Optimisation logbook

A list of optimisation decisions and their effect on performances.

2011/09/02

• [885] [886] [888] Started importing Ryan Sammartino's PS3 trigonometry library in the core ().
• [890] Do the Taylor series on x / π instead of x: results in a measurable speed increase in sin()
  • 2% with ppu-lv2-g++ 4.1.1 on a PS3
  • almost 1% with SNC 360.1 on a PS3
  • 30% with g++ 4.5.2 on mingw32 running on a Core2Quad
• [891] Do the Taylor series in two steps, for better pipelining at the cost of one extra multiplication: a notable speed increase in sin()
  • 11% with ppu-lv2-g++ 4.1.1 on a PS3
  • 2% with SNC 360.1 on a PS3
  • 7% with g++ 4.5.2 on mingw32 running on a Core2Quad

Note: there is currently no way to build all the engine using SNC. However, individual files can be built using a command along the following one:

touch trig.cpp
ps3ppusnc.exe -c trig.cpp -O3 -Xreorder=1 -Xfastmath=1 -o liblol_a-trig.o -I.
make

2011/10/11

• [1016] Encode reals using uint32_t instead of uint16_t.
  • additions are 2.1 times faster
  • multiplications are 2.4 times faster
  • divisions are 2.4 times faster (actually fixed in [1017])

2011/10/12

• [1020] Replace divisions with multiplications in Taylor series expansions: instead of x+=a/b we use x=bx+a and divide the end result by b
  • sin() is 3.8 times faster
  • exp() is 4.1 times faster