Hey raster,
Here is the non intrusive patch I talked to you about. Please apply it as it
introduces some fixes, some improvements and mostly and underlying
infrastructure for future RTL improvements.
(note hebrew & yiddish seem fine, but things expedite test seems to show are
wrong (why i don't know as i dont speak the langs- just comparing to pango /
gtk output):
arabic seems lsightl wrong (maybe composition chars not working?)
gujarati - also seems wrong
malayam - also looks wrong
persian - looks wrong
sinhala - looks wrong
tamil - looks wrong
these are what, appear to me, to look wrong. why they look wrong, i don't
know. i'm guessing its compositiong not being handled. but i dont's peak,
read or write any of these languages so i am unsure of what it really should
be like, why and how to fix it.
anyone want to put up a hand? (everything else is displaying fine as best i
can tell - the langauges i read/speak/somewhat understand are working fine).
SVN revision: 42814
be way too big to ever allocate. probably code can do with other fixes too.
also make jpeg loader rudametarily understand load regions. very brute-force.
but enough for just this moment to do testing.
SVN revision: 42507
Subject: [E-devel] [evas] Add RGBA -> grayscale 64 entries palette conversion
This is needed for E-Ink devices outta there. Names of new files,
configure.ac variables and macros are awful, suggestions are welcome.
SVN revision: 41825
evas_common_font_utf8_get_prev currently works correctly only on an
ASCII symbols. For non-ASCII it just returns random garbage somehow
constructed from a string.
Attached patch makes it work according to the comment at the start.
SVN revision: 41292
* handle error code and print out error message, this should remove
invalid access traces from valgrind.
* give "0" as pid to get self affinity, this is documented in the man
page.
SVN revision: 41118
none (ie default) and the engines actually understand it and use it.
2. fixes to scalecache and cserver too. more toto's done and its now been
stress tested by me - and i think cserve is ready to go gold. just enable it
with export EVAS_CSERVE=1 in your env for any eflapps - and run evas_cserve
(cmd-line options avalable plus cmd-line tol to query settings change on the
fly and query statsitics and state)
SVN revision: 40536
is it ok?
1. it can be --disabled in evas's configure, but i think it works WITHOUT
disabling it (runtime) as it falls back to the old way of loading
2. it may cause build problems on some platforms - without it being enabled
we won't find out, so enable.
3. it needs enabling runtime to make use of it so it should be safe for now
until you enable it.
what is it?
it is a SHARED cache server - that means images loaded are loaded BY the
cache server (not by the actual process using evas). images are shared via
shared memory segments (shm_open + mmap). this means only 1 copy is in all
ram at any time - no matter how many processes need it , and its only loaded
once. also if another app has already loaded the same data - and its in the
cache or active hash, then another process needing the same stuff will avoid
the loads as it will just get instant replies from the cache of "image already
there". as it runs in its own process it can also time-out images from the
cache too.
right now you enable it by doing 2 things
1. run evas_cserve (it has cmd-line options to configure cache etc.
2. export EVAS_CSERVE=1 (im the environment of apps that should use the cache
server).
it works (for me) without crashes or problems. except for the following:
1. preloading doesnt work so its disabled if cserve is enabled. thisis
because the load threads interfere withthe unix comms socket causing
problems. this need to really change and have the cserve know about/do
preload and let the select() on the evas async events fd listen for the
unsolicited reply "load done". but it's not broken - simple preloads are
syncronous and forced if cserve is enabled (at build time).
2. if cserve is killed/crashes every app using it will have a bad day. baaad
day. so dont do it. also cserve may be vulnerable to apps crashing on it - it
may also exit with sigpipe. this needs fixing.
3. if the apps load using relative paths - this will break as it doesnt
account for the CWD of the client currently. will be fixed.
4. no way to change cache config runtime (yet)
5. no way to get internal cache state (yet).
6. if cache server exist - it wont clean up the shmem file nodes in /dev/shm
- it will clean on restart (remove the old junk). this needs fixing.
if you fine other issues - let me know.
things for the future:
1. now its a separate server.. the server could do async http etc. loads too
2. as a server it could monitor history of usage of files and images and
auto-pre-load files it knows historically are loaded then whose data is
immediately accessed.
3. the same infra could be used to share font loads (freetype and/or
fontconfig data).
4. ultimately being able to share rendered font glyphs will help a lot too.
5. it could, on its own, monitor "free memory" and when free memory runs
load, reduce cache size dynamically. (improving low memory situations).
6. it should get a gui to query cache state/contents and display visually.
this would be awesome to have a list of thumbnails that show whats in the
cache, how many referencesa they have, last active timestamps etc.
blah blah.
please let me know if the build is broken asap though as i will vanish
offline for a bit in about 24hrs...
