Platforms: Win32 Use Message Time for WinTab events #116114

Harley Acheson · 2023-12-12T23:26:30+01:00

Harley Acheson commented

2023-12-12 23:26:30 +01:00

WinTab events are all created in result of windows messages, so for
these use the last message time, rather than supplied time, to ensure
the times are consistent with other event times.

In the Wacom reference page the say that the "pkTime" is "In absolute mode, specifies the system time at which the event was posted. In relative mode, specifies the elapsed time in milliseconds since the last packet". This second format is not compatible with blender event times. https://developer-docs.wacom.com/docs/icbt/windows/wintab/wintab-reference/

WinTab events are all created in result of windows messages, so for these use the last message time, rather than supplied time, to ensure the times are consistent with other event times. --- In the Wacom reference page the say that the "pkTime" is "In absolute mode, specifies the system time at which the event was posted. In relative mode, specifies the elapsed time in milliseconds since the last packet". This second format is not compatible with blender event times. https://developer-docs.wacom.com/docs/icbt/windows/wintab/wintab-reference/

Harley Acheson added 1 commit 2023-12-12 23:26:38 +01:00

446b2c18e1 Platforms: Win32 Use Message Time for WinTab events

WinTab events are all created in result of windows messages, so for
these use the last message time, rather than supplied time, to ensure
the times are consistent with other event times.

Harley Acheson requested review from Ray molenkamp 2023-12-12 23:44:34 +01:00

Harley Acheson added this to the Platforms, Builds Tests & Devices project 2023-12-12 23:44:43 +01:00

Ray molenkamp commented

2023-12-13 01:54:46 +01:00

I Vaguely remember @PrototypeNM1 fighting with this, and i'd like to get his opinion before making changes here

Brecht Van Lommel commented

2023-12-13 02:07:37 +01:00

The GetMessageTime() API only has a resolution of 16ms (62.5 fps). For smooth strokes from tablets we want as much resolution as we can get. And the potential issue is not just that we'd only get 16ms resolution, but that by rounding the native tablet frequency to another frequency that points would be poorly spaced as well.

I think we want to get the time from the tablet API. If some offsets needs to be applied to make it match other times, perhaps that offset could be computed at the start of the stroke (i.e. pen/mouse down) and stay fixed for the duration.

The `GetMessageTime()` API only has a resolution of 16ms (62.5 fps). For smooth strokes from tablets we want as much resolution as we can get. And the potential issue is not just that we'd only get 16ms resolution, but that by rounding the native tablet frequency to another frequency that points would be poorly spaced as well. I think we want to get the time from the tablet API. If some offsets needs to be applied to make it match other times, perhaps that offset could be computed at the start of the stroke (i.e. pen/mouse down) and stay fixed for the duration.

Nicholas Rishel commented

2023-12-13 02:09:02 +01:00

I vaguely remember that we can limit reading the Wintab packets constrained to those associated to the current WT_PACKET Windows message, currently I believe we drain the whole Wintab packet queue. If we're going to use the message origination time we will probably want to limit what we read to have closer timestamps.

I also wonder if e.g. stroke smoothing or similar operations take time between INBETWEEN_MOUSEMOVEs into consideration? If so this could cause a regression from timestamps not differing.

I vaguely remember that we can limit reading the Wintab packets constrained to those associated to the current `WT_PACKET` Windows message, currently I believe we drain the whole Wintab packet queue. If we're going to use the message origination time we will probably want to limit what we read to have closer timestamps. I also wonder if e.g. stroke smoothing or similar operations take time between `INBETWEEN_MOUSEMOVE`s into consideration? If so this could cause a regression from timestamps not differing.

Harley Acheson commented

2023-12-13 04:02:03 +01:00

The GetMessageTime() API only has a resolution of 16ms (62.5 fps)....

Its more complicated than that. Our general "get tick count" routine uses a high performance timer with a resolution of <1us. The getMilliSeconds might appear to fallback to using GetTickCount if there is no high res timer, but that can no longer happen as such a timer is guaranteed since Windows XP. We convert this value to milliseconds though so the resolution is 1ms.

