I am getting an unusually slow performance in gcc 4.8.2 compared to clang 3.4.2 for the following piece of code:

Code: Select all

#include <Eigen/Geometry>
#include <iostream>

int main(int argc, char *argv[]) {
  unsigned long int acc = 0; 

  for (int z = 0; z < 2000; ++z)
   for (int y = 0; y < 2000; ++y)
     for (int x = 0; x < 2000; ++x) {
       acc += Eigen::Vector3i(1, 1337, 345637).dot(Eigen::Vector3i(x, y, z));
     }

  if (acc != 2774412100000000000ul)
    std::cout << "Incorrect\n"; 
  return 0;
}

Compile Flags: -O3 -DNDEBUG -march=native

Clang 3.4.2 runtime : real 0m0.003s user 0m0.003s sys 0m0.000s
Gcc 4.8.2 runtime : real 0m2.247s user 0m2.248s sys 0m0.000s

Any insight into why there is such a big difference?

This has nothing to do with Eigen, it's simply that in your example the compiler can aggressively remove all the loops. For instance, you can remove the inner loop as follow:

Code: Select all

for (unsigned long z = 0; z < 2000; ++z)
{
  acc += (345637ul*2000ul*2000ul) * z + 1999000ul * 2000ul;
   for (unsigned long y = 0; y < 2000; ++y)
     acc += (1337ul*2000ul) * y;
}


and this version is obviously 2000x faster! gcc 4.8 does not perform this optimization. Of course, you can go even further and remove all the loops but apparently gcc 4.9 does not go that far and only remove the most inner one. Clang is able to go that far and compute 'acc' at compile time, thus leading to the following generate code which is essentially a no-op:

Code: Select all

   .section   __TEXT,__text,regular,pure_instructions
   .globl   _main
   .align   4, 0x90
_main:                                  ## @main
   .cfi_startproc
## BB#0:
   pushq   %rbp
Ltmp2:
   .cfi_def_cfa_offset 16
Ltmp3:
   .cfi_offset %rbp, -16
   movq   %rsp, %rbp
Ltmp4:
   .cfi_def_cfa_register %rbp
   xorl   %eax, %eax
   popq   %rbp
   ret
   .cfi_endproc


.subsections_via_symbols


Thanks ggael.