What conversion did you use for GPU to x86 FLOPS?

For native GPU FLOPS, we tried to count each operation as just one FLOP, so a sqrt is one FLOP and a reciprocal sqrt is two FLOPS, etc., which leads to native GPU FLOP counts of the form:

  • Sqrt = 1
  • Rsqrt = 2
  • Log = 1
  • Exp = 1
  • Acos =1
  • Cos =1
  • Sin =1
  • Dot3 = 5
  • Cross3 = 9

For x86 FLOPS, we followed

http://ai.stanford.edu/~paskin/slam/javadoc/javaslam/util/Flops.html

Which, for example, leads to x86 FLOP counts:

  • Sqrt = 15
  • Rsqrt = 16
  • Log = 20
  • Exp = 20