Open moyix opened 3 years ago
Turns out this was because we weren't checking for kernel launch failures and kernel launch was failing with too many resources requested for launch
. Reducing the number of threads per block fixes this example, but how can we detect the right number of threads dynamically?
On some SMT files, the fuzzer reports success with an all zero input vector. This may be correct (haven't checked), but it should be astronomically unlikely that we ever see an all-zero input with a large input. The fuzzer should be initializing each block to
i
and then AES-encrypting it with a random key before ever runningLLVMFuzzerTestOneInput
, which should (almost) never produce all zeroes. So... what's going on here?Test file: https://clc-gitlab.cs.uiowa.edu:2443/SMT-LIB-benchmarks/QF_FP/-/blob/master/griggio/fmcad12/div.c.30.smt2
Output: