Two possible things could be causing this.
1. Did you modify the kernel code for the benchmark in any way? This could be the syntax error you're seeing (unlikely, though).
2. Make sure you're using the exact versions of compilers we list in the google doc linked from the wiki. It could be a problem with compiler versions.
2a. Make sure the correct version of the CUDA compiler is in your PATH when you run gem5 (not just build the workload). ptxas is run *at runtime* as well as at compile time.