Re: [PATCH next] crypto: qat - replace avg_array() with a better function

From: Giovanni Cabiddu

Date: Tue Feb 24 2026 - 13:14:36 EST


On Fri, Feb 06, 2026 at 09:09:40PM +0000, david.laight.linux@xxxxxxxxx wrote:
> From: David Laight <david.laight.linux@xxxxxxxxx>
>
> avg_array() is defined as a 'type independant' #define.
> However the algorithm is only valid for unsigned types and the
> implementation is only valid for u64.
> All the callers pass temporary kmalloc() allocated arrays of u64.
>
> Replace with a function that takes a pointer to a u64 array.
>
> Change the implementation to sum the low and high 32bits of each
> value separately and then compute the average.
Thanks David, this is a great optimization.

I also reviewed the algorithm and confirmed it is functionally equivalent
to the previous version. I tested it on a platform with QAT and it
behaves as expected.

Some minor comments below.

> This will be massively faster as it does two divisions rather than
> one for each element.
NIT: probably not `massively faster` as the maximum value for len in the
current implementation is 4.

> Also removes some very pointless __unqual_scalar_typeof().
> They could be 'auto _x = 0 ? x + 0 : 0;' even if the types weren't fixed.
>
> Only compile tested.
>
> Signed-off-by: David Laight <david.laight.linux@xxxxxxxxx>
> ---
> .../intel/qat/qat_common/adf_tl_debugfs.c | 38 ++++++++-----------
> 1 file changed, 15 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/crypto/intel/qat/qat_common/adf_tl_debugfs.c b/drivers/crypto/intel/qat/qat_common/adf_tl_debugfs.c
> index b81f70576683..a084437a2631 100644
> --- a/drivers/crypto/intel/qat/qat_common/adf_tl_debugfs.c
> +++ b/drivers/crypto/intel/qat/qat_common/adf_tl_debugfs.c
> @@ -77,32 +77,24 @@ static int tl_collect_values_u64(struct adf_telemetry *telemetry,
> * @len: Number of elements.
> *
> * This algorithm computes average of an array without running into overflow.
> + * (Provided len is less than 2 << 31.)
Should this be 2^31 or 1 << 31?
Alternatively: `Provided len fits in u32`?

> *
> * Return: average of values.
> */
> -#define avg_array(array, len) ( \
> -{ \
> - typeof(&(array)[0]) _array = (array); \
> - __unqual_scalar_typeof(_array[0]) _x = 0; \
> - __unqual_scalar_typeof(_array[0]) _y = 0; \
> - __unqual_scalar_typeof(_array[0]) _a, _b; \
> - typeof(len) _len = (len); \
> - size_t _i; \
> - \
> - for (_i = 0; _i < _len; _i++) { \
> - _a = _array[_i]; \
> - _b = do_div(_a, _len); \
> - _x += _a; \
> - if (_y >= _len - _b) { \
> - _x++; \
> - _y -= _len - _b; \
> - } else { \
> - _y += _b; \
> - } \
> - } \
> - do_div(_y, _len); \
> - (_x + _y); \
> -})
> +static u64 avg_array(const u64 *array, size_t len)
Shall size_t len be u32 len?

> +{
> + u64 sum_hi = 0, sum_lo = 0;
> + size_t i;
> +
> + for (i = 0; i < len; i++) {
> + sum_hi += array[i] >> 32;
> + sum_lo += (u32)array[i];
> + }
> +
> + sum_lo += (u64)do_div(sum_hi, len) << 32;
> +
> + return (sum_hi << 32) + div_u64(sum_lo, len);
> +}
>
> /* Calculation function for simple counter. */
> static int tl_calc_count(struct adf_telemetry *telemetry,

Thanks,

--
Giovanni