[patch,libgfortran] PR51119 - MATMUL slow for large matrices

[Jerry DeLisle]

On 11/14/2016 11:22 PM, Thomas Koenig wrote:
> Hi Jerry,
>
>> With these changes, OK for trunk?
>
> Just going over this with a fine comb...
>
> One thing just struck me:   The loop variables should be index_type, so
>
>       const index_type m = xcount, n = ycount, k = count;
>
> [...]
>
>    index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset, i1, i2,
>           i3, i4, i5, i6;
>
>       /* Local variables */
>       GFC_REAL_4 t1[65536], /* was [256][256] */
>          f11, f12, f21, f22, f31, f32, f41, f42,
>          f13, f14, f23, f24, f33, f34, f43, f44;
>       index_type i, j, l, ii, jj, ll;
>       index_type isec, jsec, lsec, uisec, ujsec, ulsec;
>
> I agree that we should do the tuning of the inline limit
> separately.
>

Several of my iterations used index_type. I found using integer gives better 
performance. The reason is that they are of type ptr_diff_t which is a 64 bit 
integer. I suspect we eliminate one memory fetch for each of these and reduce 
the register loading by reducing the number of registers needed, two for one 
situation. I will change back and retest.

and Paul commeneted "-ftree-vectorize turns on -ftree-loop-vectorize and
-ftree-slp-vectorize already."

I will remove those to options and keep -ftree-vectorize

I will report back my findings.

Thanks, and a fine tooth comb is a very good thing.

Jerry