Add C++ adjoint diff methods #136

mlxd · 2021-08-19T17:01:13Z

Context: This PR will add C++ support for adjoint diff, based on arXiV: arXiV:2009.02823.

Description of the Change:

StateVectorManaged class added as a subclass of StateVector, where memory management is handled internally by the class.
AdjointDiff.hpp file added, with AdjointJacobian and helper classes to prepare and calculate the Jacobian for a given set of parameters and observables. For multiple observables, the calculations are handled independently by OpenMP threads and populate the Jacobian.
std::variant based functionality used to distinguish between parametric gates, non parametric gates, and arbitrary gates defined by a list of complex parameters.
MacOS wheelbuilder compiler updated to use brew-install clang for x86, as default compiler uses an older version of the Xcode SDK not supporting std::variant.
Repo refactored into subdirectories for multiple separate components:
algorithms for algorithms using and to be used by the simulator; bindings for source files directly related to binding the C++ code with Python; simulator for source-files directly related to the simulator; util for utility functions to be used across the package;
Each subdirectory has now been compartmentalized with CMake, allowing separable compilation of each translation unit, enabling C++ system binaries as well as Python bindings to be built.
Extensive testing for all of the above.
Bug-fix for the CRY gate, which had an incorrect index.

Benefits: C++ backed adjoint method is much faster (4x-6x) than Python.

Possible Drawbacks: Requires modern C++17 knowledge to maintain. Design may require some later refactor.

Related GitHub Issues:

github-actions · 2021-08-19T17:01:29Z

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 · 2021-08-19T17:02:32Z

Codecov Report

Merging #136 (e1d569f) into master (9627d7f) will not change coverage.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##            master      #136    +/-   ##
==========================================
  Coverage   100.00%   100.00%            
==========================================
  Files            3         4     +1     
  Lines           71       182   +111     
==========================================
+ Hits            71       182   +111

Impacted Files	Coverage Δ
pennylane_lightning/_serialize.py	`100.00% <100.00%> (ø)`
pennylane_lightning/lightning_qubit.py	`100.00% <100.00%> (ø)`

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 9627d7f...e1d569f. Read the comment docs.

github-actions · 2021-08-19T17:02:36Z

Test Report (C++) on Ubuntu

      1 files ±0       1 suites ±0 0s ⏱️ ±0s
  327 tests ±0   327 ✔️ ±0 0 💤 ±0 0 ❌ ±0
1 185 runs ±0 1 185 ✔️ ±0 0 💤 ±0 0 ❌ ±0

Results for commit 6112aec. ± Comparison against base commit 6112aec.

♻️ This comment has been updated with latest results.

* Update tests * Add problematic statements * Fix tests * Fix type hinting * Make function as expected * Add quick return for no jacobian * Move position * Add raise for non supported ops * Ensure jacobian shape formatted as expected * Return jacobian directly as numpy array * Ensure correct return type of jacobian * Ensure param number tracking is correct for adjjac * Fix CRY gate index * All tests passing * Calculate parametric ops when creating opsdata object * Apply formatting * Ensure MacOS wheel builder uses a more modern compiler toolchain * Fix mac wheel builder Co-authored-by: Lee J. O'Riordan <loriordan@gmail.com>

mlxd · 2021-09-03T16:42:00Z

Just a note, CodeFactor seems to be caching old runs and using those --- please ignore its Action failure.

trbromley

Thanks @mlxd! Thanks for putting in so much awesome work to get this done! Everything looks great to me, my only question is the failing tests in #142.

trbromley · 2021-09-07T19:42:22Z

.github/workflows/wheel_macos_x86_64.yml

+          CC: /usr/local/opt/llvm/bin/clang
+          CXX: /usr/local/opt/llvm/bin/clang++


Is the CC option for C (i.e., the language)? Do we need it?

Good point, and you are right. Removing CC should be fine (though I'll test it out). There may be a way to avoid use of this too, though I must check this out first.

trbromley · 2021-09-07T19:48:33Z

pennylane_lightning/_serialize.py

+        wires_map (dict): a dictionary mapping input wires to the device's backend wires
+
+    Returns:
+        Tuple[list, list, list, list, list]: A serialization of the operations, containing a list


Suggested change

Tuple[list, list, list, list, list]: A serialization of the operations, containing a list

Tuple[Tuple[list, list, list, list, list], bool]: A serialization of the operations as a

tuple of lists and a flag indicating whether non-default state preparation is used in the

tape.

trbromley · 2021-09-07T19:48:51Z

pennylane_lightning/_serialize.py

+
+def _serialize_ops(
+    tape: QuantumTape, wires_map: dict
+) -> Tuple[List[List[str]], List[np.ndarray], List[List[int]], List[bool], List[np.ndarray]]:


