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

[Jerry DeLisle]

Hi all,

Attached patch implements a fast blocked matrix multiply. The basic algorithm is 
derived from netlib.org tuned blas dgemm. See matmul.m4 for reference.

The matmul() function is compiled with -Ofast -funroll-loops. This can be 
customized further if there is an undesired optimization being used. This is 
accomplished using #pragma optimize ( string ).

My results on 3.8 GHz machine with single core:

$ gfc -Ofast -funroll-loops -finline-matmul-limit=32 compare.f90
$ ./a.out
  =========================================================
  ================            MEASURED GIGAFLOPS          =
  =========================================================
                  Matmul                           Matmul
                  fixed                 Matmul     variable
  Size  Loops     explicit   refMatmul  assumed    explicit
  =========================================================
     2  2000     23.810      0.058      0.116      0.191
     4  2000      1.979      0.294      0.437      0.421
     8  2000      3.089      0.826      0.928      0.993
    16  2000      4.115      3.262      2.600      3.381
    32  2000      6.066      5.201      3.008      4.873
    64  2000      6.596      4.847      6.624      6.603
   128  2000      8.389      5.965      8.370      8.375
   256   477      9.520      6.003      9.449      9.452
   512    59      8.563      2.783      8.359      8.500
  1024     7      8.672      1.537      8.457      8.604
  2048     1      8.586      1.753      8.371      8.511

Results may vary, but I found 32 is the right place to set the limit on inlining.

Regression tested on x86-64-linux power8 (gcc112)

Special thanks to Thomas for helping me test and debug. An additional test case 
as well.

OK for trunk?

Best regards,

Jerry


2016-11-09  Jerry DeLisle  <jvdelisle@gcc.gnu.org>
	    Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR libgfortran/51119
	* m4/matmul.m4: For the case of all strides = 1, implement a
	fast blocked matrix multiply. Fix some whitespace.
	* generated/matmul_c10.c: Regenerate.
	* generated/matmul_c16.c: Regenerate.
	* generated/matmul_c4.c: Regenerate.
	* generated/matmul_c8.c: Regenerate.
	* generated/matmul_i1.c: Regenerate.
	* generated/matmul_i16.c: Regenerate.
	* generated/matmul_i2.c: Regenerate.
	* generated/matmul_i4.c: Regenerate.
	* generated/matmul_i8.c: Regenerate.
	* generated/matmul_r10.c: Regenerate.
	* generated/matmul_r16.c: Regenerate.
	* generated/matmul_r4.c: Regenerate.
	* generated/matmul_r8.c: Regenerate.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: pr51119.diff
Type: text/x-patch
Size: 220149 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20161114/33129268/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: matmul_12.f90
Type: text/x-fortran
Size: 625 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20161114/33129268/attachment-0001.bin>