Updating of a Metal texture in every frame

Question

Created Sep ’20

Replies 6

Boosts 0

Views 1.7k

Participants 4

Hi,

We need to update some managed texture from system memory to GPU memory in every frame. Metal.

Problem: it seems that a texture still can be busy for rendering of queued frames and when we update it for next frame a previous frame can be rendered incorrectly (texture looks wrong).

How I can guess when a texture is being completely processed and can be re-used again? I don't know how long a queue of frames in Metal rendering.

It seems that DirectX can automatically create internal copies of a texture which we sent for rendering and we can start to modify the same texture for a next frame. How to do the same in Metal?

Boost

Answer 1

S-D-Dev OP

Sep ’20

I thought the same thing...
Is "MTLTexture.replaceRegion" asynchronous or synchronous, or if it is asynchronous, is there a delegate?
but...
Due to the machine's convenience, I thought I had to wait until the next drawing timing, so I did it with a double buffer.

0

Answer 2

wnsoft OP

Sep ’20

We call replaceRegion:mipmapLevel:withBytes:bytesPerRow:

Documentation: "This method immediately copies the pixel data into the texture. It does not synchronize against any GPU accesses to this texture. For example, if a command buffer includes read or write operations on a given texture, you must ensure that these operations complete before calling the replaceRegion:mipmapLevel:withBytes:bytesPerRow: method on the given texture. You can use the addCompletedHandler: method, waitUntilCompleted method, or custom semaphores to signal that a command buffer has completed execution."

According Apple's documentation, this function is synchronous. In a current frame we don't have anothr command buffer which has other operations with the same texture.

But we have previous frames which still may be in a queue for rendering and this texture can be busy (potentially).
But how we should guess that this texture is "busy"?

0

Answer 3

wnsoft OP

Oct ’20

I would grateful for any advise.

0

Answer 4

Ceylo OP

Oct ’20

Did you check https://developer.apple.com/documentation/metal/synchronization/synchronizing_events_between_a_gpu_and_the_cpu ?

1

Answer 5

darknoon OP

Oct ’20

Usually, you want to double- or triple- buffer so you're not waiting for the memory region to become available again. ie, make 3 textures and rotate between them so one is free to be updated, one is being enqueued, and one is being used for rendering at any given time.

1

Answer 6

wnsoft OP

Nov ’20

Hi Ceylo, darknoon,

Thanks for your replies!

I think that double/tripple buffering could be a real explanation of this issue.

Also an article "Synchronizing Events Between a GPU and the CPU" also is a good idea. It's difficult to understand if we really should use these sync. events. I need to think more.

Right now, our app works fine with 3 textures in a series.

0