If I understand you properly, you are going to test some cryptographical primitives by running them on some plaintext, and then taking the resulting ciphertext and giving it to a randomness test suite; your question is "what plaintext should I use? If I pick a random plaintext, then the test results might reflect the randomness of the plaintext, and no anything the cipher is doing"
In which case, you want to select a maximally nonrandom plaintext. For example, if you're testing CBC-mode encryption, a plaintext that consists of a long string of 0's would work; if you're testing a hash function (and so need a long series of plaintexts), using plaintexts of the form "0", "1", "2", "3", etc (for some simple encoding of those integers, say, in binary form) would work.
That said, unless the cryptographical algorithms you're testing are severely broken, you're not going to find anything. An explicit design goal as that an attacker who can submit plaintexts of his choosing is unable to distinguish the output from randomness; it is quite unlikely that something as mindless as the NIST test suite would manage where a clever cryptanalyst cannot.