About Me

Michael Zucchi

 B.E. (Comp. Sys. Eng.)

  also known as Zed
  to his mates & enemies!

notzed at gmail >
fosstodon.org/@notzed >

Tags

android (44)
beagle (63)
biographical (104)
blogz (9)
business (1)
code (77)
compilerz (1)
cooking (31)
dez (7)
dusk (31)
esp32 (4)
extensionz (1)
ffts (3)
forth (3)
free software (4)
games (32)
gloat (2)
globalisation (1)
gnu (4)
graphics (16)
gsoc (4)
hacking (459)
haiku (2)
horticulture (10)
house (23)
hsa (6)
humour (7)
imagez (28)
java (231)
java ee (3)
javafx (49)
jjmpeg (81)
junk (3)
kobo (15)
libeze (7)
linux (5)
mediaz (27)
ml (15)
nativez (10)
opencl (120)
os (17)
panamaz (5)
parallella (97)
pdfz (8)
philosophy (26)
picfx (2)
players (1)
playerz (2)
politics (7)
ps3 (12)
puppybits (17)
rants (137)
readerz (8)
rez (1)
socles (36)
termz (3)
videoz (6)
vulkan (3)
wanki (3)
workshop (3)
zcl (4)
zedzone (26)
Friday, 26 April 2019, 02:10

dez 2.0 for DEZ1 is here!

I just released dez 2.0 for the DEZ1 binary delta format.

Consider this a 1.0 release! The format is now frozen.

It was just going to be 1.3 but I noticed that 1.2 came out 4 years ago so I thought 2.0 was in order. I did prepare 1.3 in the repository but then I just tagged it again with 2.0.

There aren't really many tunables left so I dropped the flags field and made the OP_SINGLE_SPLIT value tunable. It doesn't make much difference anyway.

This is now a modular Java 11 project using my common java.make. I'm also using it to work on adding junit(4) testing to the java.make system. I may split out java.make into jni.make, java.make, junit.make, and so on. tuning it so haven't rolled it back into the base yet.

As usual the Delta-Z home page has download links and other snot. code.zedzone.au houses a browsable copy of the repository; which now uses git. q

Update: I decided to update the javadocs, as they are. The interesting page is the DEZFormat class which defines and describes the format.

These initial links may not be long-term stable as I will possibly look at a unified javadoc build across all my java projects. And I really need to do something about the blinding default stylesheet. My eyes! My EYES!

Update: I couldn't put up with it so I fixed the stylesheet. This time I used some Workbench 2.0 colours but without the 3D borders. Came out quite nice. I filled out the api documents somewhat as well.

Tagged dez, java.
Wednesday, 24 April 2019, 12:39

Source Code Browser

I've just installed a source code browser on code.zedzone.au and re-added Code back to the main menu of most pages.

I'm using a slightly hacked up version of gitweb. I needed to add a bit more structure and style stuff to get the look I wanted and disable some junk I didn't need. Some more Commodore colourations for a bit of fun.

This is still work in progress and i'm still trying to decide how i'll manage the repositories but hey it works. I'll be slowly migrating all my projects over to it over time. Any java projects will become modular projects first. I'm not bothering to keep any history because I don't need it.

Tagged jjmpeg, nativez, opencl, zedzone.
Wednesday, 24 April 2019, 04:10

libeze 2.0, 1.1

I had a few fixes I did ages ago I thought I should push out. Mainly some stupid list errors.

I also redid the tree api such that a version bump was required. The changes are sore of either/or and add an extra argument to several functions but it was for some consitency with the list api (no allocations, static init) and I found some intersting uses for changing the comparison function at search time while working on playerz.

The software is on the libeze page as usual. I also added a CHANGES file with the release notes/changelog.

Weeks ago I also did a lot of other experiments and played with ideas but they never fully formed. Support for more sophsiticated serialisation, elf stuff, and lots of experiemnts with memory and pool allocators. Just can't make a decision on those and I forgot where they were at so who knows.

I haven't done it for libeze yet but i've started moving my projects to using git. I don't really like it (the commands are often obtuse and the documentation isn't user oriented) but CVS support is starting to wane quite markedly in modern tools. Sigh.

