When overriding skinning in sw transform mode, don't write/restore member variables

GPU/Common/DrawEngineCommon.cpp

@@ -133,7 +133,6 @@ void DrawEngineCommon::NotifyConfigChanged() {
 
 	useHWTransform_ = g_Config.bHardwareTransform;
 	useHWTessellation_ = UpdateUseHWTessellation(g_Config.bHardwareTessellation);
-	applySkinInDecode_ = g_Config.bSoftwareSkinning;
 }
 
 u32 DrawEngineCommon::NormalizeVertices(u8 *outPtr, u8 *bufPtr, const u8 *inPtr, int lowerBound, int upperBound, u32 vertType, int *vertexSize) {
@@ -1038,7 +1037,14 @@ bool DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	return true;
 }
 
+void DrawEngineCommon::BeginFrame() {
+	applySkinInDecode_ = g_Config.bSoftwareSkinning;
+}
+
 void DrawEngineCommon::DecodeVerts(VertexDecoder *dec, u8 *dest) {
+	if (!numDrawVerts_) {
+		return;
+	}
 	// Note that this should be able to continue a partial decode - we don't necessarily start from zero here (although we do most of the time).
 	int i = decodeVertsCounter_;
 	int stride = (int)dec->GetDecVtxFmt().stride;

GPU/Common/DrawEngineCommon.h

@@ -77,6 +77,8 @@ public:
 
 	void Init();
 
+	virtual void BeginFrame();
+
 	void SetGPUCommon(GPUCommon *gpuCommon) {
 		gpuCommon_ = gpuCommon;
 	}

GPU/D3D11/DrawEngineD3D11.cpp

@@ -239,6 +239,8 @@ ID3D11InputLayout *DrawEngineD3D11::SetupDecFmtForDraw(D3D11VertexShader *vshade
 }
 
 void DrawEngineD3D11::BeginFrame() {
+	DrawEngineCommon::BeginFrame();
+
 	pushVerts_->Reset();
 	pushInds_->Reset();
 
@@ -346,12 +348,15 @@ void DrawEngineD3D11::Flush() {
 		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");
-		if (!applySkinInDecode_) {
-			applySkinInDecode_ = true;
-			lastVType_ |= (1 << 26);
-			dec_ = GetVertexDecoder(lastVType_);
+		VertexDecoder *swDec = dec_;
+		if (swDec->nweights != 0) {
+			u32 withSkinning = lastVType_ | (1 << 26);
+			if (withSkinning != lastVType_) {
+				swDec = GetVertexDecoder(withSkinning);
+			}
 		}
-		DecodeVerts(dec_, decoded_);
+
+		DecodeVerts(swDec, decoded_);
 		int vertexCount = DecodeInds();
 
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
@@ -400,7 +405,7 @@ void DrawEngineD3D11::Flush() {
 		const Lin::Vec3 scale(gstate_c.vpWidthScale, -gstate_c.vpHeightScale, gstate_c.vpDepthScale * 0.5f);
 		swTransform.SetProjMatrix(gstate.projMatrix, gstate_c.vpWidth < 0, gstate_c.vpHeight < 0, trans, scale);
 
-		swTransform.Transform(prim, dec_->VertexType(), dec_->GetDecVtxFmt(), numDecodedVerts_, &result);
+		swTransform.Transform(prim, swDec->VertexType(), swDec->GetDecVtxFmt(), numDecodedVerts_, &result);
 		// 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 (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
@@ -416,7 +421,7 @@ void DrawEngineD3D11::Flush() {
 		ApplyDrawState(prim);
 
 		if (result.action == SW_NOT_READY)
-			swTransform.BuildDrawingParams(prim, vertexCount, dec_->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
+			swTransform.BuildDrawingParams(prim, vertexCount, swDec->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
 		if (result.setSafeSize)
 			framebufferManager_->SetSafeSize(result.safeWidth, result.safeHeight);
 
