SDL2

@kas1e

Indeed!

I've made static builds of everything till now.

@IamSONIC

Quote:

As you wrote the versions of the programs compiled under clib4 are faster, more stable in a way .

Here is an example of how I used to check clib4 , SDL2 + ffmpeg compiled under clib4 under qemu

https://youtu.be/2PPHdHPbqhw


I keep my fingers crossed for this project because it is great.
I've already written about it before, and once again I'd like to thank everyone involved, thank you!!!!.

one of the docker from walkero included an SDK ‘local’ with a very large number of libraries compiled under clib4.
For example SDL2 gl4es etc.
They are compiled under gcc 11.3

Here is an example of the contents of some libraries under clib4

https://clib4pkg.amigasoft.net/ubuntu/pool/main/

I don't know if gcc 11.3 in the latest version supporting clib4 is available natively under AOS4 without hacks.
Maybe the guys will compile if it is not.

@smarkusg

Quote:

Yes, regarding to the SDL2 Benchmark Results it seems to be faster. The point about stability was written by someone else. But till now i have not recognized stability problems.

Quote:

Definitely! Also due to it's Open Source Nature missing functions, problems, etc... can be WAY faster resolved!

Quote:

I've seen these too. These .deb files can be extracted and used everywhere

@IamSONIC

I did a little test today. I am currently trying to check if the soft float for QEMU.
I managed to compile gcc + clib4 + lsoft-fp and the needed libraries for the test.
Surely this will work very badly on real hardware.
In theory it should take some of the pressure off FPU emulation on QEMU. So far the results are poor/comparable....
left sdl2 with newlib right sdl2+clib4+lsoft-fp

The topic is about SDL2 so I am posting these clib4 results

@smarkusg
Quote:On hardware with a FPU: Yes.
But the soft-float SDL2, and all libraries used by it, might be faster on the A1222+ than emulated FPU code.
Additionally -msoft-float compiled code and libraries can be linked with SPE compiled code and libraries (compatible ABI), which isn't the case for the default hard-float code and libraries (ABI incompatible to the SPE one).
If everything is statically linked you could upload both of your sdl2benchmark executables (hard-float newlib and soft-float clib4) somewhere and A1222+ users can test which one is faster.

@joerg

I'm quite sure avoiding all exceptions handler of FPU code, soft floats should be significantly faster than hard floats.

@smarkusg

If you want to see how your "SDL2Benchmark" behaves under real conditions on the A1222 I could run a few tests and compare both versions so you might be able to determine how much faster (hard-float newlib and soft-float clib4) is on hardware without a real FPU.

@joerg
In general, I have seen it like this:
Recompiled once on my computer with gcc + patches from adtools with default settings.
Recompiled the second version by changing ./adtools/gcc-build/Makefile adding options to soft-float
after recompiling on this version I recompiled with the new libsoft-fp compiler with flags under soft-float
then you have to recompile clib4 with flags under soft-float. There will be bugs, but you have to change it to make it compileable like the code under SPE.
You still need to change "fabs" to "soft" code and link with libsoft-fp

I wanted both versions to be gcc done on the same machine.
The compilation under newlin and clib4 were without libraries from the SDK(only the include - *.h part)
zlib and SDL2 were compiled on my machine under their gcc versions and additional flags.
SDL2Benchmark compiled statically. SDL2Benchmark under clib4 requires compilation of clib4.library beforehand.


If I have done something stupid and wrong please let me know.


@Maijestro
I will send you the files when I am at the machine where I have them.

@smarkusg
Quote:No, what you are doing is correct: You need a complete soft float build of clib4 incl. libgcc, libstdc++, etc.

For the A1222+ it should be a SPE build instead of course to get best performance, but soft float code at least works on the A1222+, hard float code doesn't (without emulation).
You can't use current GCC versions for SPE but have to use GCC 6, in later versions SPE support is broken (GCC 8) or was completely removed (GCC 9+), which may be more work than using current versions of GCC.

What others are doing instead, for example in
https://www.amigans.net/modules/newbb/ ... id=149878#forumpost149878
can't work: Mixing a only partial soft float (or SPE) build of clib4 with hard float parts like libgcc.

@smarkusg @joerg

smarkusg seems to have understood exactly how to create complete soft-float builds. I ran the SDL benchmark on my real machine and the results are very impressive in parts.

Most of the values doubled, others even quintupled.

Newlib SDL2 Benchmark

FillRects (mode:none) ---- 130.8 frames/s
FillRects (mode:Blend) ----127.8 frames/s

Clib4Soft-Float SDL2 Benchmark

FillRects (mode:none) ---- 1326.7 frames/s
FillRects (mode:Blend) ----1290.4 frames/s

But I'm not sure exactly which SDL2 function it needs.

SDL2 Benchmark Newlib:



Clib4 Soft-Float SDL Benchmark



I have placed both benchmarks next to each other for a better comparison. All in all impressive, thanks for letting me do this test, it could also be interesting for others.

@Maijestro

If this was done additionally for SPE as described by @joerg the difference would be much bigger to the disadvantage of newlib under A1222.
This test was compiled on GCC-11