Injection list kernel instruction value scales linearly with injection time

[TylerMG26]

I've come to notice that when injecting faults there seems to be some correlation to the value of the kernel instruction and injection run latency.

For example, if the injection list is listed per injection site as:
kernel name/parameters, kernel index, kernel instruction, op seed, bit seed.

The injection between the following two injection is about 2x on the first one compare to the bottom one.
matmul_kernel 2 500000 0.234123 0.843278
matmul_kernel 2 250000 0.234123 0.843278

I'm curious if there is a reason this happens? I've dug through a lot of the code and can't figure out why. It seems odd that the number makes a difference when I assume the binary is only instrumented in one location in both cases.

[TylerMG26]

Upon further investigation, I've noticed that the thread ID (tid) listed in nvbitfi-injection-log-temp.txt for injections where the kernel instruction value is higher will have a higher thread ID and lower values have lower thread IDs. However, this does not make any sense to me. I would expect the execution time to be roughly the same if injecting into the same kernel. This behavior makes it difficult to instrument large applications.

[fernandoFernandeSantos]

I think this happens because NVBit has to instrument all the possible instructions in the kernel and then skip the ones that are not the target of the injection.