[Public WebGL] Re: WEBGL_get_buffer_sub_data_async

Kenneth Russell [email protected]
Wed Jan 11 18:23:09 PST 2017


The reason Promises were chosen for this extension is that they're the
current best practice on the web platform for expressing asynchronous
computation. The syntax is fluent, and it would have been difficult for me
to argue with Chrome's leadership that there was a significant enough
performance implication that a raw callback should be passed, contrary to
the design of all other current web specs.

Kai and Jaume Sanchez ( https://www.clicktorelease.com/ ) are preparing
another test case based on one of Jaume's global illumination examples.
Once that's ready we'll post more data. At a high level, using this
asynchronous API lets Chrome's deeply pipelined WebGL implementation match
the performance of single-threaded and/or single-process WebGL
implementations, though the code has to be changed to know about and take
advantage of the extension.

-Ken


On Fri, Jan 6, 2017 at 9:33 AM, Maksims Mihejevs <[email protected]> wrote:

> That is very good point too.
> Most common case of similar pattern is mouse, keyboard, touch input - it
> is in callback, and you end up collecting own state information about
> input, so to access it in main loop of your app in requestAnimationFrame.
>
> Developers coming from real-time applications and other platforms get
> confused a lot by this all the time. That is based on many users on
> PlayCanvas from hobbyists to commercial clients do mention it time to times.
>
> On 6 January 2017 at 16:36, Florian Bösch <[email protected]> wrote:
>
>> The problem isn't just the promise allocation, it's also the closure
>> allocation. And while you can avoid that in some cases, you can't avoid it
>> in all cases. And beyond that, the further problem of registering callbacks
>> to do stuff would be that it lifts rendering logic out of the
>> requestAnimationFrame loop and occurs... whenever. Which isn't a proper way
>> to render. So in practice you'd end up attaching a callback just to collect
>> the result and put it in a variable you'll check for not being null at the
>> start of the next requestAnimationFrame invocation. Of course when you do
>> that, the entire exercise of providing a callback is moot. Since that'd be
>> the most common usage pattern that everybody would recommend based on sound
>> reasoning about the predictability of rendering in a requestAnimationFrame
>> loop, that makes promises/callbacks entirely moot and the wrong thing to do.
>>
>> On Fri, Jan 6, 2017 at 1:52 PM, Maksims Mihejevs <[email protected]>
>> wrote:
>>
>>> Why not to use classic callbacks with (err, response) arguments?
>>>
>>> On 6 January 2017 at 00:01, Kenneth Russell <[email protected]> wrote:
>>>
>>>> On Thu, Jan 5, 2017 at 2:55 AM, Maksims Mihejevs <[email protected]>
>>>> wrote:
>>>>
>>>>> Thank you for answer, this examples make sense.
>>>>>
>>>>> The only concern as mentioned before are: Promises. There is no a
>>>>> single use of promises in WebGL, and there is a reason why they should not
>>>>> be used in real-time applications if they can be avoided (which is the
>>>>> case).
>>>>>
>>>>> If possible Fence (Sync Objects) were involved in design process?
>>>>> Could something like this be achieved by using more generic functionality
>>>>> in GL such as Sync Objects?
>>>>>
>>>>
>>>> Promises were chosen for a particular reason. Even if a primitive like
>>>> mapBufferRange were exposed in WebGL, a typed array would still need to be
>>>> allocated as the return result. There's no provision in the ECMAScript
>>>> specification for "re-pointing" a typed array at a new backing store,
>>>> although Chrome's V8 team proposed this a while back. Mapping a buffer's
>>>> range at the OpenGL ES level returns a void*, and that would need to be
