State synthesis for quantum devices

[annagrin]

Synthesize state pointer for quantum devices:

SimulationState:

• Add hasData API that returns true iff the vector data exists on the state or can be computed.
• Add getKernelInfo API that returns optional kernel name and a list of arguments for the kernel that generated the state.

QuantumState:

• Add QuantumState class that inherits from simulation state and keeps kernel information only.

ArgumentConversion:

• Substitute state pointer argument with quake.get_state "callee.num_qubits_N" "callee.init_N" instruction.
• Add callee.num_qubits_N and callee.init_N functions that are created from the callee code and compute allocation size and initialize the allocation, respectively.
• Add a list of converters for the callees in ArgumentConverter.
• Generate substitutions for modified callee functions, recursively.

Passes:

• Add a post-synthesis ReplaceStateWithKernel pass that
  • replaces quake.get_num_qubits instructions by a call to callee.num_qubits_N()
  • replaces quake.init_state instructions by a call to callee.init_N()

Synthesis:

• Collect all substitutions for the call tree at the synthesis time for quantum devices.

Tests:

• Add fake quantum state and a test for quantum state argument conversion to test_argument_conversion
• Add quake tests for ReplaceStateWithKernel pass
• Add quake tests for ArgumentSynthesis pass with state pointer substitutions
• Add targettests/execution/qvector_init_from_state.cpp test that runs for various quantum backends emulation

Notes

Currently state pointer synthesis is only supported for kernels implemented as operator() inside a struct.

TODO (in subsequent PRs)

• c++ tests for quantum backend integration
• Use cudaq::qkernel to support c-like kernels in state pointer synthesis
• Tests for synthesis of vectors of pauli words (needs rework of pauli word support)
• python tests for quantum platforms (needs support for new synthesis in python)

Requires: #2354

[1tnguyen]

LGTM 👍

[schweitzpgi]

Still reviewing...

[schweitzpgi]

Still reviewing...

[schweitzpgi]

Suggested change

	StrAttr:$numQubitsFuncName,
	StrAttr:$initFuncName
	FlatSymbolRefAttr:$numQubitsFuncName,
	FlatSymbolRefAttr:$initFuncName

If these are truly artifacts that shall be present in the IR, let's make them Symbol attrs.

[annagrin]

If the signatures of those functions are changed by the synthesis, would this instruction get updated with symbols with new signatures during application of the substitution? I can try that and see if the synthesis works.

[schweitzpgi]

We probably want to force this to be a ptr<state>, no?

[annagrin]

Will do, thanks

[schweitzpgi]

Why do we need this? ArrayRef should subsume rigid std::vector here. I don't think we need this overload?

[annagrin]

std::vector<string> does not get auto-converted to ArrayRef<StringRef>... I can try ArrayRef<std::string> instead

[schweitzpgi]

Use SmallVector.

[annagrin]

I think i tried and something went wrong, will try again

[schweitzpgi]

Looks like you got in trouble here by moving the strings into another function, where they go out of scope and vanish.

Let's leave the string creation here, where it is. To change it might look a little nicer, but it is far less efficient. Instead of a single copy of the data, we're building vectors of copies of the data and passing those around. The LLVM Way (tm) is preferred in the compiler code.

[annagrin]

The nested argument converters (for kernels used in get_state calls) have their new kernel names, created during state argument conversion, so I moved the kernel name storage to the ArgumentConverter for the kernel.

I thought about an alternative of having a special "storage" for the new names collection and passing that storage by reference to ArgumentConverter and its children, but that does not seem to solve the problem of efficiency of copying the name for the entry kernel. I appreciate any ideas here!