Depth target gets cleared before compute pass when previously used as attachement without storing values

[AnotherNathan]

Description
Occurs with the following sequence

• render pass that clears targets and stores depth values
• render pass that does not clear targets and does not store depth values
• (wgpu clears depth target internally)
• depth target gets used as texture_depth_2d in compute pass (and has now all 0 values) -> I would have expected to still have the depth values from the first render pass available here

Repro steps
see the attached example

Expected vs observed behavior
I would expect that the depth target does not get cleared when running the compute pass.

Extra materials
The attached zip contains a minimal example of the issue.

Platform
Windows 10
wgpu 0.17 and master on Vulkan backend
RX 570 driver version 21.50.21.11

wgpu-depth-clear-bug.zip

[Wumpf]

Thank you so much for the nice repro sample!

The issue here is a (I think re-occuring) miss-understanding with the store bool flag: If it's false then the content of the buffer may be discarded. This means that after any store false operation, the texture is in an undetermined state, thus forcing wgpu-core to clear it out the next time a read operation on it occurs.

Imho this is super unclear from the API as-is. I think wgpu should expose the wgpu-core Store-Op as is instead of putting a boolean there:

pub enum StoreOp {
    /// Discards the content of the render target.
    ///
    /// If you don't care about the contents of the target, this can be faster.
    Discard = 0,
    /// Store the result of the renderpass.
    Store = 1,
}

See also https://www.w3.org/TR/webgpu/#load-and-store-ops