In C or assembler, (or in Perl using pack) you can trivially reduce the size of the tables to 4MB instead of 40MB.

With a little ingenuity, you could reduce that to 1MB by expoiting the fact that each 32-bit value has 12 'spare' bits in the simple packed encoding.

Ie. For each of the 256k 32-bit values in a 1MB table, you reuse each 4 times:

...
   0    1    2    3    4    5    6    7
0010 1100 1100 1001 1111 0000 0110 1101 
...
[download]

Say:

1. input value 0 would be encoded as nybbles 0,1,2,3,4 of the 32-bit value at offset 0.
2. input value 256k as nybbles 1, 2, 3, 5, 7.
3. input value 512k as nybbles 0, 2, 4, 6, 7.
4. input value 784k as nybbles 1, 3, 4, 5, 6.

If you're really up against it you could get it down to a bare 73k by expoiting that there are 56 combinations of 5 nybbles in each 32-bit word.