Hallo,
I've compared three ghdl setups
- ghdl 0.33 with llvm backend (from Pete Gavins .deb package)
- ghdl 0.33 with gcc backend (compiled from SourceForge source kit)
- ghdl 0.34dev with gcc backend (compiled from git master)
with two benchmark cases
- behavioral simulation (766 processes, many quite complex)
- post-synthesis functional simulation (6090 processes, mostly quite slim)
the code was generated by vivado with 'write_vhdl'
and I tried different optimize levels
-O0 (is default for gcc!)
-Og (does some optimizations which don't interfere with debugging)
-O1
-O2 (is default for llvm)
For the behavioral simulation I got:
compile time simulation time
real/ user real/ user/ sys
ghdl-0.33-llvm -O0 0m10.910s/0m09.635s 1m48.917s/1m49.576s/0m0.686s
ghdl-0.33-llvm -O2 0m18.319s/0m17.164s 1m17.047s/1m15.265s/0m0.720s
ghdl-0.33-gcc -O0 0m19.614s/0m17.161s 1m15.832s/1m16.470s/0m0.676s
ghdl-0.33-gcc -Og 0m23.892s/0m21.514s 0m51.699s/0m52.320s/0m0.669s
ghdl-0.33-gcc -O2 0m43.271s/0m40.674s 0m51.209s/0m49.033s/0m0.624s
ghdl-0.33-gcc -O3 0m57.981s/0m55.133s 0m47.617s/0m48.107s/0m0.696s
ghdl-0.34dev-gcc -O0 0m19.777s/0m16.965s 1m00.541s/1m01.267s/0m0.587s
ghdl-0.34dev-gcc -Og 0m23.839s/0m20.713s 0m44.975s/0m45.450s/0m0.805s
ghdl-0.34dev-gcc -O1 0m28.972s/0m26.259s 0m42.085s/0m42.782s/0m0.601s
ghdl-0.34dev-gcc -O2 0m40.997s/0m38.160s 0m42.661s/0m43.340s/0m0.618s
ghdl-0.34dev-gcc -O3 0m53.743s/0m50.605s 0m42.550s/0m43.156s/0m0.684s
For the post-synthesis functional simulation I got (the UNISIM lib was
compiled with -O2):
compile time simulation time
real/ user real/ user/ sys
ghdl-0.34dev-gcc -O0 0m31.871s/0m30.831s 0m58.664s/0m58.636s/0m0.055s
ghdl-0.34dev-gcc -Og 0m38.098s/0m37.066s 0m58.109s/0m58.078s/0m0.063s
ghdl-0.34dev-gcc -O1 0m45.404s/0m44.433s 0m58.373s/0m58.330s/0m0.058s
ghdl-0.34dev-gcc -O2 1m06.171s/1m04.976s 0m58.541s/0m58.514s/0m0.035s
and finally the dependence on UNISIM library optimization level (the model
was always compiled with -Og):
compile time simulation time
real/ user real/ user/ sys
UNISIM -O0 0m38.079s/0m37.075s 1m11.720s/1m11.679s/0m0.050s
UNISIM -O1 0m38.095s/0m37.198s 0m58.729s/0m58.692s/0m0.050s
UNISIM -O2 0m38.153s/0m37.152s 0m58.451s/0m58.431s/0m0.034s
UNISIM -O3 0m38.665s/0m37.625s 0m58.388s/0m58.358s/0m0.057s
From this, certainly very slim base, I observe
1. gcc -O0 and llvm -O2 give similar compile and execution speed
2. gcc -O2 gives higher compile time, but also better execution speed
3. gcc -O2 clearly outperforms llvm -O2
4. gcc -Og seems a good compromise between compile and execution speed
5. for gate level models impact of optimization level is quite low
Is that in line with what others have seen ?
With best regards, Walter
_______________________________________________
Ghdl-discuss mailing list
Ghdl-discuss@gna.org
https://mail.gna.org/listinfo/ghdl-discuss
