Re: Intent to Experiment: CSS Houdini Paint API Level 1
Le jeudi 5 janvier 2017 11:01:04 UTC-5, Jet Villegas a écrit : > Spec: https://drafts.css-houdini.org/css-paint-api/ > > Summary: The CSS Paint API is the first of several Web Rendering proposals > from the CSS Houdini Task Force. The CSS Paint API allows Web authors to > define and register a custom Paint method to be executed by the Layout > engine as other elements are rendered. > > Bug: https://bugzilla.mozilla.org/show_bug.cgi?id=1302328 > > Link to standard: https://drafts.css-houdini.org/css-paint-api/ > > Platform coverage: Android, Desktop > > Estimated or target release: TBD > > Preference behind which this will be implemented: TBD > > DevTools bug: TBD > > Tests - TBD > > Implementation Details: > > CSS Paint API depends on the implementation of the following features: > CSS Houdini Properties & Values - https://bugzilla.mozilla.org/ > show_bug.cgi?id=1273706 > Houdini "Worklets" - https://bugzilla.mozilla.org/show_bug.cgi?id=1290021 > > The planned implementation in Gecko builds upon the HTML5 Canvas2D API to > provide the rendering surface for CSS Paint. > > Risks: > > The experimental implementation has progressed enough to warrant this > public Intent to Implement. However, significant risks will need to be > mitigated by careful design and execution if these features are to pass the > experimental stage and ship in Firefox, including: > > 1. Use Cases. It's not clear that the use cases proposed in the > specification warrant the additional Rendering System complexity. Apart > from conical gradients, we haven't seen many author requests for the other > use cases. If the existing Canvas2D feature set is lacking, what are the > compelling use cases and maximally useful API for such use cases? It's not > clear that the proposed Canvas2D-based API is desirable over a different > API design (eg., WebGL) if most use cases need to directly manipulate > pixels. > > 2. Rendering Performance. The planned Canvas2D backing store approach may > be too slow for real-world usage of the API. In the future, we may replace > the Canvas2D approach to have the custom paint methods create Layout > (displayList) nodes for direct rendering by the Layout engine, bypassing > the need for a Canvas2D backing store. It's worth noting that the Paint API > isn't directly compatible with existing displayList nodes (e.g., support > for raw paths, funny shapes, & pixel manipulation.) > > There may also be other performance issues that arise with the API's usage > in combination with existing CSS features (e.g., CSS Masking, Filters, > etc.) The displayList vs. canvas bitmap implementation would probably look > a good bit different in WebRender. It's also worth noting that multiple > implementations shipping a bitmap-based version can create dependencies > that prevent us from switching to a faster alternative version in the > future. > > 3. Integration with Gecko Architecture. The Quantum Project < > https://wiki.mozilla.org/Quantum> is a major overhaul of the Firefox > Rendering Engine. Implementing the CSS Paint API while that effort is in > progress may add significant impedance. However, a counter-argument is that > we should design Quantum to allow for such extensibility in the future. > Duplication of work ( writing code that would need to be rewritten for > Quantum ) is not desirable and should be avoided. > > For WebRender/Quantum, we could initially push this through the same path > that SVG will go through (which will be rasterized on the CPU and then > cached in a GPU texture atlas). It does seem like Houdini Paint could > reduce the amount of acceleration we can do on the GPU (at least in the > short term), but we won't be any worse off than other browsers in that > regard. > > 4. Dependency on incomplete implementations/specifications. The dependency > creates a chicken/egg scenario where we can't sufficiently evaluate the > dependent specifications (e.g, Worklets, and Properties & Values) without > also implementing an initial key use case (e.g., CSS Paint or Worklets for > Web Audio.) This is somewhat mitigated by getting the implementation far > enough along to formulate informed opinions on all the specifications. > > The Houdini Task Force is meeting next week in Seattle to discuss this and > other specifications. ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
On Fri, Jan 6, 2017 at 9:19 PM, Boris Zbarsky wrote: > On 1/6/17 2:34 PM, Bill McCloskey wrote: > >> Are in-thread worklets a thing? >> > > Unclear. > Right. So far, I'm concentrating on having Worklet able to run out of main-thread. My goal is to have something ready for audioWorklet, then, I want to have something able to run on any thread. The code is not ready yet, but the main idea is to have worklet thread agnostic. ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
On Fri, Jan 6, 2017 at 12:19 PM, Boris Zbarsky wrote: > On 1/6/17 2:34 PM, Bill McCloskey wrote:The patch that allows tab switch > painting > >> while content JS runs isn't going to be backed out. If the tab switch >> paints allow JS to run, and that JS is from the same context as the >> content >> JS that was interrupted, then that would be bad. >> > > What does "same context" mean here? An in-thread worklet would be > same-JSContext/JSRuntime/whatever, but different-global and > shared-nothing (at least observably from JS) with any web pages. > > Is the badness you refer to page-observable JS-level reentry, or something > deeper in our engine internals? Because the former is not an issue with > worklets, I expect. The badness is just page-observable JS-level reentry (that I know of at least). So that sounds good. Thanks. -Bill ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
