Add __cirq_debug__ flag and conditionally disable qid validations in gates and operations

[tanujkhattar]

This is the first PR in a series of PRs to provide cirq users to conditionally disable validations in cirq to boost runtime performance. The proposal for adding a __cirq_debug__ flag is presented in #5988

This PR adds:

1. A cirq.__debug__ flag, which is thread safe and implemented using contextvars. The default value is set to __debug__ global constant in python.
2. Conditions the qid shape validations in Gate.validate_args(), Operation.validate_args and Gate.on_each on the value of cirq.__debug__ flag.

Benchmarks on my laptop show improvement of about ~20%.

$> asv run --python=same --bench XOnAllQubitsCircuit
· Discovering benchmarks
· Running 2 total benchmarks (1 commits * 2 environments * 1 benchmarks)
[  0.00%] ·· Benchmarking existing-py_Users_tanujkhattar_opt_anaconda3_envs_cirq_bin_python-PYTHONOPTIMIZE-O
[ 25.00%] ··· Running (circuit_construction.XOnAllQubitsCircuit.time_circuit_construction--).
[ 50.00%] ··· circuit_construction.XOnAllQubitsCircuit.time_circuit_construction                                                                                                                          ok
[ 50.00%] ··· ===================== ============ ============= ============= ============
              --                                           Depth(D)
              --------------------- -----------------------------------------------------
               Number of Qubits(N)       1             10           100          1000
              ===================== ============ ============= ============= ============
                        1             23.6±3μs      151±3μs     1.39±0.07ms   14.4±0.3ms
                        10            89.4±3μs      790±50μs     7.68±0.2ms   77.7±0.9ms
                       100            722±20μs    7.14±0.09ms    70.8±0.7ms    731±5ms
                       1000          7.05±0.1ms     71.2±1ms      748±10ms    7.69±0.2s
              ===================== ============ ============= ============= ============

[ 50.00%] ·· Benchmarking existing-py_Users_tanujkhattar_opt_anaconda3_envs_cirq_bin_python-PYTHONOPTIMIZE
[ 75.00%] ··· Running (circuit_construction.XOnAllQubitsCircuit.time_circuit_construction--).
[100.00%] ··· circuit_construction.XOnAllQubitsCircuit.time_circuit_construction                                                                                                                          ok
[100.00%] ··· ===================== ============ ============= ============= ============
              --                                           Depth(D)
              --------------------- -----------------------------------------------------
               Number of Qubits(N)       1             10           100          1000
              ===================== ============ ============= ============= ============
                        1             25.5±2μs      223±60μs    1.75±0.03ms   16.2±0.4ms
                        10            109±4μs       978±20μs     9.80±0.2ms   99.7±0.5ms
                       100            939±30μs    8.95±0.09ms     91.1±1ms     954±20ms
                       1000          9.42±0.2ms     89.9±1ms      949±10ms    9.63±0.03s
              ===================== ============ ============= ============= ============

This should also make the cirq.contrib.hacks.disable_validation obsolete.

cc @maffoo @zchen088 @dabacon @dstrain115

[maffoo]

I think it'd be nice to add a context manager for changing the __cirq_debug__ flag and making that the default way of using it. This ensures that state is restored even if an exception happens, which makes it much easier to know the state of the flag when looking at code.

@contextlib.contextmanager
def with_debug(value: bool) -> Iterator[None]:
    token = __cirq_debug__.set(value)
    try:
        yield
    finally:
        token.reset()

Then can use that in these tests:

with cirq.with_debug(False):
    op = cirq.H(q0, q1)
    assert op.qubits == (q0, q1)
    h_op.valudate_args([q0, q1])

I think we should encourage using a mechanism like this and discourage setting cirq.__cirq_debug__ manually (other than maybe for interactive use).

[tanujkhattar]

I think adding the context manager is fine, but I'd still like to set the default value to __debug__ so that users have a way to disable validations in existing codebase without major modifications in the code.

What do you think?

[tanujkhattar]

added a context manager, PTAL

[mpharrigan]

semi-serious question: did you benchmark whether the new if statements measurably slow down existing code with debug turned on?

I don't know how common running python with -O is, but I guess one can set this flag independently of __debug__.

[tanujkhattar]

@mpharrigan I did some more measurements as per your suggestion and here are the findings.

Setup

I benchmark different versions of Cirq for the time it takes to construct 300_000 single / two qubit operations using snakeviz. The code is as follows:

def constuct_operations(q: List[cirq.Qid]) -> Iterator[cirq.Operation]:
    for q1, q2 in zip(q[:-1], q[1:]):
        yield cirq.X(q1)
        yield cirq.Y(q2)
        yield cirq.CNOT(q1, q2)
n = 100_000
q = cirq.LineQubit.range(n)
%snakeviz [*constuct_operations(q)]

Cirq master without __cirq_debug__

This PR with __cirq_debug__=True

This PR with __cirq_debug__=False

Conclusion

There is a minor cost for adding the if statement when __cirq_debug__=True but we shave off a large percent of time spent inside validate_args when __cirq_debug__=False.

@maffoo The PR is ready for another look.

[maffoo]

LGTM with one minor comment.

[maffoo]

nit: remove blank line and sort imports since contextvars is part of the stdlib.