@Elwood

I liked the idea of amidevsupport (exept I think they were wrong wanting to hire low cost developers) so I like this idea ! If I were unemployed, I would like to take such a job.

$6500 is a very high value for french engineers but it is certainly what a company pay for an engineer for one month.

In my opinion, to start a project like that, we must have a big list of things to do (short tasks), so the developer can choose what he is able to do.
We also must trust the engineers or the structure that hires them (and fire them if they fail or doesn't really work).
Last, we must be informed : for each full day worked, the developer could be able to describe shortly what he did, the problems he met, how he sees things for the day after, ...

I remember that at a time in Haiku there was a paid job, Axel found many bugs and reported regularly to explain what he was doing, what he fixed, ...

About what could be done, this is an example of one of my development. I wished to see someone else porting libwmf while I could work on the core of PointRider. Of course, I had to port the library myself and of course it was not so easy. Read the first entry of this blog. If it was done externally, I would expect a serious job, with tests made, bugs or workarounds reported, diffs provided, etc.

@acefnq (and others) : Does it run better than the previous version (available on os4depot) looking at a CPU meter ?

Thanks for the tests.

No, I didn't use ibm_perflibs because in dosbox almost all the time is spent in the core (decoding x86 and doing huge computing) : I tried to reorganize some things to avoid cache misses (access data puting it on the same cache line, prefetch on data), improve branch prediction, add ... but I am not sure the effect is visible. It is only possible to optimize some peripheral parts (what means few percents).

The big improvement would be to write a dynamic core (called recompiler in a previous post) producing PPC code. Do you know why dosbox is so slow on PPC ? Not because PPC is bad, it is even very good (I was impressed : sometimes I thought I had found something to optimize and the compiler already did it !).
It is slow compared to x86 versions because it does not use JIT technology.

I will look again at the code to see how hard it is but nothing is documented and the code is a mess even if at the end the program is very good ...

Elwood : About Hieronymus, I didn't port it to OS4, I created it And yes, again with dosbox, I can confirm it list the same slow functions than oprofile on Linux.

kas1e : About Hugi and x-ray, I tried of course and even on Linux PPC there is a problem with colors. I checked SDL surfaces and their masks, they are good. I thing these programs access hardware in a different way ... I don't know if we could fix that.

Hi all,

I've spent again time trying to optimize dosbox. It is diffult because the core is almost impossible to change. I did some things that can bring improvements but I need people to test it, mostly on Sam on which performance difference will be easier to see.

You can download it here. I will only finalize it and put it on os4depot if it shows improvements.

Thanks !

@Rigo

Thanks ! Released at the same time than the OS update, we can't expect more. You promised, you do !

Just a note about the link in your post : it does not work because of an unwanted dot at the end.

Aaargh !

I see no reasons that will make it fail on Sam ... The only part that is only related to G3/G4 CPUs is the way to count the cache misses but I've just checked and this part is not called if another CPU is detected.

There is one thing I see I could do badly ...

I unfortunately did not test on Sam. I will see what I can do to fix that.

Does it work on Pegasos 2 ?

@ChrisH
- About C++, it does not seem to cause so many problems. Here you have the example of DosBox that is written in C++ (ahem, kind of). The function names are not so nice, though. This is why there is an option in oprofile to "demangle" C++ stuff. I could have a look.
- gcc has a profiler but to use it, it is necessary to compile your program with option -pg and you need gprof. But that way, you can't profile shared libraries ...

@Varthall
- Hum .. I am not sure to understand ... When you run Hieronymus without a big program running, all processes seem to be IDLE (100% or almost in the kernel function HAL_TaskPostTerm. Hieronymus is IDLE except when an exception occur, so it has never happened that the current process was Hieronymus itself.
- A view of the functions that cause cache misses would be great ! But it is not possible at the moment to get the performance monitor interrupts
- If you localize a instruction that has many hits, with its offset, we could get it in the disasm source. Another information I would like to get is the source line but I need to understand and implement the support of the DWARF format (used for debug sections)

@afxgroup
Thanks ! I tested OWB but I had no symbols I tried to see where time is spent when launching the executable, displaying pages, ...

@All : Thanks for wishing me an happy birthday !

And let me know the results of your tests.

