[rt] Remove ExecutionManager's qudit tracker. #1541
Conversation
boschmitt
force-pushed
the
remove_qudit_management_from_execution_manager
branch
3 times, most recently
from
0c9bd69
to
bb1acf2
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
remove_qudit_management_from_execution_manager
branch
from
bb1acf2
to
f04c511
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
| requestedAllocations.emplace_back(2, q.id); | ||
| } | ||
|
|
||
| void allocateQudits(const std::vector<cudaq::QuditInfo> &qudits) override { |
There was a problem hiding this comment.
It was useful from a performance perspective to observe (in some way) that we were really allocating a qvector, and to queue them up and request the allocation all at once (means we can create the correct GPU pointer once, and not once + N-1 changes to the data on device). Is this type of functionality retained somewhere else? maybe i'll see it further down
There was a problem hiding this comment.
It is not present anywhere else. I can implement it on the CircuitSimulator layer, so that JITed kernels would also benefit from it.
boschmitt
force-pushed
the
remove_qudit_management_from_execution_manager
branch
from
f04c511
to
5a6c996
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
remove_qudit_management_from_execution_manager
branch
from
5a6c996
to
130f5e9
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
boschmitt
force-pushed
the
remove_qudit_management_from_execution_manager
branch
from
130f5e9
to
70c802f
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
This commit achieves two goals: 1. Deduplicate the qudit tracking functionality, which is present in both the `ExecutionManager` and the `CircuitSimulator. The `ExecutionManager` does not need to keep track of qudit's allocation and deallocation, it just needs to forward requests. Indeed, it is the simulators' job to manage the qudits and the quantum state representation. Note: We should even consider moving qudit tracking from the `CircuitSimulator` base class to the subtypes since each simulator might want to handle memory differently. For example: eager deallocation of qudits might be desired when this deallocation can be done with a simple resize of the state vector. This deallocation would improve the performance of applying gates but requires specific information from the simulators themselves. 2. Completely moves the `Tracer` to the `CircuitSimulator`. Before this commit, the `Tracer` implementation was divided between the `ExecutionManager` and the `CircuitSimulator`. Note: This move required adding another level of indirection for qudit reset within the simulators. Before, this indirection was handled by the `ExecutionManager` alone. We probably had an unknown bug since the execution of JITed kernels bypass the `ExecutionManager`, and thus it is quite likely the execution of JITed kernels in the `tracer` execution context would have exploded in the presence of qudit resets.
boschmitt
force-pushed
the
remove_qudit_management_from_execution_manager
branch
from
70c802f
to
bd06fdb
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
Description
This commit achieves two goals:
ExecutionManagerand theCircuitSimulator.The
ExecutionManagerdoes not need to keep track of qudit's allocation and deallocation, it just needs to forward requests. Indeed, it is the simulators' job to manage the qudits and the quantum state representation.Note: We should even consider moving qudit tracking from the
CircuitSimulatorbase class to the subtypes since each simulator might want to handle memory differently. For example: eager deallocation of qudits might be desired when this deallocation can be done with a simple resize of the state vector. This deallocation would improve the performance of applying gates but requires specific information from the simulators themselves.Tracerto theCircuitSimulator. Before this commit, theTracerimplementation was divided between theExecutionManagerand theCircuitSimulator.Note: This move required adding another level of indirection for qudit reset within the simulators. Before, this indirection was handled by the
ExecutionManageralone. We probably had an unknown bug since the execution of JITed kernels bypass theExecutionManager, and thus it is quite likely the execution of JITed kernels in thetracerexecution context would have exploded in the presence of qudit resets.