Direct Support to probabilities, expectation values and variances in PL-Lightning #185

AmintorDusko · 2021-12-03T22:58:06Z

Context:
Pennylane Lightning doesn't offer direct support for the calculation of probabilities, expectation values and variances.
Description of the Change:
This PR introduces a C++ class template Measures that implements such calculations.
Calculations can be performed in terms of the already defined gates, or with any operator provided as a matrix.
Benefits:
A direct implementation in C++ expands the optimization options, and this class allows to perform such calculations for any operator provided in matrix form.
Possible Drawbacks:
I'm not aware of.
Related GitHub Issues:

github-actions · 2021-12-03T22:58:22Z

Hello. You may have forgotten to update the changelog!
Please edit .github/CHANGELOG.md with:

A one-to-two sentence description of the change. You may include a small working example for new features.
A link back to this PR.
Your name (or GitHub username) in the contributors section.

codecov · 2021-12-03T22:59:43Z

Codecov Report

Merging #185 (5c030be) into master (5b65f2a) will not change coverage.
The diff coverage is n/a.

❗ Current head 5c030be differs from pull request most recent head 701ea6c. Consider uploading reports for the commit 701ea6c to get more accurate results

@@           Coverage Diff           @@
##           master     #185   +/-   ##
=======================================
  Coverage   99.64%   99.64%           
=======================================
  Files           4        4           
  Lines         278      278           
=======================================
  Hits          277      277           
  Misses          1        1

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5b65f2a...701ea6c. Read the comment docs.

github-actions · 2021-12-03T23:00:53Z

Test Report (C++) on Ubuntu

      1 files ±  0       1 suites ±0 0s ⏱️ ±0s
  378 tests +10   378 ✔️ +10 0 💤 ±0 0 ❌ ±0
2 233 runs +28 2 233 ✔️ +28 0 💤 ±0 0 ❌ ±0

Results for commit 701ea6c. ± Comparison against base commit 5b65f2a.

♻️ This comment has been updated with latest results.

mlxd

Thanks for putting this together @AmintorDusko
I have some comments and questions:

pennylane_lightning/src/simulator/Measures.hpp

pennylane_lightning/src/simulator/StateVector.hpp

pennylane_lightning/src/tests/Test_Measures.cpp

pennylane_lightning/src/util/Util.hpp

AmintorDusko · 2021-12-06T13:45:32Z

I will work on every request and suggestion. Thanks for your feedback @mlxd .

AmintorDusko · 2021-12-06T22:33:44Z

Please let me know your thoughts.

AmintorDusko · 2021-12-09T19:06:49Z

Hi @mlxd. Following our discussion, I updated the probabilities calculations.
Please let me know your thoughts.

mlxd

Thanks for this @AmintorDusko.
I have left some comments and queries.

pennylane_lightning/src/simulator/Measures.hpp

pennylane_lightning/src/util/Util.hpp

mlxd · 2021-12-10T10:10:12Z

pennylane_lightning/src/util/Util.hpp

+template <typename T>
+auto transpose_state_tensor( 
+                            const std::vector<T> &tensor, 
+                            const std::vector<size_t> &new_axes)
+                            -> std::vector<T> {
+    std::vector<T> transposed_tensor(tensor.size());
+    for (size_t ind=0; ind<tensor.size(); ind++){
+        transposed_tensor[ind]=tensor[transposed_state_index(ind, new_axes)];
+    }
+  return transposed_tensor;
+}


I worry this may be slow for large data sets. Can you run some numbers on this and see how it performs for various numbers of qubits.

I explored some alternative strategies for this transposition, and at least for now, this seems to be the more efficient. In the other hand, It doesn't seem to be too slow.
I checked the performance for 100 runs, up to 22 qubits.
I applied the transpose_state_tensor function to vectors with the size of the equivalent state vector (or tensor). This shows exactly the information we are interested in. For 22 qubits, we have a vector with 8,388,608 elements, and the mean time is 0.89 seconds.

A loop like that, with no data dependency, is quite the low-hanging fruit for openMP parallelisation, but for the sizes I'm presenting here quite of an overkill. If you think we should proceed with some parallelisation strategy, I think we could impose a threshold based in the number of qubits.
The code was compiled in GCC 11 with the -Ofast flag.

Thanks Amintor. I do think we should ensure this is fast for <=30 qubits (for large node usage). We can add parallelism as a next pass though.

Hi @mlxd ,
I extended the benchmark to 26 qubits (memory is limiting this range).
I optimised the code to determine the transposed index, and the function to transpose the tensor, accordingly.
The best method so far is still based in transpose the tensor after the calculation of the probabilities.
Anyway, in this new approach for the tensor transposition, I could observe a huge increase in speed, for larger number of qubits. I'm comparing the two methods here:

I also plotted the ration between the two times:

For 26 qubits, the previous algorithm is taking 18.4 seconds, while the new one is taking 2.3 seconds. The new code is approximately 8 times faster. This qubit state vector has 67,108,864 elements. I'm performing calculation with double precision, and compiling with the -Ofast flag.

We can observe an outlier in 12 qubits, where the previous algorithm seems to outperform the new one. The ration here is 0.85, and this may be because of the number of samples (100). Also, we are talking about small number and this could be strongly affected by fluctuations.

In other run, the 12 qubits performed better in the new algorithm. I think it is reasonable to assume that the measure of times in this fast cases is specially prone to fluctuations. It seems that starting in 20 qubits would be a good choice.

AmintorDusko · 2021-12-10T22:26:21Z

Hi @mlxd. Let me know what you think.

mlxd · 2021-12-14T12:43:38Z

Thanks Amintor. I think I am happy to merge this in as the initial pass. We can add the Python bindings in a follow up PR.
Also, I think it best to never use -Ofast and favour use of -O2 or -O3 instead. We never intend to ship anything with -Ofast so we should ensure all numbers taken do not use it.

mlxd

After the change listed here, I think this is good to go.

mlxd · 2021-12-14T12:40:52Z

pennylane_lightning/src/util/Util.hpp

+        // transposed_tensor[ind] = tensor[transposed_state_index(ind,
+        // new_axes)];


Suggested change

// transposed_tensor[ind] = tensor[transposed_state_index(ind,

// new_axes)];

AmintorDusko · 2021-12-14T13:30:14Z

Thanks Amintor. I think I am happy to merge this in as the initial pass. We can add the Python bindings in a follow up PR. Also, I think it best to never use -Ofast and favour use of -O2 or -O3 instead. We never intend to ship anything with -Ofast so we should ensure all numbers taken do not use it.

Makes sense. Thanks!

…ightning into feature/measures

Implement the Measures Class

d07926a

AmintorDusko added 3 commits December 3, 2021 18:02

Fixes format

70ce88d

Tries to appease codecov

5dbbf1f

Tries to appease codecov

84d75d2

mlxd requested changes Dec 6, 2021

View reviewed changes

AmintorDusko added 2 commits December 6, 2021 17:04

improves Measures class, and tests

3635c69

Updata Measures class, and tests.

f9e5f05

AmintorDusko requested a review from mlxd December 6, 2021 22:22

AmintorDusko added 6 commits December 9, 2021 11:35

Updates probs definition.

fb4af75

Optimizes transposition

f0b2728

Improves testing

6a7d580

Merge branch 'master' into feature/measures

42cbfb8

Formats code

8344585

Appaises clang-tify formatting check.

f20d041

mlxd requested changes Dec 10, 2021

View reviewed changes

Update to Measures class and some Utilities

f91d253

AmintorDusko added 3 commits December 13, 2021 16:44

Improves efficiency of tensor transposition

51e8caa

Merge branch 'master' into feature/measures

d49fcb2

clang-tidy

6756e72

mlxd approved these changes Dec 14, 2021

View reviewed changes

Merge branch 'master' into feature/measures

8de12da

AmintorDusko and others added 6 commits December 15, 2021 08:43

Merge branch 'master' into feature/measures

dfaca94

Update CHANGELOG.md

3aaa6f2

Update CHANGELOG.md

c435afe

Merge branch 'master' into feature/measures

97e4844

Trigger CI

0b15d96

Merge branch 'feature/measures' of github.com:PennyLaneAI/pennylane-l…

701ea6c

…ightning into feature/measures

AmintorDusko merged commit ab533a0 into master Dec 15, 2021

AmintorDusko deleted the feature/measures branch December 15, 2021 17:47

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Direct Support to probabilities, expectation values and variances in PL-Lightning #185

Direct Support to probabilities, expectation values and variances in PL-Lightning #185

AmintorDusko commented Dec 3, 2021 •

edited

Loading

github-actions bot commented Dec 3, 2021

codecov bot commented Dec 3, 2021 •

edited

Loading

github-actions bot commented Dec 3, 2021 •

edited

Loading

mlxd left a comment

AmintorDusko commented Dec 6, 2021

AmintorDusko commented Dec 6, 2021

AmintorDusko commented Dec 9, 2021

mlxd left a comment

mlxd Dec 10, 2021

AmintorDusko Dec 10, 2021

AmintorDusko Dec 10, 2021 •

edited

Loading

mlxd Dec 13, 2021

AmintorDusko Dec 13, 2021

AmintorDusko Dec 13, 2021

AmintorDusko commented Dec 10, 2021

mlxd commented Dec 14, 2021

mlxd left a comment

mlxd Dec 14, 2021

AmintorDusko commented Dec 14, 2021

		// transposed_tensor[ind] = tensor[transposed_state_index(ind,
		// new_axes)];

Direct Support to probabilities, expectation values and variances in PL-Lightning #185

Direct Support to probabilities, expectation values and variances in PL-Lightning #185

Conversation

AmintorDusko commented Dec 3, 2021 • edited Loading

github-actions bot commented Dec 3, 2021

codecov bot commented Dec 3, 2021 • edited Loading

Codecov Report

github-actions bot commented Dec 3, 2021 • edited Loading

Test Report (C++) on Ubuntu

mlxd left a comment

Choose a reason for hiding this comment

AmintorDusko commented Dec 6, 2021

AmintorDusko commented Dec 6, 2021

AmintorDusko commented Dec 9, 2021

mlxd left a comment

Choose a reason for hiding this comment

mlxd Dec 10, 2021

Choose a reason for hiding this comment

AmintorDusko Dec 10, 2021

Choose a reason for hiding this comment

AmintorDusko Dec 10, 2021 • edited Loading

Choose a reason for hiding this comment

mlxd Dec 13, 2021

Choose a reason for hiding this comment

AmintorDusko Dec 13, 2021

Choose a reason for hiding this comment

AmintorDusko Dec 13, 2021

Choose a reason for hiding this comment

AmintorDusko commented Dec 10, 2021

mlxd commented Dec 14, 2021

mlxd left a comment

Choose a reason for hiding this comment

mlxd Dec 14, 2021

Choose a reason for hiding this comment

AmintorDusko commented Dec 14, 2021

AmintorDusko commented Dec 3, 2021 •

edited

Loading

codecov bot commented Dec 3, 2021 •

edited

Loading

github-actions bot commented Dec 3, 2021 •

edited

Loading

AmintorDusko Dec 10, 2021 •

edited

Loading