename T, size_t N>
 struct BinQueue {
+	BinQueue(const BinQueue &) = delete;
+	BinQueue &operator=(const BinQueue &) = delete;
 	BinQueue() {
 		Reset();
 	}
@@ -185,6 +187,8 @@ struct BinDirtyRange {
 class StringWriter;
 class BinManager {
 public:
+	BinManager(const BinManager &) = delete;
+	BinManager &operator=(const BinManager &) = delete;
 	BinManager();
 	~BinManager();
 

GPU/Software/SoftGpu.cpp

@@ -826,8 +826,8 @@ void SoftGPU::Execute_BlockTransferStart(u32 op, u32 diff) {
 
 	// Need to flush both source and target, so we overwrite properly.
 	if (Memory::IsValidRange(src, srcSize) && Memory::IsValidRange(dst, dstSize)) {
-		drawEngine_->transformUnit.FlushIfOverlap(this, "blockxfer", false, src, srcStride, width * bpp, height);
+		drawEngine_->transformUnit.FlushIfOverlap(this, "blockxfer", false, src, srcStride * bpp, width * bpp, height);
-		drawEngine_->transformUnit.FlushIfOverlap(this, "blockxfer", true, dst, dstStride, width * bpp, height);
+		drawEngine_->transformUnit.FlushIfOverlap(this, "blockxfer", true, dst, dstStride * bpp, width * bpp, height);
 	} else {
 		drawEngine_->transformUnit.Flush(this, "blockxfer_wrap");
 	}

GPU/Software/TransformUnit.h

@@ -111,6 +111,8 @@ class StringWriter;
 
 class TransformUnit {
 public:
+	TransformUnit(const TransformUnit &) = delete;
+	TransformUnit &operator=(const TransformUnit &) = delete;
 	TransformUnit();
 	~TransformUnit();
 

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -550,6 +550,7 @@ void DrawEngineVulkan::ResetAfterSkippedDraw() {
 	numDrawVerts_ = 0;
 	numDrawInds_ = 0;
 	vertexCountInDrawCalls_ = 0;
+	numVertsToDecode_ = 0;
 	decodeIndsCounter_ = 0;
 	decodeVertsCounter_ = 0;
 	gstate_c.vertexFullAlpha = true;

GPU/Vulkan/GPU_Vulkan.cpp

@@ -90,7 +90,7 @@ GPU_Vulkan::GPU_Vulkan(GraphicsContext *gfxCtx, Draw::DrawContext *draw)
 	if (discID.size()) {
 		File::CreateFullPath(GetSysDirectory(DIRECTORY_APP_CACHE));
 		shaderCachePath_ = GetSysDirectory(DIRECTORY_APP_CACHE) / (discID + ".vkshadercache");
-		LoadCache(shaderCachePath_);
+		LoadCache(shaderCachePath_, true);
 	}
 
 	InitDeviceObjects();
@@ -101,7 +101,7 @@ void GPU_Vulkan::FinishInitOnMainThread() {
 	framebufferManagerVulkan_->Init(msaaLevel_);
 }
 
-void GPU_Vulkan::LoadCache(const Path &filename) {
+void GPU_Vulkan::LoadCache(const Path &filename, bool waitForPipelines) {
 	_dbg_assert_(draw_);
 	if (!g_Config.bShaderCache) {
 		WARN_LOG(Log::G3D, "Shader cache disabled. Not loading.");
@@ -134,13 +134,15 @@ void GPU_Vulkan::LoadCache(const Path &filename) {
 	}
 	fclose(f);
 
-	// Now, since we're on the loader thread, we can just block here until all pipelines are actually created.
+	if (waitForPipelines) {
-	// This makes it so that the on-screen spinner keeps spinning until we are done.
+		// Now, since we're on the loader thread, we can just block here until all pipelines are actually created.
-	double start = time_now_d();
+		// This makes it so that the on-screen spinner keeps spinning until we are done.
-	VulkanRenderManager *rm = (VulkanRenderManager *)draw_->GetNativeObject(Draw::NativeObject::RENDER_MANAGER);
+		double start = time_now_d();
-	int maxTasksSeen = rm->WaitForPipelines();
+		VulkanRenderManager *rm = (VulkanRenderManager *)draw_->GetNativeObject(Draw::NativeObject::RENDER_MANAGER);
-	double seconds = time_now_d() - start;
+		int maxTasksSeen = rm->WaitForPipelines();
-	INFO_LOG(Log::G3D, "Waited %0.1fms for at least %d pipeline tasks to finish compiling.", seconds * 1000.0, maxTasksSeen);
+		double seconds = time_now_d() - start;
+		INFO_LOG(Log::G3D, "Waited %0.1fms for at least %d pipeline tasks to finish compiling.", seconds * 1000.0, maxTasksSeen);
+	}
 
 	if (!result) {
 		WARN_LOG(Log::G3D, "Incompatible Vulkan pipeline cache - rebuilding.");
@@ -402,6 +404,12 @@ void GPU_Vulkan::DeviceRestore(Draw::DrawContext *draw) {
 	pipelineManager_->DeviceRestore(vulkan);
 
 	InitDeviceObjects();
+
+	// DeviceLost saved the cache and then threw everything away. Load it again, or the next save would
+	// overwrite the file with only what gets drawn from now on.
+	if (shaderCachePath_.Valid()) {
+		LoadCache(shaderCachePath_, false);
+	}
 }
 
 void GPU_Vulkan::GetStats(StringWriter &w) {

GPU/Vulkan/GPU_Vulkan.h

@@ -69,7 +69,7 @@ private:
 	void InitDeviceObjects();
 	void DestroyDeviceObjects();
 