@@ -425,7 +430,7 @@ void DrawEngineD3D11::Flush() {
 		if (result.action == SW_DRAW_INDEXED) {
 			D3D11VertexShader *vshader;
 			D3D11FragmentShader *fshader;
-			shaderManager_->GetShaders(prim, dec_, &vshader, &fshader, pipelineState_, false, false, decOptions_.expandAllWeightsToFloat, true);
+			shaderManager_->GetShaders(prim, swDec, &vshader, &fshader, pipelineState_, false, false, decOptions_.expandAllWeightsToFloat, true);
 			context_->PSSetShader(fshader->GetShader(), nullptr, 0);
 			context_->VSSetShader(vshader->GetShader(), nullptr, 0);
 			shaderManager_->UpdateUniforms(framebufferManager_->UseBufferedRendering());
@@ -478,7 +483,6 @@ void DrawEngineD3D11::Flush() {
 				framebufferManager_->ApplyClearToMemory(scissorX1, scissorY1, scissorX2, scissorY2, clearColor);
 			}
 		}
-		applySkinInDecode_ = g_Config.bSoftwareSkinning;
 	}
 
 	ResetAfterDrawInline();

GPU/D3D11/DrawEngineD3D11.h

@@ -75,7 +75,7 @@ public:
 	void InitDeviceObjects();
 	void DestroyDeviceObjects();
 
-	void BeginFrame();
+	void BeginFrame() override;
 
 	void Flush() override;
 

GPU/Directx9/DrawEngineDX9.cpp

@@ -219,6 +219,7 @@ static uint32_t SwapRB(uint32_t c) {
 }
 
 void DrawEngineDX9::BeginFrame() {
+	DrawEngineCommon::BeginFrame();
 	lastRenderStepId_ = -1;
 }
 
@@ -301,12 +302,14 @@ void DrawEngineDX9::Flush() {
 			}
 		}
 	} else {
-		if (!applySkinInDecode_) {
-			applySkinInDecode_ = true;
-			lastVType_ |= (1 << 26);
-			dec_ = GetVertexDecoder(lastVType_);
+		VertexDecoder *swDec = dec_;
+		if (swDec->nweights != 0) {
+			u32 withSkinning = lastVType_ | (1 << 26);
+			if (withSkinning != lastVType_) {
+				swDec = GetVertexDecoder(withSkinning);
+			}
 		}
-		DecodeVerts(dec_, decoded_);
+		DecodeVerts(swDec, decoded_);
 		int vertexCount = DecodeInds();
 
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
@@ -361,7 +364,7 @@ void DrawEngineDX9::Flush() {
 		const Lin::Vec3 scale(gstate_c.vpWidthScale, gstate_c.vpHeightScale, gstate_c.vpDepthScale * 0.5f);
 		swTransform.SetProjMatrix(gstate.projMatrix, gstate_c.vpWidth < 0, gstate_c.vpHeight > 0, trans, scale);
 
-		swTransform.Transform(prim, dec_->VertexType(), dec_->GetDecVtxFmt(), numDecodedVerts_, &result);
+		swTransform.Transform(prim, swDec->VertexType(), swDec->GetDecVtxFmt(), numDecodedVerts_, &result);
 		// 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 (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
@@ -376,13 +379,13 @@ void DrawEngineDX9::Flush() {
 		ApplyDrawState(prim);
 
 		if (result.action == SW_NOT_READY)
-			swTransform.BuildDrawingParams(prim, vertexCount, dec_->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
+			swTransform.BuildDrawingParams(prim, vertexCount, swDec->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
 		if (result.setSafeSize)
 			framebufferManager_->SetSafeSize(result.safeWidth, result.safeHeight);
 
 		ApplyDrawStateLate();
 
-		VSShader *vshader = shaderManager_->ApplyShader(false, false, dec_, decOptions_.expandAllWeightsToFloat, true, pipelineState_);
+		VSShader *vshader = shaderManager_->ApplyShader(false, false, swDec, decOptions_.expandAllWeightsToFloat, true, pipelineState_);
 
 		if (result.action == SW_DRAW_INDEXED) {
 			if (result.setStencil) {
@@ -414,7 +417,6 @@ void DrawEngineDX9::Flush() {
 				framebufferManager_->ApplyClearToMemory(scissorX1, scissorY1, scissorX2, scissorY2, clearColor);
 			}
 		}
-		applySkinInDecode_ = g_Config.bSoftwareSkinning;
 	}
 