On 1/6/17 2:34 PM, Bill McCloskey wrote: Are in-thread worklets a thing? Unclear. The spec is agnostic on where worklets run. I know Chrome was running some of them in-thread (but not the details). For us, it's basically an implementation decision. The patch that allows tab switch painting while content JS runs isn't going to be backed out. If the tab switch paints allow JS to run, and that JS is from the same context as the content JS that was interrupted, then that would be bad. What does "same context" mean here? An in-thread worklet would be same-JSContext/JSRuntime/whatever, but different-global and shared-nothing (at least observably from JS) with any web pages. Is the badness you refer to page-observable JS-level reentry, or something deeper in our engine internals? Because the former is not an issue with worklets, I expect. -Boris ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
On Fri, Jan 6, 2017 at 6:20 AM, Boris Zbarsky wrote: > On 1/6/17 5:53 AM, Xidorn Quan wrote: > >> I guess that may also mean billm's bug 1308039 [1] which tried to do >> painting during GC (which will be backed out soon [2] because of lots of >> breakage) may never get chance to land anymore, since this API means it >> is expected to run JS code during painting. >> > > That should be OK, I think, as long as the paint worklets are on a > separate thread (and hence in a separate runtime) from the JS which is > doing GC when we decide to paint. > > In-thread worklets would be a problem for paint-during-gc, though. Are in-thread worklets a thing? The patch that allows tab switch painting while content JS runs isn't going to be backed out. If the tab switch paints allow JS to run, and that JS is from the same context as the content JS that was interrupted, then that would be bad. -Bill ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
On 1/6/17 5:53 AM, Xidorn Quan wrote: I guess that may also mean billm's bug 1308039 [1] which tried to do painting during GC (which will be backed out soon [2] because of lots of breakage) may never get chance to land anymore, since this API means it is expected to run JS code during painting. That should be OK, I think, as long as the paint worklets are on a separate thread (and hence in a separate runtime) from the JS which is doing GC when we decide to paint. In-thread worklets would be a problem for paint-during-gc, though. -Boris ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
On 01/05/2017 06:00 PM, Jet Villegas wrote: Spec: https://drafts.css-houdini.org/css-paint-api/ Summary: The CSS Paint API is the first of several Web Rendering proposals from the CSS Houdini Task Force. The CSS Paint API allows Web authors to define and register a custom Paint method to be executed by the Layout engine as other elements are rendered. Bug: https://bugzilla.mozilla.org/show_bug.cgi?id=1302328 Link to standard: https://drafts.css-houdini.org/css-paint-api/ Platform coverage: Android, Desktop Estimated or target release: TBD Preference behind which this will be implemented: TBD DevTools bug: TBD Tests - TBD Implementation Details: CSS Paint API depends on the implementation of the following features: CSS Houdini Properties & Values - https://bugzilla.mozilla.org/ show_bug.cgi?id=1273706 As you probably know, baku is working on to get support for proper worklets, not just a prototype, supporting running in different threads and so (AudioWorklets really need that). https://bugzilla.mozilla.org/show_bug.cgi?id=1328964 Houdini "Worklets" - https://bugzilla.mozilla.org/show_bug.cgi?id=1290021 The planned implementation in Gecko builds upon the HTML5 Canvas2D API to provide the rendering surface for CSS Paint. Risks: The experimental implementation has progressed enough to warrant this public Intent to Implement. However, significant risks will need to be mitigated by careful design and execution if these features are to pass the experimental stage and ship in Firefox, including: 1. Use Cases. It's not clear that the use cases proposed in the specification warrant the additional Rendering System complexity. Apart from conical gradients, we haven't seen many author requests for the other use cases. If the existing Canvas2D feature set is lacking, what are the compelling use cases and maximally useful API for such use cases? It's not clear that the proposed Canvas2D-based API is desirable over a different API design (eg., WebGL) if most use cases need to directly manipulate pixels. 2. Rendering Performance. The planned Canvas2D backing store approach may be too slow for real-world usage of the API. In the future, we may replace the Canvas2D approach to have the custom paint methods create Layout (displayList) nodes for direct rendering by the Layout engine, bypassing the need for a Canvas2D backing store. It's worth noting that the Paint API isn't directly compatible with existing displayList nodes (e.g., support for raw paths, funny shapes, & pixel manipulation.) There may also be other performance issues that arise with the API's usage in combination with existing CSS features (e.g., CSS Masking, Filters, etc.) The displayList vs. canvas bitmap implementation would probably look a good bit different in WebRender. It's also worth noting that multiple implementations shipping a bitmap-based version can create dependencies that prevent us from switching to a faster alternative version in the future. Couldn't we add some new types of worklets for different APIs? Or, has it been discussed to have a method to access the worklet? So, not window.paintWorklet but window.getPaintWorklet("2d") which could be extended later, similar to canvas context, to support different types of paint worklets. 