Faster benchmarks

[fboemer]

Removes the overhead from due to pausing/resuming in Google benchmarks, which allows removing the iteration count 10. This results in more stable benchmarks and easier control of timings. Note, the EvaluateModSwitch and EvaluateRelin benchmarks are therefore out-of-place. Also avoids using plaintext/ciphertext references, to make benchmarks independent of each other.

Interestingly, this change the performance significantly on some benchmarks. I compiled SEAL with and without HEXL on ICX with clang-10.

E.g., with HEXL=OFF, I'm seeing ~1.4x speedup on CKKS EncodeDouble/DecodeDouble.

With HEXL=ON, I'm seeing significant speedups in CKKS EvaluateMulCt, e.g. on N=8192, (~1.7x), EvaluteMutPt (~2.4x) and EvaluateSquare (~2.1x) speedups compared to before this PR.

I'm not sure why the difference, perhaps reducing the number of cache misses. Anyhow, both PALISADE, Lattigo, and HElib all seem to use the same approach as this PR.

See attached for more details: seal-bench-remove-random.xlsx

[WeiDaiWD]

There could be variance in execution time. We can guarantee a constant flow, meaning no dependency on secret data or just data, but cannot guarantee a constant time, since processor and system interrupts behaves mysteriously.

When the inputs are fixed in the current SEAL, they could fall on the faster or slower side. Then randomizing the inputs can cause big changes. I don't see how this PR will reduce cache misses though, the same values were used repeated before.

[WeiDaiWD]

Why did you change xxx_inplace to xxx? We can also keep both.

[fboemer]

Thanks for the tip; I'll close this PR until I have time to investigate further.