Tagged hacking, libeze.
Tuesday, 23 April 2019, 03:14

Big Easter Software Bundle!

After basically spending the entirety of Easter (a 4 day holiday here in Australia) plus half of today I've packaged up and released a big ol bunch of stuff.

NativeZ 1.0 is an updated version of CObject which is portable and modular. I've moved my other projects to use it rather than their own copies which had evolved and diverged over time.

jjmpeg 4.0 is updated to use NativeZ as described on the previous post. There are also a bunch of bug fixes and small improvements. It now includes a module for some demos.

zcl 1.0 has been converted to a modular project and mostly to using NativeZ apis. The tools package has moved to a demo module.

They are all now capable of cross platform builds and they all generate platform specific .jmod files as one of the outputs. To do this they all use a new common java.make which is a modular and updated version of an earlier java.make as covered in the last post (and I made further improvements). Apart from modules it supports parallel builds, generated java sources, native code libraries, and cross platform builds using a very simple configuration mechanism.

They're also a bit smaller due to compression used for some initialisation tables.

Another project I have gone quite far on is an incremental java compiler. By scanning the .class files and .java timestamps i'm able to generate complete and accurate lists of both files that need to be rebuilt and files that need to be deleted across a multi-module source compilation. I have all the parts and I think they should work but I need to tie them together into a usable system, and then extensive testing.

Sigh, but I need a holiday. I might take most of the week off - like pretty much everyone else is anyway.

Tagged java, jjmpeg, nativez, zcl.
Monday, 22 April 2019, 00:39

Development Ongoing

I've spent all of Easter so far just hacking away at a bunch of stuff. Today looks nice so I think I will try to get out of the house, but who knows. My sleep apnoea has been pretty terrible lately and when I foget to use my MAD (mandicular advancement device) i've been wiped out all day, which has been most of this week. Sigh.

I've primarily been working toward a new modular and portable release of jjmpeg and zcl, but i've been distracted along the way and also had some difficulty just making basic decisions.

java.make

I spent way too long on trying to perfect the build system and things like where intermediate files are stored. There are still trade-offs that might mean another change, sigh.

But it does work pretty well. The main makefile isn't particularly complicated even if some of the macro quoting makes it messy to read. The driver makefiles are trivial - like a simpler version of automake.

It works by the master makefile including config.make, defining some variables, and then including java.make.

This for example is the complete Makefile for nativez.

dist_VERSION=1.0
dist_NAME=nativez
dist_EXTRA=README                               \
 build.xml                                      \
 nbproject/build-impl.xml                       \
 nbproject/genfiles.properties                  \
 nbproject/project.properties                   \
 nbproject/project.xml

include config.make

java_MODULES=notzed.nativez

include java.make

It's a trivial case so it has a trivial makefile. I added a make dist target to build up an archive of the source so that meta-data is really most of the file.

The makefile also builds the native libraries and that is covered by another makefile fragment src/<module>/jni/jni.make which is automatically included if detected.

This is likewise quite trivial for such a trivial example.

notzed.nativez_JNI_LIBRARIES = nativez

nativez_SOURCES = nativez-jni.c nativez-$(TARGET:-amd64=).c
nativez_CFLAGS = -Os -Wmissing-prototypes
nativez_HEADERS = nativez.h

This includes the main .c file and either nativez-windows.c or nativez-linux.c depending on the current target.

So make does this:

  1. Compiles all .java under src/notzed.nativez/classes to .class files.
  2. Generates all native binding headers (from javac -h).
  3. Automagically creates .d dependency files for all .o files.
  4. Compiles all .c into .o objects.
  5. Copies nativez.h file into a staging area for the .jmod.
  6. Links the .o files into libnativez.so or nativez.dll depending on the target.
  7. Copies all the .class files into a modular .jar file.
  8. Copies all the .class files, the .so file, the .h files and license files into a .jmod file.

And only does each step if necessary. For example step 2 doesn't just take the output of javac which would force every .c file to be recompiled any time any .java file changes regardless of what it included. So I use install -C to only copy changed headers from a staging area. I also use some tricks to avoid building all the .c dependencies (which requires first compiling all the java) when you do make clean or make dist.

