dix: fix up coordinate scaling when external monitors are present

The goal of all this is to get an x/y motion reflecting the motion
on the device, i.e. a circle on the device is a circle on the screen.

This is currently done by scaling the y coordinate depending on the screen
ratio vs device ratio. Depending on that ratio the movement on the y axis may
be accelerated (ratio < 1) or slowed (ratio > 1). This leads to the weird
effect that changing the screen ratio by plugging a new monitor changes the
speed of the touchpad.

Use a different algorithm: calculate the physical movement on the device, map
that to the same-ish distance on the screen, then convert that back into a
device-specific vector. This way we get the same mapping regardless of the
current screen dimensions.

Since the pointer accel code doesn't take device resolution into account, make
sure we apply our crazy mapping before we accelerate. This way we accelerate
resolution-independent.

Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
Reviewed-by: Hans de Goede <hdegoede@redhat.com>

dix/getevents.c

@@ -770,27 +770,65 @@ add_to_scroll_valuator(DeviceIntPtr dev, ValuatorMask *mask, int valuator, doubl
 }
 
 
+/* FIXME: relative events from devices with absolute axis ranges is
+   fundamentally broken. We map the device coordinate range into the screen
+   range, but don't really account for device resolution in that.
+
+   what we do here is a hack to make touchpads usable. for a given relative
+   motion vector in device coordinates:
+   1. calculate physical movement on the device in metres
+   2. calculate pixel vector that is the same physical movement on the
+      screen (times some magic number to provide sensible base speed)
+   3. calculate what percentage this vector is of the current screen
+      width/height
+   4. calculate equivalent vector in % on the device's min/max axis range
+   5. Use that device vector as the actual motion vector
+
+   e.g. 10/50mm on the device, 10/50mm on the screen are 30/100 pixels,
+   30/100 pixels are 1/3% of the width, 1/3% of the device is a vector of
+   20/80 -> use 20/80 as dx/dy.
+
+   dx/dy is then applied to the current position in device coordinates,
+   mapped to screen coordinates and thus the movement on the screen reflects
+   the motion direction on the device.
+ */
 static void
 scale_for_device_resolution(DeviceIntPtr dev, ValuatorMask *mask)
 {
-    double y;
+    double x, y;
     ValuatorClassPtr v = dev->valuator;
     int xrange = v->axes[0].max_value - v->axes[0].min_value + 1;
     int yrange = v->axes[1].max_value - v->axes[1].min_value + 1;
 
-    double screen_ratio = 1.0 * screenInfo.width/screenInfo.height;
+    /* Assume 100 units/m for devices without resolution */
-    double device_ratio = 1.0 * xrange/yrange;
+    int xres = 100000, yres = 100000;
-    double resolution_ratio = 1.0;
-    double ratio;
 
-    if (!valuator_mask_fetch_double(mask, 1, &y))
+    /* If we have multiple screens with different dpi, it gets complicated:
-        return;
+       we have to map which screen we're on and then take the dpi of that
+       screen to be somewhat accurate.  */
+    const ScreenPtr s = screenInfo.screens[0];
+    const double screen_res = 1000.0 * s->width/s->mmWidth; /* units/m */
 
-    if (v->axes[0].resolution != 0 && v->axes[1].resolution != 0)
+    /* some magic multiplier, so unaccelerated movement of x mm on the
-        resolution_ratio = 1.0 * v->axes[0].resolution/v->axes[1].resolution;
+       device reflects x * magic mm on the screen */
+    const double magic = 4;
 
-    ratio = device_ratio/resolution_ratio/screen_ratio;
+    if (v->axes[0].resolution != 0 && v->axes[1].resolution != 0) {
-    valuator_mask_set_double(mask, 1, y / ratio);
+        xres = v->axes[0].resolution;
+        yres = v->axes[1].resolution;
+    }
+
+    if (valuator_mask_isset(mask, 0)) {
+        x = valuator_mask_get_double(mask, 0);
+        x = magic * x/xres * screen_res/screenInfo.width * xrange;
+        valuator_mask_set_double(mask, 0, x);
+    }
+
+    if (valuator_mask_isset(mask, 1)) {
+        y = valuator_mask_get_double(mask, 1);
+        y = magic * y/yres * screen_res/screenInfo.height * yrange;
+        valuator_mask_set_double(mask, 1, y);
+    }
 }
 
 /**
@@ -804,15 +842,6 @@ moveRelative(DeviceIntPtr dev, int flags, ValuatorMask *mask)
 {
     int i;
     Bool clip_xy = IsMaster(dev) || !IsFloating(dev);
-    ValuatorClassPtr v = dev->valuator;
-
-    /* for abs devices in relative mode, we've just scaled wrong, since we
-       mapped the device's shape into the screen shape. Undo this. */
-    if ((flags & POINTER_ABSOLUTE) == 0 && v && v->numAxes > 1 &&
-        v->axes[0].min_value < v->axes[0].max_value &&
-        v->axes[1].min_value < v->axes[1].max_value) {
-        scale_for_device_resolution(dev, mask);
-    }
 
     /* calc other axes, clip, drop back into valuators */
     for (i = 0; i < valuator_mask_size(mask); i++) {
@@ -1441,10 +1470,21 @@ fill_pointer_events(InternalEvent *events, DeviceIntPtr pDev, int type,
             set_raw_valuators(raw, &mask, raw->valuators.data);
     }
     else {
+        ValuatorClassPtr v = pDev->valuator;
+
         transformRelative(pDev, &mask);
 
+        /* for abs devices in relative mode, we've just scaled wrong, since we
+           mapped the device's shape into the screen shape. Undo this. */
+        if (v && v->numAxes > 1 &&
+            v->axes[0].min_value < v->axes[0].max_value &&
+            v->axes[1].min_value < v->axes[1].max_value) {
+            scale_for_device_resolution(pDev, &mask);
+        }
+
         if (flags & POINTER_ACCELERATE)
             accelPointer(pDev, &mask, ms);
+
         if ((flags & POINTER_NORAW) == 0 && raw)
             set_raw_valuators(raw, &mask, raw->valuators.data);