Copy color to depth, defer depth buffer copies

GPU/Common/FramebufferManagerCommon.cpp

@@ -377,6 +377,11 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 		//
 		// We use a special fragment shader flag to convert color to depth.
 		vfb = GetLatestDepthBufferAt(params.fb_address /* !!! */, params.fb_stride);
+		if (vfb) {
+			vfb->depthBindSeq = GetBindSeqCount();
+		}
+		// Avoid causing another depth copy on top.
+		gstate_c.usingDepth = true;
 	}
 
 	gstate_c.SetFramebufferRenderMode(mode);
@@ -416,6 +421,8 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 
 	// None found? Create one.
 	if (!vfb) {
+		gstate_c.usingDepth = false;  // reset depth buffer tracking
+
 		vfb = new VirtualFramebuffer{};
 		vfb->fbo = nullptr;
 		vfb->fb_address = params.fb_address;
@@ -442,17 +449,6 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 		ResizeFramebufFBO(vfb, drawing_width, drawing_height, true);
 		NotifyRenderFramebufferCreated(vfb);
 
-		// Looks up by z_address, so if one is found here and not have last pointers equal to this one,
-		// there is another one.
-		VirtualFramebuffer *prevDepth = GetLatestDepthBufferAt(vfb->z_address, vfb->z_stride);
-
-		// We might already want to copy depth, in case this is a temp buffer.  See #7810.
-		if (prevDepth != vfb) {
-			if (!params.isClearingDepth && prevDepth) {
-				BlitFramebufferDepth(prevDepth, vfb);
-			}
-		}
-
 		SetColorUpdated(vfb, skipDrawReason);
 
 		INFO_LOG(FRAMEBUF, "Creating FBO for %08x (z: %08x) : %d x %d x %s", vfb->fb_address, vfb->z_address, vfb->width, vfb->height, GeBufferFormatToString(vfb->format));
@@ -510,18 +506,19 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 		VirtualFramebuffer *prev = currentRenderVfb_;
 		currentRenderVfb_ = vfb;
 		NotifyRenderFramebufferSwitched(prev, vfb, params.isClearingDepth);
+		gstate_c.usingDepth = false;  // reset depth buffer tracking
 	} else {
+		// Something changed, but we still got the same framebuffer we were already rendering to.
+		// Might not be a lot to do here, we check in NotifyRenderFramebufferUpdated
 		vfb->last_frame_render = gpuStats.numFlips;
 		frameLastFramebufUsed_ = gpuStats.numFlips;
 		vfb->dirtyAfterDisplay = true;
 		if ((skipDrawReason & SKIPDRAW_SKIPFRAME) == 0)
 			vfb->reallyDirtyAfterDisplay = true;
-
 		NotifyRenderFramebufferUpdated(vfb, vfbFormatChanged);
 	}
 
 	vfb->colorBindSeq = GetBindSeqCount();
-	vfb->depthBindSeq = GetBindSeqCount();
 
 	gstate_c.curRTWidth = vfb->width;
 	gstate_c.curRTHeight = vfb->height;
@@ -530,6 +527,26 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 	return vfb;
 }
 
+// Called on the first use of depth in a render pass.
+void FramebufferManagerCommon::SetDepthFrameBuffer() {
+	if (!currentRenderVfb_) {
+		return;
+	}
+
+	// Looks up by z_address, so if one is found here and not have last pointers equal to this one,
+	// there is another one.
+	VirtualFramebuffer *prevDepth = GetLatestDepthBufferAt(currentRenderVfb_->z_address, currentRenderVfb_->z_stride);
+
+	if (prevDepth != currentRenderVfb_) {
+		if (!gstate_c.clearingDepth && prevDepth) {
+			BlitFramebufferDepth(prevDepth, currentRenderVfb_);
+		}
+		prevDepth = currentRenderVfb_;
+	}
+
+	currentRenderVfb_->depthBindSeq = GetBindSeqCount();
+}
+
 void FramebufferManagerCommon::DestroyFramebuf(VirtualFramebuffer *v) {
 	// Notify the texture cache of both the color and depth buffers.
 	textureCache_->NotifyFramebuffer(v, NOTIFY_FB_DESTROYED);
@@ -669,17 +686,6 @@ void FramebufferManagerCommon::NotifyRenderFramebufferSwitched(VirtualFramebuffe
 	textureCache_->ForgetLastTexture();
 	shaderManager_->DirtyLastShader();
 
-	// Copy depth between the framebuffers, if the z_address is the same (checked inside.)
-	VirtualFramebuffer * prevDepth = GetLatestDepthBufferAt(vfb->z_address, vfb->z_stride);
-
-	// We might already want to copy depth, in case this is a temp buffer.  See #7810.
-	if (prevDepth != vfb) {
-		if (!isClearingDepth && prevDepth) {
-			BlitFramebufferDepth(prevDepth, vfb);
-		}
-		prevDepth = vfb;
-	}
-
 	if (vfb->drawnFormat != vfb->format) {
 		ReinterpretFramebuffer(vfb, vfb->drawnFormat, vfb->format);
 	}

GPU/Common/FramebufferManagerCommon.h

@@ -266,6 +266,8 @@ class FramebufferManagerCommon {
 			return vfb;
 		}
 	}
+	void SetDepthFrameBuffer();
+
 	void RebindFramebuffer(const char *tag);
 	std::vector<FramebufferInfo> GetFramebufferList() const;
 

GPU/GPUCommon.cpp

@@ -1685,6 +1685,16 @@ void GPUCommon::Execute_Prim(u32 op, u32 diff) {
 		return;
 	}
 
+	if (!gstate_c.usingDepth) {
+		bool isClearingDepth = gstate.isModeClear() && gstate.isClearModeDepthMask();;
+
+		if ((gstate.isDepthTestEnabled() || isClearingDepth)) {
+			gstate_c.usingDepth = true;
+			gstate_c.clearingDepth = isClearingDepth;
+			framebufferManager_->SetDepthFrameBuffer();
+		}
+	}
+
 	const void *verts = Memory::GetPointerUnchecked(gstate_c.vertexAddr);
 	const void *inds = nullptr;
 	u32 vertexType = gstate.vertType;

GPU/GPUState.h

@@ -572,6 +572,9 @@ struct GPUStateCache {
 
 	uint64_t dirty;
 
+	bool usingDepth;  // For deferred depth copies.
+	bool clearingDepth;
+
 	bool textureFullAlpha;
 	bool vertexFullAlpha;