340 lines
10 KiB
Python
340 lines
10 KiB
Python
|
import sys
|
||
|
|
import os
|
||
|
|
import gc
|
||
|
|
import threading
|
||
|
|
|
||
|
|
import numpy as np
|
||
|
|
from numpy.testing import assert_equal, assert_, assert_allclose
|
||
|
|
from scipy.sparse import (_sparsetools, coo_matrix, csr_matrix, csc_matrix,
|
||
|
|
bsr_matrix, dia_matrix)
|
||
|
|
from scipy.sparse._sputils import supported_dtypes
|
||
|
|
from scipy._lib._testutils import check_free_memory
|
||
|
|
|
||
|
|
import pytest
|
||
|
|
from pytest import raises as assert_raises
|
||
|
|
|
||
|
|
|
||
|
|
def int_to_int8(n):
|
||
|
|
"""
|
||
|
|
Wrap an integer to the interval [-128, 127].
|
||
|
|
"""
|
||
|
|
return (n + 128) % 256 - 128
|
||
|
|
|
||
|
|
|
||
|
|
def test_exception():
|
||
|
|
assert_raises(MemoryError, _sparsetools.test_throw_error)
|
||
|
|
|
||
|
|
|
||
|
|
def test_threads():
|
||
|
|
# Smoke test for parallel threaded execution; doesn't actually
|
||
|
|
# check that code runs in parallel, but just that it produces
|
||
|
|
# expected results.
|
||
|
|
nthreads = 10
|
||
|
|
niter = 100
|
||
|
|
|
||
|
|
n = 20
|
||
|
|
a = csr_matrix(np.ones([n, n]))
|
||
|
|
bres = []
|
||
|
|
|
||
|
|
class Worker(threading.Thread):
|
||
|
|
def run(self):
|
||
|
|
b = a.copy()
|
||
|
|
for j in range(niter):
|
||
|
|
_sparsetools.csr_plus_csr(n, n,
|
||
|
|
a.indptr, a.indices, a.data,
|
||
|
|
a.indptr, a.indices, a.data,
|
||
|
|
b.indptr, b.indices, b.data)
|
||
|
|
bres.append(b)
|
||
|
|
|
||
|
|
threads = [Worker() for _ in range(nthreads)]
|
||
|
|
for thread in threads:
|
||
|
|
thread.start()
|
||
|
|
for thread in threads:
|
||
|
|
thread.join()
|
||
|
|
|
||
|
|
for b in bres:
|
||
|
|
assert_(np.all(b.toarray() == 2))
|
||
|
|
|
||
|
|
|
||
|
|
def test_regression_std_vector_dtypes():
|
||
|
|
# Regression test for gh-3780, checking the std::vector typemaps
|
||
|
|
# in sparsetools.cxx are complete.
|
||
|
|
for dtype in supported_dtypes:
|
||
|
|
ad = np.array([[1, 2], [3, 4]]).astype(dtype)
|
||
|
|
a = csr_matrix(ad, dtype=dtype)
|
||
|
|
|
||
|
|
# getcol is one function using std::vector typemaps, and should not fail
|
||
|
|
assert_equal(a.getcol(0).toarray(), ad[:, :1])
|
||
|
|
|
||
|
|
|
||
|
|
@pytest.mark.slow
|
||
|
|
@pytest.mark.xfail_on_32bit("Can't create large array for test")
|
||
|
|
def test_nnz_overflow():
|
||
|
|
# Regression test for gh-7230 / gh-7871, checking that coo_toarray
|
||
|
|
# with nnz > int32max doesn't overflow.
|
||
|
|
nnz = np.iinfo(np.int32).max + 1
|
||
|
|
# Ensure ~20 GB of RAM is free to run this test.
|
||
|
|
check_free_memory((4 + 4 + 1) * nnz / 1e6 + 0.5)
|
||
|
|
|
||
|
|
# Use nnz duplicate entries to keep the dense version small.
|
||
|
|
row = np.zeros(nnz, dtype=np.int32)
|
||
|
|
col = np.zeros(nnz, dtype=np.int32)
|
||
|
|
data = np.zeros(nnz, dtype=np.int8)
|
||
|
|
data[-1] = 4
|
||
|
|
s = coo_matrix((data, (row, col)), shape=(1, 1), copy=False)
