[MSL] Support sparse image residency operations in MSL 2.2

[cdavis5e]

MSL 2.2 on iOS at least adds new methods for reading and sampling textures that additionally return whether or not the texture memory was resident. These correspond to OpImageSparseXxx in SPIR-V. It should now be possible to implement those operations.

[cdavis5e]

It looks like Apple removed this incomplete feature from the MSL 2.2 headers. Maybe the next version will support it.

[oscarbg]

yep.. but we should reopen as it should have returned today in Xcode 11 GM as it's a new feature of A13 GPU..
see (https://twitter.com/gavkar/status/1171494278545018882)

[cdavis5e]

They are indeed restored to the MSL 2.2 headers. I guess Apple didn't want anybody poking at this while they finished up this super-secret feature.

[Degerz]

Any format limitations with sparse textures on A13 ?

Also, it seems to be implied that the tile mappings are updated from the GPU rather than the CPU ? Is that true ?

[cdavis5e]

Any format limitations with sparse textures on A13 ?

There might be. See -[MTLDevice sparseTileSizeWithTextureType:pixelFormat:sampleCount:].

Also, it seems to be implied that the tile mappings are updated from the GPU rather than the CPU ? Is that true ?

That's true. It's actually true of Vulkan as well--see vkQueueBindSparse().

[Degerz]

There might be. See -[MTLDevice sparseTileSizeWithTextureType:pixelFormat:sampleCount:].

I wish Apple would just publicly document the allowed pixel formats like Microsoft does with D3D12 instead. It'd be way more useful this way to see if Metal supports sparse volume textures or sparse MSAA textures ...

I see that there's an individual sampleCount parameter for sparseTileSize so there might be hope after all to support MSAA textures on at least Intel devices but it would be a big disappointment if all configurations returned 1 ...

That's true. It's actually true of Vulkan as well--see vkQueueBindSparse().

I don't mean being able to update the page table into device memory since that's really ugly performance wise. I want the ability to specifically update tile mappings via from the GPU but it looks like that'll remain exclusive to Intel HW so I guess I misinterpreted the tweet ...

Edit: Same question was asked on twitter ...

[cdavis5e]

I want the ability to specifically update tile mappings via from the GPU

Metal can kinda sorta do GPU-driven sparse updates. See -[MTLResourceStateCommandEncoder updateTextureMapping:mode:indirectBuffer:indirectBufferOffset:]. I guess it's not what you were hoping for, which in Metal terms would be an indirect command buffer that supported resource state commands.

[oscarbg]

any update on this?

[Wangsh09]

any update on this?