Our new getMessageTime uses this timer's value and subtracts the difference between the 32-bit GetMessageTime and GetTickCount. So we end up with a 64 timer with good accuracy that doesn't roll over. But the portion of this time value from when a windows message is delivered to when we collect it has a resolution of 16ms.

but that by rounding the native tablet frequency to another frequency that points would be poorly spaced as well.

Do we really space the points using time and not just position?

AFAIK with this change we will receive the same number of messages from Wintab and send the same number of events to Blender at the same time they always have. With the only change being our own timestamp in the Blender events will be of a format that matches the format of other event times. We use detect double-click by comparing subsequent values, but not sure we use it anywhere else yet.

> The `GetMessageTime()` API only has a resolution of 16ms (62.5 fps).... Its more complicated than that. Our general "get tick count" routine uses a high performance timer with a resolution of <1us. The `getMilliSeconds` might appear to fallback to using `GetTickCount` if there is no high res timer, but that can no longer happen as such a timer is [guaranteed since Windows XP](https://learn.microsoft.com/en-us/windows/win32/api/profileapi/nf-profileapi-queryperformancefrequency). We convert this value to milliseconds though so the resolution is 1ms. Our new `getMessageTime` uses this timer's value and subtracts the difference between the 32-bit `GetMessageTime` and `GetTickCount`. So we end up with a 64 timer with good accuracy that doesn't roll over. But the portion of this time value from when a windows message is delivered to when we collect it has a resolution of 16ms. > but that by rounding the native tablet frequency to another frequency that points would be poorly spaced as well. Do we really space the points using _time_ and not just position? AFAIK with this change we will receive the same number of messages from Wintab and send the same number of events to Blender at the same time they always have. With the only change being our own timestamp in the Blender events will be of a format that matches the format of other event times. We use detect double-click by comparing subsequent values, but not sure we use it anywhere else yet.

Brecht Van Lommel commented

2023-12-13 11:03:29 +01:00

Its more complicated than that. Our general "get tick count" routine uses a high performance timer with a resolution of <1us. The getMilliSeconds might appear to fallback to using GetTickCount if there is no high res timer, but that can no longer happen as such a timer is guaranteed since Windows XP. We convert this value to milliseconds though so the resolution is 1ms.

GetTickCount is the time the event is being processed, not the time the event happened. Its accuracy does not help because it's not measuring the right thing.

With high frequency tablet events you will end up processing a few events together every time, and then you'll have clumps of events that have nearly the same time because they happened to be processed together.

Do we really space the points using time and not just position?

We indeed do not seem to be using it in the brush code. If it was accurate we could use it for the airbrush, to properly match position and time.

If it's not trivial to keep the accuracy I guess it's to lose it. But I would like to see a comment noting the loss of accuracy and that it's okay because it's only used for double click checking.

> Its more complicated than that. Our general "get tick count" routine uses a high performance timer with a resolution of <1us. The `getMilliSeconds` might appear to fallback to using `GetTickCount` if there is no high res timer, but that can no longer happen as such a timer is [guaranteed since Windows XP](https://learn.microsoft.com/en-us/windows/win32/api/profileapi/nf-profileapi-queryperformancefrequency). We convert this value to milliseconds though so the resolution is 1ms. `GetTickCount` is the time the event is being processed, not the time the event happened. Its accuracy does not help because it's not measuring the right thing. With high frequency tablet events you will end up processing a few events together every time, and then you'll have clumps of events that have nearly the same time because they happened to be processed together. > Do we really space the points using _time_ and not just position? We indeed do not seem to be using it in the brush code. If it was accurate we could use it for the airbrush, to properly match position and time. If it's not trivial to keep the accuracy I guess it's to lose it. But I would like to see a comment noting the loss of accuracy and that it's okay because it's only used for double click checking.

This pull request can be merged automatically.

This branch is out-of-date with the base branch

You are not authorized to merge this pull request.

View command line instructions.

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u WinTabMessageTimes:Harley-WinTabMessageTimes

git checkout Harley-WinTabMessageTimes

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Platforms: Win32 Use Message Time for WinTab events #116114

Checkout