Suggested change

) -> Tuple[List[List[str]], List[np.ndarray], List[List[int]], List[bool], List[np.ndarray]]:

) -> Tuple[Tuple[List[List[str]], List[np.ndarray], List[List[int]], List[bool], List[np.ndarray]], bool]:

While this should be fine, I wonder will it mess with the conversion to C++ (though I'd imagine pybind is clever enough that we do need to be explicit).

trbromley · 2021-09-07T19:53:12Z

pennylane_lightning/lightning_qubit.py

+        if not CPP_BINARY_AVAILABLE:
+            return super().adjoint_jacobian(tape, starting_state, use_device_state)


Do we need this? Since LightningQubit is reset to DefaultQubit if CPP_BINARY_AVAILABLE (see lines 205 and after)

.github/workflows/tests.yml

pennylane_lightning/src/algorithms/AdjointDiff.hpp

pennylane_lightning/src/bindings/Bindings.cpp

pennylane_lightning/src/simulator/StateVectorManaged.hpp

pennylane_lightning/src/util/Error.hpp

pennylane_lightning/lightning_qubit.py

tests/test_adjoint_jacobian.py

* Add test for dev pl * Add test * Move to original tests * Add additional ops creation method * Fix QFT errors * Reformat * Revert "Fix QFT errors" This reverts commit f62a763. * Preclude certain gates * Update test * Add double comments * Update test * Add tests for standard and custom wire labels * Remove print * Extend tests for RX,RY * Prevent trainable params from going negative * Ensure correct coef applied when using generator * Update Ops repr * Reenable tests * Update CI for non compiled builds * Add exception for Projector observable * Add Hermitian in tensor product * Ensure data is passed as complex128 to cpp * Run black * Ensure serialisation of numpy data is cast correctly * Remove redundant methods * Add support to get obs data * Fix serialize test array structures * Add hermitian checks * Ensure testing complete for current requirements * Reformat bindings * Skip test for Hermitian on default * Ensure bin available passed as class param * Skip for non binary tests Co-authored-by: Lee J. O'Riordan <loriordan@gmail.com>

trevor-vincent

Awesome work Lee, just two minor comments.

pennylane_lightning/src/tests/Test_AdjDiff.cpp

pennylane_lightning/src/algorithms/AdjointDiff.hpp

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

mlxd · 2021-09-13T09:20:36Z

Thanks for the reviews all. As per our chat on Friday I'm ignoring the codefactor report and merging this, as it still fails to remove the cached version.

trbromley · 2021-09-13T12:02:58Z

.github/CHANGELOG.md

+* PennyLane-Lightning now has support for the adjoint Jacobian method of http://arxiv.org/abs/2009.02823.
+  [(136)](https://github.com/PennyLaneAI/pennylane-lightning/pull/136)


For me this is a major new feature 🚀 🎉 We can probably move it up during release 😁

