Merge pull request #68 from qiboteam/sparse

Support sparse matrices in cast methods

src/qibojit/custom_operators/__init__.py

@@ -114,6 +114,7 @@ def set_platform(self, name): # pragma: no cover
         self.Tensor = xp.ndarray
         self.random = xp.random
         self.newaxis = xp.newaxis
+        self.sparse = self.platform.sparse
         # enable multi-GPU if no macos
         import sys
         if sys.platform != "darwin":
@@ -142,15 +143,16 @@ def set_threads(self, nthreads):
     def to_numpy(self, x):
         if isinstance(x, self.np.ndarray):
             return x
-        elif self.platform.name in ("cupy", "cuquantum")  and isinstance(x, self.platform.cp.ndarray):  # pragma: no cover
-            return x.get()
-        return self.np.array(x)
+        return self.platform.to_numpy(x)
 
     def cast(self, x, dtype='DTYPECPX'):
         if isinstance(dtype, str):
             dtype = self.dtypes(dtype)
         return self.platform.cast(x, dtype=dtype)
 
+    def issparse(self, x):
+        return self.platform.issparse(x)
+
     def check_shape(self, shape):
         if self.platform.name in ("cupy", "cuquantum")  and isinstance(shape, self.Tensor): # pragma: no cover
             shape = shape.get()
@@ -176,18 +178,19 @@ def expm(self, x):
             return self.backend.asarray(super().expm(x))
         return super().expm(x)
 
-    def eigh(self, x):
-        if self.platform.name == "cupy" and self.platform.is_hip: # pragma: no cover
-            # FIXME: Fallback to numpy because eigh is not implemented in rocblas
-            result = self.np.linalg.eigh(self.to_numpy(x))
-            return self.cast(result[0]), self.cast(result[1])
-        return super().eigh(x)
+    def eigh(self, x, k=6):
+        return self.platform.eigh(x, k)
 
-    def eigvalsh(self, x):
-        if self.platform.name == "cupy" and self.platform.is_hip: # pragma: no cover
+    def eigvalsh(self, x, k=6):
+        if self.issparse(x):
+            log.warning("Calculating sparse matrix eigenvectors because "
+                        "sparse modules do not provide ``eigvals`` method.")
+            return self.eigh(x, k=k)[0]
+        elif self.platform.name == "cupy" and self.platform.is_hip: # pragma: no cover
             # FIXME: Fallback to numpy because eigvalsh is not implemented in rocblas
             return self.cast(self.np.linalg.eigvalsh(self.to_numpy(x)))
-        return super().eigvalsh(x)
+        else:
+            return self.np.linalg.eigvalsh(x)
 
     def unique(self, x, return_counts=False):
         if self.platform.name in ("cupy", "cuquantum"):  # pragma: no cover
@@ -339,14 +342,6 @@ def transpose_state(self, pieces, state, nqubits, order):
         state = self._constructed_platforms.get("numba").transpose_state(pieces, state, nqubits, order)
         return self.reshape(state, original_shape)
 
-    def assert_allclose(self, value, target, rtol=1e-7, atol=0.0):
-        if self.platform.name in ("cupy", "cuquantum"): # pragma: no cover
-            if isinstance(value, self.backend.ndarray):
-                value = value.get()
-            if isinstance(target, self.backend.ndarray):
-                target = target.get()
-        self.np.testing.assert_allclose(value, target, rtol=rtol, atol=atol)
-
     def apply_gate(self, state, gate, nqubits, targets, qubits=None):
         return self.platform.one_qubit_base(state, nqubits, *targets, "apply_gate", gate, qubits)
 

src/qibojit/custom_operators/platforms.py

@@ -7,13 +7,22 @@ def __init__(self): # pragma: no cover
         self.name = "abstract"
         self.gates = None
         self.ops = None
+        self.sparse = None
         self.test_regressions = {}
         self.supports_multigpu = False
 
     @abstractmethod
     def cast(self, x, dtype=None, order=None): # pragma: no cover
         raise NotImplementedError
 
