Extension for CUDA.jl

Introduction

This is an extension to support QuantumObject.data conversion from standard dense and sparse CPU arrays to GPU (CUDA.jl) arrays.

This extension will be automatically triggered if QuantumToolbox.jl, CUDACore, and cuSPARSE are all loaded:

using QuantumToolbox
using CUDACore, cuSPARSE

User can also trigger the extension by importing the entire CUDA package, since it internally loads both libraries (CUDACore and cuSPARSE):

using QuantumToolbox
using CUDA, CUDA.cuSPARSE

We wrapped several functions in CUDACore and cuSPARSE in order to not only converting QuantumObject.data into GPU arrays, but also changing the element type and word size (32 and 64) since some of the GPUs perform better in 32-bit. The functions are listed as follows (where input A is a QuantumObject):

• cu(A; word_size=64): return a new QuantumObject with CUDA arrays and specified word_size.

• CuArray(A): If A.data is a dense array, return a new QuantumObject with CUDACore.CuArray.

• CuArray{T}(A): If A.data is a dense array, return a new QuantumObject with CUDACore.CuArray under element type T.

• CuSparseVector(A): If A.data is a sparse vector, return a new QuantumObject with cuSPARSE.CuSparseVector.

• CuSparseVector{T}(A): If A.data is a sparse vector, return a new QuantumObject with cuSPARSE.CuSparseVector under element type T.

• CuSparseMatrixCSC(A): If A.data is a sparse matrix, return a new QuantumObject with cuSPARSE.CuSparseMatrixCSC.

• CuSparseMatrixCSC{T}(A): If A.data is a sparse matrix, return a new QuantumObject with cuSPARSE.CuSparseMatrixCSC under element type T.

• CuSparseMatrixCSR(A): If A.data is a sparse matrix, return a new QuantumObject with cuSPARSE.CuSparseMatrixCSR.

• CuSparseMatrixCSR{T}(A): If A.data is a sparse matrix, return a new QuantumObject with cuSPARSE.CuSparseMatrixCSR under element type T.

We suggest to convert the arrays from CPU to GPU memory by using the function cu because it allows different data-types of input QuantumObject.

Here are some examples:

Converting dense arrays

V = fock(2, 0) # CPU dense vector

Quantum Object:   type=Ket()   dims=[2]   size=(2,)
2-element Vector{ComplexF64}:
 1.0 + 0.0im
 0.0 + 0.0im

cu(V)

Quantum Object:   type=Ket()   dims=[2]   size=(2,)
2-element CuArray{ComplexF64, 1, CUDA.DeviceMemory}:
 1.0 + 0.0im
 0.0 + 0.0im

cu(V; word_size = 32)

Quantum Object:   type=Ket()   dims=[2]   size=(2,)
2-element CuArray{ComplexF32, 1, CUDA.DeviceMemory}:
 1.0 + 0.0im
 0.0 + 0.0im

M = Qobj([1 2; 3 4]) # CPU dense matrix

Quantum Object:   type=Operator()   dims=[2]   size=(2, 2)   ishermitian=false
2×2 Matrix{Int64}:
 1  2
 3  4

cu(M)

Quantum Object:   type=Operator()   dims=[2]   size=(2, 2)   ishermitian=false
2×2 CuArray{Int64, 2, CUDA.DeviceMemory}:
 1  2
 3  4

cu(M; word_size = 32)

Quantum Object:   type=Operator()   dims=[2]   size=(2, 2)   ishermitian=false
2×2 CuArray{Int32, 2, CUDA.DeviceMemory}:
 1  2
 3  4

Converting sparse arrays

V = fock(2, 0; sparse=true) # CPU sparse vector

Quantum Object:   type=Ket()   dims=[2]   size=(2,)
2-element SparseVector{ComplexF64, Int64} with 1 stored entry:
  [1]  =  1.0+0.0im

cu(V)

Quantum Object:   type=Ket()   dims=[2]   size=(2,)
2-element CuSparseVector{ComplexF64, Int32} with 1 stored entry:
  [1]  =  1.0+0.0im

cu(V; word_size = 32)

Quantum Object:   type=Ket()   dims=[2]   size=(2,)
2-element CuSparseVector{ComplexF32, Int32} with 1 stored entry:
  [1]  =  1.0+0.0im

M = sigmax() # CPU sparse matrix

Quantum Object:   type=Operator()   dims=[2]   size=(2, 2)   ishermitian=true
2×2 SparseMatrixCSC{ComplexF64, Int64} with 2 stored entries:
     ⋅      1.0+0.0im
 1.0+0.0im      ⋅

cu(M)

Quantum Object:   type=Operator()   dims=[2]   size=(2, 2)   ishermitian=true
2×2 CuSparseMatrixCSC{ComplexF64, Int32} with 2 stored entries:
     ⋅      1.0+0.0im
 1.0+0.0im      ⋅

cu(M; word_size = 32)

Quantum Object:   type=Operator()   dims=[2]   size=(2, 2)   ishermitian=true
2×2 CuSparseMatrixCSC{ComplexF32, Int32} with 2 stored entries:
     ⋅      1.0+0.0im
 1.0+0.0im      ⋅