I updated Hieronymus yesterday (birthday release !). It is available here on os4depot.

Hieronymus is a statistical profiler, that means it samples periodically what is running, then it decodes information and find in which functions was spent the CPU time.

It will be useful for big programs to know which parts (functions) to optimize, to see if the time is spent in the program itself or in libraries (and which ones).

It is really easy to use ! Just be sure your program has its symbols and run "Hieronymus duration=10" for example.

I improved the output (factorizing by function names) and added the number of cache misses on G3/G4 (this information is harder to use but at least we get it). I also extended the maximum duration time to a much bigger value.

Here is an example of output :



What do we see ?
- DosBox consumes 57% of the CPU time, that confirms what the CPU docky used to show
- Else the kernel is in HAL_TaskPostTerm (I suppose it is related to IDLE)
- In DosBox, heavy functions are : Core_Normal_Run, mem_readb, mem_readw and UpdateOneEiPs (audio), that is the same kind of results that oprofile gives on Linux.
- About cache misses, it would be necessary, to run Hieronymus again and look at the values. By the way, that has less interest until we don't know which functions caused the cache misses ...

Just to mention that I have the same DSI error when I launch OWB 3.22.

And I see nothing related to tabs ...

Thanks for this update, though. And I will update to 3.23 as soon as it will be available.

orgin : I tried it yesterday and it works well as far as I tested. Drag'n'drop is ok.
The only thing that I find confusing is the source and destination toggle ... but with drag'n'drop it becomes useless.

I had a good experience with Filer, there are prefs, no crash and it has advanced features : very good work !

Hans : This is great news ! A big step forward. It is very motivating to see improvements like that even if the development was hard.
I thank you so much ! And I hope one day we will get 3D support.

I've written an example to use a timer (I will add it on gurumed.net) but I have 2 questions regarding what was said in this thread :

Chris : You wrote "ioreq must be allocated with AllocSysObjectTags". What is wrong with CreateIORequest even if I understand using AllocSysObjectTags is more compliant with the updated API of OS4 ?
CreateIORequest is not marked as deprecated in autodocs.

broadblues : You wrote "SetSignal(0L,m_timersigflag)", is it mandatory ? What is the exact behaviour of this call ?

@Rigo

I use Snoopy but I have to say it has a strange behavior sometimes : you want to use the scroller to see what happened at the beginning but some flood can prevent scrolling to the last line. In this case I have to save the log in emergency and quit Snoopy.

Nice to ear that it is developper by a core member. I should suggest him some improvements. When developping, this kind of tool is essential.

@VooDoo

I will do my best. I shew the current version at the Alchimie Show at the beginning of November. I did nothing on it since unfortunately. But I've worked on Sonix (webcam tool) and a new secret tool that will be out soon.

@jackepc

Thank you for this list and for chosing PointRider but its description is wrong : PointRider is a viewer of PowerPoint files.
In the same archive, there is another tool called MeltingPoint that converts PPS files to Hollywood projects.

For information, they will be updated at the beginning of the next year.

Quote:

I second that. For years, I used SDI headers that make hooks really simple to use and the source is readable and it supports all systems.
They perfectly fill my needs.

I just wanted to understand the reason (it seems that my suggestion about 64kb page size is right, due to the implementation with mmap in mind, as Rogue said).

I don't care about few dozens of additionnal kilobytes in executable. Tiny programs will have a size of 100kb instead of 20 or 40 kb, there is no problem. No complaint to address.

Thanks for the explanation.

Optimizing with -O2 is still useful. Even if a section (data section I suppose) is padded up to 64kb, that does not mean the code will execute slower. It is just a waste of room, no more.
And the data cache won't be impacted because the additionnal padding zone will not be used.

But I just still do not understand why this 64kb alignment is required in the binary file ...

If it is a deliberate change in binutils, I don't understand why. I see no reason for a 64kb alignment on sections ... maybe except for usage of 64kb pages on some architectures.
But again, the section could be copied in a 64kb page when loading.

But this is not related to FPU alignment on PPC440.

First I should test ssh on Amiga.

What I need in fact :
- ssh usable in a shell to access remote stations (my Linux server for example)
- a graphical program to manage terminal communication for example to get the serial line (debug, control console of bootloader on a remote board, ...)

I don't know exactly what putty does ...