Add validation for GPUDevice.createTexture() #799

litherum · 2020-05-21T23:21:44Z

kvark

Thank you for writing this! It looks great overall, just have a few concerns.

kvark · 2020-05-22T14:40:18Z

spec/index.bs

+
+  1. If device is lost, or if this call doesn't follow the [$createTexture Valid Usage$], return an error texture.
+  1. Let |t| be a new {{GPUTexture}} object.
+  1. Set |t|.{{GPUTexture/[[textureSize]]}} to |descriptor|.{{GPUTextureDescriptor/size}}.


did you consider just making the whole |descriptor| to be in the internal slot?

Should it be? I was just copying how buffers did it.

Also, that reminds me, are slots copied by value or by reference? It would be bad if the spec said that modifying a descriptor after creating the texture caused the texture to be modified.

It would be just easier to specify if you say the descriptor is copied by value into the internal slot.

Unfortunately right now copying patterns from other parts of the spec doesn't work well because they haven't been copyedited.

I want to change more of the spec to use this pattern (just copying the whole descriptor directly into an internal slot).

spec/index.bs

kvark · 2020-05-22T15:31:36Z

spec/index.bs

+      1. If |descriptor|.{{GPUTextureDescriptor/sampleCount}} > 1:
+
+        1. |descriptor|.{{GPUTextureDescriptor/mipLevelCount}} must be 1.
+        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/depth=] must be 1.


we should also say that the dimension has to be "2d", i.e. disallow 1D multisampled textures (for now)

It is done in that dimension's list of validation checks, but I think it would be good to have all sample count constraints in the same place.

I couldn't get Bikeshed to give me nested "if"s :(

I think it works if you add enough indentation. But sometimes 2 spaces is enough while other times you need 4 spaces. It's a bit weird.

Generally it requires 4 spaces; I think it usually treats 2 spaces similarly to 0 spaces.

spec/index.bs

Kangz · 2020-05-22T15:46:06Z

spec/index.bs

+      1. If |descriptor|.{{GPUTextureDescriptor/sampleCount}} > 1:
+
+        1. |descriptor|.{{GPUTextureDescriptor/mipLevelCount}} must be 1.
+        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/depth=] must be 1.


It is done in that dimension's list of validation checks, but I think it would be good to have all sample count constraints in the same place.

spec/index.bs

Kangz · 2020-05-22T15:48:10Z

spec/index.bs

+      1. If |descriptor|.{{GPUTextureDescriptor/format}} is a [=depth or stencil format=]:
+
+        1. |descriptor|.{{GPUTextureDescriptor/dimension}} must be "2d".
+        1. |descriptor|.{{GPUTextureDescriptor/sampleCount}} must be 1.


nit: Could go with the rest of the multisampled texture validation.

I couldn't get Bikeshed to give me nested "if"s :(

spec/index.bs

Kangz · 2020-05-22T15:52:10Z

spec/index.bs

+        1. |descriptor|.{{GPUTextureDescriptor/sampleCount}} must be 1.
+
+      1. |descriptor|.{{GPUTextureDescriptor/usage}} must be a combination of {{GPUTextureUsage}} values.
+


There is some missing validation:

first mip level width and height is a multiple of the compressed block size.

compressed textures can only be copySrc / copyDst / Sampled.

other usage constraints (we have non-renderable types), or add an issue for it. Dawn's Format.cpp summarizes what we have found by investigating the various APIs, and all usages of all formats are tested.

Unextended WebGPU doesn't have any compressed textures, and we haven't added any extensions to add any yet, and our implementation doesn't support compressed textures yet, so I'm not comfortable adding rules for compressed textures. I can open an issue for it.

I'll open a new issue for usage constraints. AFAICT Metal doesn't have any constraints here, so I'm not sure what to restrict.

"texture-compression-bc" is already there although the formats enums aren't in the spec yet. We're pretty comfortable adding text for compressed textures because @Jiawei-Shao did extensive investigation and testing in Dawn so we know they work in all APIs with the current validation constraints.

There's a bit a of chicken and egg issues because you've voiced concern about adding the extension because we didn't have validation for it, and now you have concerns adding the validation because we don't have the extension yet ^^.

That said we can do compressed texture validation in a follow-up if it feels more comfortable.

Metal does have a large number of usage constraints on texture that you can find starting page 6 of the Metal Feature Set document. We can add that independently since it will likely be a large PR in itself.

BTW though there are no compressed textures in the spec, there is a lot of setup for them (block based formats/texel blocks). You can rely on that ("block size > 1") to talk about compressed formats.

