Add validation rules for copyBufferToTexture #648

Jiawei-Shao · 2020-03-26T08:23:40Z

spec/index.bs

Kangz · 2020-03-26T09:26:49Z

Thanks, it looks very good overall, I just wish the usage rules where grouped semantically, as right now it's difficult to make sure everything is handled. Maybe it could be:

the following validation rules apply.

For {{encoder}}:

 - encoder must be a valid GPUCommandEncoder.
 - ...

For source.buffer:

For source.texture:

For the copy being in-bounds:

For the texel block alignment:

For other source members:

For other destination members:

kvark

Looks great! Just have a few notes

spec/index.bs

kvark · 2020-03-26T22:26:38Z

spec/index.bs

+  - If |source|.{{GPUBufferCopyView/bytesPerRow}} is 0, the number of |bytes in a complete row| must be a multiple of 256.
+  - |destination|.{{GPUTextureCopyView/mipLevel}} must be less than the {{GPUTexture/[[mipLevelCount]]}} of |destination|.{{GPUTextureCopyView/texture}}.
+  - |destination|.{{GPUTextureCopyView/arrayLayer}} must be less than the {{GPUTexture/[[textureSize]]}}.{{GPUExtent3DDict/depth}} of |destination|.{{GPUTextureCopyView/texture}}.
+  - {{GPUExtent3DDict/depth}} must be 1 when the {{GPUTexture/[[dimension]]}} of {{GPUTextureCopyView/texture}} is {{GPUTextureDimension/2d}}.


also, for GPUTextureDimension/1d

Actually not true for 2d textures, because all 2d textures are semantically 2d texture arrays. Instead, this should just say the depth must be <= the texture's depth.

Metal only supports copying into one layer of a 2d texture:
https://developer.apple.com/documentation/metal/mtlblitcommandencoder/1400752-copyfrombuffer?language=objc
"when you copy to a 2D texture, depth must be 1."

I think we want to expose multi-layer copies, we can implement them as multiple copies when necessary

"when you copy to a 2D texture, depth must be 1."

I'm pretty sure this is talking about 2D non-array textures specifically.

That documentation is kind of confusing though. On sourceBytesPerImage it says:

For 3D textures and 2D array textures, the stride in bytes between 2D images of the source buffer memory. The value must be a multiple of the destination texture's pixel size, in bytes.

which strongly implies that copies into multiple layers of a 2D array texture is allowed. But destinationSlice doesn't say anything about it. I think destinationSlice is actually destinationFirstSlice?

Hi @JusSn @litherum, what do you think?

I also notice that in MoltenVK the multi-layer copy is implemented by an iteration over every layers in a for-loop.

spec/index.bs

kainino0x · 2020-03-27T00:21:26Z

spec/index.bs

+  - If |source|.{{GPUBufferCopyView/bytesPerRow}} is 0, the number of |bytes in a complete row| must be a multiple of 256.
+  - |destination|.{{GPUTextureCopyView/mipLevel}} must be less than the {{GPUTexture/[[mipLevelCount]]}} of |destination|.{{GPUTextureCopyView/texture}}.
+  - |destination|.{{GPUTextureCopyView/arrayLayer}} must be less than the {{GPUTexture/[[textureSize]]}}.{{GPUExtent3DDict/depth}} of |destination|.{{GPUTextureCopyView/texture}}.
+  - {{GPUExtent3DDict/depth}} must be 1 when the {{GPUTexture/[[dimension]]}} of {{GPUTextureCopyView/texture}} is {{GPUTextureDimension/2d}}.


Actually not true for 2d textures, because all 2d textures are semantically 2d texture arrays. Instead, this should just say the depth must be <= the texture's depth.

spec/index.bs

Jiawei-Shao · 2020-03-30T08:40:51Z

spec/index.bs

+  - |copySize|.[=Extent3D/width=] cannot be 0.
+  - |copySize|.[=Extent3D/height=] cannot be 0.
+  - |copySize|.[=Extent3D/depth=] cannot be 0.


I found on D3D12 it is not allowed to specify copySize.width, copySize.height or copySize.depth to 0 when I was writing SPEC for copy commands, so here I suggest we not allow copySize.width, copySize.height or copySize.depth be 0.

Here are the error message reported from D3D12 validation layer:

D3D12 ERROR: ID3D12CommandList::CopyTextureRegion: D3D12_SUBRESOURCE_FOOTPRINT::Width, D3D12_SUBRESOURCE_FOOTPRINT::Height, and D3D12_SUBRESOURCE_FOOTPRINT::Depth cannot be 0. The minimum value is 1. Width = 64, Height = 0, and Depth = 1. [ RESOURCE_MANIPULATION ERROR #854: COPYTEXTUREREGION_INVALIDDSTDIMENSIONS]

Imho empty copies should be allowed, because making them noops is as much work as validating them out.