I'm still debating whether i blow away some of the javac output directories to ensure an absolutely correct build, or let that be a separate step and just rely on using the per-module compile flag which is much faster if less accurate.

I'm also still debating whether it builds the .jar and .jmod files by default or as a seprate target. For these trivial projects it doesn't really take any time but if it did it could become a pain as the code can be executed without either.

jjmpeg is a somewhat more complex case which includes generated java sources, external dependencies, and (now) generated c sources and it isn't much more complex to define.

Initially I had everything in the same project but decided to split it out. The former did allow me to hone down the details on the build system though and i'm pretty happy with it now; it's very accurate, simple, flexible, and easy to use.

nativez

I mostly just moved over to the new build system (trivial) and added a few new functions.

However due to some of the latest in jjmpeg I will probably revisit some of the underlying ideas here.

zcl

I converted this into a modular project which mostly meant moving some stuff around. I also changed it to use NativeZ as it's GC/binding mechanism and to use nativez utility functions where appropriate. Being modular lets me split of auxilliary demo code more cleanly so i've done that too.

zcl is all pretty much implemented in a single large (5KLOC) file. I actually spent a couple of hours splitting it apart into multiple files because I thought it was getting a bit unwieldy (and modern emacs struggles with the size of the file unfortunately). But then I changed my mind and threw it all away.

zcl uses a generated pointer table and typedef casts to manage the OpenCL function linkage and while it works i'm thinking of revisiting it again.

I haven't used OpenCL for a while and don't need it at work at the moment and I can never really think of anything useful to do with it. So it hasn't been tested much.

jjmpeg

I already had a modular jjmpeg so this wasn't a big change, until I decided to keep going. First I moved to the new build system. I already had moved it to NativeZ but there were some lingering functions that I hadn't cleaned up.

Then I wrote some demos. A simple example to read video frames. One that saves frames to image files via Java2D functions. Another that displays video frames in a resizable window at about the right speed (i.e. a mute video player). And then a music player, which lets you adjust output parameters.

The music player needed some more bindings and converted enums so I added that. And it needed a way to get to the samples in a format I could play.

First a bit of background. In earlier versions of jjmpeg I just exposed the data pointer array of an AVFrame as a ByteBuffer and provided entry points into libswscale to allow one to convert frame data formats. This is very messy and error prone and really quite pants to use. So copying the ideas from the PixelReader from JavaFX I've created a similar AVPixelReader which is flexible enough to implement MOST of the PixelReader api and hide all the details of libswscale. The current implementation works by creating an AVPixelReader targeting the size you want and then you can write it various Java targets in whatever pixel format you want. I've got some nice helpers that wrap this to write to some Java2D BufferedImage (via the DataBuffer) or JavaFX WritableImage (via a PixelReader).

So I decided to do the same thing with libswresample and samples. But I don't need to support other pixel readers/writers and only just dump out arrays of bytes so I made a much simpler interface.

Although libswresample can stream I basically enforced that a whole buffer needs to be converted at a go otherwise the logic of tracking the frame gets a bit messy.

At this point I was cleaning things up and thinking about a release but got side-tracked again ...

jjmpeg - bindings

jjmpeg has a couple of different ways that things are semi-automated. All simple parameter accessors are created using C macros. Java JNI handles are loaded via a table which describes the type, name, and signature. Library functions from libavcodec.so et al are defined by a table and loaded by a simple platform-specific function. The tables are all const'd, the variables are mostly static, so everything should go in read-only memory and be fairly efficient to utilise.

The tables are handrolled. As are the variables which hold the function pointers - which means i'm copying them directly from header files and turning them into variables. This is error prone and provides no compile time safety confirming that the function matches the prototype.

And with 64-bit pointers the tables themselves are quite large. Each entry has pointers to strings, pointers to pointers, and some flags.

So I thought i'd revisit that.

With a simple code generator I create a struct with all function variables in it, extracted from the headers at compile time. I have a compact descriptor - essentially a stream of bytes - to encode enough information in order to fill out said structure.

It's a mess of shit perl but it works.

