search

jasonlarkin / p3dfft

Automatically exported from code.google.com/p/p3dfft
GNU General Public License v3.0
0 stars 0 forks source link

Stride1 single-precision PGI crashes on Kraken #52

Closed GoogleCodeExporter closed 9 years ago

GoogleCodeExporter commented 9 years ago 
Stride1 single-precision PGI crashes on Kraken, but runs fine on Ranger.
2048^3, 1024 cores.

Original issue reported on code.google.com by dmitry...@gmail.com on 10 Aug 2010 at 1:50

GoogleCodeExporter commented 9 years ago 
Program seg-faults.
FFTW version 3.2.1, PGI versions 10.5 and 9.0.4

Original comment by dmitry...@gmail.com on 10 Aug 2010 at 4:48

GoogleCodeExporter commented 9 years ago 
It appears that single precision inverse tests with small/normal sized yields 
incorrect results

Single precision
 (128 128 128) grid
 2 proc. dimensions
1 repetitions
Reading proc. grid from file dims
Using processor grid 1 x 4
Iteration 0
Results of inverse transform: 
(128,3,4) -524287.968750
(128,127,4) 524287.937500
(128,3,126) 524287.937500
(128,127,126) -524287.937500
Results are correct
max diff =0.0625
Time per loop=0.01543
timer[1] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[2] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[3] (avg/max/min): 1.002359E-11 4.513025E-03 4.263878E-03
timer[4] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[5] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[6] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[7] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[8] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[9] (avg/max/min): 7.681420E-12 3.546953E-03 3.263950E-03
timer[10] (avg/max/min): 6.844079E-12 3.022909E-03 2.998829E-03
timer[11] (avg/max/min): 0.000000E+00 0.000000E+00 0.000000E+00
timer[12] (avg/max/min): 7.039508E-12 3.147125E-03 3.058910E-03

This is regardless of whether or not it is stride1, even, or neither.

Original comment by dan.djc...@gmail.com on 11 Aug 2010 at 12:15

GoogleCodeExporter commented 9 years ago 
These resutls are correct. Due to single precision 
there is some roundoff error but in this case it 
is small relative to the magnitude of the largest 
numbers.

Original comment by dmitry...@gmail.com on 14 Aug 2010 at 11:30

GoogleCodeExporter commented 9 years ago 
The problem has disappeared on its own. Programs don't crash 
and produce right answers.

Original comment by dmitry...@gmail.com on 31 Aug 2010 at 11:09