Rounding error with Intrinsics?

Dear Intel developers,

I'm using Intel Intrinsics with Intel 13.1.3. I have some rounding error by using _mm_mul_ps routine. This is the piece of code:

       __m128 denom_tmp = _mm_setzero_ps();
       __m128  sample_tmp = sample_r1;
       __m128  sample_tmp2;
​        for(j = 0; j < tot_iter; j+=UNROLL_STEP) {
            sample_tmp2 = _mm_mul_ps(*sample_tmp, *sample_tmp);
            denom_tmp = _mm_add_ps(sample_tmp2, denom_tmp);
            sample_tmp++;
        }

sample_r1 is filled before that code. My goal is simply to do a square of each elements of sample_r1.These are the first four values of sample_r1: (-570,911  -236,614   36,6958  27,5522).

After the loop above, the first four results of denom_tmp are: (325940  55986,2  1346,58  759,121).

The correct results done by hand and compared with scalar version is:( 325939,369921   55986,184996  1346,58173764,   759,12372484). So, I have an incredible rounding error. Maybe I'm using the multiplication routine in a wrong manner? Can I increase the precision in some way?

Thanks in advance.