-
Notifications
You must be signed in to change notification settings - Fork 137
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add robustness to GPU shaders #537
Conversation
Make each stage quit early if a previous stage has failed. The CPU shaders are minimally changed to be layout compatible. For the most part, they'll panic on a bounds check if sizes are exceeded. That's arguably useful for debugging, but a case can be made they should have the same behavior as the GPU shaders. Work towards #366
Possibly this can be merged as-is, but perhaps more work should be done so that clients can handle the failures reasonably. For now, rendering appears to freeze (zoom into longpathdash, then zoom back out; rendering will resume). |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
A few nits, but generally this looks like a decent improvement
I think the CPU shaders probably do need some rethinking, but I'm not sure the best way to do so
@@ -63,7 +64,18 @@ fn main( | |||
for (var i = 0u; i < N_SLICE; i += 1u) { | |||
atomicStore(&sh_bitmaps[i][local_id.x], 0u); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I know this isn't related to this PR, but as far as I can tell, this is already guaranteed to be zeroed. If this is to work around a driver/naga bug, we should have a comment here
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ah, I didn't realize that was a strong guarantee. In WebGPU world, it's probably worth skipping this explicit zeroing, but in native world it might be worth compiling with zeroing by infrastructure disabled, in which case we would need this.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Good point - for us it's not impactful, but e.g. before #363 this would have mattered for the MSL conversion
// This sets us up for technical UB, as lots of threads will be writing | ||
// to the same locations. But I think it's fine, and predicating the | ||
// writes would probably slow things down. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Would it be reasonable to have new_cmd=cmd_offset
here? I think that would avoid the UB - instead we'd just overwrite in the same location in each loop
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Intriguing idea! However, that won't quite avoid UB, as cmd_offset will edge into the allocation following this one. Setting it to cmd_limit - (PTCL_INCREMENT - PTCL_HEADROOM)
almost works, but only if it's not in its initial segment. I can't think of a good solution in that case, as we still ideally want the limit where it is so it accurately allocates as if there were enough memory, for the purposes of reporting the size back. My gut feeling is that if we were very concerned about the technical UB, we should in fact predicate the writes.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ah, because the allocations are variably sized? I can't say I'm that happy about adding UB, but I do agree that it's unlikely to cause a problem in practise.
I wonder how bad the cost of writing to the same location is in terms of memory bandwidth/cache coherency?
I agree that this is probably fine as-is, though
shader/coarse.wgsl
Outdated
if (failed & (STAGE_BINNING | STAGE_TILE_ALLOC | STAGE_PATH_COARSE)) != 0u { | ||
if failed != 0u { | ||
if wg_id.x == 0u && local_id.x == 0u { | ||
atomicOr(&bump.failed, STAGE_COARSE); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
What does this represent? I would think this should be STAGE_{BEFORE_COARSE}
, and then only if the check on seg_counts
failed
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'll make the change, now that I'm thinking these flags will be interpreted by the client, not just to early-out downstream.
let lines = atomicLoad(&bump.lines); | ||
indirect.count_x = (lines + (WG_SIZE - 1u)) / WG_SIZE; | ||
if atomicLoad(&bump.failed) != 0u { | ||
indirect.count_x = 0u; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm impressed that this works. Reading the specs suggest it's fine.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, this works and it's the only way I'm aware of that allows you to "abort" this type of indirect dispatch (there are more sophisticated ways with bindless, see for example: https://developer.apple.com/documentation/metal/indirect_command_encoding/encoding_indirect_command_buffers_on_the_gpu?language=objc).
Interestingly, I couldn't find any explicit wording in the WebGPU, Metal, Vulkan, or D3D12 docs that this is the expected behavior but "0" falls within the accepted range for all of them. See also this past discussion: gpuweb/gpuweb#1045
Clarify some nits, and also make a distinction between reporting failure in path_count and coarse.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
In its current state, the fact that there's no feedback at all when we miss a frame isn't ideal. I guess for debugging we could have a feature which slightly increases the effective alpha (/applies a red tinged overlay?) when rendering fails
Although that would have to increase each frame (which might actually be ideal, to give a view on the latency we get in dealing with under-allocation in #366.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Approved with one question.
@@ -152,11 +155,19 @@ fn main( | |||
// We need to check only prior stages, as if this stage has failed in another workgroup, | |||
// we still want to know this workgroup's memory requirement. | |||
if local_id.x == 0u { | |||
var failed = atomicLoad(&bump.failed) & (STAGE_BINNING | STAGE_TILE_ALLOC | STAGE_FLATTEN); | |||
if atomicLoad(&bump.seg_counts) > config.seg_counts_size { | |||
failed |= STAGE_PATH_COUNT; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why not set this in path_count?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Basically because path_count doesn't bind config. I'm also a bit wary of divergence but there's probably no meaningful impact on performance. I think it basically comes down to a style issue whether you tick the flag there or later.
If you can really rely on buffer robustness, then maybe at some point you can drop the write predication and just look at the read after the fact. One thing at the back of my head is the possibility of wrapping u32, but I think I'll choose not to worry about that too much right now.
let lines = atomicLoad(&bump.lines); | ||
indirect.count_x = (lines + (WG_SIZE - 1u)) / WG_SIZE; | ||
if atomicLoad(&bump.failed) != 0u { | ||
indirect.count_x = 0u; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, this works and it's the only way I'm aware of that allows you to "abort" this type of indirect dispatch (there are more sophisticated ways with bindless, see for example: https://developer.apple.com/documentation/metal/indirect_command_encoding/encoding_indirect_command_buffers_on_the_gpu?language=objc).
Interestingly, I couldn't find any explicit wording in the WebGPU, Metal, Vulkan, or D3D12 docs that this is the expected behavior but "0" falls within the accepted range for all of them. See also this past discussion: gpuweb/gpuweb#1045
Following #537 it is possible for the flatten stage to fail and flag a failure. In some cases this can cause invalid / corrupt bounding box data to propagate downstream, leading to a hang in the per-tile backdrop calculation loop. Triggering this is highly subtle, so I don't have a test case as part of vello scenes that can reliably reproduce this. Regardless, it makes sense to check for the upstream failures and terminate the work in general.
Following #537 it is possible for the flatten stage to fail and flag a failure. In some cases this can cause invalid / corrupt bounding box data to propagate downstream, leading to a hang in the per-tile backdrop calculation loop. Triggering this is highly subtle, so I don't have a test case as part of vello scenes that can reliably reproduce this. Regardless, it makes sense to check for the upstream failures and terminate the work in general.
Following #537 it is possible for the flatten stage to fail and flag a failure. In some cases this can cause invalid / corrupt bounding box data to propagate downstream, leading to a hang in the per-tile backdrop calculation loop. Triggering this is highly subtle, so I don't have a test case as part of vello scenes that can reliably reproduce this. Regardless, it makes sense to check for the upstream failures and terminate the work in general. I made backdrop_dyn check for any upstream failure and I didn't make it signal its own failure flag. I also didn't change the logic in the CPU shader since the other stages I checked (flatten, coarse) do not implement error signaling in their CPU counterparts. Let me know if you'd like me to work on those.
Make each stage quit early if a previous stage has failed.
The CPU shaders are minimally changed to be layout compatible. For the most part, they'll panic on a bounds check if sizes are exceeded. That's arguably useful for debugging, but a case can be made they should have the same behavior as the GPU shaders.
Work towards #366