In this case: In practice this should be correct, but if we want we could eventually add an explicit flag to a texture format capability table that says "supports multiple samples".

Kangz · 2020-05-27T09:00:00Z

spec/index.bs

-        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/width=] must be less than or equal to 2048.
-        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/height=] must be less than or equal to 2048.
-        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/depth=] must be less than or equal to 2048.
+        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/width=] must be less than or equal to the owning {{GPUDevice}}'s {{GPULimits/maxTextureLayers}}.


For 3D textures I think each dimension should use maxTextureSize instead of maxTextureLayers, or maybe we could have split limits for 1D, 2D and 3D (3D's default limit can't be more than 2048 because that's what the majority of Vulkan devices support)

It seems that we have to split here:

maxTextureSize1D = 8192

maxTextureSize2D = 8192

maxTextureSize3D = 2048

maxTextureLayers = 256

kainino0x · 2020-05-27T21:17:52Z

spec/index.bs

+    ::
+        The maximum number of layers in an array texture.
+
+        Note: This is also used for the size limit of 3D textures.


All three dimensions are tied to maxTextureLayers?

Is there a fundamental reason this is true (i.e. coming from the native APIs)? If not, we should probably split up the limits in case there is hardware out there where this would be overly restrictive.

Also would you mind updating https://github.com/gpuweb/gpuweb/blob/master/design/Limits.md to document where these numbers come from?

kainino0x · 2020-05-27T21:27:15Z

spec/index.bs

+      1. Let |d| = |size|.[=Extent3D/depth=].
+
+    1. Let |m| = the maximum value of |w|, |h|, and |d|.
+    1. Return one plus the greatest integral value of |x| for which 2^|x| <= |m|.


Pretty sure this will work:

Suggested change

1. Return one plus the greatest integral value of |x| for which 2^|x| <= |m|.

1. Return one plus the greatest integral value of |x| for which 2<sup>|x|</sup> ≤ |m|.

kainino0x · 2020-05-27T21:28:49Z

spec/index.bs

+      1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/width=] must be nonzero.
+      1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/height=] must be nonzero.
+      1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/depth=] must be nonzero.
+      1. |descriptor|.{{GPUTextureDescriptor/mipLevelCount}} must be nonzero.
+      1. |descriptor|.{{GPUTextureDescriptor/sampleCount}} must be nonzero.


nit: I know this is expressed by the types, but "greater than zero" would be slightly clearer here than "nonzero"

kainino0x · 2020-05-27T21:31:34Z

spec/index.bs

+    <dfn abstract-op>createTexture Valid Usage</dfn>
+      Given a {{GPUDevice}} |this| and a {{GPUTextureDescriptor}} |descriptor|


nit: indent level changes between these two lines.

Also I'd put a paragraph break between the dfn and "Given".

Suggested change

<dfn abstract-op>createTexture Valid Usage</dfn>

Given a {{GPUDevice}} |this| and a {{GPUTextureDescriptor}} |descriptor|

<dfn abstract-op>createTexture Valid Usage</dfn>

Given a {{GPUDevice}} |this| and a {{GPUTextureDescriptor}} |descriptor|

kainino0x · 2020-05-27T21:38:34Z

spec/index.bs

+      1. |descriptor|.{{GPUTextureDescriptor/sampleCount}} must be nonzero.
+      1. If |descriptor|.{{GPUTextureDescriptor/dimension}} is "1d":
+
+        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/width=] must be less than or equal to the owning {{GPUDevice}}'s {{GPULimits/maxTextureSize}}.


nit: The "owning GPUDevice" is |this|

kainino0x · 2020-05-27T21:44:56Z

spec/index.bs

+        1. |descriptor|.{{GPUTextureDescriptor/sampleCount}} must be 1.
+
+      1. |descriptor|.{{GPUTextureDescriptor/usage}} must be a combination of {{GPUTextureUsage}} values.
+


BTW though there are no compressed textures in the spec, there is a lot of setup for them (block based formats/texel blocks). You can rely on that ("block size > 1") to talk about compressed formats.

In this case: In practice this should be correct, but if we want we could eventually add an explicit flag to a texture format capability table that says "supports multiple samples".

Kangz · 2020-07-03T16:50:41Z

This was almost ready to be merged, can we iterate on it and get it in the spec?

Kangz · 2020-07-17T14:04:24Z

Up! This PR should be most non-controversial, and we can just add Issue: to remember things to come back to (like compressed format constraints)

kvark · 2020-07-17T18:29:12Z

spec/index.bs

+    1. Calculate the values of |w|, |h|, and |d|. If the |dimension| is "1d":
+
+      1. Let |w| = |size|.[=Extent3D/width=].
+      1. Let |h| = 1.