3. Integration with Gecko Architecture. The Quantum Project < https://wiki.mozilla.org/Quantum> is a major overhaul of the Firefox Rendering Engine. Implementing the CSS Paint API while that effort is in progress may add significant impedance. However, a counter-argument is that we should design Quantum to allow for such extensibility in the future. Duplication of work ( writing code that would need to be rewritten for Quantum ) is not desirable and should be avoided. For WebRender/Quantum, we could initially push this through the same path that SVG will go through (which will be rasterized on the CPU and then cached in a GPU texture atlas). It does seem like Houdini Paint could reduce the amount of acceleration we can do on the GPU (at least in the short term), but we won't be any worse off than other browsers in that regard. 4. Dependency on incomplete implementations/specifications. The dependency creates a chicken/egg scenario where we can't sufficiently evaluate the dependent specifications (e.g, Worklets, and Properties & Values) without also implementing an initial key use case (e.g., CSS Paint or Worklets for Web Audio.) This is somewhat mitigated by getting the implementation far enough along to formulate informed opinions on all the specifications. The Houdini Task Force is meeting next week in Seattle to discuss this and other specifications. ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
Re: Intent to Experiment: CSS Houdini Paint API Level 1
Great news! I guess that may also mean billm's bug 1308039 [1] which tried to do painting during GC (which will be backed out soon [2] because of lots of breakage) may never get chance to land anymore, since this API means it is expected to run JS code during painting. [1] https://bugzilla.mozilla.org/show_bug.cgi?id=1308039 [2] https://bugzilla.mozilla.org/show_bug.cgi?id=1328423 - Xidorn On Fri, Jan 6, 2017, at 03:00 AM, Jet Villegas wrote: > Spec: https://drafts.css-houdini.org/css-paint-api/ > > Summary: The CSS Paint API is the first of several Web Rendering > proposals > from the CSS Houdini Task Force. The CSS Paint API allows Web authors to > define and register a custom Paint method to be executed by the Layout > engine as other elements are rendered. > > Bug: https://bugzilla.mozilla.org/show_bug.cgi?id=1302328 > > Link to standard: https://drafts.css-houdini.org/css-paint-api/ > > Platform coverage: Android, Desktop > > Estimated or target release: TBD > > Preference behind which this will be implemented: TBD > > DevTools bug: TBD > > Tests - TBD > > Implementation Details: > > CSS Paint API depends on the implementation of the following features: > CSS Houdini Properties & Values - https://bugzilla.mozilla.org/ > show_bug.cgi?id=1273706 > Houdini "Worklets" - https://bugzilla.mozilla.org/show_bug.cgi?id=1290021 > > The planned implementation in Gecko builds upon the HTML5 Canvas2D API to > provide the rendering surface for CSS Paint. > > Risks: > > The experimental implementation has progressed enough to warrant this > public Intent to Implement. However, significant risks will need to be > mitigated by careful design and execution if these features are to pass > the > experimental stage and ship in Firefox, including: > > 1. Use Cases. It's not clear that the use cases proposed in the > specification warrant the additional Rendering System complexity. Apart > from conical gradients, we haven't seen many author requests for the > other > use cases. If the existing Canvas2D feature set is lacking, what are the > compelling use cases and maximally useful API for such use cases? It's > not > clear that the proposed Canvas2D-based API is desirable over a different > API design (eg., WebGL) if most use cases need to directly manipulate > pixels. > > 2. Rendering Performance. The planned Canvas2D backing store approach may > be too slow for real-world usage of the API. In the future, we may > replace > the Canvas2D approach to have the custom paint methods create Layout > (displayList) nodes for direct rendering by the Layout engine, bypassing > the need for a Canvas2D backing store. It's worth noting that the Paint > API > isn't directly compatible with existing displayList nodes (e.g., support > for raw paths, funny shapes, & pixel manipulation.) > > There may also be other performance issues that arise with the API's > usage > in combination with existing CSS features (e.g., CSS Masking, Filters, > etc.) The displayList vs. canvas bitmap implementation would probably > look > a good bit different in WebRender. It's also worth noting that multiple > implementations shipping a bitmap-based version can create dependencies > that prevent us from switching to a faster alternative version in the > future. > > 3. Integration with Gecko Architecture. The Quantum Project < > https://wiki.mozilla.org/Quantum> is a major overhaul of the Firefox > Rendering Engine. Implementing the CSS Paint API while that effort is in > progress may add significant impedance. However, a counter-argument