Simplify circuit first before rebase

[yao-cqc]

from pytket.circuit import Circuit
from pytket.passes import SynthesiseTK, CommuteThroughMultis, RemoveRedundancies
c = Circuit(2).Rz(0.3,0).CX(0,1).Rz(-0.3,0).CX(0,1)
SynthesiseTK().apply(c)
# [TK1(0.5, 0.5, 3.5) q[0]; TK1(0, 3, 3) q[1]; TK2(0.5, 0, 0) q[0], q[1]; TK2(0.5, 0, 0) q[0], q[1]; TK1(0.5, 0.5, 0.5) q[0]; TK1(0, 0, 1) q[1]; ]

But original c should be easily reduced to identity with:

CommuteThroughMultis().apply(c)
RemoveRedundancies().apply(c)

The problem is that SynthesiseTK (i.e. synthesise_tk) does the TK2 rebase first, which makes it harder for the subsequent optimisations to work. Same thing applies to synthesise_tket.

I think this behaviour might have a considerable impact on performance since these rebase passes are often ran first. For example, SynthesiseTK in FullPeepholeOptimise and the pytket-quantinuum level 1 pass.

[sjdilkes]

I've found similar issues with compiling to TK2 as a target. As an example, SynthesiseTK or the common pytket-quantinuum passes aren't able to clock that XXPhase(a,0,1).YYPhase(b,0,1).ZZPhase(c,0,1) == TK2(a,b,c,0,1).
Or, running FullPeepholeOptimise on YYPhase(b,0,1) returns a TK2(b,0,0,0,1) with a bunch of TK1 wrappers, rather than TK2(0,b,0,0,1). In the FullPeepholeOptimise case, it loses track when converting circuits to CX & TK1 when calling CliffordSimp and is unable to notice and recover the reduced gates.

I agree with the observation that these scenarios could be considerably impacting gate count performance. I'm wondering how far we can get with a new pass that:

1. Is able to merge TK1 with TK2 by changing the basis of rotations in TK2
2. Is able to merge TK2 with TK2

[yao-cqc]

Not sure how far we can get, but the pass you suggested is definitely worth to have.

[github-actions]

This issue has been automatically marked as stale.