The resolution of java objects for jni use is also likewise table driven, so I did the same for that. It makes the assumption that jclass, jmethodID, and jfieldID are all pointer sized but that seems a pretty safe bet.

In short I shaved off about 20% of the size of libjjmpeg.so doing this.

This is an example of the driver file which defines things of interest.

header avutil libavutil/opt.h { 
        av_free
        av_dict_free

        av_set_options_string
        av_opt_serialize
        av_opt_copy
        av_opt_next
        av_opt_find

        av_opt_set
        av_opt_get

        ...
}

java AVOption au/notzed/jjmpeg/AVOptions$AVOption {
        <init>, (JLjava/lang/String;Ljava/lang/String;IILjava/lang/String;)V
        p, J
}

I'm not particularly enamoured of the Java definitions but it simplifies the generator. Hmm, looking now I can use javap to get this info so maybe I can simplify this somewhat whilst performing additional verification checks at the same time.

The generator creates a couple of static anonymous structs and a descrption string, and as a bonus includes the required header files. There are also some #defines so I didn't have to change any of the code that uses them.

/* This file was autogenerated on Monday 22 April  09:35:47 ACST 2019: */
/*  src/notzed.jjmpeg/jni/extract-proto.pl src/notzed.jjmpeg/jni/extract-proto.pl -f /opt/ffmpeg/4.0 src/notzed.jjmpeg/jni/jj-avoptions.def */
#include <libavutil/opt.h>
static struct functable {
        /* libavutil */
        void (*av_free)(void *ptr);
        void (*av_dict_free)(AVDictionary **m);
        int (*av_set_options_string)(void *ctx, const char *opts, const char *key_val_sep, const char *pairs_sep);
        int (*av_opt_serialize)(void *obj, int opt_flags, int flags, char **buffer, const char key_val_sep, const char pairs_sep);
        int (*av_opt_copy)(void *dest, const void *src);
        const AVOption *(*av_opt_next)(const void *obj, const AVOption *prev);
        const AVOption *(*av_opt_find)(void *obj, const char *name, const char *unit, int opt_flags, int search_flags);
        int (*av_opt_set)(void *obj, const char *name, const char *val, int search_flags);
        int (*av_opt_get)(void *obj, const char *name, int search_flags, uint8_t **out_val);
} fn;
static const char *fn_names =
        "#avutil\0"
        "av_free\0"
        "av_dict_free\0"
        "av_set_options_string\0"
        "av_opt_serialize\0"
        "av_opt_copy\0"
        "av_opt_next\0"
        "av_opt_find\0"
        "av_opt_set\0"
        "av_opt_get\0"
        ;
static struct {
        // au/notzed/jjmpeg/AVOptions$AVOption
        jclass AVOption_classid;
        jmethodID AVOption_new_jlliil;
        jfieldID AVOption_p;
} java;
#define AVOption_classid java.AVOption_classid
#define AVOption_new_jlliil java.AVOption_new_jlliil
#define AVOption_p java.AVOption_p
static const char *java_names =
        "#au/notzed/jjmpeg/AVOptions$AVOption\0"
        ".<init>\0(JLjava/lang/String;Ljava/lang/String;IILjava/lang/String;)V\0"
        ",p\0J\0"
        ;

Rather than have a single global table I have one per file (roughly on per class). This leads to a small amount of duplication but keeps everything simple and isolated and also aids incremental compilation. By having each table static to the compilation unit also allows the optimiser to remove any address calculation - it just goes to the linker via the .bss section.

I'm still thinking of other ways to reduce the binding size and/or overhead. For example the accessors are pretty small ... but there's many of them. The problem is if I use tables I need to keep them in sync between Java (e.g. field ids), or if I use strings then I need table lookups. ffmpeg seems to be moving toward the 'libavutil/opt.h' mechanism for accessing everything anyway (probably because of python influence?) so perhaps I should just use that or at least evaluate it.

But I will first have to move these changes to nativez and perhaps zcl.

