CS257-coursework

Advanced Computer Architecture Coursework - C Code Optimisation

Initial values on Fedora i5-2430m

CS257 Optimisation Coursework

Starting simulation of 1000 stars for 1000 timesteps.

Loop 0 = 0.004863 seconds.

Loop 1 = 16.031893 seconds.

Loop 2 = 0.003379 seconds.

Loop 3 = 0.005438 seconds.

Total = 16.045573 seconds.

GFLOP/s = 1.245203

GB/s = 0.002493

Answer = 86672752.000000

Loop 1 Unrolling by a factor 4 results (1000 1000 0)

Loop 0 = 0.009664 seconds.

Loop 1 = 16.067531 seconds.

Loop 2 = 0.003277 seconds.

Loop 3 = 0.008118 seconds.

Total = 16.088589 seconds.

GFLOP/s = 1.241874

GB/s = 0.002486

Answer = 13772848.000000 (answer is way off, need to take a look at this)

Observation: The timings for the loop did not improve at all, the answer has dramatically changed, GFLOP stays the same.

Using SSE first attempt

__m128 r2inv_v = _mm_div_ps(_mm_set1_ps(1.0f),_mm_sqrt_ps(r2_v)); loop interchanged everything inside the inner loop

Loop 0 = 0.004085 seconds. (loop fission)

Loop 1 = 3.475009 seconds.

Loop 2 = 0.004975 seconds.

Loop 3 = 0.007287 seconds.

Total = 3.491356 seconds.

GFLOP/s = 5.722705 GB/s = 0.011457

Answer = 90729344.000000 (answer is off by 4.57% increase)

---

Attempt at improving locality by using loads in the outer loop, but the speed decreased by at least 0.6s

__m128 r2_v = _mm_set1_ps(eps) + rx_vrx_v + ry_vry_v + rz_v*rz_v; // GNU extension gives answer 84007608.000000 but decreased speed from 3.47 to 3.7/4 and gflop from 5.72 to 5.27/4.9 compared to __m128 r2_v = _mm_mul_ps(rx_v, rx_v) + _mm_mul_ps(ry_v, ry_v) + _mm_mul_ps(rz_v, rz_v) + _mm_set1_ps(eps);

Moving all the lines that are only dependant on j outside the inner loop improved speed and did not affect the answer.

rsqrt + Newton-Raphson has better precision than rsqrt but is a bit slower than it.