Hi all,<br><br> I've been studying and twisting the tuneup program in GMP4.2.4 recently. And I got a question about the fft_step_size().<br><br>mp_size_t<br>fft_step_size (int k)<br>
{<br> mp_size_t step;<br><br> <b>step = MAX ((mp_size_t) 1 << (k-1), BITS_PER_MP_LIMB) / BITS_PER_MP_LIMB; //what's the special purpose of this line?</b><br><b> step *= (mp_size_t) 1 << k; //</b><b> Is it not good enough just to set step to 2^k? </b><br>
...<br> return step;<br>}<br><br>In
addition, I noticed that if I half the step size in fft_step_size(), I
can gain better performance around the point where Toom3 is replaced by
SSA. A plot is provided in attachment, where you can see the threshold
generated by default tuning program (green line) causes a loss since it
sets the fft threshold greater than what it actually is. The one
labeled 'tune2' is what I got by halving the step size. All the values
in labels correspond to FFT thresholds in each case. <br clear="all"><br>-- <br>Best regards,<br><br>Lingchuan Meng<br>