Why is vectorization so much more effective with floats than with doubles?

Question

I noticed that when vectorizing a loop in a C program, the speedup achieved is much greater when using operands of the type float compared to double operands.

Example:

for (int i = 0; i < N; i++) {
    a[i] += b[i] * c[i];
}    

When a, b and c arrays of size 20,000 each and I repeat this loop 1,000,000:

• Without vectorization it takes around 24 seconds with both floats and doubles

• With auto vectorization (compiling with -O1 -ftree-vectorize) it takes 7 seconds with floats and 21 seconds with doubles

• With OpenMP (#pramga omp simd) it is similar to the above bullet point.

  What could be the reason for this?

Edit: Further information:

1. Processor: Intel Core i7-2677M CPU @ 1.80GHz
2. The surrounding code is nothing but array allocations (using calloc) and a loop where the arrays b and c are filled with constant values.

Answer 1

With many of these operations SIMD (single-instruction, multiple data) instructions come into play. Floats are half the size of doubles and so twice as many of them can be dealt with in a single instruction. I am,however, surprised that the speed is triple using floats rather than simply doubled. I suspect, but don't know for sure, that is due to floats being much easier to operate on - the actual mantissa extractions etc.