Other

I fixed a few bugs along the way. I wasn't creating a few objects correctly so they weren't being tracked by the recovery process - mainly a holdout from earlier implementations which tracked objects differently. I wasn't freeing an AVFormatContext properly - these can be created in two different ways and need to be freed differently. Sigh, even the free functions take different argumements (some of the avcodec api is such a mess). Anyway I use the opaque pointer to track this so I could do it properly. There are probably other leaks but it's rather much harder to test than a standalone C program like playerz.

I also looked very briefly at writing a Vulkan binding as well. Hmm, i'm not sure on this one, it's a pretty big API although as I learn more it will probably shrink in apparent size. But regardless it's a bit more than a weekend project that OpenCL was.

It will probably need a code generator to make it managable, at least for certain objects like the construction parameters. I'm mostly concerned with the overheads so will be looking into different mechanisms.

And these might trigger a reworking of everything else along the way.

Still, i'm not sure yet on it, I don't really have a need for it or know it well enough to do a good job either.

Tagged hacking, java, jjmpeg, nativez, opencl.
Monday, 08 April 2019, 23:23

Modular Java, Make. FTW!

I've been working on packaging up the modular versions of various bits of software I have. One big change is that due to the way javac (and to a lesser extent netbeans) works I had to rearrange the source tree, and it also completely changes the way javac is invoked. So my previous java.make is no longer useful.

As i've now added some inter-module dependencies (e.g. nativez) there's a another question: do I keep them as separate packages, or do I just combine them into a single modular project?

For work I went through the same question previously and ended up just putting everything in the one project.

So ... for similar reasons i'll probably do the same. One strong reason not to is that the projects aren't as closely related - jjmpeg, zcl, and so on.

So what to do ...

Recursive Makefiles

I thought one solution is to just have a build system which isolates the modules so that you can only build the ones you're interested in. i.e. if you don't have external dependencies for one you're not interested in, just build the rest.

So ... based on the structure I have already I spent an inordinate amount of time trying to get the separate makefiles to work. In short: I fucking couldn't. I got 95% of it working, and the dependencies were relatively simple (yet very fragile!) but I just couldn't get make to rebuild some targets on a single run. A second run would always work. Just not good enough.

The problem is that as the toplevel makefile doesn't know when the sub-makefiles need to run it has to try to run them every time using phony targets. But it has to run them at the correct time. Generators are run first, then the module is compiled (which generates any native .h files), then jni is built if it exists, resources are copied, and jmod or jars created as appropriate. I could almost get it to work but some of the dependencies weren't recognised as updated; make would say 'this target needs rebuilding' then the next line 'this target is updated'. I think it has something to do with the way it handles phony targets and because I had multiple references to a status file from different levels. It would recognise one update but not another. I could get accurate builds, but not always get the jar/jmod from 'make jar'. I also had to do some other weird tricks like dummy actions on rules which where updated by dependencies via phony rules. I mean it worked but it was very obtuse.

I actually got it very close and it even supported parallel make quite well, but fuck what a pain to get working.

One Makefile To Rule Them All

So I did some searching and came across a chapter of a book that suggested using a single makefile, but using includes to make the parts more manageable.

So I took the per-module makefiles and stripped out some common stuff, renamed some variables and redid the toplevel makefile. After a little mucking about with the generated depenedencies a little bit I had something working within an hour. And parallel make works even better.

As an example this is what the toplevel makefile currently looks like. It automatically finds various things like java source, generators, and native sub-systems and runs them all at the correct time via the dependency graph.

nclude config.make

java_MODULES=notzed.jjmpeg notzed.nativez notzed.jjmpeg.fx notzed.jjmpeg.awt

notzed.jjmpeg_JDEPMOD=notzed.nativez
notzed.jjmpeg.awt_JDEPMOD=notzed.jjmpeg
notzed.jjmpeg.fx_JDEPMOD=notzed.jjmpeg

notzed.jjmpeg.fx_JAVACFLAGS=--module-path /usr/local/javafx-sdk-11/lib

include java.make
  

Provided targets include '{module}', 'jar', '{module}-jar'.

So now I have that working I can think about separating the projects into separate ones rather than modules in way that lets them all be built together. Probably in that case I would add an {module}_XDEPMOD variable which references the location and module(s) and just builds them before doing anything else. On the other hand this isn't the way you do it with autoconf (and other build systems, although some handle it) for external dependencies so maybe it's not worth it.