is > that > we should design Quantum to allow for such extensibility in the future. > Duplication of work ( writing code that would need to be rewritten for > Quantum ) is not desirable and should be avoided. > > For WebRender/Quantum, we could initially push this through the same path > that SVG will go through (which will be rasterized on the CPU and then > cached in a GPU texture atlas). It does seem like Houdini Paint could > reduce the amount of acceleration we can do on the GPU (at least in the > short term), but we won't be any worse off than other browsers in that > regard. > > 4. Dependency on incomplete implementations/specifications. The > dependency > creates a chicken/egg scenario where we can't sufficiently evaluate the > dependent specifications (e.g, Worklets, and Properties & Values) without > also implementing an initial key use case (e.g., CSS Paint or Worklets > for > Web Audio.) This is somewhat mitigated by getting the implementation far > enough along to formulate informed opinions on all the specifications. > > The Houdini Task Force is meeting next week in Seattle to discuss this > and > other specifications. > ___ > dev-platform mailing list > dev-platform@lists.mozilla.org > https://lists.mozilla.org/listinfo/dev-platform ___ dev-platform mailing list dev-platform@lists.mozilla.org https:/
Intent to Experiment: CSS Houdini Paint API Level 1
Spec: https://drafts.css-houdini.org/css-paint-api/ Summary: The CSS Paint API is the first of several Web Rendering proposals from the CSS Houdini Task Force. The CSS Paint API allows Web authors to define and register a custom Paint method to be executed by the Layout engine as other elements are rendered. Bug: https://bugzilla.mozilla.org/show_bug.cgi?id=1302328 Link to standard: https://drafts.css-houdini.org/css-paint-api/ Platform coverage: Android, Desktop Estimated or target release: TBD Preference behind which this will be implemented: TBD DevTools bug: TBD Tests - TBD Implementation Details: CSS Paint API depends on the implementation of the following features: CSS Houdini Properties & Values - https://bugzilla.mozilla.org/ show_bug.cgi?id=1273706 Houdini "Worklets" - https://bugzilla.mozilla.org/show_bug.cgi?id=1290021 The planned implementation in Gecko builds upon the HTML5 Canvas2D API to provide the rendering surface for CSS Paint. Risks: The experimental implementation has progressed enough to warrant this public Intent to Implement. However, significant risks will need to be mitigated by careful design and execution if these features are to pass the experimental stage and ship in Firefox, including: 1. Use Cases. It's not clear that the use cases proposed in the specification warrant the additional Rendering System complexity. Apart from conical gradients, we haven't seen many author requests for the other use cases. If the existing Canvas2D feature set is lacking, what are the compelling use cases and maximally useful API for such use cases? It's not clear that the proposed Canvas2D-based API is desirable over a different API design (eg., WebGL) if most use cases need to directly manipulate pixels. 2. Rendering Performance. The planned Canvas2D backing store approach may be too slow for real-world usage of the API. In the future, we may replace the Canvas2D approach to have the custom paint methods create Layout (displayList) nodes for direct rendering by the Layout engine, bypassing the need for a Canvas2D backing store. It's worth noting that the Paint API isn't directly compatible with existing displayList nodes (e.g., support for raw paths, funny shapes, & pixel manipulation.) There may also be other performance issues that arise with the API's usage in combination with existing CSS features (e.g., CSS Masking, Filters, etc.) The displayList vs. canvas bitmap implementation would probably look a good bit different in WebRender. It's also worth noting that multiple implementations shipping a bitmap-based version can create dependencies that prevent us from switching to a faster alternative version in the future. 3. Integration with Gecko Architecture. The Quantum Project < https://wiki.mozilla.org/Quantum> is a major overhaul of the Firefox Rendering Engine. Implementing the CSS Paint API while that effort is in progress may add significant impedance. However, a counter-argument is that we should design Quantum to allow for such extensibility in the future. Duplication of work ( writing code that would need to be rewritten for Quantum ) is not desirable and should be avoided. For WebRender/Quantum, we could initially push this through the same path that SVG will go through (which will be rasterized on the CPU and then cached in a GPU texture atlas). It does seem like Houdini Paint could reduce the amount of acceleration we can do on the GPU (at least in the short term), but we won't be any worse off than other browsers in that regard. 4. Dependency on incomplete implementations/specifications. The dependency creates a chicken/egg scenario where we can't sufficiently evaluate the dependent specifications (e.g, Worklets, and Properties & Values) without also implementing an initial key use case (e.g., CSS Paint or Worklets for Web Audio.) This is somewhat mitigated by getting the implementation far enough along to formulate informed opinions on all the specifications. The Houdini Task Force is meeting next week in Seattle to discuss this and other specifications. ___ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform