MPS: Performance improvement when measuring all qubits #1277

merav-aharoni · 2021-06-20T19:14:29Z

Summary

Performance improvement for the case where all qubits are measured.

Details and comments

In the previous implementation, after every qubit is measured, its effect is propagated to all other qubits. After that, the next qubit is measured. In the new implementation, if we are measuring all qubits, then it is not necessary to propagate to effect to all the remaining qubits. We only propagate to the nearest neighbors, because then we measure these neighbors next.
It is possible to implement for cases where a subset of the qubits is measured, but this will require additional effort to decide when propagation is necessary and when not.

…hmarking

merav-aharoni · 2021-06-21T14:00:08Z

Also should review the heuristics that chooses between algorithms for measurement - the parameters should probably be changed.

merav-aharoni · 2021-06-21T14:12:12Z

Here are the results of the performance of the previous apply_measure as compared to the current implementation. The circuit is RealAmplitudes with 10000 shots.

merav-aharoni · 2021-06-21T14:19:44Z

@yaelbh , @chriseclectic - please review

yaelbh

Very very nice

releasenotes/notes/mps-performance-of-measure-9991ac98da35150b.yaml

src/simulators/matrix_product_state/matrix_product_state_internal.cpp

yaelbh · 2021-06-22T08:57:59Z

src/simulators/matrix_product_state/matrix_product_state_internal.cpp

+    left_qubit = min_qubit == 0 ? 0 : min_qubit - 1;
+  }
+  if (num_qubits_ == 1)
+    right_qubit = 0;


Same comment as in apply_measure_internal. There is an unclear separation in the order of qubits according to apply_meausre, which is perhaps not necessary.

In the case of measure_all, since we are measuring in order from 0 to n-1, this will not really be necessary - there will be no propagation to the left. But I think it is better to cover this, for the sake of completeness. Let me know if you think I should do this differently.

yaelbh · 2021-06-22T08:59:46Z

src/simulators/matrix_product_state/matrix_product_state_internal.cpp

@@ -1698,11 +1715,11 @@ void MPS::measure_reset_update_internal(const reg_t &qubits,
 mps_container_t MPS::copy_to_mps_container() {
  move_all_qubits_to_sorted_ordering();
  mps_container_t ret;
-  for (auto i=0; i<num_qubits(); i++) {
+  for (uint_t i=0; i<num_qubits(); i++) {


Why not auto? By the way, when you used uint before, perhaps it should be uint_t?

I changed all of these, because the compiler was giving warnings on them and I wanted to eliminate the warnings.

src/simulators/matrix_product_state/matrix_product_state_internal.hpp

… algorithm for measure_all depends on the qubit ordering

…neighbors_internal

chriseclectic

Looks good. I'm not sure if doc string for "mps_sample_measure_algorithm" needs updating for this change or if it is an improvement to the existing options.

releasenotes/notes/mps-performance-of-measure-9991ac98da35150b.yaml

chriseclectic · 2021-06-22T21:30:45Z

test/terra/backends/qasm_simulator/qasm_measure.py

+    # ---------------------------------------------------------------------
+    # Test MPS algorithms for measure
+    # ---------------------------------------------------------------------
+    def test_mps_measure_alg_qv(self):


It might be better to add these tests to the new AerSimulator tests which are a lot cleaner for specifying that they run on specific simulation methods or with custom simulation options.

Eg see https://github.com/Qiskit/qiskit-aer/blob/main/test/terra/backends/aer_simulator/test_algorithms.py#L68-L73 for example of running test on specific method with custom simulator options.

I am not sure where would be most suitable for this test. Perhaps under https://github.com/Qiskit/qiskit-aer/blob/main/test/terra/backends/aer_simulator/test_options.py? on the other hand, it is testing measurement, so that's why I put it in test_measure.
It's your call where to put it.

Your last comment somehow disappeared. You mentioned putting the option apply_measure in AerSimulator and in QasmSimulator. This option is not new. It is already in both these files.

… into improve-measure

omarshehab · 2021-07-09T10:16:48Z

@chriseclectic , am I correct to understand the following?

Previously, there were three options for the value to be passed on for "mps_sample_measure_algorithm" option. They are "mps_probabilities", "mps_apply_measure", and “mps_heuristic”.
This fix adds one more option which is "mps_measure_all".
We are using the QSVM API for 20 qubits. To take advantage of this improvement we will have to pass backend_options = {"mps_sample_measure_algorithm": "mps_measure_all"}.

merav-aharoni · 2021-07-09T10:49:27Z

Hi @omarshehab, we didn't add another option mps_measure_all - that was something we did temporarily just to compare performance. The new improvement works by default, so you don't have to set anything, except for selecting the mps simulation method. It works as part of mps_apply_measure. Note however, that this improvement applies only when you measure all the qubits and not a subset.

omarshehab · 2021-07-09T13:37:43Z

@merav-aharoni , I understand that the default is "mps_sample_measure_algorithm": "mps_heuristic". In that case, if I don't set "mps_sample_measure_algorithm": "mps_apply_measure", the new improvements will not be used, right?

mishmash · 2021-07-09T15:24:09Z

@omarshehab , it may still be used, but that just depends on whether or not the heuristic chooses the "mps_apply_measure" option. But, as mentioned by @merav-aharoni above, the heuristic is now out of date with these improvements. I would personally be in favor of removing the heuristic altogether, default to "mps_apply_measure", and provide a more detailed description of the inner workings of the two options in the docs; e.g., "mps_apply_measure" generates samples directly from the MPS each taking O(n \chi^2) time [or O(n^2 \chi^2) if measuring only a subset], while "mps_probabilities" first reconstructs the full statevector from the MPS in O(2^n n \chi^2) time and then generates samples each in O(1) time -- from this information, the user should be able to make an informed decision as to which option to use.

merav-aharoni · 2021-07-10T09:28:44Z

Thanks, @mishmash for adding the clarification. The intent is indeed to implement this algorithm also when measuring a subset of the qubits, and to either remove or improve the heuristic accordingly. I am not sure I will get to this in the near future, so perhaps it is best to update the documentation, as you suggest.

Performance improvement when measuring all qubits

877d1eb

merav-aharoni requested review from chriseclectic, hhorii, mtreinish and vvilpas as code owners June 20, 2021 19:14

merav-aharoni added 6 commits June 20, 2021 22:16

Increase delta in the test

3902b46

Increased preciseness of test_mps_measure_alg_qv

761f7cd

Removed the option measure_all that was inserted temporarily for benc…

a091ea2

…hmarking

Fixed documentation and added release note

707f8ce

Inserted else that was accidentally deleted

8459cb4

Fixed bug when circuit has only 1 qubit

3207e19

merav-aharoni changed the title ~~[WIP] MPS: Performance improvement when measuring all qubits~~ MPS: Performance improvement when measuring all qubits Jun 21, 2021

yaelbh suggested changes Jun 22, 2021

View reviewed changes

merav-aharoni added 5 commits June 22, 2021 15:46

Changed 0/1 to false/true. Added an explanation to emphasize that the…

15e9b1e

… algorithm for measure_all depends on the qubit ordering

Added mention of mps in the release note

0396f0f

Move the case of a single qubit to outside the function propagate_to_…

1720201

…neighbors_internal

Merged with main

352c754

Increase delta for test_MPS_measure_alg_qv because of failure

f38dddf

chriseclectic reviewed Jun 23, 2021

View reviewed changes

chriseclectic and others added 4 commits June 24, 2021 14:17

Merge branch 'main' into improve-measure

f721cde

Changed subject to other

a7ca67d

Merge branch 'improve-measure' of github.com:merav-aharoni/qiskit-aer…

2b6b0e3

… into improve-measure

Merge branch 'main' into improve-measure

3bf7aff

chriseclectic approved these changes Jun 25, 2021

View reviewed changes

chriseclectic merged commit 172f8b3 into Qiskit:main Jun 25, 2021

omarshehab mentioned this pull request Jul 11, 2021

Allow measuring subset of qubits when QuantumKernel.evaluate() is run on MPS simulator qiskit-community/qiskit-machine-learning#157

Closed

merav-aharoni mentioned this pull request Jul 12, 2021

MPS: Improve performance of apply_measure for any subset of qubits #1290

Merged

chriseclectic pushed a commit to chriseclectic/qiskit-aer that referenced this pull request Aug 12, 2021

MPS: Performance improvement when measuring all qubits (Qiskit#1277)

d567e22

mishmash mentioned this pull request Nov 3, 2021

MPS: New measurement sampling method #1377

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

MPS: Performance improvement when measuring all qubits #1277

MPS: Performance improvement when measuring all qubits #1277

merav-aharoni commented Jun 20, 2021

merav-aharoni commented Jun 21, 2021

merav-aharoni commented Jun 21, 2021 •

edited

Loading

merav-aharoni commented Jun 21, 2021

yaelbh left a comment

yaelbh Jun 22, 2021

merav-aharoni Jun 22, 2021

yaelbh Jun 22, 2021

merav-aharoni Jun 22, 2021

chriseclectic left a comment

chriseclectic Jun 22, 2021

merav-aharoni Jun 23, 2021

merav-aharoni Jun 23, 2021

omarshehab commented Jul 9, 2021 •

edited

Loading

merav-aharoni commented Jul 9, 2021

omarshehab commented Jul 9, 2021

mishmash commented Jul 9, 2021

merav-aharoni commented Jul 10, 2021

MPS: Performance improvement when measuring all qubits #1277

MPS: Performance improvement when measuring all qubits #1277

Conversation

merav-aharoni commented Jun 20, 2021

Summary

Details and comments

merav-aharoni commented Jun 21, 2021

merav-aharoni commented Jun 21, 2021 • edited Loading

merav-aharoni commented Jun 21, 2021

yaelbh left a comment

Choose a reason for hiding this comment

yaelbh Jun 22, 2021

Choose a reason for hiding this comment

merav-aharoni Jun 22, 2021

Choose a reason for hiding this comment

yaelbh Jun 22, 2021

Choose a reason for hiding this comment

merav-aharoni Jun 22, 2021

Choose a reason for hiding this comment

chriseclectic left a comment

Choose a reason for hiding this comment

chriseclectic Jun 22, 2021

Choose a reason for hiding this comment

merav-aharoni Jun 23, 2021

Choose a reason for hiding this comment

merav-aharoni Jun 23, 2021

Choose a reason for hiding this comment

omarshehab commented Jul 9, 2021 • edited Loading

merav-aharoni commented Jul 9, 2021

omarshehab commented Jul 9, 2021

mishmash commented Jul 9, 2021

merav-aharoni commented Jul 10, 2021

merav-aharoni commented Jun 21, 2021 •

edited

Loading

omarshehab commented Jul 9, 2021 •

edited

Loading