Added validation for both setBindGroup() variants #998

toji · 2020-08-12T22:16:50Z

Also fixes an error I made with a previous PR where I mistakenly compared GPUDevices against GPURenderPassEncoders instead of the GPUDevice the render pass encoder was created with.

Preview | Diff

kainino0x

LGTM

kvark

Thank you, good stuff!

spec/index.bs

kvark · 2020-08-13T14:31:46Z

spec/index.bs

+                        - |bindGroup|.{{GPUObjectBase/[[device]]}} is |this|.{{GPUObjectBase/[[device]]}}.
+                        - |index| &lt; {{GPULimits/maxBindGroups|GPULimits.maxBindGroups}}.
+                        - dynamicOffsets.length is
+                            |bindGroup|.{{GPUBindGroup/[[layout]]}}.{{GPUBindGroupLayout/[[dynamicOffsetCount]]}}.


we also need to check that the binding offset + dynamic offset + minimumBufferSize is <= buffer size, for each binding with dynamic offset.

Right, forgot that we can validate this here rather than during the draw call.

I took a stab at it but... wow. There's just no clean way to document this validation! Feedback on how to help it read better is very welcome.

Since this is a newly written rule, added it to the test plan: https://hackmd.io/@webgpu/H1MwoqqAU#encodingcmdsprogrammable_passspects

kvark · 2020-08-13T14:33:34Z

spec/index.bs

+                |index|: The index to set the bind group at.
+                |bindGroup|: Bind group to use for subsequent render or compute commands.
+                |dynamicOffsetsData|: Array containing buffer offsets in bytes for each entry in
+                    |bindGroup| with marked as {{GPUBindGroupLayoutEntry/hasDynamicOffset}}.


ok, here comes the fun part: what is the order?
In Vulkan, dynamic offsets need to be provided in the order of binding, which can be different from the order of either BindGroupLayoutEntry or BindGroupEntry. I suggest to follow this route and specify that the offsets are taken in the order of increased binding number

Oof, I have no idea what's going to be best here. Gonna let the editors work this one out. 😅

I definitely think it should be the binding number order. BindGroup(Layout)Descriptor.entries should be treated as an unordered map from binding => entry.

toji · 2020-08-13T17:04:45Z

Oops! Forgot to push when I made my previous changes. Code is updated now.

spec/index.bs

kvark

Thank you!

kvark · 2020-08-13T19:38:18Z

CI is sad

toji · 2020-08-13T20:01:59Z

Weird. I had fixed that, wonder what sort of git-goof I did to lose it. 🤷 Fixed (again) now.

toji · 2020-08-13T21:34:52Z

Well crumb, my ability to git properly is wildly off today. I accidentally added code that was intended for a separate PR to the final commit of this one. Specifically: aeaf777

The change was applying a suggestion from @kainino0x in a previous PR to move the repeated checks for |object| is [=valid=] && |object|.device is |this|.device into a separate algorithm that can be reused everywhere. The change was "done", and I was going to upload it for review just now what I noticed the error.

I apologize for the mistake, as this was certainly not something I intended to slip in unnoticed. Depending on what the editors would prefer I'm happy to:

Revert this PR entirely and re-post it without the accidental addition.
Submit a change to only revert the accidentally added code and post it again as a separate PR.
Leave it be if the code looks good. (It's not a particularly complicated change.)

kvark · 2020-08-13T21:45:35Z

I think we should leave it. The spec is still heavily a draft, let's fix it as we go.

kainino0x · 2020-08-14T23:51:52Z

spec/index.bs

+                        - |bindGroup|.{{GPUObjectBase/[[device]]}} is |this|.{{GPUObjectBase/[[device]]}}.
+                        - |index| &lt; {{GPULimits/maxBindGroups|GPULimits.maxBindGroups}}.
+                        - dynamicOffsets.length is
+                            |bindGroup|.{{GPUBindGroup/[[layout]]}}.{{GPUBindGroupLayout/[[dynamicOffsetCount]]}}.


