This vectorised version of Marsaglia's XORSHIFT random number generator (RNG) is implemented using Intel's streaming SIMD extensions (SSE) version 4.1. The RNG generates 32-bit random samples. The width of the vector registers is 128 bit giving a vector length of 4 32-bit integers.

To be written.

For a description of Marsaglia's XORSHIFT RNG, please refer to <> or the documentation of the TinyRNG library .

First iteration of the SSE implementation.

Second iteration of the SSE implementation. Decrese the granularity

Make it more fine grained again

gcc test.c xorshiftSSE.c -o xorshiftSSE -O3 -msse4.1 -I.

== Timing results ==

guido@linux-9cge:~/scratch/xorshiftSSE/remake> ./xorshiftSSE xorshift32 1000000000 samples in 4.162631 sec. xorshift32SSE 1000000000 samples in 1.987162 sec. speedup 2.095 xorshift32SSE unrolled 1000000000 samples in 1.094806 sec. speedup 3.802 xorshift function pointers 1000000000 samples in 1.286171 sec.speedup 3.236

guido@linux-9cge:~/scratch/xorshiftSSE/remake> ./xorshiftSSE xorshift32 1000000000 samples in 3.843894 sec. xorshift32SSE 1000000000 samples in 1.756420 sec. speedup 2.188 xorshift32SSE unrolled 1000000000 samples in 0.918752 sec. speedup 4.184 xorshift function pointers 1000000000 samples in 1.282542 sec. speedup 2.997

guido@linux-9cge:~/scratch/xorshiftSSE/remake> ./xorshiftSSE xorshift32 1000000000 samples in 3.957081 sec. xorshift32SSE 1000000000 samples in 1.987700 sec. speedup 1.991 xorshift32SSE unrolled 1000000000 samples in 1.095842 sec. speedup 3.611 xorshift function pointers 1000000000 samples in 1.285295 sec. speedup 3.079

Average speed-up:

xorshift32SSE 2.091 xorshift32SSE unrolled 3.866 xorshift function pointers 3.104