Hi there,
I'm running a custom image built with Yocto on my Akita. Currently i'm
running the 3.2 kernel. I suspect that the ads7846 driver takes quite a bit
more cpu time than the older, simpler corgi_ts driver. For instance, simply
running "top" and keeping constant pressure on the touchpanel will start
off the ksoftirqd and the irq/107-ads7846 process will take up to 3% of cpu
time. This does not happen on the older kernels with corgi_ts driver. I
know this is greatly oversimplified but i do have oprofile stats to make a
better point.
I use kdrive 1.7.99 and gtk 2.16.6 in my image. Here's another simple test
i did:
I open a text file in leafpad (~1000 lines). Scroll to the bottom and then
back up again two times using the touchscreen to drag the scrollbar. This
is done in portrait mode to factor out the fb rotation performance penalty.
Scrolling is mostly blitting so most of the action should be in Xfbdev and
Pixman. Here are some abbreviated oprofile stats on a 3.2 kernel for this
simple test:
System wide
==================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
CPU_CYCLES:100000|
samples| %|
------------------
80225 30.6922 vmlinux
35017 13.3967 Xfbdev
25602 9.7947 libpixman-1.so.0.23.6
24123 9.2289 libc-2.13.so
18613 7.1209 libglib-2.0.so.0.3000.0
18328 7.0119 libpango-1.0.so.0.2800.4
10324 3.9497 libgobject-2.0.so.0.3000.0
8641 3.3058 libgdk-x11-2.0.so.0.1200.7
6890 2.6359 libgtk-x11-2.0.so.0.1200.7
6163 2.3578 oprofiled
Kernel - 3.2
===================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
12809 15.9663 __do_softirq
9602 11.9688 __sched_text_start
3620 4.5123 cpu_xscale_switch_mm
3569 4.4487 giveback
2900 3.6148 spitz_ads7846_wait_for_hsync
2657 3.3119 rcu_bh_qs
1793 2.2350 tasklet_action
1720 2.1440 pump_messages
1631 2.0330 ring_buffer_consume
1587 1.9782 spi_async_locked
1510 1.8822 pump_transfers
1287 1.6042 input_event
1118 1.3936 run_ksoftirqd
1073 1.3375 worker_thread
982 1.2241 process_one_work
972 1.2116 get_signal_to_deliver
959 1.1954 rcu_sched_qs
783 0.9760 complete
Xfbdev
==================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
2135 6.2480 dixLookupPrivate
1292 3.7810 miGlyphs
1280 3.7459 Dispatch
1090 3.1898 CompositePicture
1087 3.1811 ReadRequestFromClient
968 2.8328 WaitForSomething
Pixman
==================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
7537 30.1854 pixman_blt_mmx.part.1
3411 13.6609 mmx_composite_over_n_8_0565
1177 4.7138 pixman_fill_mmx
1128 4.5176 pixman_image_composite32
1074 4.3013 mmx_composite_add_8_8
1039 4.1612 _pixman_bits_image_setup_accessors
832 3.3321 _pixman_image_validate
765 3.0638 pixman_region32_selfcheck
654 2.6192 pixman_region32_init
652 2.6112 _pixman_lookup_composite_function
561 2.2468 pixman_compute_composite_region32
531 2.1266 bits_image_property_changed
It seems that most of the time is spent in the kernel, processing ts
interrupts. This visibly impacts performance as you can clearly see text
redrawing as you scroll. Here's oprofile output of that same little test,
only now i use a usb mouse to drag the scrollbar:
System wide
==================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
CPU_CYCLES:100000|
samples| %|
------------------
39003 19.8228 vmlinux
35678 18.1329 libpixman-1.so.0.23.6
25286 12.8513 Xfbdev
22998 11.6885 libc-2.13.so
13126 6.6711 libglib-2.0.so.0.3000.0
10189 5.1784 libpango-1.0.so.0.2800.4
9336 4.7449 libgobject-2.0.so.0.3000.0
7392 3.7569 libgtk-x11-2.0.so.0.1200.7
6290 3.1968 libgdk-x11-2.0.so.0.1200.7
4160 2.1143 oprofiled
Kernel
=================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
3672 9.4147 __sched_text_start
3474 8.9070 cpu_xscale_switch_mm
3143 8.0584 __do_softirq
1132 2.9023 ring_buffer_consume
1069 2.7408 __hrtimer_start_range_ns
927 2.3767 do_select
588 1.5076 __copy_to_user_std
581 1.4896 __wake_up_sync_key
556 1.4255 unix_stream_recvmsg
556 1.4255 vector_swi
455 1.1666 default_idle
455 1.1666 mc_copy_user_page
432 1.1076 ktime_get_ts
430 1.1025 core_sys_select
Xfbdev
=================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
1696 6.8197 dixLookupPrivate
1124 4.5197 Dispatch
827 3.3254 ReadRequestFromClient
752 3.0238 miGlyphs
638 2.5654 CompositePicture
617 2.4810 dixLookupResourceByType
582 2.3403 XYToWindow
542 2.1794 WaitForSomething
509 2.0467 image_from_pict
487 1.9583 XaceHook
460 1.8497 damageComposite
379 1.5240 dixLookupPrivateAddr
359 1.4436 getDrawableDamageRef
333 1.3390 fbValidateGC
325 1.3068 fbBltOne
Pixman
=================
CPU: ARM/XScale PMU2, speed 416000 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
23715 67.0029 pixman_blt_mmx.part.1
2656 7.5041 mmx_composite_over_n_8_0565
1099 3.1050 pixman_fill_mmx
812 2.2942 mmx_composite_add_8_8
712 2.0116 _pixman_bits_image_setup_accessors
619 1.7489 pixman_image_composite32
487 1.3759 pixman_region32_selfcheck
455 1.2855 pixman_region32_init
397 1.1217 _pixman_lookup_composite_function
353 0.9973 _pixman_image_validate
Scrolling with a mouse is much faster and the oprofile data shows it. It is
also visually faster and screen redraws are not so visible.
