Mail Archives: geda-user/2018/05/08/03:36:50
--2KPJUYRELGREHTBQUEREVnhgwp
Content-Transfer-Encoding: quoted-printable
Content-Type: text/plain; charset=UTF-8
Hello Igor, Dnia 8 maja 2018 04:48 michalwd1979 <gedau AT igor2 DOT repo DOT hu&g=
t; napisa=C5=82(a): 2. load the single layer version (4096.lht); do not zo=
om or pan, leave it as it is after load, and run benchmark() a few times -=
we need to average the value of multiple runs, they show some variation h=
ere (I got values between 50.5 and 53.9 for this file) For me it is betw=
een 23.5 and 21.5, averaging at ~22.6. What is interesting, running benchma=
rk on this file does not cause 100% load on cpu, what suggest that the limi=
ting factor is GPU or GPU bus speed. Here I have PCIE x1 socket so it might=
be the case. 3. within the same pcb-rnd session, please load the multi=
-layer version (4096_lay.lht), again leave it as it is and run benchmark()=
a few times (I got values (I got values between 49.5 and 54) Here it is=
from 21.5 to 22.7 with 22.1 average. Very similar to single layer version.=
=C2=A0 This tests one specific rendering feature of pcb-rnd that is dif=
ferent compared to geda/pcb's gl rendering. If we get about the same v=
alues, we know it's not our "new" layer compositing code. If w=
e see the multi-layer version being much slower, we know we need to implem=
ent the same optimization in the gl HID that we have in the software rende=
r HIDs. One more notice: With my board loaded even moving a cursor causes=
instant 100% cpu load for a few seconds, with benchmark() it is the same. =
Mainline or opengl pcb never causes high cpu usage even with furious pannin=
g or rotating. While sometimes the movement is not super smooth, GUI is alw=
ays very responsive. This is not the case with pcb-rnd, here I have to wait=
for a menu to open after for example cursor movement. Also turning ON/OFF =
"subcircuits" layer makes a huge difference - maybe we are looking =
in the wrong place?? Best Regards, Michael Widlok=0D
--2KPJUYRELGREHTBQUEREVnhgwp
Content-Transfer-Encoding: quoted-printable
Content-Type: text/html; charset=UTF-8
<div>Hello Igor, <br></div><div class=3D"nh_extra"><p>Dnia 8 maja 2018 04:4=
8 michalwd1979 <gedau AT igor2 DOT repo DOT hu> napisa=C5=82(a):<br></p><blockqu=
ote class=3D"nh_quote" style=3D"border-left: 2px solid #999; padding-left: =
8px; margin: 0;"><div id=3D"gwp59fd49e5"><div>2. load the single layer vers=
ion (4096.lht); do not zoom or pan, leave it<br></div><div>as it is after l=
oad, and run benchmark() a few times - we need to average<br></div><div>the=
value of multiple runs, they show some variation here (I got values<br></d=
iv><div>between 50.5 and 53.9 for this file)<br></div></div></blockquote></=
div><div><br></div><div>For me it is between 23.5 and 21.5, averaging at ~2=
2.6. What is interesting, running benchmark on this file does not cause 100=
% load on cpu, what suggest that the limiting factor is GPU or GPU bus spee=
d. Here I have PCIE x1 socket so it might be the case. <br></div><div><br><=
/div><div class=3D"nh_extra"><blockquote class=3D"nh_quote" style=3D"border=
-left: 2px solid #999; padding-left: 8px; margin: 0;"><div id=3D"gwp59fd49e=
5"><div><br></div><div>3. within the same pcb-rnd session, please load the =
multi-layer version<br></div><div>(4096_lay.lht), again leave it as it is a=
nd run benchmark() a few times (I<br></div><div>got values (I got values be=
tween 49.5 and 54)<br></div></div></blockquote></div><div><br></div><div>He=
re it is from 21.5 to 22.7 with 22.1 average. Very similar to single layer =
version. <br></div><div><br></div><div class=3D"nh_extra"><blockquote=
class=3D"nh_quote" style=3D"border-left: 2px solid #999; padding-left: 8px=
; margin: 0;"><div id=3D"gwp59fd49e5"><div><br></div><div>This tests one sp=
ecific rendering feature of pcb-rnd that is different<br></div><div>compare=
d to geda/pcb's gl rendering. If we get about the same values, we<br></div>=
<div>know it's not our "new" layer compositing code. If we see the multi-la=
yer<br></div><div>version being much slower, we know we need to implement t=
he same<br></div><div>optimization in the gl HID that we have in the softwa=
re render HIDs.<br></div></div></blockquote></div><div><br></div><div>One m=
ore notice: With my board loaded even moving a cursor causes instant 100% c=
pu load for a few seconds, with benchmark() it is the same. Mainline or ope=
ngl pcb never causes high cpu usage even with furious panning or rotating. =
While sometimes the movement is not super smooth, GUI is always very respon=
sive. This is not the case with pcb-rnd, here I have to wait for a menu to =
open after for example cursor movement. Also turning ON/OFF "subcircuits" l=
ayer makes a huge difference - maybe we are looking in the wrong place??<br=
></div><div>Best Regards,<br></div><div>Michael Widlok<br></div><div><br></=
div>
--2KPJUYRELGREHTBQUEREVnhgwp--
- Raw text -