GridDevice gateset, gate_duration, and compilation_target_gateset support #5315

verult · 2022-04-29T22:08:54Z

Part of #5050

Adds the remaining major components to GridDevice.

Displays a warning about old Cirq version if DeviceSpecification contains gates not recognized by the client.
Added compilation target gateset example in GridDevice docstring

@dstrain115 @MichaelBroughton @tanujkhattar

…port * Displays a warning about old Cirq version if DeviceSpecification contains gates not recognized by the client. * Added compilation target gateset example in GridDevice docstring

verult · 2022-05-10T19:38:04Z

Pulled in the latest changes. This is now ready for review!

tanujkhattar

Left a first round of comments. Main concern is around using the existing target gatesets as it is based on which 2q gate is present in the proto. We tighten the checks to make sure that the constructed compilation target is a strict subset of the constructed gateset.

tanujkhattar · 2022-05-10T20:49:48Z

cirq-google/cirq_google/devices/grid_device.py

+            cirq_gates = [ops.SYC]
+            fsim_gates.append(ops.SYC)
+        elif gate_name == 'sqrt_iswap':
+            cirq_gates = [cirq.SQRT_ISWAP]


We should not push cirq.SQRT_ISWAP to cirq_gates since gates_list.extend(cirq_gates) below will add it to gates_list; and we will again add the equivalent gates_list.append(ops.FSimGateFamily(gates_to_accept=fsim_gates)).

The gate durations should also be specified for the corresponding fsim gate family; since we are accepting all equivalents operations across types; instead of specifying it only for cirq.GateFamily(cirq.SQRT_ISWAP) which is what is being done right now.

Different FSimGates have different durations today:

Cirq/cirq-google/cirq_google/devices/known_devices.py

Lines 304 to 306 in 9459df7

'fsim_pi_4': 32_000,

'inv_fsim_pi_4': 32_000,

'syc': 12_000,

And in general, each gate type in DeviceSpecification can have a separate duration. IIUC including the gate, while redundant, doesn't change gate validation behavior.

Since gate duration is purely informational, IMO it's OK to not have all the equivalent forms of a gate as part of the key. The alternative solution of having duration under the FSimGateFamily and taking the min/max/some other aggregate of the durations of all the different gates necessarily loses some information.

IIUC including the gate, while redundant, doesn't change gate validation behavior.

The purpose of a gateset is description and validation. Having redundant gates makes description more confusing and verbose, which should be avoided.

Different FSimGates have different durations today:

We can insert a separate fsim gate family instance for each fsim gate type in DeviceSpecification; so we can associate gate duration with the corresponding fsim gate family. i.e.

# gate durations dict should contain this. {cirq.FSimGateFamily(gates_to_accept=[cirq.SQRT_ISWAP]): 32_000, cirq.FSimGateFamily(gates_to_accept=[cg.SYC]): 12_000} # instead of . {cirq.GateFamily(cirq.SQRT_ISWAP): 32_000, cirq.GateFamily(cg.SYC): 12_000}

In general, we should be consistent with the gate families that we are using for description and validation, so that inconsistencies like the following cannot occur:

$> sqrt_iswap_gate = cirq.FSimGate(-np.pi/4, 0) # Returns True because of validation across types by FSimGateFamily. $> assert sqrt_iswap_gate in sqrt_iswap_metadata.gateset # Should return True but will return False right now because `sqrt_iswap_gate in cirq.GateFamily(cirq.SQRT_ISWAP)` is False $> assert any(sqrt_iswap_gate in gf for gf in sqrt_iswap_metadata.gate_durations)

As discussed offline, this makes a lot of sense, and I'll push for a change in the GridDeviceMetadata.gate_duration. Didn't end up using a function, but the proposal captures the same spirit. Will change to the approach you suggested, and leave a TODO to update it once #5427 moves forward.

Actually your approach currently fails GridDeviceMetadata validation because all gate duration keys are expected to be in the gateset. Let's discuss in #5427 since it potentially involves a change in GridDeviceMetadata. Are you onboard with keeping the gate duration the way it is and leaving a TODO for followup?

Synced offline. Will change the gatesets to have a separate FSimGateFamily for each 2q gate instead of a single FSimGateFamily.

Updated. I considered changing the test to a more black-box approach of checking whether certain gates belong in the gateset rather than asserting via gateset equality, but decided against it because we do want to test the exact gateset elements to verify its string representation.

cirq-google/cirq_google/devices/grid_device.py

tanujkhattar · 2022-05-10T21:03:00Z

cirq-google/cirq_google/devices/grid_device.py

+                cirq.PhasedXZGate,
+                cirq.XPowGate,
+                cirq.YPowGate,
+                cirq.ZPowGate,


Should we add a cirq.ZPowGate gate here given that we have special identifiers for virtual_zpow and physical_zpow ? Note that if cirq.GateFamily(cirq.ZPowGate) is present in the gateset, it will accept all (i.e. both tagged and untagged) instances of ZPowGate; in which case the physical_zpow and virtual_zpow gate families will be irrelevant.

