We discovered a strange phenomenon in some code for solving tridiagonal
system.
The following code highlights the problem:
#include <stdlib.h>
int main(int argc, char** argv) {
int i, j;
float temp;
const int N = 1024;
const int Nruns = 100000;
float * x = (float * ) malloc(N*sizeof(float));
for( j = 0; j < Nruns; ++j ) {
for( i = 0; i < N; ++i ) x[i] = i+1;
// temp = x[N-1]; // Variant 2
for( i = N-2; i >= 0; --i ) {
// temp = x[i] = 1.0f + 2.0f*temp; // Variant 2
x[i] = 1.0f + 2.0f*x[i+1]; // Variant 1
}
}
}
Look at the timings:
gcc -O2 test.c
time ./a.out
real 0m14.347s
user 0m14.301s
sys 0m0.002s
After commenting out Variant 1 and activating both lines from Variant 2:
gcc -O2 test.c
time ./a.out
real 0m11.541s
user 0m11.466s
sys 0m0.008s
Variant 1 again:
gcc -O2 -msse -mfpmath=sse test.c
time ./a.out
real 0m0.676s
user 0m0.672s
sys 0m0.002s
Variant 2:
gcc -O2 -msse -mfpmath=sse test.c
time ./a.out
real 0m0.426s
user 0m0.424s
sys 0m0.001s
Why does the introduction of the temporary variable give a performance
increase of 20-30% ?
Why is the SSE version so much faster in this case?
I thought that GCC >4 uses SSE per default for FP math?
