PWM backlight interpolation adjustments

I was trying to adjust the brightness-levels for the trogdor boards:
https://chromium-review.googlesource.com/c/chromiumos/third_party/kernel/+/2291209
Like on a lot of panels, trogdor's low end needs to be cropped,
and now that we have the interpolation stuff I wanted to make use of it
and bake in even the curve that's customary to have on chromebooks.

I found the current behavior of the pwm_bl driver a little unintuitive
and non-linear. See patch 1 for a suggested fix for this.

A few veyron dts files were relying on this (perhaps weird) behavior.
Those devices also want a minimum brightness like trogdor, so changed
them to use the new way.

Finally, given that trogdor's dts is part of linux-next now, add the
brightness-levels to it, since that's the original reason I was looking at
this.

Changes in v2:
- Fixed type promotion in the driver
- Removed "backlight: pwm_bl: Artificially add 0% during interpolation",
userspace works just fine without it because it already knows how to use
bl_power for turning off the display.
- Added brightness-levels to trogdor as well, now the dts is upstream.


Alexandru Stan (3):
  backlight: pwm_bl: Fix interpolation
  ARM: dts: rockchip: veyron: Remove 0 point from brightness-levels
  arm64: dts: qcom: trogdor: Add brightness-levels

 arch/arm/boot/dts/rk3288-veyron-jaq.dts      |  2 +-
 arch/arm/boot/dts/rk3288-veyron-minnie.dts   |  2 +-
 arch/arm/boot/dts/rk3288-veyron-tiger.dts    |  2 +-
 arch/arm64/boot/dts/qcom/sc7180-trogdor.dtsi |  9 +++
 drivers/video/backlight/pwm_bl.c             | 70 +++++++++-----------
 5 files changed, 43 insertions(+), 42 deletions(-)

-- 
2.28.0

Whenever num-interpolated-steps was larger than the distance
between 2 consecutive brightness levels the table would get really
discontinuous. The slope of the interpolation would stick with
integers only and if it was 0 the whole line segment would get skipped.

Example settings:
	brightness-levels = <0 1 2 4 8 16 32 64 128 256>;
	num-interpolated-steps = <16>;

The distances between 1 2 4 and 8 would be 1, and only starting with 16
it would start to interpolate properly.

Let's change it so there's always interpolation happening, even if
there's no enough points available (read: values in the table would
appear more than once). This should match the expected behavior much
more closely.

Signed-off-by: Alexandru Stan <redacted>
---

 drivers/video/backlight/pwm_bl.c | 70 ++++++++++++++------------------
 1 file changed, 31 insertions(+), 39 deletions(-)

diff --git a/drivers/video/backlight/pwm_bl.c b/drivers/video/backlight/pwm_bl.c
index dfc760830eb9..3e77f6b73fd9 100644
--- a/drivers/video/backlight/pwm_bl.c
+++ b/drivers/video/backlight/pwm_bl.c

@@ -230,8 +230,7 @@ static int pwm_backlight_parse_dt(struct device *dev,
 				  struct platform_pwm_backlight_data *data)
 {
 	struct device_node *node = dev->of_node;
-	unsigned int num_levels = 0;
-	unsigned int levels_count;
+	unsigned int num_levels;
 	unsigned int num_steps = 0;
 	struct property *prop;
 	unsigned int *table;

@@ -260,12 +259,11 @@ static int pwm_backlight_parse_dt(struct device *dev,
 	if (!prop)
 		return 0;
 
-	data->max_brightness = length / sizeof(u32);
+	num_levels = length / sizeof(u32);
 
 	/* read brightness levels from DT property */
-	if (data->max_brightness > 0) {
-		size_t size = sizeof(*data->levels) * data->max_brightness;
-		unsigned int i, j, n = 0;
+	if (num_levels > 0) {
+		size_t size = sizeof(*data->levels) * num_levels;
 
 		data->levels = devm_kzalloc(dev, size, GFP_KERNEL);
 		if (!data->levels)

@@ -273,7 +271,7 @@ static int pwm_backlight_parse_dt(struct device *dev,
 
 		ret = of_property_read_u32_array(node, "brightness-levels",
 						 data->levels,
-						 data->max_brightness);
+						 num_levels);
 		if (ret < 0)
 			return ret;
 

@@ -298,7 +296,13 @@ static int pwm_backlight_parse_dt(struct device *dev,
 		 * between two points.
 		 */
 		if (num_steps) {
-			if (data->max_brightness < 2) {
+			unsigned int num_input_levels = num_levels;
+			unsigned int i;
+			u32 x1, x2, x, dx;
+			u32 y1, y2;
+			s64 dy;
+
+			if (num_input_levels < 2) {
 				dev_err(dev, "can't interpolate\n");
 				return -EINVAL;
 			}

