Speed up circuit construction when contents are a list of moments

[tanujkhattar]

When the input is a list of moment, we can just assign that list to self._moments instead of using either append or _load_contents_with_earliest_strategy.

• Circuit.append: This is already essentially O(D) if the the input is a list of moments because of the special case isinstance check:

  Cirq/cirq-core/cirq/circuits/circuit.py

  Line 2102 in ff81f80

  self._moments.insert(k, moment_or_op)

• Circuit._load_contents_with_earliest_strategy: has complexity O(D * Q) instead of O(D), even if input is all Moments because it iterates on all the qubits that the moment acts on, for each moment; to keep track of the latest used index in case it encounters an operation which is not part of a moment and needs to be inserted using the earliest strategy. See
  

  Cirq/cirq-core/cirq/circuits/circuit.py

  Line 1825 in ff81f80

  for q in mop_qubits:

The fact that introducing _load_contents_with_earliest_strategy slowed down moment-by-moment circuit construction to O(D * Q) instead of O(D) is also visible from the regressions in circuit construction asv benchmarks:

After this change, the moment-by-moment circuit construction time for a distance 25 surface code is again ~30ms instead of > 3 seconds. The output for asv run on a local copy is as follows:

$ asv run --python=same --bench SurfaceCodeRotatedMemoryZ.time_circuit_construction_moment_by_moment --cpu-affinity 10 
· Discovering benchmarks
· Running 1 total benchmarks (1 commits * 1 environments * 1 benchmarks)
[  0.00%] ·· Benchmarking existing-py_usr_local_google_home_tanujkhattar_anaconda3_envs_cirq_bin_python
[ 50.00%] ··· Running (circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_moment_by_moment--).
[100.00%] ··· circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_moment_by_moment                                                                                                                                                                                 ok
[100.00%] ··· ========== =============
               distance               
              ---------- -------------
                  3         497±10μs  
                  5       1.39±0.08ms 
                  7        2.70±0.1ms 
                  9        4.23±0.2ms 
                  11       6.64±0.1ms 
                  13       9.17±0.4ms 
                  15       11.5±0.2ms 
                  17       15.7±0.6ms 
                  19       19.7±0.7ms 
                  21        24.9±1ms  
                  23       27.6±0.6ms 
                  25       32.8±0.4ms 
              ========== =============

compared to the current master version:

$ asv run  --bench SurfaceCodeRotatedMemoryZ.time_circuit_construction_moment_by_moment --cpu-affinity 8
· Creating environments
· Discovering benchmarks
· Running 1 total benchmarks (1 commits * 1 environments * 1 benchmarks)
[  0.00%] · For Cirq commit ff81f807 <master>:
[  0.00%] ·· Benchmarking virtualenv-py3.8
[ 50.00%] ··· Running (circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_moment_by_moment--).
[100.00%] ··· circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_moment_by_moment                                                                                                                                                                                 ok
[100.00%] ··· ========== ============
               distance              
              ---------- ------------
                  3       2.03±0.1ms 
                  5       8.02±0.2ms 
                  7        26.2±3ms  
                  9        69.1±4ms  
                  11       137±8ms   
                  13       246±9ms   
                  15       455±7ms   
                  17       774±20ms  
                  19      1.14±0.01s 
                  21      1.67±0.04s 
                  23      2.50±0.08s 
                  25      3.57±0.04s 
              ========== ============

cc @dabacon @dstrain115 @maffoo @dkothari11

[maffoo]

I'd probably go for list instead, since we don't need immutability here, but it'd be interesting to know if there is any performance difference one way or the other. If not, then up to you whether to use list or tuple.

[tanujkhattar]

Tuples are usually slightly faster than lists because they don't overallocate like lists (to make append faster) so the memory footprint is more compact. Also, tuples don't need to go through an additional level of indirection when accessing elements like lists; see https://stackoverflow.com/a/22140115 for more details

As an example, here are numbers for running the operation by operation construction benchmark (in which the flattened_contents would contain all the operations and hence have a large size).

Using tuples for flattened_contents:

[  0.00%] ·· Benchmarking existing-py_usr_local_google_home_tanujkhattar_anaconda3_envs_cirq_bin_python
[ 50.00%] ··· Running (circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_operations_by_operation--).    
[100.00%] ··· circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_operations_by_operation                                                                                                                                                                          ok
[100.00%] ··· ========== ============
               distance              
              ---------- ------------
                  3       2.68±0.1ms 
                  5       18.2±0.7ms 
                  7        72.2±2ms  
                  9        198±6ms   
                  11       464±10ms  
                  13       933±40ms  
                  15      1.68±0.06s 
                  17      2.95±0.09s 
                  19      4.73±0.2s  
                  21      6.39±0.02s 
                  23      9.50±0.03s 
                  25      13.7±0.4s  
              ========== ============

Using lists for flattened_contents:

[  0.00%] ·· Benchmarking existing-py_usr_local_google_home_tanujkhattar_anaconda3_envs_cirq_bin_python
[ 50.00%] ··· Running (circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_operations_by_operation--).
[100.00%] ··· circuit_construction.SurfaceCodeRotatedMemoryZ.time_circuit_construction_operations_by_operation                                                                                                                                                                          ok
[100.00%] ··· ========== ============
               distance              
              ---------- ------------
                  3       2.81±0.2ms 
                  5       20.3±0.8ms 
                  7        79.9±3ms  
                  9        204±4ms   
                  11       467±10ms  
                  13       934±10ms  
                  15      1.75±0.1s  
                  17      2.92±0.08s 
                  19      4.42±0.1s  
                  21      6.65±0.02s 
                  23      9.85±0.04s 
                  25      13.8±0.2s  
              ========== ============