 	ResetAfterDrawInline();

GPU/Directx9/DrawEngineDX9.h

@@ -65,7 +65,7 @@ public:
 	void InitDeviceObjects();
 	void DestroyDeviceObjects();
 
-	void BeginFrame();
+	void BeginFrame() override;
 
 	// So that this can be inlined
 	void Flush() override;

GPU/GLES/DrawEngineGLES.cpp

@@ -147,6 +147,8 @@ void DrawEngineGLES::ClearInputLayoutMap() {
 }
 
 void DrawEngineGLES::BeginFrame() {
+	DrawEngineCommon::BeginFrame();
+
 	FrameData &frameData = frameData_[render_->GetCurFrame()];
 	frameData.pushIndex->Begin();
 	frameData.pushVertex->Begin();
@@ -328,12 +330,14 @@ void DrawEngineGLES::Flush() {
 		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");
-		if (!applySkinInDecode_) {
-			applySkinInDecode_ = true;
-			lastVType_ |= (1 << 26);
-			dec_ = GetVertexDecoder(lastVType_);
+		VertexDecoder *swDec = dec_;
+		if (swDec->nweights != 0) {
+			u32 withSkinning = lastVType_ | (1 << 26);
+			if (withSkinning != lastVType_) {
+				swDec = GetVertexDecoder(withSkinning);
+			}
 		}
-		DecodeVerts(dec_, decoded_);
+		DecodeVerts(swDec, decoded_);
 		int vertexCount = DecodeInds();
 
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
@@ -387,7 +391,7 @@ void DrawEngineGLES::Flush() {
 		const bool invertedY = gstate_c.vpHeight * (params.flippedY ? 1.0 : -1.0f) < 0;
 		swTransform.SetProjMatrix(gstate.projMatrix, gstate_c.vpWidth < 0, invertedY, trans, scale);
 
-		swTransform.Transform(prim, dec_->VertexType(), dec_->GetDecVtxFmt(), numDecodedVerts_, &result);
+		swTransform.Transform(prim, swDec->VertexType(), swDec->GetDecVtxFmt(), numDecodedVerts_, &result);
 		// 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 (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
@@ -403,7 +407,7 @@ void DrawEngineGLES::Flush() {
 		ApplyDrawState(prim);
 
 		if (result.action == SW_NOT_READY)
-			swTransform.BuildDrawingParams(prim, vertexCount, dec_->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
+			swTransform.BuildDrawingParams(prim, vertexCount, swDec->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
 		if (result.setSafeSize)
 			framebufferManager_->SetSafeSize(result.safeWidth, result.safeHeight);
 
@@ -447,7 +451,6 @@ void DrawEngineGLES::Flush() {
 			}
 			gstate_c.Dirty(DIRTY_BLEND_STATE);  // Make sure the color mask gets re-applied.
 		}
-		applySkinInDecode_ = g_Config.bSoftwareSkinning;
 	}
 
 bail:

GPU/GLES/DrawEngineGLES.h

@@ -81,7 +81,7 @@ public:
 
 	void ClearTrackedVertexArrays() override {}
 
-	void BeginFrame();
+	void BeginFrame() override;
 	void EndFrame();
 
 	// So that this can be inlined

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -164,6 +164,8 @@ void DrawEngineVulkan::DeviceRestore(Draw::DrawContext *draw) {
 }
 
 void DrawEngineVulkan::BeginFrame() {
+	DrawEngineCommon::BeginFrame();
+
 	lastPipeline_ = nullptr;
 
 	// These will be re-bound if needed, let's not let old bindings linger around too long.
@@ -381,15 +383,16 @@ void DrawEngineVulkan::Flush() {
 		}
 	} else {
 		PROFILE_THIS_SCOPE("soft");
-
-		if (!applySkinInDecode_) {
-			applySkinInDecode_ = true;
-			lastVType_ |= (1 << 26);
-			dec_ = GetVertexDecoder(lastVType_);
+		VertexDecoder *swDec = dec_;
+		if (swDec->nweights != 0) {
+			u32 withSkinning = lastVType_ | (1 << 26);
+			if (withSkinning != lastVType_) {
+				swDec = GetVertexDecoder(withSkinning);
+			}
 		}
 		int prevDecodedVerts = numDecodedVerts_;
 
