…n OpenCL code.

…pecification.

…code.

Updated maxBpw tables.  Changed -tune to handle new FFT spec code.
Added TABMUL_CHAIN option (it did not deserve to be a WMH code because it has little impact on Z).

…at most 500 kernels, a marker, and 500 more kernels. When prpll tries to add another kernel to the queue, we loop checking if the marker has been reached else performing a lengthy sleep (knowing there are 500 kernels to execute after the marker).

…ktodo files.

Autoprimenet needs this change.

…useful FFT spec could be found in tune.txt

…rence)

…pported by nVidia GPUs.

…ce condition were launched thread accesses uninitialized variables.

…nup work remains.

…mentations that are not faster (at least on TitanV).

…word / little word range. Carryutil sloppy routines may or may not use this feature in the future.

…r more sloppy carries which gives a tiny performance boost for M31+M61 NTTs.

…TitanV. Explored alternate weakMul and csq implementations.

…need to expose the MIDDLE_CHAINMUL option to the end user.

…DE was not set.

… command line argument -smallest

…good size L2 cache.

… an RTX 5080 but 1% slower on a Titan V.

…_grouping.

…HAIN31=1 which is rarely set.

…e I'll figure out a may to use them profitably in the future.

…han unsigned long

…ong is 64 bits.

…t. Now we just need to automatically detect the GPU's actual level of CUDA support.

…ltins.

…e fails if file is open.

… does not support the builtins required for variant zero.

As a poor workaround, this change let's the user specify NO_ASM to bypass tuning FP64 variant zero.

…ve roundoff errors.

…ng on use of fma function on floats).

Some minor changes on wording of tune output.