possibly serious bug affecting msieve 1.48 & 1.49

[jrk]

I should warn of a bug I introduced in the GPU version of polynomial selection in msieve 1.48 and 1.49 that might affect some of you.

The problem: If the sieve size used exceeds 127 bits, a problem occurs that results in every possible stage1 polynomial being recorded. If you invoke msieve with -np1, or -np with redirected output, the output files will fill up the hard disk.

The stock builds should be fine to use. However, if the internal parameters in poly_skew.c are changed to accommodate larger input numbers and a large stage1 norm or large leading coefficient is used for polynomial selection, you might encounter this problem.

The cause of this problem is an errant conversion of a double-precision floating point value to single-precision in the GPU code. This was necessary since use of double-precision is not always available in CUDA. However, I forgot to consider the case where the exponent in the double-precision value is too large to fit in single-precision, in which case the value is converted to infinity.

I fixed this a while ago on trunk, by always forcing the sieve size to be small enough to fit, but didn't realize until now what effect the original bug would have for the user. So I thought I should mention it here since filling up disks is not a nice problem to have.

The upcoming version 1.50 and the current branch of new GPU code do not have this problem.