Faster ... Faster Builds

Whilst doing this I noticed early on that javac has a '-m' flag for building modules that performs dependency checking. Either I didn't notice this when I first wrote the modular build system or I didn't think it important: while it only rebuilds what is necessary it doesn't remove stale files (deleted inner/anon classes or colocated classes).

Slightly stale builds are not the end of the world so it's probably not a big ask to require a make clean occasionally anyway as the improved build time is significant.

But there is a better way to do it!

Using the jdk.compiler one can wrap access to javac and do some interesting things. To cut a long story short I wrote a javac server that monitors which source files are recompiled and automatically deletes any stale files. This works if all compiles go through the server but can result in stale files if files are compiled otherwise.

So I wrote some more code that scans the compiled class files once and then does a reverse lookup based on the SourceFile, this_class, and module-source-path to discover stale files and prime the database. This is done at server startup. Actually it should be able to be run at any time to post-fix the class tree, so it could be incorporated into the build in other ways.

I wrote a simple .class parser (almost trivial infact) to be able to retrieve the class metadata and some other simple code to do the source lookup.

I wrapped this up in a simple REST-like-service and a script 'sjavac' that compiles and runs the server on demand and then passes requests to it. The server just shuts itself down if it's idle for too long. The compile times for jjmpeg+nativez isn't very long anyway but using the compile server alone speeds up complete builds significantly, and incremental builds are now sub-second.

Another issue is concurrency from parallel builds. javac will recompile any out of date classes it encounters even if it isn't in the module specified by -m - which can lead to two separate compilers working on the same files. This conflict can break a build run. But I think this is due to incorrect makefile dependency definitions as if they are done properly this shouldn't happen.

sjavac serialises compiles at the module level anyway because the compiler uses multiple threads already (afaict, it's a bit too fast to find out on this project) and I haven't set the code up to handle concurrent compile requests anyway, but I want it to be a drop-in replacement for javac for a project build.

Now that i've moved to a single toplevel makefile it's actually a very simple setup. The java.make referenced above is only 100 lines long. sjavac is 100 lines of bash and 600 lines of stand-alone java. And together these provide faster, more accurate, and repetable builds than a special tool like ant or maven. With significantly less setup.

Update 10/4/19: Oh for fucks sake!

It turns out that javac -m doesn't recompile dependent classes if they are out of date, just the first-level dependencies or any missing classes they require. This doesn't make a lot of sense because there are many changes where this behaviour would lead to an inconsistent build so it makes this option fairly useless.

It's a trivial change to the makefile to fix it but it means that it only builds incrementally at the module level.

I guess this gives me the next problem to solve. I think I have enough scaffolding to be able to extract the dependency tree from the .class files. Well I'll see on that.

Tagged java, jjmpeg, nativez.
Tuesday, 12 March 2019, 06:20

modular cross-platform jjmpeg

So well this isn't an announcement of a new release because i'm basically too lazy to package it up at the moment.

But I have created a modular jjmpeg for Java 11 and added support for cross compilation. I'm going to be playing with some videos for work so I thought it might be useful if I generate any tools for other users.

As part of it i've moved it's use of CObject over to using the previously mentioned nativez stuff. I also moved a bunch of the JNI utility functions over to nativez because they're generally useful elsewhere.

Who knows when i'll get to publishing it.

Tagged hacking, jjmpeg, nativez.
Saturday, 02 February 2019, 02:08

ez-alloc, libeze, that music player thing

So it's been a pretty long week of hacking. I went deep on the memory allocator and eventually got one debugged and working. I made some very silly mistakes which made what is fairly straightforward code a lot harder to get working than it should have been but that's the nature of the beast.

It grew into about 1000 lines w/ comments, but then again GNU libc malloc is 5000 in just the main file.

I'm kinda bummed out about the performance after all that though, which is leading down another hole ...

ez-alloc

So basically I started by taking a few ideas from GNU libc malloc:

Plus some other ideas and changes.

Before I got this working I decided to drop the per-size small lists. It was cluttering the code and I wanted to see how it worked without it.

