Re: Pil vs Pil21 performances
hi all, September 5, 2020 12:28 AM, "Davide BERTOLOTTO" wrote: > Ok, I managed to turn on the optimization (opt -O3) and the results improved > > *** recursive > > -- picolisp 3.007 sec > > -- pil21 4.249 sec > > *** iterative > > -- picolisp 0.515 sec > > -- pil21 0.368 sec > > Apparently with the optimization cache does not segfault... 1. There are several unofficial mirrors to get sources: daily: https://github.com/picolisp/pil21 hourly: https://git.envs.net/mpech/pil21 https://hub.darcs.net/tankf33der/pil21 2. There is unofficial test suite: https://git.envs.net/mpech/pil21-tests Eventually running on: Arch, Void, Alpine, Centos, Oracle Linux, Fedora, Debian. 3. This is required steps to compile on Solaris and Macos, please try to repeat on others OS. LLVM7+ required plus ecosystem. https://git.envs.net/mpech/pil21-tests/src/branch/master/INSTALL-.md 4. for running pil21 I copied pil to pil21 script, and have easy separation between implementations: [mpech@lambda pil21]$ ls -l pil* -rwxr-xr-x 1 mpech mpech 62 aug 30 22:47 pil -rwxr-xr-x 1 mpech mpech 62 aug 30 22:47 pil21 $ 5. p.s. "ulimit -s 16000" is good start point for pil21. p.s.s. To make some noise and fun try to run your picolisp's code on pil21. p.s.s.s. Comparing LLVM and Assembler backend is not fair, and pil21 will be always slower in 5-15% range. (mike) -- UNSUBSCRIBE: mailto:picolisp@software-lab.de?subject=Unsubscribe
Re: Pil vs Pil21 performances
On Fri, Sep 04, 2020 at 11:17:44PM +0200, Davide BERTOLOTTO wrote: > Ok, I managed to turn on the optimization (opt -O3) and the results improved > > *** recursive > -- picolisp 3.007 sec > -- pil21 4.249 sec > > *** iterative > -- picolisp 0.515 sec > -- pil21 0.368 sec > > Apparently with the optimization cache does not segfault... Good to know. Perhaps the optimizer also improves stack usage somehow? ☺/ A!ex -- UNSUBSCRIBE: mailto:picolisp@software-lab.de?subject=Unsubscribe
Re: Pil vs Pil21 performances
Hi Davide, > I compiled pil21 with llvm on termux and then I ran some simple fibonacci > benchmark. Here the results > > *** recursive > n 37 fib 24157817 > -- picolisp 3.013 sec > -- pil21 5.332 sec > -- python 21.247 sec > > *** iterative > n 8 fib 4589178984+ > -- picolisp 0.518 sec > -- pil21 0.409 sec > -- python 0.124 sec Thanks for testing! We found similar relations too. > I got also a segmentation fault while trying to use cache to speed up the > calculation with pil21 (don't know if Alex is aware of that) Yes. The reason is that pil21 needs more stack space, due to a more limited control of the stack (as opposed to asm in pil64). So it gives a stack overflow. Both pil64 and pil21 do runtime stack checks only if one ore more coroutines are active. Otherwise they delegate the check to the hardware, resulting a segmentation fault. You can run it the test you set a higher stack limit, e.g. in .bashrc with ulimit -s 262144 # 256 MiB ☺/ A!ex -- UNSUBSCRIBE: mailto:picolisp@software-lab.de?subject=Unsubscribe
Re: Pil vs Pil21 performances
Ok, I managed to turn on the optimization (opt -O3) and the results improved *** recursive -- picolisp 3.007 sec -- pil21 4.249 sec *** iterative -- picolisp 0.515 sec -- pil21 0.368 sec Apparently with the optimization cache does not segfault... Cheers, Davide On Fri, Sep 4, 2020, 22:50 Davide BERTOLOTTO wrote: > Hi all, > > I compiled pil21 with llvm on termux and then I ran some simple fibonacci > benchmark. Here the results > > *** recursive > n 37 fib 24157817 > -- picolisp 3.013 sec > -- pil21 5.332 sec > -- python 21.247 sec > > *** iterative > n 8 fib 4589178984+ > -- picolisp 0.518 sec > -- pil21 0.409 sec > -- python 0.124 sec > > I got also a segmentation fault while trying to use cache to speed up the > calculation with pil21 (don't know if Alex is aware of that) > > From the results it seems that performance of llvm are slightly better for > iterative solutions and fairly worse for recursive solutions > > Question: I used the makefile in the pil21.tgz to compile pil21. Is there > any optimization in? I didn't dig (yet) into llvm compiler flags :) > > Regards, > Davide >
Pil vs Pil21 performances
Hi all, I compiled pil21 with llvm on termux and then I ran some simple fibonacci benchmark. Here the results *** recursive n 37 fib 24157817 -- picolisp 3.013 sec -- pil21 5.332 sec -- python 21.247 sec *** iterative n 8 fib 4589178984+ -- picolisp 0.518 sec -- pil21 0.409 sec -- python 0.124 sec I got also a segmentation fault while trying to use cache to speed up the calculation with pil21 (don't know if Alex is aware of that) >From the results it seems that performance of llvm are slightly better for iterative solutions and fairly worse for recursive solutions Question: I used the makefile in the pil21.tgz to compile pil21. Is there any optimization in? I didn't dig (yet) into llvm compiler flags :) Regards, Davide