For example: What happens if the proto specification includes phased_xz and doesn't contain virtual_zpow ? Should a cirq.Z(q) be accepted or not? What if it contains phased_xz and physical_zpow but doesn't contain virtual_zpow ?

Yeah the acceptance of ZPowGates should definitely be controlled entirely by virtual_zpow and physical_zpow. Thanks for the catch!

For the case where the spec includes phased_xz but not virtual_zpow, because PhysicalZTag() on a PhasedXZGate is ignored, a Z gate specified as a PhasedXZGate should be accepted and will be applied by the device in some way.

cirq-google/cirq_google/devices/grid_device.py

I avoided mentioning the alternative of inserting no_compile tags because transforming a circuit which contains WaitGates is the wrong approach as all delays would be scrambled by the transformation.

dstrain115

Aside from some nits, it is fine with me. I think Tanuj should also approve since he is more familiar with compilation target gatesets.

dstrain115 · 2022-05-20T22:13:14Z

cirq-google/cirq_google/devices/grid_device.py

@@ -77,6 +83,85 @@ def _validate_device_specification(proto: v2.device_pb2.DeviceSpecification) ->
                    )


+def _build_gateset_and_gate_durations(
+    proto: v2.device_pb2.DeviceSpecification,
+) -> Tuple[cirq.Gateset, Dict[cirq.GateFamily, cirq.Duration]]:


This is a private method, but I would still add a docstring since it's a little long.

dstrain115 · 2022-05-20T22:28:02Z

cirq-google/cirq_google/devices/grid_device.py

+def _build_compilation_target_gatesets(
+    gateset: cirq.Gateset,
+) -> Sequence[cirq.CompilationTargetGateset]:
+    """Detects compilation target gatesets based on what gates are inside the gateset."""


So, if we support multiple gatesets, do we have a way to decompose to a combination of CZ and sqrt-iswap for instance? Or do we have to pick one?

I think picking one is probably fine, since most grids will likely only support one type of gate, but we should make this clear in the documentation.

Yep we would have to pick one. If there's a use case to compile to both CZ and sqrt-iswap in the future we could revisit; doing so now for every target would probably create a combinatorial explosion of target gatesets.

On the other hand, if a device supports both CZ and sqrt-iswap, and someone wants to compile a circuit containing sqrt-iswap + some other arbitrary 2q gates using a CZ target gateset, the intended behavior is to leave sqrt-iswap gates untouched and compile other 2q gates to CZ + 1q gates. This isn't supported yet by existing target gatesets, but I'm planning to change that.

Will clarify this behavior in documentation.

tanujkhattar · 2022-05-20T23:17:42Z

cirq-google/cirq_google/devices/grid_device.py

+            cirq_gates = [ops.SYC]
+            fsim_gates.append(ops.SYC)
+        elif gate_name == 'sqrt_iswap':
+            cirq_gates = [cirq.SQRT_ISWAP]


IIUC including the gate, while redundant, doesn't change gate validation behavior.

The purpose of a gateset is description and validation. Having redundant gates makes description more confusing and verbose, which should be avoided.

Different FSimGates have different durations today:

We can insert a separate fsim gate family instance for each fsim gate type in DeviceSpecification; so we can associate gate duration with the corresponding fsim gate family. i.e.

# gate durations dict should contain this. {cirq.FSimGateFamily(gates_to_accept=[cirq.SQRT_ISWAP]): 32_000, cirq.FSimGateFamily(gates_to_accept=[cg.SYC]): 12_000} # instead of . {cirq.GateFamily(cirq.SQRT_ISWAP): 32_000, cirq.GateFamily(cg.SYC): 12_000}

In general, we should be consistent with the gate families that we are using for description and validation, so that inconsistencies like the following cannot occur:

$> sqrt_iswap_gate = cirq.FSimGate(-np.pi/4, 0) # Returns True because of validation across types by FSimGateFamily. $> assert sqrt_iswap_gate in sqrt_iswap_metadata.gateset # Should return True but will return False right now because `sqrt_iswap_gate in cirq.GateFamily(cirq.SQRT_ISWAP)` is False $> assert any(sqrt_iswap_gate in gf for gf in sqrt_iswap_metadata.gate_durations)

cirq-google/cirq_google/devices/grid_device.py

tanujkhattar · 2022-05-20T23:20:43Z

cirq-google/cirq_google/devices/grid_device.py

+    A note about CompilationTargetGatesets:
+
+    A circuit which contains `cirq.WaitGate`s will be dropped if it is transformed using
+    CompilationTargetGatesets generated by GridDevice. To better control circuit timing, insert
+    WaitGates after the circuit has been transformed.


I think this note should go to the docstring of compilation_target_gatesets property. This is a specific detail of compilation targets we are using today, and should not be tied to the description of device.

