[Pixman] [PATCH 05/12] vmx: implement fast path vmx_composite_copy_area

[Oded Gabbay]

On Tue, Jul 14, 2015 at 11:41 AM, Siarhei Siamashka
<siarhei.siamashka at gmail.com> wrote:
> It is a good idea to have at least one benchmark result in the commit
> message. Or provide a convincing explanation why this particular code
> is beneficial. The "no changes were observed" commit message does not
> do a good job justifying the need for this patch.
>
> If none of the traces of cairo applications can show any improvements,
> then we can at least use "lowlevel-blt-bench src_8888_8888". I get
> the following results on my Playstation3:
>
> == before ==
>
>            src_8888_8888 =  L1: 437.68  L2: 277.55  M:159.34 (242.58%)
>            HT: 90.31  VT: 50.77  R: 50.67  RT: 17.64 ( 148Kops/s)
>
> == after ==
>
>            src_8888_8888 =  L1: 850.60  L2: 453.91  M:174.26 (265.31%)
>            HT:105.68  VT: 54.17  R: 54.88  RT: 18.72 ( 154Kops/s)
>
>
> Assuming that the commit message gets updated,
> Acked-by: Siarhei Siamashka <siarhei.siamashka at gmail.com>

I went and tested the "lowlevel-blt-bench src_8888_8888" with this
patch and to my surprise, the fast path made the results much worse
(Except from L1):

reference memcpy speed = 25058.0MB/s (6264.5MP/s for 32bpp fills)

            Before         After           Change
L1          6475.97       7430.21         +14.74%
L2          6019.82       4284.12         -28.83%
M           3004.82       2901.99         -3.42%
HT          1666.21       1278.17         -23.29%
VT          1719.05       1481.61         -13.81%
R           759.46           581.2           -23.47%
RT          218.86        181.07          -17.27%
Kops/s      1464          1292            -11.75%

So I thought maybe I will see the improvement only in ppc, so I run it
on POWER7 ppc 64 bit:

reference memcpy speed = 10651.0MB/s (2662.7MP/s for 32bpp fills)

                Before          After           Change
L1              4165.37         4228.5          +1.52%
L2              4337.16         4241.26         -2.21%
M               1678.75         1607.69         -4.23%
HT              886.92          808.52          -8.84%
VT              781.87          742.73          -5.01%
R               483.37          445.87          -7.76%
RT              175.08          165.24          -5.62%
Kops/s          1193            1135            -4.86%

Nope, don't see the difference there as well.

How did you check it on your machine ? Did you check with only
blt+copy area patches, or with other patches applied as well ?

Currently, because both these patches don't show improvement in
lowlevel-blt-bench AND in cairo, I'm willing to drop them from this
patch-set until further investigation.

    Oded