So you are suggesting that "it is no-op if any of the copySize.width, copySize.height or copySize.depth is 0"?

Yes, but that can be non-normative ("Note: ...") because when we actually define the copies they'll probably be done as for-loops which describe that behavior.

spec/index.bs

Jiawei-Shao · 2020-03-30T08:42:42Z

spec/index.bs

+  The following validation rules apply:
+  - |textureCopyView|.{{GPUTextureCopyView/texture}} must be a [=valid=] {{GPUTexture}}.
+  - The {{GPUTexture/[[sampleCount]]}} of |textureCopyView|.{{GPUTextureCopyView/texture}} must be 1.
+  - |textureCopyView|.{{GPUTextureCopyView/mipLevel}} must be less than the {{GPUTexture/[[mipLevelCount]]}} of
+    |textureCopyView|.{{GPUTextureCopyView/texture}}.
+  - If the |textureCopyView|.{{GPUTextureCopyView/texture}} is {{GPUTextureDimension/1d}} or {{GPUTextureDimension/3d}}:
+    - |textureCopyView|.{{GPUTextureCopyView/arrayLayer}} must be zero.
+  - If the {{GPUTexture/[[dimension]]}} of |textureCopyView|.{{GPUTextureCopyView/texture}} is {{GPUTextureDimension/2d}}:
+    - |textureCopyView|.{{GPUTextureCopyView/arrayLayer}} must be less than the {{GPUTexture/[[textureSize]]}}.[=Extent3D/depth=] of |textureCopyView|.{{GPUTextureCopyView/texture}}.
+  - |textureCopyView|.{{GPUTextureCopyView/origin}}.{{GPUOrigin3DDict/x}} must be a multiple of |blockWidth|.
+  - |textureCopyView|.{{GPUTextureCopyView/origin}}.{{GPUOrigin3DDict/y}} must be a multiple of |blockHeight|.


These rules are the same for both buffer-to-texture and texture-to-buffer copies so I decide to put it here.

spec/index.bs

kainino0x · 2020-03-30T22:19:47Z

nit: Please mark comment threads as "resolved" when you resolve them so they get minimized; you can do that instead of commenting "Done". :)

kainino0x

Sorry for another huge review!! This is looking really good overall!

spec/index.bs

Jiawei-Shao · 2020-04-03T08:58:57Z

OK. There are totally 5 TODOs:

Always allow bytesPerRow == 0?
Allow CopySize == 0? (D3D12 doesn't allow this)
Allow copying multiple 2D array layers in one copy command? (Metal doesn't support this)
Copy with 1D and 3D textures (need more investigation)
Copy with depth stencil textures (Copying of depth24plus formats #652)

kainino0x · 2020-04-03T23:07:01Z

There was agreement in the F2F here (based on my recollection - the minutes don't seem to capture it) that rowsPerImage=0 would only be allowed for 2d copies (there's no "tightly packed" mode for it). I thought we also agreed there would be no "tightly packed" mode for bytesPerRow but maybe @kvark @JusSn could give explicit approval on that (and the other issues).

Note that it would be a nonbreaking change to allow "0 = tightly packed" later on if we really find it's useful.

kainino0x · 2020-04-06T19:49:06Z

Resolutions from the meeting:

There are totally 5 TODOs:

Always allow bytesPerRow == 0?

bytesPerRow = 0 has no special meaning. Have a rule saying two rows of the copy may not overlap, making 0 invalid if height>1.

rowsPerImage = 0 has no special meaning. Have a rule saying two images of the copy may not overlap, making 0 invalid if depth > 1.

Allow CopySize == 0? (D3D12 doesn't allow this)

Yes, but don't explicitly call it out (except maybe in non-normative text). Just don't disallow CopySize=0.

Allow copying multiple 2D array layers in one copy command? (Metal doesn't support this)

Yes.

Copy with 1D and 3D textures (need more investigation)

Copy with depth stencil textures (Copying of depth24plus formats #652)

Leave TODOs in the spec for now.

Jiawei-Shao · 2020-04-07T08:55:55Z

Hi @kainino0x, I have updated all these changes in the latest commit. PTAL, thanks!

kainino0x

latest commit lgtm with nits, but I still have a bunch of open comments that I'd like to get resolved at some point.

I think we should merge this without those. I think I'll go ahead and take care of some syntax stuff right now.

spec/index.bs

kainino0x · 2020-04-07T19:07:56Z

@kvark @JusSn would one of you like to do a quick review before we merge this and handle the remaining comments in followup?