My understanding is that we want to call out WaitGate because Google devices all support this gate. Is this concern general enough to all devices to be kept in the compilation_target_gateset property in cirq.GridDeviceMetadata? If so, I would suggest to keep it in CompilationTargetGateset instead, since all CompilationTargetGatesets would have this issue. If not, I would keep it where it is now.

Also, if we were to keep this comment in cirq-core, are there other identity-like gates that may not be negligible that we should call out?

I think this is mainly used by the Google devices and we don't need to add it to the CompilationTargetGateset class.

Sounds good, will keep the docstring here then since GridDevice doesn't have a compilation_target_gatesets property.

…factor/device-gateset

tanujkhattar

LGTM! Thanks for all the changes!

…port (quantumlib#5315) * GridDevice gateset, gate_duration, and compilation_target_gateset support * Displays a warning about old Cirq version if DeviceSpecification contains gates not recognized by the client. * Added compilation target gateset example in GridDevice docstring * Addresses Tanuj's comments * Added GridDevice docstring comment about WaitGates. I avoided mentioning the alternative of inserting no_compile tags because transforming a circuit which contains WaitGates is the wrong approach as all delays would be scrambled by the transformation. * Set gate_duration to None if the spec doesn't have duration info * Addressed Doug's comments * Addressed Tanuj's 2nd round of comments * Break FSimGateFamily into one for each 2q gate type

verult requested review from MichaelBroughton, tanujkhattar and dstrain115 April 29, 2022 22:08

verult requested review from wcourtney, a team, vtomole and cduck as code owners April 29, 2022 22:08

verult force-pushed the cg-device-refactor/device-gateset branch 3 times, most recently from d0e90df to 5b592d7 Compare April 29, 2022 23:12

CirqBot added the size: XL lines changed >1000 label Apr 30, 2022

tanujkhattar self-assigned this May 3, 2022

verult force-pushed the cg-device-refactor/device-gateset branch from 5ca5795 to 1ee848d Compare May 10, 2022 19:32

GridDevice gateset, gate_duration, and compilation_target_gateset sup…

ff8aebc

…port * Displays a warning about old Cirq version if DeviceSpecification contains gates not recognized by the client. * Added compilation target gateset example in GridDevice docstring

verult force-pushed the cg-device-refactor/device-gateset branch from 1ee848d to ff8aebc Compare May 10, 2022 19:36

tanujkhattar requested changes May 10, 2022

View reviewed changes

verult added 3 commits May 18, 2022 01:15

Addresses Tanuj's comments

486376e

Added GridDevice docstring comment about WaitGates.

84bdfe5

I avoided mentioning the alternative of inserting no_compile tags because transforming a circuit which contains WaitGates is the wrong approach as all delays would be scrambled by the transformation.

Set gate_duration to None if the spec doesn't have duration info

17262f0

verult mentioned this pull request May 18, 2022

DeviceSpecification serialization implementation #5375

Merged

dstrain115 approved these changes May 20, 2022

View reviewed changes

tanujkhattar requested changes May 20, 2022

View reviewed changes

Addressed Doug's comments

288a4dc

verult mentioned this pull request May 31, 2022

Added GateDurationTable to simplify duration lookup of equivalent gates. #5427

Closed

verult added 3 commits June 1, 2022 00:10

Addressed Tanuj's 2nd round of comments

8135a8c

Break FSimGateFamily into one for each 2q gate type

7a1bcd9

Merge branch 'master' of github.com:quantumlib/Cirq into cg-device-re…

e9de7ad

…factor/device-gateset

tanujkhattar approved these changes Jun 3, 2022

View reviewed changes

tanujkhattar merged commit 2948e18 into quantumlib:master Jun 3, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

GridDevice gateset, gate_duration, and compilation_target_gateset support #5315

GridDevice gateset, gate_duration, and compilation_target_gateset support #5315

verult commented Apr 29, 2022 •

edited

verult commented May 10, 2022

tanujkhattar left a comment

tanujkhattar May 10, 2022

verult May 17, 2022

tanujkhattar May 20, 2022

verult May 31, 2022 •

edited

verult Jun 1, 2022

verult Jun 1, 2022

verult Jun 1, 2022

tanujkhattar May 10, 2022

verult May 17, 2022

dstrain115 left a comment

dstrain115 May 20, 2022

dstrain115 May 20, 2022

verult May 31, 2022

tanujkhattar May 20, 2022

tanujkhattar May 20, 2022

verult May 31, 2022

tanujkhattar Jun 1, 2022

verult Jun 2, 2022

tanujkhattar left a comment

GridDevice gateset, gate_duration, and compilation_target_gateset support #5315

GridDevice gateset, gate_duration, and compilation_target_gateset support #5315

Conversation

verult commented Apr 29, 2022 • edited

verult commented May 10, 2022

tanujkhattar left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

verult May 31, 2022 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dstrain115 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tanujkhattar left a comment

Choose a reason for hiding this comment

verult commented Apr 29, 2022 •

edited

verult May 31, 2022 •

edited