+    @abstractmethod
+    def to_numpy(self, x): # pragma: no cover
+        raise NotImplementedError
+
+    @abstractmethod
+    def issparse(self, x): # pragma: no cover
+        raise NotImplementedError
+
     @abstractmethod
     def one_qubit_base(self, state, nqubits, target, kernel, gate, qubits=None): # pragma: no cover
         raise NotImplementedError
@@ -65,13 +74,16 @@ def __init__(self):
             4: self.gates.apply_four_qubit_gate_kernel,
             5: self.gates.apply_five_qubit_gate_kernel
             }
+        from scipy import sparse
+        self.sparse = sparse
 
     def cast(self, x, dtype=None, order='K'):
+        if dtype is None:
+            dtype = x.dtype
         if isinstance(x, self.np.ndarray):
-            if dtype is None:
-                return x
-            else:
-                return x.astype(dtype, copy=False, order=order)
+            return x.astype(dtype, copy=False, order=order)
+        elif self.sparse.issparse(x):
+            return x.astype(dtype, copy=False)
         else:
             try:
                 x = self.np.array(x, dtype=dtype, order=order)
@@ -81,6 +93,22 @@ def cast(self, x, dtype=None, order='K'):
                 x = self.np.array(x.get(), dtype=dtype, copy=False, order=order)
             return x
 
+    def to_numpy(self, x):
+        if self.sparse.issparse(x):
+            return x.toarray()
+        return self.np.array(x, copy=False)
+
+    def issparse(self, x):
+        return self.sparse.issparse(x)
+
+    def eigh(self, x, k=6):
+        if self.issparse(x):
+            if k < x.shape[0]:
+                from scipy.sparse.linalg import eigsh
+                return eigsh(x, k=k, which='SA')
+            x = self.to_numpy(x)
+        return self.np.linalg.eigh(x)
+
     def one_qubit_base(self, state, nqubits, target, kernel, gate, qubits=None):
         ncontrols = len(qubits) - 1 if qubits is not None else 0
         m = nqubits - target - 1
@@ -172,6 +200,9 @@ def __init__(self):
         self.is_hip = cupy_backends.cuda.api.runtime.is_hip
         self.KERNELS = ("apply_gate", "apply_x", "apply_y", "apply_z", "apply_z_pow",
                         "apply_two_qubit_gate", "apply_fsim", "apply_swap")
+        from scipy import sparse
+        self.npsparse = sparse
+        self.sparse = cp.sparse
         if self.is_hip:  # pragma: no cover
             self.test_regressions = {
                 "test_measurementresult_apply_bitflips": [
@@ -246,13 +277,47 @@ def calculate_blocks(self, nstates, block_size=DEFAULT_BLOCK_SIZE):
         return nblocks, block_size
 
     def cast(self, x, dtype=None, order='C'):
+        if dtype is None:
+            dtype = x.dtype
         if isinstance(x, self.cp.ndarray):
-            if dtype is None:
-                return x
+            return x.astype(dtype, copy=False, order=order)
+        elif self.sparse.issparse(x):
+            if dtype != x.dtype:
+                return x.astype(dtype)
             else:
-                return x.astype(dtype, copy=False, order=order)
+                return x
+        elif self.npsparse.issparse(x):
+            cls = getattr(self.sparse, x.__class__.__name__)
+            return cls(x, dtype=dtype)
         return self.cp.asarray(x, dtype=dtype, order=order)
 
+    def to_numpy(self, x):
+        if isinstance(x, self.cp.ndarray):
+            return x.get()
+        elif self.sparse.issparse(x):
+            return x.toarray().get()
+        elif self.npsparse.issparse(x):
+            return x.toarray()
+        return self.np.array(x, copy=False)
+
+    def issparse(self, x):
+        return self.sparse.issparse(x) or self.npsparse.issparse(x)
+
+    def eigh(self, x, k=6):
+        if self.issparse(x):
+            if k < x.shape[0]:
+                # Fallback to numpy because cupy's ``sparse.eigh`` does not support 'SA'
+                from scipy.sparse.linalg import eigsh  # pylint: disable=import-error
+                result = eigsh(x.get(), k=k, which='SA')
+                return self.cast(result[0]), self.cast(result[1])
+            x = x.toarray()
+        if self.is_hip:
+            # Fallback to numpy because eigh is not implemented in rocblas
+            result = self.np.linalg.eigh(self.to_numpy(x))
+            return self.cast(result[0]), self.cast(result[1])
+        else:
+            return self.cp.linalg.eigh(x)
+
     def get_kernel_type(self, state):
         if state.dtype == self.cp.complex128:
             return "double"

src/qibojit/tests/test_platforms.py

@@ -34,6 +34,22 @@ def test_cast(platform, array_type):
     K.assert_allclose(final, target)
 
