davbjor-simd

A benchmarking test for davbjor-raytracer to benchmark the usage of SIMD instructions vs single threaded cpu instructions.

Result

Through benchmarking with C++ standard library , testing the time for the pathtracer to run, the program failed to produce a faster version using SIMD than using compiler optimized version of standard C++.

Example of path tracing a 1200px wide image with a depth of 6.

Without optimization   |  SIMD optimization
Iterations:
0/15
time: 1186 ms          | time: 1696 ms
1/15
time: 1199 ms          | time: 1702 ms
2/15
time: 1196 ms          | time: 1769 ms
3/15
time: 1197 ms          | time: 1770 ms
4/15
time: 1207 ms          | time: 1924 ms
5/15
time: 1196 ms          | time: 1895 ms
6/15
time: 1212 ms          | time: 2040 ms
7/15
time: 1210 ms          | time: 1847 ms
8/15 
time: 1202 ms          | time: 1741 ms
9/15 
time: 1197 ms          | time: 1830 ms
10/15
time: 1224 ms          | time: 1959 ms
11/15
time: 1223 ms          | time: 1797 ms
12/15
time: 1255 ms          | time: 1763 ms
13/15
time: 1224 ms          | time: 1746 ms
14/15
time: 1199 ms          | time: 1765 ms


Total
time: 18132 ms         | time: 27251ms

Summary:

SIMD is ca: 50% slower

Example of SIMD code:

An example of the code changed, in this case vector dot-product:

double dot(const vec3 &u, const vec3 &v){
    double r = 0;
    r += u.e[0] * v.e[0];
    r += u.e[1] * v.e[1];
    r += u.e[2] * v.e[2];
    return r;
}

Is replaced for a SIMD function:

float simd_dot(vec4 a, vec4 b){
    const __m128 xmm_a = _mm_load_ps(&a.x);
    const __m128 xmm_b = _mm_load_ps(&b.x);
    const __m128 xmm_res = _mm_dp_ps(xmm_a, xmm_b, 0xff);
    return _mm_cvtss_f32(xmm_res);
}

Lessons

Simply doing SIMD functions instead of standard arithmetic operations is slower.

To create a faster result, you need to utilize the parallelization in SIMD, for example tracing 4 or 8 paths at the same time using SIMD instructions.