input: Restrict synthetic touch gesture to view bounds.

input: Restrict synthetic touch gesture to view bounds.

BUG=686390

[sunnyps, 2017-01-30 23:45:20]

sunnyps@chromium.org changed reviewers:
+ bokan@chromium.org

[sunnyps, 2017-01-30 23:45:21]

This fixes the crash that happens in the smoothness.tough_pinch_zoom_cases
benchmark on android. WDYT of this approach?

[bokan, 2017-02-01 04:43:54]

On 2017/01/30 23:45:21, sunnyps wrote:
> This fixes the crash that happens in the smoothness.tough_pinch_zoom_cases
> benchmark on android. WDYT of this approach?

This means the size of the zoom will depend on the screen size and so introduce
variation between devices. I think a better approach would just be to make the
pinch distance small enough so that the gesture touches never leave the screen
on any devices (and DCHECK if they leave the viewport). i.e. If we want to zoom
in to 3X, rather than using 3 large pinches, use 6 pinch gestures of half size.

[sunnyps, 2017-02-02 21:36:41]

On 2017/02/01 04:43:54, bokan wrote:
> On 2017/01/30 23:45:21, sunnyps wrote:
> > This fixes the crash that happens in the smoothness.tough_pinch_zoom_cases
> > benchmark on android. WDYT of this approach?
> 
> This means the size of the zoom will depend on the screen size and so
introduce
> variation between devices. I think a better approach would just be to make the
> pinch distance small enough so that the gesture touches never leave the screen
> on any devices (and DCHECK if they leave the viewport). i.e. If we want to
zoom
> in to 3X, rather than using 3 large pinches, use 6 pinch gestures of half
size.

My "fix" doesn't work because it only clamps at one end.

The real problem is caused by mixing up CSS pixels and DIPs in telemetry and
then trying to correct that in GpuBenchmarking::PinchBy/SmoothScrollBy/etc.

1. GpuBenchmarking::VisualViewportWidth/Height is in CSS pixels.
2. gesture_common.js getWindowWidth/Height converts this into DIP by scaling
with page scale factor.
3. gesture_common.js getBoundingRect returns the element's bounding rect
relative to the root frame in CSS pixels.
4. gesture_common.js getBoundingVisibleRect clamps the bounding rect in CSS
pixels to window width/height in DIP to get the anchor point.
5. GpuBenchmarking::PinchBy converts anchor point in CSS pixels to DIP by
scaling with page scale factor.

[sunnyps, 2017-02-02 22:36:00]

On 2017/02/02 21:36:41, sunnyps wrote:
> On 2017/02/01 04:43:54, bokan wrote:
> > On 2017/01/30 23:45:21, sunnyps wrote:
> > > This fixes the crash that happens in the smoothness.tough_pinch_zoom_cases
> > > benchmark on android. WDYT of this approach?
> > 
> > This means the size of the zoom will depend on the screen size and so
> introduce
> > variation between devices. I think a better approach would just be to make
the
> > pinch distance small enough so that the gesture touches never leave the
screen
> > on any devices (and DCHECK if they leave the viewport). i.e. If we want to
> zoom
> > in to 3X, rather than using 3 large pinches, use 6 pinch gestures of half
> size.
> 
> My "fix" doesn't work because it only clamps at one end.
> 
> The real problem is caused by mixing up CSS pixels and DIPs in telemetry and
> then trying to correct that in GpuBenchmarking::PinchBy/SmoothScrollBy/etc.
> 
> 1. GpuBenchmarking::VisualViewportWidth/Height is in CSS pixels.
> 2. gesture_common.js getWindowWidth/Height converts this into DIP by scaling
> with page scale factor.
> 3. gesture_common.js getBoundingRect returns the element's bounding rect
> relative to the root frame in CSS pixels.
> 4. gesture_common.js getBoundingVisibleRect clamps the bounding rect in CSS
> pixels to window width/height in DIP to get the anchor point.
> 5. GpuBenchmarking::PinchBy converts anchor point in CSS pixels to DIP by
> scaling with page scale factor.

I made a patch which converts to DIP inside telemetry (scale getBoundingRect by
page scale factor) and that seems to work fine. The anchor point is at the
middle of the screen which is expected. But I wonder if it's better to use CSS
pixels in telemetry and convert to DIP (and clamp) inside GpuBenchmarking. WDYT?