@@ -308,14 +312,7 @@ static int pwm_backlight_parse_dt(struct device *dev,
 			 * taking in consideration the number of interpolated
 			 * steps between two levels.
 			 */
-			for (i = 0; i < data->max_brightness - 1; i++) {
-				if ((data->levels[i + 1] - data->levels[i]) /
-				   num_steps)
-					num_levels += num_steps;
-				else
-					num_levels++;
-			}
-			num_levels++;
+			num_levels = (num_input_levels - 1) * num_steps + 1;
 			dev_dbg(dev, "new number of brightness levels: %d\n",
 				num_levels);
 

@@ -327,24 +324,25 @@ static int pwm_backlight_parse_dt(struct device *dev,
 			table = devm_kzalloc(dev, size, GFP_KERNEL);
 			if (!table)
 				return -ENOMEM;
-
-			/* Fill the interpolated table. */
-			levels_count = 0;
-			for (i = 0; i < data->max_brightness - 1; i++) {
-				value = data->levels[i];
-				n = (data->levels[i + 1] - value) / num_steps;
-				if (n > 0) {
-					for (j = 0; j < num_steps; j++) {
-						table[levels_count] = value;
-						value += n;
-						levels_count++;
-					}
-				} else {
-					table[levels_count] = data->levels[i];
-					levels_count++;
+			/*
+			 * Fill the interpolated table[x] = y
+			 * by draw lines between each (x1, y1) to (x2, y2).
+			 */
+			dx = num_steps;
+			for (i = 0; i < num_input_levels - 1; i++) {
+				x1 = i * dx;
+				x2 = x1 + dx;
+				y1 = data->levels[i];
+				y2 = data->levels[i + 1];
+				dy = (s64)y2 - y1;
+
+				for (x = x1; x < x2; x++) {
+					table[x] = y1 +
+						div_s64(dy * ((s64)x - x1), dx);
 				}
 			}
-			table[levels_count] = data->levels[i];
+			/* Fill in the last point, since no line starts here. */
+			table[x2] = y2;
 
 			/*
 			 * As we use interpolation lets remove current

@@ -353,15 +351,9 @@ static int pwm_backlight_parse_dt(struct device *dev,
 			 */
 			devm_kfree(dev, data->levels);
 			data->levels = table;
-
-			/*
-			 * Reassign max_brightness value to the new total number
-			 * of brightness levels.
-			 */
-			data->max_brightness = num_levels;
 		}
 
-		data->max_brightness--;
+		data->max_brightness = num_levels - 1;
 	}
 
 	return 0;

-- 
2.28.0

On Tue, Oct 13, 2020 at 01:01:01AM -0700, Alexandru Stan wrote:

Whenever num-interpolated-steps was larger than the distance
between 2 consecutive brightness levels the table would get really
discontinuous. The slope of the interpolation would stick with
integers only and if it was 0 the whole line segment would get skipped.

Example settings:
	brightness-levels = <0 1 2 4 8 16 32 64 128 256>;
	num-interpolated-steps = <16>;

The distances between 1 2 4 and 8 would be 1, and only starting with 16
it would start to interpolate properly.

Both comments a perilously close to nitpicking but enough that I wanted
to reply...

I'd suggest that the current behaviour as having two properties.

1. It was designed to generate strictly increasing tables (no repeated
   values).

2. It's implementation contains quantization errors when calculating the
   step size. This results in both the discards of some interpolated
   steps you mentioned (it is possible to insert extra steps between 4
   and 8 whilst retaining a strictly increasing table). It also
   results in a potentially large undershoot when multiplying a step
   size (64 interpolated steps and a gap of 127 is likely to get a visual
   jump as we hop through 63 physical steps in one go).

#1 can is a policy that can be changed. #2 is a bug that could be fixed.

To be clear I don't object to generating a monotonically increasing
table but I'd prefer the policy change to be explicitly described in
the description.

75 more quoted lines

I don't think it is possible for x - x1 to be negative (e.g. what is the
s64 for). Obviously it makes little functional difference whether the
cast is there or not but I don't like fixed point code that has been
written with "just in case" casts.


Daniel.

Hi Alexandru,

On Tue, Oct 13, 2020 at 1:57 PM Alexandru Stan [off-list ref] wrote:

Whenever num-interpolated-steps was larger than the distance
between 2 consecutive brightness levels the table would get really
discontinuous. The slope of the interpolation would stick with
integers only and if it was 0 the whole line segment would get skipped.

