Better control of load/store actions for Camera

If you are using "legacy" rendering pipeline (not SRP) then Camera component is the one making all the choices about load/store flags of its target texture. If you render to actual RenderTexture you can use RenderTexture.DiscardContents to tweak load/store actions [1].
But, alas, there is no way to do so for backbuffer (and you might also say that doing the right combination of DiscardContents is akin some magic). So what is preferred is for you to setup render target yourself, but this is also not that straightforward. The "natural" place would be in OnPreRender, but after that and before actually rendering scene, camera will render shadows (or depth texture https://docs.unity3d.com/Manual/SL-CameraDepthTexture.html in general). So you need to setup render target after shadows but before actual rendering.
Enter CommandBuffer [https://docs.unity3d.com/Manual/GraphicsCommandBuffers.html] API. Assuming you use forward rendering you need BeforeForwardOpaque camera event. But there is yet another caveat: if camera has clear flags set - clear will happen BEFORE BeforeForwardOpaque event (and that will setup render target). So the full solution would be:
1. make camera "Don't clear" [CameraClearFlags.Nothing]
2. add command buffer for BeforeForwardOpaque event
3. in there set rendertarget to the one you want with appropriate load/store flags
4. do an explicit clear yourself. Be warned: you *must* clear depth. [2]

bringing this all together, this is sample code (setting backbuffer, discarding depth after rendering, clearing color/depth):

Code (CSharp):

1. using System.Collections;
2. using System.Collections.Generic;
3. using UnityEngine;
4. using Rendering   = UnityEngine.Rendering;
5. using LoadAction  = UnityEngine.Rendering.RenderBufferLoadAction;
6. using StoreAction = UnityEngine.Rendering.RenderBufferStoreAction;
7. public class TestSetRenderTarget : MonoBehaviour {
8.     void Start() {
9.         Rendering.CommandBuffer cb = new Rendering.CommandBuffer();
10.         cb.SetRenderTarget(Rendering.BuiltinRenderTextureType.CameraTarget,
11.             LoadAction.DontCare, StoreAction.Store, LoadAction.DontCare, StoreAction.DontCare
12.         );
13.         cb.ClearRenderTarget(true,true, new Color(1,0,0,1), 1.0f);
14.         GetComponent<Camera>().AddCommandBuffer(Rendering.CameraEvent.BeforeForwardOpaque, cb);
15.     }
16. }

[1] RenderTexture.DiscardContents will do different things depending on whether it is current render target: if it is, storeAction will be DontCare, otherwise loadAction will be DontCare.
[2] Internally unity will delay setting render target until actual draw call, so those two calls (set render target and clear) will be combined, i.e. MTLRenderPassAttachmentDescriptor.loadAction will be set to MTLLoadActionClear.

 

for 2020.1 we decided that we had enough and we fixed this "end of frame facepalm" - if you have MSAA color it will be (simply, w/o store) resolved to drawable, and depth/stencil will be discarded. We will backport to 2019 LTS too after 2019.3 release
@johnfrog is it what you asked about? (i will recheck interop with memoryless just in case)

 

>>We're still on 2018.4.8f1
hm, i guess we might look into backporting (but that's equally scary for us 8))