>>>> exposed to ECMAScript. Passing in an already-allocated typed array wouldn't
>>>> work in the current form of the APIs.
>>>>
>>>> Given that an allocation would be needed anyway, returning a Promise is
>>>> much more web-friendly. It avoids the need for the application to deal with
>>>> sync objects, and potentially inconsistent states of the returned typed
>>>> array. getBufferSubDataAsync lets the application specify where it wants
>>>> the returned data to be copied.
>>>>
>>>> -Ken
>>>>
>>>>
>>>>
>>>>> Kind Regards,
>>>>> Max
>>>>>
>>>>> On 4 January 2017 at 18:55, Kai Ninomiya <[email protected]> wrote:
>>>>>
>>>>>> Max,
>>>>>>
>>>>>> In my demo [1] there are 3 different possible readback paths:
>>>>>> * readPixels to CPU [2]
>>>>>> * readPixels to PBO + getBufferSubData [3]
>>>>>> * readPixels to PBO + getBufferSubDataAsync [4]
>>>>>>
>>>>>> As you said, getBufferSubData(/Async) can be used for reading back
>>>>>> any buffer data (such as transform feedback or GPGPU shader results). PBO
>>>>>> is necessary (AFAIK) for async readback from framebuffer data (note: an
>>>>>> async readPixels wouldn't be as useful as it would block any operation
>>>>>> which writes to the framebuffer).
>>>>>>
>>>>>> -Kai
>>>>>>
>>>>>> [1] https://github.com/kainino0x/getBufferSubDataAsync-Demo/
>>>>>> [2] https://github.com/kainino0x/getBufferSubDataAsync-Demo/
>>>>>> blob/master/index.html#L245
>>>>>> [3] https://github.com/kainino0x/getBufferSubDataAsync-Demo/
>>>>>> blob/master/index.html#L261
>>>>>> [4] https://github.com/kainino0x/getBufferSubDataAsync-Demo/
>>>>>> blob/master/index.html#L280
>>>>>>
>>>>>> On Wed, Jan 4, 2017 at 5:09 AM Maksims Mihejevs <[email protected]>
>>>>>> wrote:
>>>>>>
>>>>>>> From PlayCanvas side, we express a need for async glReadPixels path
>>>>>>> too. We and our users have been using it in many ways, some of the ways:
>>>>>>> 1. GPU picking: ID encoded in unique colour, reading pixel under
>>>>>>> mouse.
>>>>>>> 2. GPU screen to world: reading pixel from depth texture, and using
>>>>>>> frustum with math reconstructing world position.
>>>>>>> 3. Render Target to another Canvas. In Editor we have thumbnail
>>>>>>> previews for materials, models, cubemaps and other assets. We render them
>>>>>>> into render target in main context and then reading pixels to create
>>>>>>> ImageData so it can be put to another canvas using putImageData.
>>>>>>> 4. Some custom algorithms to generate large amounts of computation
>>>>>>> heavy data saved into texture, then read on CPU - this depends per case.
>>>>>>> Sometimes async approach is viable there, sometimes it is not.
>>>>>>>
>>>>>>> In many cases glReadPixels is called per each frame, like for
>>>>>>> picking, and easily can drop frame rate due to blocking nature.
>>>>>>>
>>>>>>> I've noticed that PBOs are mentioned in WebGL 2.0 spec, but not much
>>>>>>> info apart of just that mention: https://www.khronos.o
>>>>>>> rg/registry/webgl/specs/latest/2.0/#4.2
>>>>>>> PBOs would allow to get render target data into buffers without
>>>>>>> stalling GPU pipeline, and then read them.
>>>>>>>
>>>>>>> How does getBufferSubDataAsync relates to PBOs?
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Max
>>>>>>>
>>>>>>> On 4 January 2017 at 05:27, Kenneth Russell <[email protected]> wrote:
>>>>>>>
>>>>>>> Apologies for not discussing this extension on public_webgl before
>>>>>>> introducing it as a draft in the WebGL extension registry.
>>>>>>>
>>>>>>> The cost of synchronous glReadPixels has been a longstanding problem
>>>>>>> in WebGL. The Chrome browser specifically has a particularly deep graphics
>>>>>>> pipeline, and draining it with a synchronous call each frame imposes a
>>>>>>> too-great performance penalty. This has forced applications to rewrite
