Pixel Shift Study

Pixel Shift Study - Problem finaly solved !!!!!!!!

1. What we now know ( 20060619 ) :

problem of pixel shift occures accidentally it can easily be seen in the first columns where characteristic shift occurs
It didn't occure on frames on pioneer2 ( 20060519 , 20060520, 20060521 are OK )
It started when we moved to DRAGON
with USB1.1 , when I did :
rmmod ehci_hcd
frames look fine !
( in order to restore USB2.0 mode run : modprobe ehci_hcd )
Today ( 20060619 ) we checked kernel without SMP and I took 30 frames with k2a and found shift on them , so that's rather
not due to fact that DRAGON is a 2 processor machine
I tried to change the only option in ehci_hcd module which is : log2_irq_thresh , I checked log2_irq_thresh=4 and log2_irq_thresh=6
and it didn't change anything, I don't see any other posibility of changing USB2.0 speed now.
some ealier checks are available here

2. Tools for testing

I wrote program checkframe , which can detect pixel shift in frames it can be used on pi1 ( or maybe also on pi2 if I will
compile it before departure ) , usage is :
checkframe k2a_060618_00025.fitc -check_frame
This program detects step in column 15 and 16
getleft! - script for cutting out left border of the image

3. Plans for testing

check if "hot pixels" are for sure correlated with pixel shift
try to plug camera to USB2.0 on motherboard , rather then on slot
try to understand what is going on ...
marking bad frames and anaylse "hot pixels" on them

4. Problem - solved

Pixel shift problem finally solved - buffer 4096 bytes used for copying was to big - according to consultation with Robert Sulej it should be
512 bytes, but in such case transfer is to slow. I used buff_size=1024 and pixel shift disappeared !!!!
Today (20060721) is the first night of tests without pixel shift and without correcting for it !!!
So on pi2 I comment out lines in custom cfg file /opt/pi/dev/pisys/daq/config/custom/ccd.cfg , lines :

# testing correction of pixel shift :
CCD_CHECK_FRAME=1
CCD_REPAIR_FRAME=1

changed to :

# testing correction of pixel shift :
# CCD_CHECK_FRAME=1
# CCD_REPAIR_FRAME=1

The soruce code change was done only in file Device2K2K.cpp in function Device2K2K::GetDataRawIOCTL , where buff_size was changed from 4096
to 1024 , maybe 512 should be used but then transfer is becoming to slow ...
Some more tests are maybe needed to see if 2048 will also be good .
Table below summarizes transfer times for different buffer sizes :

Buffer Size	Transfer Time	Pixel Shift
512	TO SLOW - I checked it >= 5 sec	NO
1024	Data transfer took : 1.05 sec	NO
2048	Data transfer took : 0.53 sec	YES
4096	Data transfer took : 0.53 sec	YES

Dodatkowo sprawdzilem , ze na pioneer2 ( czyli obecnie pi3 ) czas transmisji wynosil okolo 2 sec, czyli po zmianie kompa w wyniku moze
jakis roznic w hardwarze mielismy przy tym samym buforze odczyt duzo szybszy ( ~0.5 ) po zmieniejszeniu bufora , zmniejszyla sie tez szybkosc do 1 sec
i jak widac to juz jest wystarczajaco wolno i nie ma pixel shitfa.
Wartosci czasow transmisji mozna znalezc w headerach fitsow , pod keywordem CROTIME.
Generalnie widac ze jak transmisja jest za szybka to nie jest dobrze i nalezy pamietac o tym zeby nie byla szybsza niz 1 sec.