PathScale Is Looking To Hire More Compiler Engineers

**haagch** · 12 February 2016, 06:10 AM

That reminded me to put the ekopath-bin package back in the new AUR.

Looks like they still have work to do:

Code:

$ g++ -O3 -fopenmp smallpt.cpp -o smallpt-gcc
$ pathCC -O3 -fopenmp smallpt.cpp -o smallpt-pathCC
smallpt.cpp:55:48: warning: add explicit braces to avoid dangling else [-Wdangling-else]
  if (++depth>5) if (erand48(Xi)<p) f=f*(1/p); else return obj.e; //R.R.
                                               ^
1 warning generated.
$ ./smallpt-gcc
Rendering (4 spp) 100.00%./smallpt-gcc  12.27s user 0.07s system 665% cpu 1.853 total
$ ./smallpt-pathCC
Rendering (4 spp) 14.08%[1]    18744 segmentation fault (core dumped)  ./smallpt-pathCC

Looks like we are super deep in the stack there when it happens: http://haagch.frickel.club/files/gdb-smallpt-pathCC.txt

As usual in these cases, testing OMP_NUM_THREADS=1 ./smallpt-pathCC is the first thing to do and it indeed runs fine, so it's an openmp issue...

But at least that speed goal is reached:

Code:

$ g++ -march=native -O3 -fopenmp smallpt.cpp -o smallpt-gcc
$ pathCC -march=ivy -O3 -fopenmp smallpt.cpp -o smallpt-pathCC
smallpt.cpp:55:48: warning: add explicit braces to avoid dangling else [-Wdangling-else]
  if (++depth>5) if (erand48(Xi)<p) f=f*(1/p); else return obj.e; //R.R.
                                               ^
1 warning generated.
$ OMP_NUM_THREADS=1 ./smallpt-pathCC
Rendering (4 spp) 100.00%OMP_NUM_THREADS=1 ./smallpt-pathCC  6.35s user 0.04s system 99% cpu 6.417 total
$ OMP_NUM_THREADS=1 ./smallpt-gcc
Rendering (4 spp) 100.00%OMP_NUM_THREADS=1 ./smallpt-gcc  7.79s user 0.06s system 99% cpu 7.874 total

By the way: What happened to path64? The github repository is gone and openhub says it didn't receive any commits for over three years.

PathScale Is Looking To Hire More Compiler Engineers

PathScale Is Looking To Hire More Compiler Engineers

Comment