[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20250406103516.53a32bca@pumpkin>
Date: Sun, 6 Apr 2025 10:35:16 +0100
From: David Laight <david.laight.linux@...il.com>
To: Nicolas Pitre <npitre@...libre.com>
Cc: Andrew Morton <akpm@...ux-foundation.org>, linux-kernel@...r.kernel.org,
Uwe Kleine-König <u.kleine-koenig@...libre.com>, Oleg
Nesterov <oleg@...hat.com>, Peter Zijlstra <peterz@...radead.org>, Biju Das
<biju.das.jz@...renesas.com>
Subject: Re: [PATCH 1/3] lib: Add mul_u64_add_u64_div_u64() and
mul_u64_u64_div_u64_roundup()
On Sat, 5 Apr 2025 21:46:25 -0400 (EDT)
Nicolas Pitre <npitre@...libre.com> wrote:
> On Sat, 5 Apr 2025, David Laight wrote:
>
> > The existing mul_u64_u64_div_u64() rounds down, a 'rounding up'
> > variant needs 'divisor - 1' adding in between the multiply and
> > divide so cannot easily be done by a caller.
> >
> > Add mul_u64_add_u64_div_u64(a, b, c, d) that calculates (a * b + c)/d
> > and implement the 'round down' and 'round up' using it.
> >
> > Update the x86-64 asm to optimise for 'c' being a constant zero.
> >
> > For architectures that support u128 check for a 64bit product after
> > the multiply (will be cheap).
> > Leave in the early check for other architectures (mostly 32bit) when
> > 'c' is zero to avoid the multi-part multiply.
> >
> > Note that the cost of the 128bit divide will dwarf the rest of the code.
> > This function is very slow on everything except x86-64 (very very slow
> > on 32bit).
> >
> > Add kerndoc definitions for all three functions.
> >
> > Signed-off-by: David Laight <david.laight.linux@...il.com>
>
> Reviewed-by: Nicolas Pitre <npitre@...libre.com>
>
> Sidenote: The 128-bits division cost is proportional to the number of
> bits in the final result. So if the result is 0x0080000000000000 then
> the loop will execute only once and exit early.
Some performance measurements for the test cases:
0: ok 50 25 19 19 19 19 19 19 19 19 mul_u64_u64_div_u64
1: ok 2 0 0 0 0 0 0 0 0 0 mul_u64_u64_div_u64
2: ok 4 4 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
3: ok 4 4 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
4: ok 4 4 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
5: ok 15 8 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
6: ok 275 225 223 223 223 223 223 224 224 223 mul_u64_u64_div_u64
7: ok 9 6 4 4 4 4 5 5 4 4 mul_u64_u64_div_u64
8: ok 241 192 187 187 187 187 187 188 187 187 mul_u64_u64_div_u64
9: ok 212 172 169 169 169 169 169 169 169 169 mul_u64_u64_div_u64
10: ok 12 6 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
11: ok 9 5 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
12: ok 6 4 4 4 4 4 4 4 4 4 mul_u64_u64_div_u64
13: ok 6 5 5 4 4 4 4 4 4 4 mul_u64_u64_div_u64
14: ok 4 4 5 5 4 4 4 4 4 5 mul_u64_u64_div_u64
15: ok 18 12 8 8 8 8 8 8 8 8 mul_u64_u64_div_u64
16: ok 18 11 6 6 6 6 6 6 6 6 mul_u64_u64_div_u64
17: ok 22 16 11 11 11 11 11 11 11 11 mul_u64_u64_div_u64
18: ok 25 18 9 9 9 9 9 10 9 10 mul_u64_u64_div_u64
19: ok 272 231 227 227 226 227 227 227 227 226 mul_u64_u64_div_u64
20: ok 198 188 185 185 185 186 185 185 186 186 mul_u64_u64_div_u64
21: ok 207 198 196 196 196 196 196 196 196 196 mul_u64_u64_div_u64
22: ok 201 189 190 189 190 189 190 189 190 189 mul_u64_u64_div_u64
23: ok 224 184 181 181 181 181 180 180 181 181 mul_u64_u64_div_u64
24: ok 238 185 179 179 179 179 179 179 179 179 mul_u64_u64_div_u64
25: ok 208 178 177 177 177 177 177 177 177 177 mul_u64_u64_div_u64
26: ok 170 146 139 139 139 139 139 139 139 139 mul_u64_u64_div_u64
27: ok 256 204 196 196 196 196 196 196 196 196 mul_u64_u64_div_u64
28: ok 227 195 194 195 194 195 194 195 194 195 mul_u64_u64_div_u64
Entry 0 is an extra test and is the test overhead - subtracted from the others.
Each column is clocks for one run of the test, but for this set I'm running
the actual test 16 times and later dividing the clock count by 16.
It doesn't make much difference though, so the cost of the
mfence; rdpmc; mfence; nop_test; mfence; rdpmc; mfence
really is about 190 clocks (between the rdpmc results).
As soon as you hit a non-trival case the number of clocks increases
dramatically.
This is on a zen5 in 64bit mode ignoring the u128 path.
(I don't have the packages installed to compile a 32bit binary).
Maybe I can compile it for arm32, it'll need the mfence and rdpmc changing.
But maybe something simple will be ok on a pi-5.
(oh and yes, I didn't need to include autoconf.h)
David
Powered by blists - more mailing lists