Open fyellin opened 2 weeks ago
Those lifetime constraints got lifted for Render & Compute pass not too long ago, see https://github.com/gfx-rs/wgpu/pull/5884 and related PRs. Very similar work has to be done for bundles. However, I thought that most of it was already in place. But obviously there's still something missing. A suite of tests similar to what was added for render & compute pass is in order, fixing those and then removing the those lifetimes from the public api
I recently wrote some test code for python that tests deleting objects in the user code even though they area still in use by wgpu. Everything worked perfectly except for RenderBundleEncoder
. Behind the scenes, I had to force RenderBundleEncoder
to retain a pointer to the objects it depended on.
Should I file a different bug about the fact that RenderBundle
and RenderBundleEncoder
panic when they find something they don't like, rather than returning an error? Or is this part of the same cleanup? Passing the wrong arguments formats to createRenderBundleEncoder()
or forgetting to call set_pipeline()
will cause a crash.
part of the same cleanup, I think this issue captures this well enough already.
Generally, violating the lifetime constraints that the rust borrow checker tries to uphold (and naturally can't be automatically enforced in a C interface) is undefined behavior. So it's less of a crash bug but more of a "those lifetimes really shouldn't be there" issue (:
(well and ofc from a different point of view, wgpu-native
is currently not correctly implementing the contract of webgpu.h
, which it could do easily if those lifetime constraints weren't there)
Although
wgpu::RenderPass
andwgpu::RenderBundleEncoder
both implement the traitwpu::util::RenderEncoder
, they have different expectations for the lifetime of their common methods. For methods such asset_pipeline
,set_bind_group
, andset_vertex_buffer
,RenderPass
has no expectations for the lifetime of the argument.RenderBundleEncoder
requires that the argument live at least as long asself
.There is no way of enforcing this lifetime restriction in the C interface or the Python interface to this code. This is documented in this bug at wgpu-native.
Attached is RenderBundleEncoderBug.zip which contains RenderBundleEncoder.py. It is run by calling
python RenderBundleEncoderBug.py <flag>
where for this bug,flag
is one ofnull-pipeline
,null-bind-group
,null-vertex-buffer
,null-index-buffer
,null-indirect-buffer
. These show that nulling any of these values has no affect on aRenderPass
but will cause theRenderBundleEncoder
to panic.My preference is that this lifetime restriction in
RenderBundleEncoder
(andRenderEncoder
) be removed. Whatever techniqueRenderPass
uses to keep track of the objects (and keep them alive?) can certainly be applied toRenderBundleEncoder
.Alternatively,
RenderBundleEncoder
shouldn't panic. It should never panic except for a serious non-recoverable error. this is an error that can easily be reported back to the user.