Implement GPU tensor network simulator by using cuTensorNet #1637

doichanj · 2022-10-28T10:05:13Z

Summary

This is implementation of a new simulation method tensor_network for device=GPU accelerated by cuQuantum SDK.
tensor_network method is tensor network contraction based simulator for general quantum circuits.

Details and comments

tensor_network method may used for quantum circuit simulation with large number of qubits, but has fewer circuit depth.
This simulation method simulates quantum circuits by contracting tensor network built from connections of gates of input circuit. Tensor network contraction can be parallelize by using slicing technique optimized in cuTensorNet SDK.

This is performance results of tensor_network method compared to statevector using QAOA circuit + save expectation value (using random connected graph, connection ratio = 0.1) accelerated by single NDIVIA Tesla V100 (32GB)

This PR requires shot branching implementation PR #1596
Currently there is no implementation for device=CPU

tensor_network method uses density matrix in tensor network to get norm of statevector.
So it also supports operations for density_matrix (can simulate noise sampling in single shot)
Because using density matrix inside, number of qubits for output of contraction is limited to memory size similar to density matrix.
(So we can not get large statevector/reduced density matrix output)

To simulate quantum circuits with large number of qubits, expectation values, measure or sampling can be used. (Maybe sampling is very slow for random results)

This simulator supports parallelization by multiple-GPUs or multiple nodes on cluster (MPI), tensor network contraction is parallelized.

hhorii · 2023-01-24T12:17:35Z

test/terra/backends/aer_simulator/test_conditional.py

@@ -78,9 +82,11 @@ def test_conditional_gates_64bit(self, method, device):
    @supported_methods(SUPPORTED_METHODS)
    def test_conditional_gates_132bit(self, method, device):
        """Test conditional gate operations on 132-bit conditional register."""
-        shots = 100


I believe that this line is necessary. shots can be used without assignment, and then tests fail.

I fixed test cases

hhorii · 2023-01-24T13:11:49Z

src/simulators/state.hpp

@@ -229,6 +229,11 @@ class Base {
  //set creg bit counts before initialize creg
  virtual void set_num_creg_bits(uint_t num_memory, uint_t num_register){}

+  void has_statevector_ops(bool flg)


This method name looks odd. State does not have operations (ops). Maybe enable_save_statevec is better (though save_amp is also enabled).

Hmmm, has_statevector_ops_ is just for AER::TensorNetwork::State. I think It should be a method of the class.

I moved this to state in tensor network and changed name enable_density_matrix

hhorii · 2023-01-24T13:25:29Z

src/simulators/tensor_network/tensor.hpp

+  int base_;
+  int rank_;
+  std::vector<int64_t> extents_;
+  int64_t size_;
+  std::vector<std::complex<data_t>> tensor_;
+  std::vector<int32_t> modes_;
+  reg_t qubits_;
+  bool sp_tensor_;


could you leave some comments on each member? Maybe tensor_ is data of multi-dimensional array defined rank_ and modes_.
Also, why int64_t and int32_t are used here? Can they be replaced with uint64_t and uint32_t? Or can they be negative?

comment added

Integer types defined in here is supported types in cuTensorNet, the same type is used to avoid data conversion.

qiskit_aer/backends/aer_simulator.py

hhorii · 2023-01-24T13:37:32Z

This is a new feature. A release note is necessary. I think the note should include some simple samples to use tensor_network method. I believe it can be available only if the system has NVIDIA GPU and cuQuantum. These information should be documented in release note also.

I'm wondering what will be happen if tensor_network is specified in no cuQuantum environment. I guess an error will be raised. Right?

…tensor_network

doichanj · 2023-01-25T07:22:57Z

I added exception when method=tensor_network, and device=CPU, but this causes issue in getting available methods in aer_simulator.py because available_methods only tests methods on CPU. So I modified to search for all the supported devices to return tensor_network if GPU is available.

hhorii · 2023-01-27T06:02:24Z

cmake/conan_utils.cmake

@@ -10,7 +10,7 @@ macro(setup_conan)
    # Right now every dependency shall be static
    set(CONAN_OPTIONS ${CONAN_OPTIONS} "*:shared=False")

-    set(REQUIREMENTS nlohmann_json/3.1.1 spdlog/1.9.2)
+    set(REQUIREMENTS nlohmann_json/3.1.1 spdlog/1.5.0)


is there a reason of this downgrade?

There is an inconsistency of version to https://github.com/doichanj/qiskit-aer/blob/24df5135cc06d9f9dd839d985b46bd91b2511232/cmake/dependency_utils.cmake#L27 and this causes build error on Power systems.

hhorii

Now this looks fine. I would like to show some tutorial to show good examples for this new method.

Implement tensor network simulator by using cuTensorNet

c7cbad0

doichanj requested a review from hhorii October 28, 2022 10:05

doichanj added 10 commits October 28, 2022 19:05

Merge branch 'main' into tensor_network

003a18d

fix aer_simulator.py

cc32976

return values in TensorNetContractorDummy

7009bb4

merge qiskit-aer/main

71ed780

Fix for cuQuantum 22.11.0

b681594

resolve conflict

0b661db

Merge branch 'main' into tensor_network

fdc7090

Merge branch 'main' into tensor_network

9b83484

remove dependencieis to shot branching

93b64c3

Merge branch 'main' into tensor_network

296499c

hhorii reviewed Jan 24, 2023

View reviewed changes

qiskit_aer/backends/aer_simulator.py Outdated Show resolved Hide resolved

doichanj added 3 commits January 25, 2023 15:07

fix test cases, remove remaining shot-branching options

e4ca42a

Merge branch 'tensor_network' of github.com:doichanj/qiskit-aer into …

92c6575

…tensor_network

add release note, fix available_methods to handle tensor_network

170dce5

fix available_methods

24df513

hhorii reviewed Jan 27, 2023

View reviewed changes

doichanj and others added 2 commits January 27, 2023 16:03

spdlog version

419d075

Merge branch 'main' into tensor_network

f6e1196

hhorii added this to the Aer 0.12.0 milestone Jan 27, 2023

hhorii approved these changes Jan 30, 2023

View reviewed changes

hhorii merged commit 3e34a24 into Qiskit:main Jan 30, 2023

wshanks mentioned this pull request Feb 12, 2023

Add qiskit-aer (and muparserx) conda-forge/staged-recipes#21404

Merged

10 tasks

HaoTy mentioned this pull request Apr 24, 2023

tensor_network method powered by cuQuantum is not available #1790

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement GPU tensor network simulator by using cuTensorNet #1637

Implement GPU tensor network simulator by using cuTensorNet #1637

doichanj commented Oct 28, 2022 •

edited

Loading

hhorii Jan 24, 2023

doichanj Jan 25, 2023 •

edited

Loading

hhorii Jan 24, 2023

hhorii Jan 24, 2023

doichanj Jan 25, 2023

hhorii Jan 24, 2023

doichanj Jan 25, 2023

hhorii commented Jan 24, 2023

doichanj commented Jan 25, 2023

hhorii Jan 27, 2023

doichanj Jan 27, 2023

hhorii left a comment

Implement GPU tensor network simulator by using cuTensorNet #1637

Implement GPU tensor network simulator by using cuTensorNet #1637

Conversation

doichanj commented Oct 28, 2022 • edited Loading

Summary

Details and comments

Choose a reason for hiding this comment

doichanj Jan 25, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hhorii commented Jan 24, 2023

doichanj commented Jan 25, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hhorii left a comment

Choose a reason for hiding this comment

doichanj commented Oct 28, 2022 •

edited

Loading

doichanj Jan 25, 2023 •

edited

Loading