Since this is a newly written rule, added it to the test plan: https://hackmd.io/@webgpu/H1MwoqqAU#encodingcmdsprogrammable_passspects

kainino0x · 2020-08-14T23:54:15Z

spec/index.bs

+                |index|: The index to set the bind group at.
+                |bindGroup|: Bind group to use for subsequent render or compute commands.
+                |dynamicOffsetsData|: Array containing buffer offsets in bytes for each entry in
+                    |bindGroup| with marked as {{GPUBindGroupLayoutEntry/hasDynamicOffset}}.


I definitely think it should be the binding number order. BindGroup(Layout)Descriptor.entries should be treated as an unordered map from binding => entry.

kainino0x · 2020-08-14T23:56:18Z

spec/index.bs

+                                |dynamicOffsetsData|[|dynamicOffsetIndex| + |dynamicOffsetsDataStart|].
+                            - |bufferBinding|.{{GPUBufferBinding/offset}} + |bufferDynamicOffset| +
+                                |minBufferBindingSize| &le;
+                                |bufferBinding|.{{GPUBufferBinding/buffer}}.{{GPUBuffer/[[size]]}}.


minBufferBindingSize is optional, and this rule should only apply if it's provided.

In the definition of the GPUBindGroupLayoutEntry it says that minBufferBindingSize defaults to 0 for buffer types. If that's the case then this works out regardless of if it's provided or not, right?

Hmmm, I think this was unintentional. It looks like it was effectively a merge conflict between #874 and #919. minBufferBindingSize should not have a default (undefined is a special/sentinel value that says the size will be checked at draw time instead of createBindGroup time).

kainino0x · 2020-08-15T00:00:30Z

spec/index.bs

        </div>
 </dl>

+<div algorithm>


Perhaps we should pre-bake the list of dynamically-offset buffer bindings in createBindGroupLayout. I think it would just be a sorted list of binding numbers like [1, 3, 4] and you could use those to index back into the bind group (layout) entries

kainino0x · 2020-08-15T00:01:09Z

spec/index.bs

      - If the [=device=] that owns an object is lost, the object becomes invalid.
 </div>

+<div algorithm>


This refactor looks excellent.

Added validation for both setBindGroup() variants

d4fac44

toji requested review from kvark and kainino0x August 12, 2020 22:16

kainino0x approved these changes Aug 13, 2020

View reviewed changes

kvark requested changes Aug 13, 2020

View reviewed changes

Feedback from @kvark

0223029

kvark reviewed Aug 13, 2020

View reviewed changes

spec/index.bs Outdated Show resolved Hide resolved

Broke iteration over dynamic buffer bindings out into it's own algorithm

270c5f0

kvark approved these changes Aug 13, 2020

View reviewed changes

Making CI happy

aeaf777

kvark merged commit cf0645f into gpuweb:main Aug 13, 2020

toji deleted the setBindGroup-validation branch August 13, 2020 22:36

kainino0x reviewed Aug 15, 2020

View reviewed changes

ben-clayton pushed a commit to ben-clayton/gpuweb that referenced this pull request Sep 6, 2022

Throw descriptive error if query doesn't match any cases (gpuweb#998)

bcac299

Added validation for both setBindGroup() variants #998

Added validation for both setBindGroup() variants #998

Uh oh!

Conversation

toji commented Aug 12, 2020 • edited by pr-preview bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kainino0x left a comment

Choose a reason for hiding this comment

Uh oh!

kvark left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

toji commented Aug 13, 2020

Uh oh!

Uh oh!

kvark left a comment

Choose a reason for hiding this comment

Uh oh!

kvark commented Aug 13, 2020

Uh oh!

toji commented Aug 13, 2020

Uh oh!

toji commented Aug 13, 2020

Uh oh!

kvark commented Aug 13, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

toji commented Aug 12, 2020 •

edited by pr-preview bot

Loading