I also added a huge allocator which uses a separate mmap for very large allocations, which is also what GNU libc does.

For accuracy testing I created a LD_PRELOAD module which replaced the memory allocation calls with my own and ran some binaries like ls or xterm. I can also use it to create a log of memory actions which I can replay so I can actually debug the thing.

For performance and overhead testing I have a small not-particularly representative test using libavl (it just allocates 1 larger block then thousands of 32-byte blocks). I read /usr/share/dict/words into an array. Then in order I insert them all into the tree, look them all up, then remove them. And another also not-particularly representative loop which frees and allocates random amounts.

Things that work well

I think the tiny allocator quite neat. It uses page alignment and a magic number to be able to allow ez_free() to quickly check whether an allocation is a tiny allocation and handle it appropriately.

It helps avoid fragmentation in the main memory and can save a lot of memory if there are many allocations of the same size.

It would be a big problem if a pointer was mis-identified as a tiny address when it wasn't. The possibility of this is not zero without additional checks, although very very unlikely.

Using ez-list and ez-tree makes the code relatively straightforward and easy to read. Having the same node pointers being used for either is a big plus - I can move nodes between trees or lists easily. List nodes can be removed without knowing the list header so I can have many lists but not care where they come from when performing a coalsecing (I use a single status bit to indicate wheather the block is in a list or the one tree). This makes the coalescing function 'just work' without having to worry about maintaining other state as mergable blocks are merged.

Things I really fucked up

Probably the worst was that I forgot to mask out the status bits of the size in the tree compare function. So ... nodes would go missing when I fiddled with the bits! These bits don't get changed very often while the node is in the tree so that lead to difficult to track down problems.

Idiot. Idiot.

gcc and weirdness

The C99+ aliasing rules broke certain sequences of ez-list calls as well. For example if you perform node.remove on the head of the list and then list.addhead, it would write the pointers back in the wrong order. Another subtle and hard to find bug and gcc didn't warn about it. I found a workaround using a soft barrier asm volatile ("" : : : "memory"); in node.remove but I don't really like it as I can't be sure other sequences might break. Using -fno-strict-aliasing works too of course. Maybe i'll just make the list and node pointers volatile, I don't think it will affect performance negatively doing this and the code pretty much requires them to behave that way.

I've had some other very strange behaviour with gcc as well, sometimes the optimiser makes radically different decisions for no obvious reason. It appears to do it even when I haven't made any code changes, but I don't think my memory is reliable to be absolutely sure on that.

Another waste of time?

The xterm test was quite interesting. With the tiny allocator enabled it quickly reaches a steady state with 21 pages of tiny allocations having over 50% occupancy. There are only 6 free large blocks, and 13 lost ones.

However pmap indicates more total memory is consumed vs glibc malloc which was a bit of a bummer.

For the performance tests I fix my CPU at a specific frequency to get more consistent and repeatable results.

The tree test runs about 10% faster than system malloc and saves roughly 1MB of maximum rss (/usr/bin/time), about 8.5MB vs 9.5MB. At least that's a winner. Actually if I remove the lookup loop so that the alloc code is more of the total time then the performance nearly equalises. So most of the performance improvement is down to allocation alignment and not using a tiny allocator. Which is quite useful all the same.

However the random test runs about 2-3x slower than system malloc and uses about the same total amount of memory. That was a real bummer and every time I compile the code it seems to get a bit slower (and use more peak memory to boot). But ... this is absolutely nothing like the allocation used in a real program such as xterm so I don't even know if it's worth worrying about. I should try with something more demanding but they likely already use pools or the like.

I suppose even if it is a bit slower it can be used as a low-overhead free-supporting pool allocator which might be useful.

I did some shitty profiling and found that the biggest cost was probably the tree manipulation during ez_alloc(). There are some simplish things I haven't tried that should help it.

For example when i split a free block I always remove the node and then re-insert it if there's any of it left (either the lost list or back into the tree). If I can determine that the new node has not changed ordering then it is possible to just replace the entry directly and save both the remove and insert.