Example settings:
        brightness-levels = <0 1 2 4 8 16 32 64 128 256>;
        num-interpolated-steps = <16>;

The distances between 1 2 4 and 8 would be 1, and only starting with 16
it would start to interpolate properly.

Let's change it so there's always interpolation happening, even if
there's no enough points available (read: values in the table would
appear more than once). This should match the expected behavior much
more closely.

Signed-off-by: Alexandru Stan <redacted>

Thanks for your patch!

Yummy, 64-by-32 divisions.
Shouldn't this use a rounded division?

Nevertheless, I think it would be worthwhile to implement this using
a (modified) Bresenham algorithm, avoiding multiplications and
divisions, and possibly increasing accuracy as well.

https://en.wikipedia.org/wiki/Bresenham%27s_line_algorithm

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

On Wed, 14 Oct 2020 at 23:55, Geert Uytterhoeven [off-list ref] wrote:

Hi Alexandru,

On Tue, Oct 13, 2020 at 1:57 PM Alexandru Stan [off-list ref] wrote:

quoted

Whenever num-interpolated-steps was larger than the distance
between 2 consecutive brightness levels the table would get really
discontinuous. The slope of the interpolation would stick with
integers only and if it was 0 the whole line segment would get skipped.

Example settings:
        brightness-levels = <0 1 2 4 8 16 32 64 128 256>;
        num-interpolated-steps = <16>;

The distances between 1 2 4 and 8 would be 1, and only starting with 16
it would start to interpolate properly.

Let's change it so there's always interpolation happening, even if
there's no enough points available (read: values in the table would
appear more than once). This should match the expected behavior much
more closely.

Signed-off-by: Alexandru Stan <redacted>

Thanks for your patch!

Thanks for your reply!

I'm sorry I haven't replied earlier. Looks like your reply was marked as spam.
Rest be assured my spam filter has been disciplined! :D

quoted

--- a/drivers/video/backlight/pwm_bl.c
+++ b/drivers/video/backlight/pwm_bl.c

@@ -327,24 +324,25 @@ static int pwm_backlight_parse_dt(struct device *dev,
                        table = devm_kzalloc(dev, size, GFP_KERNEL);
                        if (!table)
                                return -ENOMEM;
-
-                       /* Fill the interpolated table. */
-                       levels_count = 0;
-                       for (i = 0; i < data->max_brightness - 1; i++) {
-                               value = data->levels[i];
-                               n = (data->levels[i + 1] - value) / num_steps;
-                               if (n > 0) {
-                                       for (j = 0; j < num_steps; j++) {
-                                               table[levels_count] = value;
-                                               value += n;
-                                               levels_count++;
-                                       }
-                               } else {
-                                       table[levels_count] = data->levels[i];
-                                       levels_count++;
+                       /*
+                        * Fill the interpolated table[x] = y
+                        * by draw lines between each (x1, y1) to (x2, y2).
+                        */
+                       dx = num_steps;
+                       for (i = 0; i < num_input_levels - 1; i++) {
+                               x1 = i * dx;
+                               x2 = x1 + dx;
+                               y1 = data->levels[i];
+                               y2 = data->levels[i + 1];
+                               dy = (s64)y2 - y1;
+
+                               for (x = x1; x < x2; x++) {
+                                       table[x] = y1 +
+                                               div_s64(dy * ((s64)x - x1), dx);

Yummy, 64-by-32 divisions.
Shouldn't this use a rounded division?

It won't hurt. But it really doesn't make much of a difference either way.

Nevertheless, I think it would be worthwhile to implement this using
a (modified) Bresenham algorithm, avoiding multiplications and
divisions, and possibly increasing accuracy as well.

https://en.wikipedia.org/wiki/Bresenham%27s_line_algorithm

Sure, it might be a little faster to use Bresenham's line algorithm.
Looks like to implement it I would have to deal with some fixed point
math and still have to do divisions occasionally.
I don't think performance is critical here, the values get calculated
only once when the driver loads, and the algorithm's accuracy
improvements might be at most 1 LSB.

Meanwhile the formula I already implemented is almost the same as the
formulas found at
https://en.wikipedia.org/wiki/Linear_interpolation#:~:text=gives
I would like to keep it as is, as straightforward as possible.

quoted

                                }
                        }
-                       table[levels_count] = data->levels[i];
+                       /* Fill in the last point, since no line starts here. */
+                       table[x2] = y2;

                        /*
                         * As we use interpolation lets remove current

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

Alexandru Stan (amstan)