-	void LoadCache(const Path &filename);
+	void LoadCache(const Path &filename, bool waitForPipelines);
 	void SaveCache(const Path &filename);
 
 	FramebufferManagerVulkan *framebufferManagerVulkan_;

GPU/Vulkan/PipelineManagerVulkan.h

@@ -63,6 +63,9 @@ private:
 
 // Simply wraps a Vulkan pipeline, providing some metadata.
 struct VulkanPipeline {
+	VulkanPipeline() = default;
+	VulkanPipeline(const VulkanPipeline &) = delete;
+	VulkanPipeline &operator=(const VulkanPipeline &) = delete;
 	~VulkanPipeline() {
 		desc->Release();
 	}

GPU/Vulkan/ShaderManagerVulkan.h

@@ -40,6 +40,8 @@ class VulkanPushPool;
 
 class VulkanFragmentShader {
 public:
+	VulkanFragmentShader(const VulkanFragmentShader &) = delete;
+	VulkanFragmentShader &operator=(const VulkanFragmentShader &) = delete;
 	VulkanFragmentShader(VulkanContext *vulkan, FShaderID id, FragmentShaderFlags flags, const char *code, SPIRVCache *cache);
 	~VulkanFragmentShader();
 
@@ -63,6 +65,8 @@ protected:
 
 class VulkanVertexShader {
 public:
+	VulkanVertexShader(const VulkanVertexShader &) = delete;
+	VulkanVertexShader &operator=(const VulkanVertexShader &) = delete;
 	VulkanVertexShader(VulkanContext *vulkan, VShaderID id, VertexShaderFlags flags, const char *code, bool useHWTransform, SPIRVCache *cache);
 	~VulkanVertexShader();
 

GPU/Vulkan/TextureCacheVulkan.cpp

@@ -143,7 +143,7 @@ VkSampler SamplerCache::GetOrCreateSampler(const SamplerCacheKey &key) {
 
 	if (key.aniso) {
 		// Docs say the min of this value and the supported max are used.
-		samp.maxAnisotropy = 1 << g_Config.iAnisotropyLevel;
+		samp.maxAnisotropy = 1 << key.anisoLevel;
 		samp.anisotropyEnable = true;
 	} else {
 		samp.maxAnisotropy = 1.0f;

GPU/Vulkan/TextureCacheVulkan.h

@@ -37,6 +37,8 @@ class StringWriter;
 
 class SamplerCache {
 public:
+	SamplerCache(const SamplerCache &) = delete;
+	SamplerCache &operator=(const SamplerCache &) = delete;
 	SamplerCache(VulkanContext *vulkan) : vulkan_(vulkan), cache_(16) {}
 	~SamplerCache();
 	VkSampler GetOrCreateSampler(const SamplerCacheKey &key);

GPU/ge_constants.h

@@ -330,7 +330,7 @@ enum GEVertexType : uint32_t {
 	GE_VTYPE_IDX_NONE  = (0<<11),
 	GE_VTYPE_IDX_8BIT  = (1<<11),
 	GE_VTYPE_IDX_16BIT = (2<<11),
-	GE_VTYPE_IDX_32BIT = (3<<11),
+	GE_VTYPE_IDX_32BIT = (3<<11),  // In name only: the hardware ignores the upper 16 bits of each index.
 	GE_VTYPE_IDX_MASK  = (3<<11),
 #define GE_VTYPE_IDX_SHIFT 11
 };

UI/ImDebugger/ImGe.cpp

@@ -96,9 +96,14 @@ void DrawFramebuffersWindow(ImConfig &cfg, FramebufferManagerCommon *framebuffer
 		ImGui::SliderFloat("Scale", &cfg.fbViewerZoom, 0.5f, 16.0f, "%.2f", ImGuiSliderFlags_Logarithmic);
 
 		// Now, draw the image of the selected framebuffer.
+		// Null without buffered rendering, or if creating it failed.
 		Draw::Framebuffer *fb = vfbs[cfg.selectedFramebuffer]->fbo;
-		ImTextureID texId = ImGui_ImplThin3d_AddFBAsTextureTemp(fb, Draw::Aspect::COLOR_BIT, ImGuiPipeline::TexturedOpaque);
+		if (fb) {
-		ImGui::Image(texId, ImVec2(fb->Width() * cfg.fbViewerZoom, fb->Height() * cfg.fbViewerZoom));
+			ImTextureID texId = ImGui_ImplThin3d_AddFBAsTextureTemp(fb, Draw::Aspect::COLOR_BIT, ImGuiPipeline::TexturedOpaque);
+			ImGui::Image(texId, ImVec2(fb->Width() * cfg.fbViewerZoom, fb->Height() * cfg.fbViewerZoom));
+		} else {
+			ImGui::TextUnformatted("(no framebuffer object)");
+		}
 	}
 
 	ImGui::End();
@@ -602,7 +607,8 @@ ImGeReadbackViewer::~ImGeReadbackViewer() {
 }
 
 VirtualFramebuffer *ImGeReadbackViewer::GetVFB(FramebufferManagerCommon *fbMan) const {
-	return fbMan->GetExactVFB(gstate.getFrameBufAddress(), gstate.FrameBufStride(), gstate.FrameBufFormat());
+	// fbMan is null with the software renderer.
+	return fbMan ? fbMan->GetExactVFB(gstate.getFrameBufAddress(), gstate.FrameBufStride(), gstate.FrameBufFormat()) : nullptr;
 }
 
 void ImGeReadbackViewer::DeviceLost() {