This document discusses performance optimizations made to the QtWebKit browser on MIPS platforms. It describes profiling the browser to identify optimization opportunities in video/audio decoding, image operations, and drawing functions. SIMD instructions on the MIPS processor were used to optimize the GStreamer multimedia framework and Qt drawing. Measurements showed up to 30% faster rendering times after optimizing Latin-1 to UTF-16 conversion, alpha blending, color conversions, and string operations using MIPS DSIMD instructions. Most changes were merged upstream but not backported to older Qt versions.

1. IMPROVING $PORT
PERFORMANCE ON $ARCH
PLATFORM-BASED PERFORMANCE TUNING OF WEBKIT
(PORT=QT ARCH=MIPS74KF)
Embedded Linux Conference
April 29 — May 1, 2014
Adrián Pérez de Castro

2. WHOAMI
aperez@igalia.com
+AdrianPerezDeCastro
@aperezdc

3. THE CHALLENGE
MAKE A QTWEBKIT-BASED BROWSER USEABLE
ON LIMITED HARDWARE
MIPS 74Kf @500 MHz
RAM: 256 MB
No GPU

4. MIPS74KF
“Classic” MIPS32
+
FPU
+
MMU
+
DSP

5. DSP?
No. Not really a DSP.
SIMD instructions suitable for signal processing.

6. CAN WE USE THIS TO IMPROVE PERFORMANCE?
CHALLENGE ACCEPTED

7. THE PLAN
PROFILE →OPTIMIZE →VALIDATE

8. WHAT TO OPTIMIZE
Video/audio decoding.
Image operations.

9. WHERE TO OPTIMIZE?
Can we improve the platform overall,
not just WebKit?
Yes!
QtWebKit uses the Qt drawing functions.
A/V decoding uses GStreamer, which uses Orc.
Good candidates for SIMD code.

10. LIMITATIONS
No Valgrind.
No GDB.
No perf.
No performance counters.
↓
qemu + gdbserver.
gperftools.
CLOCK_PROCESS_CPUTIME_ID

11. ROLL YOUR OWN TOOLS
(WITH HELP FROM EXISTING ONES)

12. GNU HAMMER^WTIME!
#Usefullpathtoavoidusingtheshell'stimebuiltin
#Onelineperrunwithuser/systemtimeandpagefaults
/usr/bin/time-a-otimings.txt
-f'%U%S%F%x%C'$COMMAND
#Forexample,measuringtheqtdemuxGStreamercomponent
/usr/bin/time-a-otimings.txt
-f'%U%S%F%x%C'gst-launch-q
filesrc=file.mp4!qtdemux!video/x-h264!fakesink

13. TIMING
Beware of CLOCK_PROCESS_CPUTIME_ID's resolution!
#defineCLOCK_MAX_RESOLUTION_DELTA (10000.0*1e-9)
boolusePosixClock(){
staticboolchecked=false;
staticbooluseposix;
if(!checked){
if(posixClockAvailable()){
doubleres_theorical=posixClockTheoricalResolution();
doubleres_empirical=posixClockEmpiricalResolution();
useposix=fabs(res_theorical-res_empirical)
<=CLOCK_MAX_RESOLUTION_DELTA;
}else{
useposix=false;
}
checked=true;
}
returnuseposix;
}
clock.cc

14. WEBSNAP
%g++-DMAIN-oclockclock.cc
%./clock
CLOCK_PROCESS_CPUTIME_IDissupported
Resolution(advertised/empirical):0.0000000010/0.0000002460s
Sampledresolution:0.0000005470s
Printingthelinesabovetook0.0000483550s
%LD_PRELOAD=/usr/lib/libprofiler.so
./websnaphttp://igalia.com1000pprof
Loading100%Layoutcompleted
Loadsuccessful
libprofile.sodetected(0x7f77468e8f90,0x7f77468e8fd0),output'pprof'
Profilingstarted,code:0x1,timeout:0
PROFILE:interrupts/evictions/bytes=634/537/22168
http://igalia.com10006.2709987870s
%mkdirout&&./runtests1000<urls.txt
github.com/aperezdc/websnap

15. ...AND BEYOND
Ad-hoc Python/Bash scripts:
Fixlibrarypathsinprofileroutput.
Datamunging.
Measurementscomparison.
GenerateCSVfiles.
Reportgeneration.
…

16. SOME RESULTS
(DETAILED)

17. LATIN-1→UTF16: V0
//"dst"array(uint16_t*)
//"src"array(uin8_t*)
while(len--)
*dst++=(uchar)*src++;

18. LATIN-1→UTF16: V1
;a0:"dst"array(uint16_t*)
;a1:"src"array(uint8_t*)
;a2:"len"
1: lbu t1,0(a1)
addiu a2,a2,-1 ;len--
sh t1,0(a0)
addiu a0,a0,2 ;dst++
bnez a2,1b
addiua1,a1,1 ;src++

19. LATIN-1→UTF16: V2
1: lw t1,(a1) ;t1=ABCD
;
;TODO:extractbytesfromt1tot2/t3,padding
; themwithzeroes:t2=0A0B,t3=0C0D
;
addiu a1,a1, 4 ;src++
addiu a2,a2,-4 ;len--
sw t2,0(a0)
sw t3,4(a0)
bnez a2,1b
addiua0,a0,8 ;dst+=2

20. LATIN-1→UTF16: V3
1: lw t1,(a1) ;t1=ABCD
srl t2,t1,24 ;t2=000A
sll t2,t2,16 ;t2=0A00
sll t1,t1, 8 ;t1=BCD0
srl t4,t1,24 ;t4=000B
or t2,t2,t4 ;t2=0A0B
sll t1,t1, 8 ;t1=CD00
srl t1,t1,16 ;t1=00CD
andi t4,t1,0xFF00;t4=00C0
sll t4,t4, 8 ;t4=0C00
or t3,t1,t4 ;t3=0C0D
addiu a1,a1, 4 ;src++
addiu a2,a2,-4 ;len--
sw t2,0(a0)
sw t3,4(a0)
bnez a2,1b
addiua0,a0,8 ;dst+=2

21. LATIN-1→UTF16: V4
;DSPinstructionscanunpackbytesdirectly :-)
1: lw t1,(a1) ;t1=ABCD
preceu.ph.qblt2,t1 ;t2=0A0B
preceu.ph.qbrt3,t1 ;t3=0C0D
addiu a1,a1, 4 ;src++
addiu a2,a2,-4 ;len--
sw t2,0(a0)
sw t3,4(a0)
bnez a2,1b
addiua0,a0,8 ;dst+=2

22. LATIN-1 →UTF-16

23. ALPHA BLENDING

24. UTF-16 STRICMP()

25. RESULTS
Speedup histogram

26. UP TO 30% FASTER RENDERING
Thanksto:
Orc backend using MIPS DSP instructions
QImagecomposition operations
Color conversion (RGB16/888→ARGB32)
Alpha premultiplication and blending
String conversions and comparisons

27. UPSTREAM STATUS
Orc backend complete upstream
Initial work based on Qt 4.8
Most of the code is already in Qt 5.2
Rest in the next release
No backport to Qt 4.8

28. THANK YOU
FOR YOUR ATTENTION
perezdecastro.org
+AdrianPerezDeCastro
@aperezdc