Finally, here's that same test, on that same image, but this time with a
2.6.26-RP kernel which uses the old corgi_ts driver. Dragging the scrollbar
with the ts.
System wide
==================
CPU: ARM/XScale PMU2, speed 622 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
CPU_CYCLES:100000|
samples| %|
------------------
131815 25.0331 libpixman-1.so.0.23.6
84564 16.0596 libc-2.13.so
82888 15.7414 vmlinux
51252 9.7333 Xfbdev
30044 5.7057 libglib-2.0.so.0.3000.2
25637 4.8688 libgobject-2.0.so.0.3000.2
19355 3.6757 libpango-1.0.so.0.2800.4
18839 3.5777 libgtk-x11-2.0.so.0.1600.6
17266 3.2790 libgdk-x11-2.0.so.0.1600.6
17112 3.2498 libcairo.so.2.11000.2
10855 2.0615 libpthread-2.13.so
9703 1.8427 oprofiled
vmlinux - 2.6.26-RP
==================
CPU: ARM/XScale PMU2, speed 622 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
11787 14.2204 ts_interrupt_main.isra.0
7505 9.0544 cpu_xscale_switch_mm
6219 7.5029 __do_softirq
4832 5.8296 ide_intr
3778 4.5580 schedule
2174 2.6228 get_signal_to_deliver
1483 1.7892 __wake_up_sync
1431 1.7264 __copy_to_user
1413 1.7047 tty_ldisc_deref
1243 1.4996 __switch_to
1158 1.3971 do_select
Xfbdev
=================
CPU: ARM/XScale PMU2, speed 622 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
3159 6.2847 dixLookupPrivate
2472 4.9179 Dispatch
2217 4.4106 miGlyphs
1947 3.8735 CompositePicture
1817 3.6148 ReadRequestFromClient
1275 2.5366 WaitForSomething
1178 2.3436 dixLookupPrivateAddr
1135 2.2580 image_from_pict
992 1.9735 getDrawableDamageRef
977 1.9437 damageRegionProcessPending
885 1.7607 dixLookupResourceByType
869 1.7288 damageComposite
818 1.6274 miSpriteSourceValidate
767 1.5259 ValidatePicture
724 1.4404 FlushIfCriticalOutputPending
621 1.2355 fbBltOne
pixman
==================
CPU: ARM/XScale PMU2, speed 622 MHz (estimated)
Counted CPU_CYCLES events (clock cycles counter) with a unit mask of 0x00
(No unit mask) count 100000
samples % symbol name
106356 81.0838 pixman_blt_mmx.part.1
5485 4.1817 mmx_composite_over_n_8_0565
2327 1.7741 pixman_fill_mmx
1613 1.2297 _pixman_bits_image_setup_accessors
1519 1.1581 mmx_composite_add_8_8
1325 1.0102 pixman_image_composite32
1273 0.9705 pixman_region32_selfcheck
1155 0.8806 pixman_region32_init
And yet again, oprofile output for the 2.6.26-RP suggests that most of the
time is spent blitting, as it should be. Cpu speed is not accurately
reported, it's fixed at 416mhz as reported by /proc/cpuinfo. Scrolling is
also visibly faster than the 3.2.kernel.
So what's going on here? Is the ads7846 driver really such a hog? Or is
this due to the pxa spi driver? Corgi_ts used a light ssp microwire
interface.
I tried to reintroduce the old corgi_ts driver into the current kernel but
it doesn't work yet.
_______________________________________________
Zaurus-devel mailing list
Zaurus-devel@lists.linuxtogo.org
http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/zaurus-devel
