Re: [PATCH 1/6] math.h: Add macros for rounding to closest value

[Dave Martin <Dave.Martin@arm.com>]

On Mon, Jul 08, 2024 at 09:29:38PM +0530, Devarsh Thakkar wrote:
> Add below rounding related macros:
> 
> round_closest_up(x, y) : Rounds x to closest multiple of y where y is a
> power of 2, with a preference to round up in case two nearest values are
> possible.
> 
> round_closest_down(x, y) : Rounds x to closest multiple of y where y is a
> power of 2, with a preference to round down in case two nearest values are
> possible.
> 
> roundclosest(x, y) : Rounds x to closest multiple of y, this macro should
> generally be used only when y is not multiple of 2 as otherwise
> round_closest* macros should be used which are much faster.
> 
> Examples:
>  * round_closest_up(17, 4) = 16
>  * round_closest_up(15, 4) = 16
>  * round_closest_up(14, 4) = 16
>  * round_closest_down(17, 4) = 16
>  * round_closest_down(15, 4) = 16
>  * round_closest_down(14, 4) = 12
>  * roundclosest(21, 5) = 20
>  * roundclosest(19, 5) = 20
>  * roundclosest(17, 5) = 15
> 
> Signed-off-by: Devarsh Thakkar <devarsht@ti.com>
> Acked-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
> ---
> NOTE: This patch is inspired from the Mentor Graphics IPU driver [1]
> which uses similar macro locally and which is updated in further patch
> in the series to use this generic macro instead along with other drivers
> having similar requirements.
> 
> Link: https://elixir.bootlin.com/linux/v6.8.9/source/drivers/gpu/ipu-v3/ipu-image-convert.c#L480 [1]
> ---
>  include/linux/math.h | 63 ++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 63 insertions(+)
> 
> diff --git a/include/linux/math.h b/include/linux/math.h
> index dd4152711de7..79e3dfda77fc 100644
> --- a/include/linux/math.h
> +++ b/include/linux/math.h
> @@ -34,6 +34,52 @@
>   */
>  #define round_down(x, y) ((x) & ~__round_mask(x, y))
>  
> +/**
> + * round_closest_up - round closest to be multiple of specified value (which is
> + *                    power of 2) with preference to rounding up
> + * @x: the value to round
> + * @y: multiple to round closest to (must be a power of 2)
> + *
> + * Rounds @x to closest multiple of @y (which must be a power of 2).
> + * The value can be either rounded up or rounded down depending upon rounded
> + * value's closeness to the specified value. If there are two closest possible
> + * values, i.e. the difference between the specified value and it's rounded up
> + * and rounded down values is same then preference is given to rounded up
> + * value.
> + *
> + * To perform arbitrary rounding to closest value (not multiple of 2), use
> + * roundclosest().
> + *
> + * Examples:
> + * * round_closest_up(17, 4) = 16
> + * * round_closest_up(15, 4) = 16
> + * * round_closest_up(14, 4) = 16
> + */
> +#define round_closest_up(x, y) round_down((x) + (y) / 2, (y))
> +
> +/**
> + * round_closest_down - round closest to be multiple of specified value (which
> + *			is power of 2) with preference to rounding down
> + * @x: the value to round
> + * @y: multiple to round closest to (must be a power of 2)
> + *
> + * Rounds @x to closest multiple of @y (which must be a power of 2).
> + * The value can be either rounded up or rounded down depending upon rounded
> + * value's closeness to the specified value. If there are two closest possible
> + * values, i.e. the difference between the specified value and it's rounded up
> + * and rounded down values is same then preference is given to rounded up
> + * value.
> + *
> + * To perform arbitrary rounding to closest value (not multiple of 2), use
> + * roundclosest().
> + *
> + * Examples:
> + * * round_closest_down(17, 4) = 16
> + * * round_closest_down(15, 4) = 16
> + * * round_closest_down(14, 4) = 12
> + */
> +#define round_closest_down(x, y) round_up((x) - (y) / 2, (y))
> +

Naming aside, is there an actual use case for having both roundclosest()
and round_closest_up() today?

(i.e., is there any potential caller that would actually care about the
rounding direction for borderline cases?)

>  #define DIV_ROUND_UP __KERNEL_DIV_ROUND_UP
>  
>  #define DIV_ROUND_DOWN_ULL(ll, d) \
> @@ -77,6 +123,23 @@
>  }							\
>  )
>  
> +/**
> + * roundclosest - round to nearest multiple
> + * @x: the value to round
> + * @y: multiple to round nearest to
> + *
> + * Rounds @x to nearest multiple of @y.
> + * The rounded value can be greater than or less than @x depending
> + * upon it's nearness to @x. If @y will always be a power of 2, consider
> + * using the faster round_closest_up() or round_closest_down().
> + *
> + * Examples:
> + * * roundclosest(21, 5) = 20
> + * * roundclosest(19, 5) = 20
> + * * roundclosest(17, 5) = 15
> + */
> +#define roundclosest(x, y) rounddown((x) + (y) / 2, (y))

Won't this go wrong if (x) + (y) / 2 overflows?  This may happen even in
some cases where the correctly rounded value would be in range.

The existing rounddown() already leaves something to be desired IIUC: if
given a negative dividend, it looks like it actually rounds up, at least
on some arches.  But maybe people don't use it that way very often.
Perhaps I'm missing something.

[...]

Cheers
---Dave