-		DecodeVerts(dec_, decoded_);
+		DecodeVerts(swDec, decoded_);
 		int vertexCount = DecodeInds();
 
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;
@@ -445,7 +448,7 @@ void DrawEngineVulkan::Flush() {
 		const Lin::Vec3 scale(gstate_c.vpWidthScale, gstate_c.vpHeightScale, gstate_c.vpDepthScale * 0.5f);
 		swTransform.SetProjMatrix(gstate.projMatrix, gstate_c.vpWidth < 0, gstate_c.vpHeight < 0, trans, scale);
 
-		swTransform.Transform(prim, dec_->VertexType(), dec_->GetDecVtxFmt(), numDecodedVerts_, &result);
+		swTransform.Transform(prim, swDec->VertexType(), swDec->GetDecVtxFmt(), numDecodedVerts_, &result);
 		// 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 (result.action == SW_CLEAR && everUsedEqualDepth_ && gstate.isClearModeDepthMask() && result.depth > 0.0f && result.depth < 1.0f)
@@ -453,7 +456,7 @@ void DrawEngineVulkan::Flush() {
 
 		if (result.action == SW_NOT_READY) {
 			// decIndex_ here is always equal to inds currently, but it may not be in the future.
-			swTransform.BuildDrawingParams(prim, vertexCount, dec_->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
+			swTransform.BuildDrawingParams(prim, vertexCount, swDec->VertexType(), inds, RemainingIndices(inds), numDecodedVerts_, VERTEX_BUFFER_MAX, &result);
 		}
 
 		if (result.setSafeSize)
@@ -481,9 +484,9 @@ void DrawEngineVulkan::Flush() {
 				VulkanFragmentShader *fshader = nullptr;
 				VulkanGeometryShader *gshader = nullptr;
 
-				shaderManager_->GetShaders(prim, dec_, &vshader, &fshader, &gshader, pipelineState_, false, false, decOptions_.expandAllWeightsToFloat, true);
+				shaderManager_->GetShaders(prim, swDec, &vshader, &fshader, &gshader, pipelineState_, false, false, decOptions_.expandAllWeightsToFloat, true);
 				_dbg_assert_msg_(!vshader->UseHWTransform(), "Bad vshader");
-				VulkanPipeline *pipeline = pipelineManager_->GetOrCreatePipeline(renderManager, pipelineLayout_, pipelineKey_, &dec_->decFmt, vshader, fshader, gshader, false, 0, framebufferManager_->GetMSAALevel(), false);
+				VulkanPipeline *pipeline = pipelineManager_->GetOrCreatePipeline(renderManager, pipelineLayout_, pipelineKey_, &swDec->decFmt, vshader, fshader, gshader, false, 0, framebufferManager_->GetMSAALevel(), false);
 				if (!pipeline || !pipeline->pipeline) {
 					// Already logged, let's bail out.
 					ResetAfterDraw();
@@ -564,7 +567,6 @@ void DrawEngineVulkan::Flush() {
 				framebufferManager_->ApplyClearToMemory(scissorX1, scissorY1, scissorX2, scissorY2, result.color);
 			}
 		}
-		applySkinInDecode_ = g_Config.bSoftwareSkinning;
 	}
 
 	ResetAfterDrawInline();
@@ -582,7 +584,6 @@ void DrawEngineVulkan::ResetAfterDraw() {
 	vertexCountInDrawCalls_ = 0;
 	decodeIndsCounter_ = 0;
 	decodeVertsCounter_ = 0;
-	applySkinInDecode_ = g_Config.bSoftwareSkinning;
 	gstate_c.vertexFullAlpha = true;
 }
 

GPU/Vulkan/DrawEngineVulkan.h

@@ -136,7 +136,7 @@ public:
 		return pipelineLayout_;
 	}
 
-	void BeginFrame();
+	void BeginFrame() override;
 	void EndFrame();
 
 	void DirtyAllUBOs();