When I coalesce nodes I find full consecutive runs and remove all the ones i've found as I go. Then insert the coalesced one back. It should be possible to do the same check here and save one of the removes and the insert in some cases but it is a bit more difficult to determine (maybe just find the largest node that belongs to the tree and try that?).

All non-tiny allocs trigger a coalesce and then search the free block tree. GNU malloc can sometimes use the most recently freed block here, which in this case would avoid the tree altogether.

Finally, I could always add back the specific-sized 'small' lists for recently freed blocks but I want to avoid that for various reasons.

I can't see these doubling the speed for this silly random test-case but I could be wrong.

I suppose I could also try other indexing structures as an alternative to the balanced tree such as a skip-list (really, a skip-index-array I guess). I've not used them before but I had a break yesterday and today it seems like it shouldn't be too difficult.

Update: 3/2/2019 I tried the skip-index idea. And lo, once I got it working the random test is now on par with GNU libc. I guess I will go with that then. I suppose then all i've really done is re-implement GNU libc malloc in the end; although the tiny allocator might benefit some usages. And well - the code is 1/5 the size and considerably more readable. And i'll wrap it up in a way that can be used as a pool allocator. I'm back to work this week so i'll have to give it a break for a while anyway; i'm also exhausted and need to get away from it. Spent some time in the garden today although I discovered the 47C head might've killed more roses out the front. Sigh.

Profiling

I haven't really looked into C profiling for years. Well apart from /usr/bin/time or gettimeofday() (although I just found clock_gettime()).

I suppose people use oprofile? I should at least look into it.

gprof is pretty much useless. Well I guess it did identify the tree manipulation as the major cost but looking at the code would tell you that anyway. It just doesn't sample often enough (i'm getting well under 1% sampling of running time).

Searches tend to turn up advice to use cachegrind, ... but I can't see anything resembling cycle counts in any of it's output. Ubuntu wont let me install kcachegrind because of some dependency issue (probably because i've blacklisted some snot like fucking-shit-audio) but I can't see how that can make up numbers which aren't there already

So ... I started working on my own as well. Yeah like that's a good idea! Ok, maybe it's really time to give it a break. It's really not something I need and i'm not sure -finstrument-functions is going to do a good enough job. I'll see i'll see.

I did snarf some more code out of ezesdk that I've added to libeze. I've got a very small ez-elf that lets you do basic but useful things like lookup function names.

That music player thing

So this has been a pretty extended foray well beyond the code I needed to implement a music player! I haven't touched that for weeks - on the other hand I have just left it running for those weeks and it's still going fine. The remote works well. I've got some tools to send commands around and it all just works.

The last time i mentioned it I was wondering how to do a menu but that night I came up with the obvious answer for a device with no display. Just make it talk!

So I poked around with espeak-ng for a while and fiddled with the player code so I can inject synthesised signal into the audio stream.

I got as far as implementing a say-this-song function. When you press the lips button it says the title and artist of the current song if they're available or just the filename if they're not. As I saw mention in a mailing list; it's quite a clear voice but unintentially quite hilarious at times! It also speaks too fast and doesn't pause anywhere nearly long enough for full stops but at least the former can be adjusted.

I then hit a bit of a design snag. The keyboard monitor takes the keys pressed and turns them into high level player commands. For example the cursor keys are fast-forward/rewind/next-track/previous-track. But during a menu these would change behaviour and be for navigating the menu.

So I wasn't sure where I should put the menu driver logic. If it was in the player then the keyboard module would just send through raw keypresses for the other keys and the player would behave differently depending on what's going on.

If it's in the keyboard module then it could route messages around to potentially other components and the player would implement higher level functions like 'enter menu mode', 'say this', 'preview that', 'rename playlist' and so on.

My initial ideas were a but muddled and the second case seemed a bit messy. But now I think about it again it makes a lot more sense and I can leave the player to playing music and it only really needs to maintain a 'playing' or 'not playing' state and do some low level stream pause management to make sure it all works with alsa.

Regardless, I do have a goal here and once I add a basic menu that would basically be a working player but i'm not sure when i'll get back to it.

Tagged hacking, libeze, playerz.
Newer Posts | Older Posts
Copyright (C) 2019 Michael Zucchi, All Rights Reserved. Powered by gcc & me!