negative loop block sizes

[GoogleCodeExporter]

I think there is, possibly, incorrect integer arithmetic in the code that 
chooses loop sizes (integer overflow?)

Reproduce:
1. Compile with IBM XL ./configure 
--prefix=/soft/apps/p3dfft-2.4/p3dfft-2.4-fftw-3.1.2-double --enable-ibm 
--enable-stride1 --enable-1d --enable-fftw 
--with-fftw=/soft/apps/fftw-3.1.2-double CC="tmpwrap mpixlc_r" CFLAGS="-O3 
-qhot" FC="tmpwrap mpixlf77_r" FCFLAGS="-O3 -qhot"
2. Use specific dimensions (see below)

What is the expected output?
positive loop sizes, correct transform

What do you see instead?
Negative last loop size, the result of the transform is all zeros

What version of the product are you using? On what operating system?
Compiled on Intrepid with IBM XL

3072^3 transforms:
Using processor grid  4  x  3072  Using loop block sizes  1 5 1 5
Using processor grid  128  x  128 Using loop block sizes  1 170 1 7
Using processor grid  8  x  1024  Using loop block sizes  1 10 1 3
Using processor grid  128  x  256 Using loop block sizes  1 170 1 14
Using processor grid  16  x  1024 Using loop block sizes  1 21 1 7
Using processor grid  128  x  512 Using loop block sizes  1 170 1 -28
Using processor grid  32  x  512  Using loop block sizes  1 42 1 7

4096^3 transforms
Using processor grid  4  x  4096   Using loop block sizes  1 4 1 4
Using processor grid  64  x  512   Using loop block sizes  1 64 1 8
Using processor grid  128  x  512  Using loop block sizes  1 128 1 -16
Using processor grid  128  x  1024 Using loop block sizes  1 128 1 1

Original issue reported on code.google.com by andrey.a...@gmail.com on 10 Nov 2011 at 5:55

[GoogleCodeExporter]

Fixed in later versions

Original comment by dmitry...@gmail.com on 16 Mar 2015 at 11:58

[GoogleCodeExporter]

Original comment by dmitry...@gmail.com on 17 Mar 2015 at 12:03

• Changed state: Done