Isn't Extent3D/height guaranteed to be 1 for 1D textures? I.e. can we just use Extent3D/width/height/depth for w/h/d instead of having this switch over the dimension?

kvark · 2020-07-17T18:30:54Z

spec/index.bs

+      1. Let |d| = |size|.[=Extent3D/depth=].
+
+    1. Let |m| = the maximum value of |w|, |h|, and |d|.
+    1. Return one plus the greatest integral value of |x| for which 2^|x| <= |m|.


shouldn't this be < instead of <=? If m is already a power of two, say 2^k, we consider the mipmap count to be k, while the current argument would produce k+1

kvark · 2020-07-17T18:33:43Z

spec/index.bs

-        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/width=] must be less than or equal to 2048.
-        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/height=] must be less than or equal to 2048.
-        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/depth=] must be less than or equal to 2048.
+        1. |descriptor|.{{GPUTextureDescriptor/size}}.[=Extent3D/width=] must be less than or equal to the owning {{GPUDevice}}'s {{GPULimits/maxTextureLayers}}.


It seems that we have to split here:

maxTextureSize1D = 8192

maxTextureSize2D = 8192

maxTextureSize3D = 2048

maxTextureLayers = 256

spec/index.bs

This PR adds validation rules for texture creation. It is based on: - PR 799 (gpuweb#799) from Myles, and the comments. - Issue 1522 (gpuweb#1522) I investigated when I was implementing 3D textures on Dawn project, and the comments.

kainino0x · 2021-04-27T01:24:58Z

superseded by #1658

* Add spec for texture creation This PR adds validation rules for texture creation. It is based on: - PR 799 (#799) from Myles, and the comments. - Issue 1522 (#1522) I investigated when I was implementing 3D textures on Dawn project, and the comments. * Update spec/index.bs Co-authored-by: Kai Ninomiya <[email protected]> * Update spec/index.bs Co-authored-by: Kai Ninomiya <[email protected]> * Address feedback from Kai and Dzmitry * Address more feedback from Kai and Dzmitry * Update spec/index.bs Co-authored-by: Kai Ninomiya <[email protected]> Co-authored-by: Kai Ninomiya <[email protected]>

Add validation for GPUDevice.createTexture()

69623e4

kvark requested changes May 22, 2020

View reviewed changes

Kangz reviewed May 22, 2020

View reviewed changes

Address first round of feedback

d6f7935

Kangz reviewed May 27, 2020

View reviewed changes

kainino0x reviewed May 27, 2020

View reviewed changes

kvark changed the base branch from master to main June 23, 2020 13:15

kvark reviewed Jul 17, 2020

View reviewed changes

Kangz mentioned this pull request Nov 4, 2020

Texture dimensions should be required to be >=1 (forbid zero-sized textures) #1200

Closed

kvark mentioned this pull request Nov 19, 2020

Add validation steps for createTexture and createTextureView. #1237

Closed

kvark mentioned this pull request Jan 5, 2021

Maximum values of Texture Dimension #1327

Closed

Richard-Yunchao mentioned this pull request Mar 16, 2021

Validation rules and issues of createTexture #1522

Closed

Richard-Yunchao mentioned this pull request Apr 13, 2021

Document texture dimension limits in Limit.md explainer #1632

Merged

Richard-Yunchao mentioned this pull request Apr 22, 2021

Add spec for texture creation #1658

Merged

kainino0x closed this Apr 27, 2021

		1. \|descriptor\|.{{GPUTextureDescriptor/sampleCount}} must be 1.

		1. \|descriptor\|.{{GPUTextureDescriptor/usage}} must be a combination of {{GPUTextureUsage}} values.

	1. Return one plus the greatest integral value of \|x\| for which 2^\|x\| <= \|m\|.
	1. Return one plus the greatest integral value of \|x\| for which 2<sup>\|x\|</sup> ≤ \|m\|.

		<dfn abstract-op>createTexture Valid Usage</dfn>
		Given a {{GPUDevice}} \|this\| and a {{GPUTextureDescriptor}} \|descriptor\|

Add validation for GPUDevice.createTexture() #799

Add validation for GPUDevice.createTexture() #799

Uh oh!

Conversation

litherum commented May 21, 2020 • edited by pr-preview bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kvark left a comment

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!

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!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

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!

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!

Kangz commented Jul 3, 2020

Uh oh!

Kangz commented Jul 17, 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!

Uh oh!

litherum commented May 21, 2020 •

edited by pr-preview bot

Loading