 
+@pytest.mark.parametrize("array_type", [None, "float32", "float64"])
+@pytest.mark.parametrize("format", ["coo", "csr", "csc", "dia"])
+def test_sparse_cast(platform, array_type, format):
+    from scipy import sparse
+    sptarget = sparse.rand(512, 512, dtype=array_type, format=format)
+    assert K.issparse(sptarget)
+    final = K.to_numpy(K.cast(sptarget))
+    target = sptarget.toarray()
+    K.assert_allclose(final, target)
+    if K.platform.name != "numba":  # pragma: no cover
+        sptarget = getattr(K.sparse, sptarget.__class__.__name__)(sptarget)
+        assert K.issparse(sptarget)
+        final = K.to_numpy(K.cast(sptarget))
+        K.assert_allclose(final, target)
+
+
 def test_to_numpy(platform):
     x = [0, 1, 2]
     target = K.to_numpy(K.cast(x))
@@ -53,19 +69,50 @@ def test_basic_matrices(platform):
     K.assert_allclose(K.expm(m), expm(m))
 
 
-def test_backend_eigh(platform):
-    m = np.random.random((16, 16))
-    eigvals2, eigvecs2 = np.linalg.eigh(m)
-    eigvals1, eigvecs1 = K.eigh(K.cast(m))
-    K.assert_allclose(eigvals1, eigvals2)
-    K.assert_allclose(K.abs(eigvecs1), np.abs(eigvecs2))
-
-
-def test_backend_eigvalsh(platform):
-    m = np.random.random((16, 16))
-    target = np.linalg.eigvalsh(m)
-    result = K.eigvalsh(K.cast(m))
-    K.assert_allclose(target, result)
+@pytest.mark.parametrize("sparse_type", [None, "coo", "csr", "csc", "dia"])
+def test_backend_eigh(platform, sparse_type):
+    if sparse_type is None:
+        m = np.random.random((16, 16))
+        eigvals1, eigvecs1 = K.eigh(K.cast(m))
+        eigvals2, eigvecs2 = np.linalg.eigh(m)
+    else:
+        from scipy.sparse import rand
+        m = rand(16, 16, format=sparse_type)
+        m = m + m.T
+        eigvals1, eigvecs1 = K.eigh(K.cast(m), k=16)
+        eigvals2, eigvecs2 = K.eigh(K.cast(m.toarray()))
+    K.assert_allclose(eigvals1, eigvals2, atol=1e-10)
+    K.assert_allclose(K.abs(eigvecs1), np.abs(eigvecs2), atol=1e-10)
+
+
+@pytest.mark.parametrize("sparse_type", [None, "coo", "csr", "csc", "dia"])
+def test_backend_eigvalsh(platform, sparse_type):
+    if sparse_type is None:
+        m = np.random.random((16, 16))
+        target = np.linalg.eigvalsh(m)
+        result = K.eigvalsh(K.cast(m))
+    else:
+        from scipy.sparse import rand
+        m = rand(16, 16, format=sparse_type)
+        m = m + m.T
+        result = K.eigvalsh(K.cast(m), k=16)
+        target, _ = K.eigh(K.cast(m.toarray()))
+    K.assert_allclose(target, result, atol=1e-10)
+
+
+@pytest.mark.parametrize("sparse_type", ["coo", "csr", "csc", "dia"])
+@pytest.mark.parametrize("k", [6, 8])
+def test_backend_eigh_sparse(platform, sparse_type, k):
+    from scipy.sparse.linalg import eigsh
+    from scipy import sparse
+    from qibo import hamiltonians
+    ham = hamiltonians.TFIM(6, h=1.0)
+    m = getattr(sparse, f"{sparse_type}_matrix")(K.to_numpy(ham.matrix))
+    eigvals1, eigvecs1 = K.eigh(K.cast(m), k)
+    eigvals2, eigvecs2 = eigsh(m, k, which='SA')
+    eigvals1 = K.to_numpy(eigvals1)
+    eigvals2 = K.to_numpy(eigvals2)
+    K.assert_allclose(sorted(eigvals1), sorted(eigvals2))
 
 
 def test_unique_and_gather(platform):