mlxd added 2 commits August 19, 2021 17:56

Add inner product routine with STL and BLAS options

fbdeedd

Add prelim adj diff implementation

a26cb1e

mlxd and others added 25 commits August 20, 2021 11:23

Refactor build-system structure to enable better extensibility

b0d49a2

Add explicit build support for algorithms

adbc267

Clean Util cmake

88a1595

Add explicit overload resolution to innerProd

ec9db68

Add build support for AdjointDiff

d47eb2c

Update testing for new utils

ca950b3

Ensure formatting is run in nested dirs

409759f

Add cout functionality for StateVector

b955bee

Add dot for conjugated and non conjugated data

140f950

Begin dev of adjoint diff tests

cf07cc9

Add additional tests to util suite

9e8ae62

Attempt debugging of multi-obs issues

aa19e98

Merge branch master into 7072_adj_diff

35ba63d

Ensure FPIC on for generated .a libs

edf6b1b

Fix native flag in CMake

4ea6c3d

Ensure propagation of flags to lower dirs

5859083

Update build flags

1d456fa

Add additional testing

3e91d07

Update tests

1a12eb9

Encapsulate ops and obs data for Python serialization

2317a94

Restructure the obs and ops data for adjoint calcs

2216b51

Lint and add defines

f5e029a

Fix testing binary location

ca706c2

Merge branch 'master' into 7072_adj_diff

81d41ba

Attempt to appease codefactor

5df74a0

mlxd and others added 5 commits September 1, 2021 15:47

Add variant support for observables from Python

61ccffc

Allow support for "Hermitian" observables

3626f2f

Add support for mixing of observable types

05bdacb

Fix compilation

380135a

mlxd marked this pull request as ready for review September 3, 2021 16:41

mlxd requested review from trbromley and trevor-vincent September 3, 2021 16:53

trbromley reviewed Sep 7, 2021

View reviewed changes

josh146 reviewed Sep 8, 2021

View reviewed changes

pennylane_lightning/lightning_qubit.py Show resolved Hide resolved

josh146 reviewed Sep 8, 2021

View reviewed changes

pennylane_lightning/lightning_qubit.py Outdated Show resolved Hide resolved

josh146 reviewed Sep 8, 2021

View reviewed changes

pennylane_lightning/lightning_qubit.py Show resolved Hide resolved

josh146 reviewed Sep 8, 2021

View reviewed changes

tests/test_adjoint_jacobian.py Show resolved Hide resolved

trevor-vincent suggested changes Sep 13, 2021

View reviewed changes

pennylane_lightning/src/tests/Test_AdjDiff.cpp Outdated Show resolved Hide resolved

pennylane_lightning/src/algorithms/AdjointDiff.hpp Show resolved Hide resolved

mlxd and others added 10 commits September 13, 2021 08:50

Update pennylane_lightning/src/algorithms/AdjointDiff.hpp

e0c4006

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

Update pennylane_lightning/src/simulator/StateVectorManaged.hpp

2ae4d7e

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

Update pennylane_lightning/src/util/Error.hpp

cf5ac07

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

Update pennylane_lightning/src/algorithms/AdjointDiff.hpp

9003139

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

Update pennylane_lightning/src/algorithms/AdjointDiff.hpp

7c51f2c

Co-authored-by: Tom Bromley <49409390+trbromley@users.noreply.github.com>

Fix typos and remove comments

e6f99d9

Update changelog

189a837

Fix codecov reduction

0441afd

Trigger build

04ee646

CodeCov fix

e1d569f

mlxd merged commit 6112aec into master Sep 13, 2021

mlxd deleted the 7072_adj_diff branch September 13, 2021 09:24

trbromley reviewed Sep 13, 2021

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add C++ adjoint diff methods #136

Add C++ adjoint diff methods #136

mlxd commented Aug 19, 2021 •

edited

Loading

github-actions bot commented Aug 19, 2021

codecov bot commented Aug 19, 2021 •

edited

Loading

github-actions bot commented Aug 19, 2021 •

edited

Loading

mlxd commented Sep 3, 2021

trbromley left a comment

trbromley Sep 7, 2021

mlxd Sep 8, 2021

trbromley Sep 7, 2021

trbromley Sep 7, 2021

mlxd Sep 8, 2021

trbromley Sep 7, 2021

trevor-vincent left a comment

mlxd commented Sep 13, 2021

trbromley Sep 13, 2021

		CC: /usr/local/opt/llvm/bin/clang
		CXX: /usr/local/opt/llvm/bin/clang++

-        Tuple[list, list, list, list, list]: A serialization of the operations, containing a list
+        Tuple[Tuple[list, list, list, list, list], bool]: A serialization of the operations as a
+        tuple of lists and a flag indicating whether non-default state preparation is used in the
+        tape.

	) -> Tuple[List[List[str]], List[np.ndarray], List[List[int]], List[bool], List[np.ndarray]]:
	) -> Tuple[Tuple[List[List[str]], List[np.ndarray], List[List[int]], List[bool], List[np.ndarray]], bool]:

		if not CPP_BINARY_AVAILABLE:
		return super().adjoint_jacobian(tape, starting_state, use_device_state)

		* PennyLane-Lightning now has support for the adjoint Jacobian method of http://arxiv.org/abs/2009.02823.
		[(136)](https://github.com/PennyLaneAI/pennylane-lightning/pull/136)

Add C++ adjoint diff methods #136

Add C++ adjoint diff methods #136

Conversation

mlxd commented Aug 19, 2021 • edited Loading

github-actions bot commented Aug 19, 2021

codecov bot commented Aug 19, 2021 • edited Loading

Codecov Report

github-actions bot commented Aug 19, 2021 • edited Loading

Test Report (C++) on Ubuntu

mlxd commented Sep 3, 2021

trbromley left a comment

Choose a reason for hiding this comment

trbromley Sep 7, 2021

Choose a reason for hiding this comment

mlxd Sep 8, 2021

Choose a reason for hiding this comment

trbromley Sep 7, 2021

Choose a reason for hiding this comment

trbromley Sep 7, 2021

Choose a reason for hiding this comment

mlxd Sep 8, 2021

Choose a reason for hiding this comment

trbromley Sep 7, 2021

Choose a reason for hiding this comment

trevor-vincent left a comment

Choose a reason for hiding this comment

mlxd commented Sep 13, 2021

trbromley Sep 13, 2021

Choose a reason for hiding this comment

mlxd commented Aug 19, 2021 •

edited

Loading

codecov bot commented Aug 19, 2021 •

edited

Loading

github-actions bot commented Aug 19, 2021 •

edited

Loading