>>>>>>> certain algorithms when porting to WebGL.
>>>>>>>
>>>>>>> getBufferSubDataAsync is a direct parallel to getBufferSubData, and
>>>>>>> solves these performance pitfalls in Chrome. We've gathered data from two
>>>>>>> test cases so far, a GPU-based picking algorithm and a GPGPU global
>>>>>>> illumination algorithm, and the results look good. We will present this
>>>>>>> data on public_webgl soon, when making a case for moving the extension
>>>>>>> forward.
>>>>>>>
>>>>>>> -Ken
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Mon, Jan 2, 2017 at 2:27 PM, Maksims Mihejevs <[email protected]
>>>>>>> > wrote:
>>>>>>>
>>>>>>> Worth mentioning that promises are extremely bad for GC and
>>>>>>> real-time applications, they do not provide a developer enough control to
>>>>>>> structure logic so to avoids any allocations.
>>>>>>>
>>>>>>> Promises - are not good for real-time at all, and lead to issues
>>>>>>> with GC. Any API in WebGL that is meant to be used in real-time
>>>>>>> applications should not be based on API's that are not real-time friendly.
>>>>>>>
>>>>>>> On 2 January 2017 at 20:49, Florian Bösch <[email protected]> wrote:
>>>>>>>
>>>>>>> Upon thinking about this extension, I don't think it should exist at
>>>>>>> all. Ideally the mapBuffer semantic would be exposed. But even if it isn't,
>>>>>>> it shall not be that an extension is required to express functionality
>>>>>>> already found in the core functionality of the underlying ES specification.
>>>>>>>
>>>>>>> Furthermore, getBufferSubDataAsync does not adequately express the
>>>>>>> reality of map/flush/unmap, and hides the fact that unmap/flush are still
>>>>>>> synchronizing calls happening. However getBufferSubDataAsync obstructs
>>>>>>> appropriate code dealing with proper insertion of synchronization points.
>>>>>>>
>>>>>>> In addition, it would lead to allocating promises once or many times
>>>>>>> per frame, and since tracking would be required in some instances, would
>>>>>>> also lead to allocating a closure once or many times a call. An issue that
>>>>>>> map buffer range does not exhibit.
>>>>>>>
>>>>>>> Due to the lack of discussion of this feature, I believe a great
>>>>>>> disservice is done to WebGL 2 by the introduction of these ideas/APIs and I
>>>>>>> strongly suggest to withdraw this from draft immediately and go back to the
>>>>>>> drawing board.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Mon, Jan 2, 2017 at 5:48 PM, Florian Bösch <[email protected]>
>>>>>>> wrote:
>>>>>>>
>>>>>>> This extension https://www.khronos.org/registry/webgl/extensions/
>>>>>>> WEBGL_get_buffer_sub_data_async/ has been introduced and elevated
>>>>>>> to draft without any public discussion.
>>>>>>>
>>>>>>> In a nutshell it proposes a new WebGL2 function called
>>>>>>> getBufferSubDataAsync which returns a promise that will be called
>>>>>>> eventually with the buffer data.
>>>>>>>
>>>>>>> I think there are several problems:
>>>>>>>
>>>>>>>    1. The extension process states that "*Extensions move through
>>>>>>>    four states during their development: proposed, draft, community approved,
>>>>>>>    and Khronos ratified**"*. This extension never moved through the
>>>>>>>    proposal stage.
>>>>>>>    2. The extension introduces promises to the WebGL API. This
>>>>>>>    requires a more fundamental discussion.
>>>>>>>    3. A discussion if this extension is required if WebWorkers can
>>>>>>>    access the same context as the main thread has not happened.
>>>>>>>
>>>>>>> This extension should be in proposal status, and the necessary
>>>>>>> discussions should happen first.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://khronos.org/pipermail/public_webgl_khronos.org/attachments/20170111/e327bc02/attachment.html>


More information about the public_webgl mailing list