SVN revision: 40478
this (apparently) makes right-to-left work for display. that means:
hebrew
yiddish
arabic
persian
...
i don't read (or write ... or speak) an of these...
so.. i'll leave it up to others to find bugs/problems and send patches.
thanks a huge amount to TAsn for the patch!
SVN revision: 39972
dunno why, but at least it does for me and some users at #e/#edevelop,
e17 freezes at start, so does canola and other evas apps, maybe due
64-bits? No time to investigate right now (at a conference).
SVN revision: 39738
enabled. the blending is not working/complete. the neon for fills and copies
isnt actually faster though currently :(
2. scalecache infra - disabled for now. working on it.
SVN revision: 39723
* mainly unused parameters
* in src/lib/imaging/evas_imaging.c, set font to NULL
* in src/lib/canvas/evas_object_gradient.c, add unititialized member
there are a *lot* of reported warnings by llvm, i'll fix them later
there are also *lots* of unused parameters (compile evas with -W). I'll
fix them later too
SVN revision: 39172
2. make gl engine able to use cutouts - in some cases its faster, some
slower. it's a mixed bag. not sure what to make of it. it's #defined to be
disabled atm.
SVN revision: 39114
* evas/src/lib/engines/common/evas_font.h,
* evas/src/lib/engines/common/evas_font_draw.c,
* evas/src/lib/engines/common/evas_font_load.c,
* evas/src/lib/engines/common/evas_font_query.c: Add cache for font kerning.
This patch give something around 2% for all tests around text in expedite,
except for Textblock Intl where it give a 3 times boost.
Regarding text rendering speed, something is strange when used by evas_pipe.
All tests using Styles are around 40% faster without evas_pipe. 30% faster
for Text Change. But Text Basic 7% slower. So it should be possible to have
faster rendering when using evas_pipe for font rendering.
SVN revision: 38993
tyo fix up some of these breaks first and there isn't a lot of time devoted
to this. so revert this. it's in svn history so we can dig it out any time we
like.
SVN revision: 37453
1. nearest scaling is now broken - it's always linear interpolation. this
will lead to slowdowns. i need to fix this - a must.
2. i think it's time i put in a transformed image cache that can cache an
image object at a transform (and share it) automatically.
3. transforms in non-software-engines will not work - broken. need to at
least do xrender and gl engines.
any volunteers to help?
SVN revision: 37447
565. what it does is it's a half-way house between no dither and "small
dither mask". what is does is change the "rounding" between odd and even
lines. the resolt is not as nice as small (or large) dither mask - but its
almost "good". it gives a horizontal-blinds like look to dithering. it's
almost as fast as no-dither in my tests on a core2 (of course making lots of
use of branch prediction or conditional instructins - whihc arm nd x86 have).
SVN revision: 36373
- some enignes break as they dont have the stubbed out functions, and
xrender/gl engines dont even implement the drawing and need to (but are
stubbed out).
SVN revision: 35677
pthread manual says it is safe to destroy them after they are used
with pthread_create: "If the attributes specified by attr are modified
later, the thread���s attributes shall not be affected."
SVN revision: 34937
Image_Entry flag structure. This fix a bug with 16 bpp software engine.
* Change image loader module API to take any Image_Entry. Same goes
for evas_common_image_premul and evas_common_image_set_alpha_sparse.
* Use new eet API: eet_data_image_read_to_surface.
SVN revision: 34728
totally done and may hand control back to a non rendering codepath - or
before u do any fp ops u are unsure of the cpu state beforehand. see the big
fat comment.
i was right. it was a missing emms.
SVN revision: 34707
2. free then malloc otherwise realloc might ALSO have to memcpy if it cant
resize the segment whihc means overhead we dont need/want.
SVN revision: 34441
* formatting
* put WIN32_CFLAGS in AM_CFLAGS and not AM_CPPFLAGS, as it is where it belongs
* rename create_shared_lib to lt_no_undefined
* pass -Wl,--enable-auto-import to libtool when compiling with cegcc
* add files to EXTRA_DIST only when they are not in _SOURCES or _include_HEADERS (they
are added anyway)
SVN revision: 34353
* Allow Windows Mobile to correctly load dll's
* Use correct scheme for EAPI on Windows and include config.h when necessary
* add -mwin32 to compiler flags when compiling with cegcc
SVN revision: 34024
every thing is not really cleaned on shutdown (It make it crash, if
you shutdown completely the font system and then restart it again).
Author: Cedric BAIL
SVN revision: 33746
As agreed on IRC, evas_hash_foreach() now takes const, to make clear
that hash shouldn't be changed. If one wants to change he must do a
cast and return 0. However this will require users to be updated in
applications.
SVN revision: 33708
* use non deprecated version of AC_INIT and AM_INIT_AUTOMAKE
and check the required minimal versions.
