Pádraig Brady <[email protected]> writes:
> On 26/06/2026 05:11, Collin Funk wrote:
>> This increases throughput on systems with unlocked stdio functions.
>> E.g., 2.5x in the following case:
>> $ timeout 1m taskset 1 shuf -r -i 0-100 \
>> | taskset 2 uniq -c \
>> | taskset 3 pv -r > /dev/null
>> [ 100MiB/s]
>> $ timeout 1m taskset 1 shuf -r -i 0-100 \
>> | taskset 2 uniq -c \
>> | taskset 3 pv -r > /dev/null
>> [ 246MiB/s]
>
> Nice speedup!
> I wonder could we refactor to a shared numeric output routine.
Good point. I pushed it yesterday, since I didn't realize you were away.
There is room to improve 'cksum', which prints each hex byte with a
separate printf call. Since the bulk of the time is spent hashing with
typical invocations, it doesn't really matter much. But we might as well
use a generic routine if a good one can be created.
There is some code in format_address_std function of 'od' that can
probably be used. The code written, written in 2000 as a separate
implementation for each base:
/* Use a special case of the code for each base. This is measurably
faster than generic code. */
switch (address_base)
{
case 8:
/* Base 8 conversion snipped. */
break;
case 10:
/* Base 10 conversion snipped. */
break;
case 16:
/* Base 16 conversion snipped. */
break;
}
I wonder if that comment is still true today. I'll have to do some
testing.
Collin