Jiawei-Shao · 2020-04-08T01:39:43Z

Hi @kainino0x, I have updated this PR. PTAL, thanks!

kvark · 2020-04-08T19:06:39Z

spec/index.bs

-texels, used when copying data from or to a {{GPUTexture}}.
-
-  - For {{GPUTextureDimension/2d}} textures, data is copied from or to one [=mipmap level=] and one [=array layer=] of the [=texture=].
+A {{GPUTextureCopyView}} is a view of a sub-region of one or multiple contiguous [=texture=] [=subresources=] with the initial


we may need to put an Issue here to clarify what "contiguous" means.
I.e. if I select mipmaps 1-2 and array layers 1-2, is that sub-region contiguous?

kainino0x · 2020-04-08T22:07:34Z

Last commit LGTM.

Jiawei, will you be able to take a followup to address the rest of the unresolved comments on this PR?

Jiawei-Shao · 2020-04-09T00:39:45Z

@kainino0x @kvark @Kangz thanks for all your suggestions!

Jiawei, will you be able to take a followup to address the rest of the unresolved comments on this PR?

Yes I will. I plan to do further investigations after finishing another two PRs about the copy commmands (texture-to-buffer and texture-to-texture).

* Add validation rules for copyBufferToTexture * Small fix * Address more reviewer's suggestions * Address more comments from the reviewers * Fix several typos

kainino0x · 2020-06-19T22:02:26Z

@Jiawei-Shao I think the open issues on this PR are still open. Will you be able to take them?

Also, I just noticed that the rendered indentation for "For other members in layout:" in the spec right now is broken.

Jiawei-Shao · 2020-06-22T01:03:02Z

@kainino0x Yes I will.

Here are the remaining two opens:

Copy with 1D and 3D textures (need more investigation, I will take this one first)
Copy with depth stencil textures (Copying of depth24plus formats #652)

Also now we don't allow bytesPerRow==0 when height > 1, this rule should also be implemented in Dawn.

kainino0x · 2020-06-22T21:43:46Z

I'm mainly talking about the copyediting issues outstanding above: indentation, naming, incorrect usage of [=valid=], redundant rules

* Depth/stencil formats should be invalid for 3D textures * Address feedback from Kai: need to revise test name or description * Address feedback from Kai

Jiawei-Shao added 2 commits March 26, 2020 16:26

Add validation rules for copyBufferToTexture

331e14f

Small fix

4e8d6b0

Kangz reviewed Mar 26, 2020

View reviewed changes

spec/index.bs Outdated Show resolved Hide resolved

Kangz reviewed Mar 26, 2020

View reviewed changes

spec/index.bs Outdated Show resolved Hide resolved

Kangz reviewed Mar 26, 2020

View reviewed changes

spec/index.bs Outdated Show resolved Hide resolved

kvark mentioned this pull request Mar 26, 2020

Specify the state machine of GPUCommandEncoder #653

Closed

kvark requested changes Mar 26, 2020

View reviewed changes

kainino0x reviewed Mar 27, 2020

View reviewed changes

Address more reviewer's suggestions

65d1f2f

Jiawei-Shao commented Mar 30, 2020

View reviewed changes

spec/index.bs Show resolved Hide resolved

Jiawei-Shao commented Mar 30, 2020

View reviewed changes

kvark reviewed Mar 30, 2020

View reviewed changes

spec/index.bs Outdated Show resolved Hide resolved

kainino0x reviewed Mar 31, 2020

View reviewed changes

Address more comments from the reviewers

2965138

kainino0x approved these changes Apr 7, 2020

View reviewed changes

spec/index.bs Outdated Show resolved Hide resolved

spec/index.bs Outdated Show resolved Hide resolved

spec/index.bs Outdated Show resolved Hide resolved

Fix several typos

abfe1a4

kvark self-requested a review April 8, 2020 19:02

kvark approved these changes Apr 8, 2020

View reviewed changes

kvark merged commit 03da857 into gpuweb:master Apr 8, 2020

Add validation rules for copyBufferToTexture #648

Add validation rules for copyBufferToTexture #648

Uh oh!

Conversation

Jiawei-Shao commented Mar 26, 2020 • edited by pr-preview bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Kangz commented Mar 26, 2020

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!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

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!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kainino0x commented Mar 30, 2020

Uh oh!

kainino0x left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Jiawei-Shao commented Apr 3, 2020

Uh oh!

kainino0x commented Apr 3, 2020

Uh oh!

kainino0x commented Apr 6, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Jiawei-Shao commented Apr 7, 2020

Uh oh!

kainino0x left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Jiawei-Shao commented Mar 26, 2020 •

edited by pr-preview bot

Loading

kainino0x commented Apr 6, 2020 •

edited

Loading