* add bzipped distribution archive
* add AC_LIBTOOL_WIN32_DLL
* forbid libtool to check fortran
* compute libtool versioning from the version of the package
* pass the directories based on ${prefix} to the preoprocessor
with the -D option
* replace INCLUDES, wich is deprecated since 2001 by AM_CPPFLAGS
* remove useless -L flags in *_la_LDFLAGS
SVN revision: 32337
software-16 uses. it works and in some cases gets massive speedups (70%+) but
in a few its slowdowns (30% down) in expedite tests - why, i don't know. it
should be the same or better in all tests. disabled for now - also not
complete. < 32bpp wont' work and not sure rotation works and masks don't work
either.
SVN revision: 31928
Due the comparions, the code worked fine, but use the correct type
size so it's cleaner.
By: Brett Nash (compare-whole-pointer.patch)
SVN revision: 31696
tuned for best performance on my core2 duo desktop - for now. will check
more. also make the yuv colorspace code be a bit more robust and fix leak in
gl engine with shaders.
SVN revision: 30192
* fix reference counting for font_int(s)
* properly backout if face doesnt load correctly, this way we dont end up
using the face again
* put some comments about fudged font cache
SVN revision: 29863
in evas_gl_texture.c i have a frag shader, and it tries to use a set of 3
textures that act as the yuv planes, BUT the u and v textures (Utex and Vtex)
are simply getting values from the Ytex - regardless of what i try. grrr.
what's up with that?
SVN revision: 27495
1. disable viewports other than 1:1 at 0,0
2. remove output space coorsds for pointer.
3. remove geom caching
4. make threaded pipelined engine a runtime detect if u have > 1 cpu.
5. pthread build default if u have pthread.h and sched.h
SVN revision: 27131
sometimes slower)
2. --enable-pthreads will enable multi-threaded rendering (current support is
for up to 4 threads so if you have a new fanled quad core or dual cpu dual
core box or whatever you will in theory be able to max moe of its cpu grunt
with the software rendering engine. this can only be done because i added the
pipelines which means almsot entirely lock-free multithreading internally in
evas. the only locks are for fonts but with a little work i might be able to
remove some/most of those too)
for now pthreaded rendering likely will be linux only (it relies on sched.h
for setting scheduler params to force the slave threads to run on separate
cpu's as linux likes to keep them on the same cpu otherwise and thus we get
no speedups at all - only slowdowns).
aso note that it is a bit of a mixed bag. complex ops (like smooth scaling
with alpha blending) get speedups, but simple ops (like blits/fills) slow down.
this all neds examination and tweaking still - but it's a start.
SVN revision: 27098
renderer gets threaded). if i thread at the simplest levels (low down in for
example the image scaler code - one of the most expensvie gfx routnes) on an
actual dual core system - performance drops by 40%. this just doesn't work
well at that level. thread creates and joins per render op are just a bad
thing (tm) :) so this really needs to go in much higher up and that presents
problems. :( i will need to clearly define entry and exit points to and from
threaded space (and thus all the locks) - remove all nested calls (where
internal code goes thru the same entry/exit points traditionally so it
deadlocks itself).. anyway - this here has all that code stripepd out i
played with - it is just the autofoo and build stuff so we can turn on/off
thread support at will in the build.
SVN revision: 26817
and found that reordering the evas list containing the font path list was
actually hurting cache performance. Frequent cache misses occurred in
evas_mempool_free in it's main loop. So removing the list node and re adding it
actually decreased cache performance in some cases. This would also cause memory
allocation bouncing if mempools were not used.
Added evas_list_promote_list to move a list node to the beginning of a list.
Changed reordering lists to use evas_list_promote_list.
SVN revision: 24387
* When fonts are added from files we will only load the details
* The first font is always loaded
* I know there is a patch for this on the list but this way seems better to me
SVN revision: 21964
* This can be disabled with --enable-fontconfig=no
* Test in e17 with "enlightenment_remote -font-default-set title_bar Serif 12"
* This is just the beginning: TODO: style searching, language reorder bypass
SVN revision: 21394
2. be able to skip a copy in certain cases when scaling - should improve
speed in several situations - evas is defintiely not optimal :)
SVN revision: 19983
debugging/fixing and likely int he end the exact same result of fixing them.
yes - we lose performance - but it actually is correct now :) if we want to
do such radical changes- i sugegst moving to premultiplied alpha and makign a
tonne of externally tested routines in a test harness first to compare
correctness and speed in an isolated environment.
SVN revision: 18947
do NOT depend on order operation precedence. it broke scaling. laos other
completely bizarre mmx things were going wrong with mm7 ending up not 0 so
i've had to force it to be 0.
SVN revision: 18811
--enable gl, qtopia and directfb enigines - they are either incomplete, buggy
or simply used so little that its not worth building unless the user REALLY
wants the support.
SVN revision: 18424