|
||
|
|
# Sums nnz duplicates to produce a 1x1 array containing 4.
|
||
|
|
d = s.toarray()
|
||
|
|
|
||
|
|
assert_allclose(d, [[4]])
|
||
|
|
|
||
|
|
|
||
|
|
@pytest.mark.skipif(
|
||
|
|
not (sys.platform.startswith('linux') and np.dtype(np.intp).itemsize >= 8),
|
||
|
|
reason="test requires 64-bit Linux"
|
||
|
|
)
|
||
|
|
class TestInt32Overflow:
|
||
|
|
"""
|
||
|
|
Some of the sparsetools routines use dense 2D matrices whose
|
||
|
|
total size is not bounded by the nnz of the sparse matrix. These
|
||
|
|
routines used to suffer from int32 wraparounds; here, we try to
|
||
|
|
check that the wraparounds don't occur any more.
|
||
|
|
"""
|
||
|
|
# choose n large enough
|
||
|
|
n = 50000
|
||
|
|
|
||
|
|
def setup_method(self):
|
||
|
|
assert self.n**2 > np.iinfo(np.int32).max
|
||
|
|
|
||
|
|
# check there's enough memory even if everything is run at the
|
||
|
|
# same time
|
||
|
|
try:
|
||
|
|
parallel_count = int(os.environ.get('PYTEST_XDIST_WORKER_COUNT', '1'))
|
||
|
|
except ValueError:
|
||
|
|
parallel_count = np.inf
|
||
|
|
|
||
|
|
check_free_memory(3000 * parallel_count)
|
||
|
|
|
||
|
|
def teardown_method(self):
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
def test_coo_todense(self):
|
||
|
|
# Check *_todense routines (cf. gh-2179)
|
||
|
|
#
|
||
|
|
# All of them in the end call coo_matrix.todense
|
||
|
|
|
||
|
|
n = self.n
|
||
|
|
|
||
|
|
i = np.array([0, n-1])
|
||
|
|
j = np.array([0, n-1])
|
||
|
|
data = np.array([1, 2], dtype=np.int8)
|
||
|
|
m = coo_matrix((data, (i, j)))
|
||
|
|
|
||
|
|
r = m.todense()
|
||
|
|
assert_equal(r[0,0], 1)
|
||
|
|
assert_equal(r[-1,-1], 2)
|
||
|
|
del r
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
@pytest.mark.slow
|
||
|
|
def test_matvecs(self):
|
||
|
|
# Check *_matvecs routines
|
||
|
|
n = self.n
|
||
|
|
|
||
|
|
i = np.array([0, n-1])
|
||
|
|
j = np.array([0, n-1])
|
||
|
|
data = np.array([1, 2], dtype=np.int8)
|
||
|
|
m = coo_matrix((data, (i, j)))
|
||
|
|
|
||
|
|
b = np.ones((n, n), dtype=np.int8)
|
||
|
|
for sptype in (csr_matrix, csc_matrix, bsr_matrix):
|
||
|
|
m2 = sptype(m)
|
||
|
|
r = m2.dot(b)
|
||
|
|
assert_equal(r[0,0], 1)
|
||
|
|
assert_equal(r[-1,-1], 2)
|
||
|
|
del r
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
del b
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
@pytest.mark.slow
|
||
|
|
def test_dia_matvec(self):
|
||
|
|
# Check: huge dia_matrix _matvec
|
||
|
|
n = self.n
|
||
|
|
data = np.ones((n, n), dtype=np.int8)
|
||
|
|
offsets = np.arange(n)
|
||
|
|
m = dia_matrix((data, offsets), shape=(n, n))
|
||
|
|
v = np.ones(m.shape[1], dtype=np.int8)
|
||
|
|
r = m.dot(v)
|
||
|
|
assert_equal(r[0], int_to_int8(n))
|
||
|
|
del data, offsets, m, v, r
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
_bsr_ops = [pytest.param("matmat", marks=pytest.mark.xslow),
|
||
|
|
pytest.param("matvecs", marks=pytest.mark.xslow),
|
||
|
|
"matvec",
|
||
|
|
"diagonal",
|
||
|
|
"sort_indices",
|
||
|
|
pytest.param("transpose", marks=pytest.mark.xslow)]
|
||
|
|
|
||
|
|
@pytest.mark.slow
|
||
|
|
@pytest.mark.parametrize("op", _bsr_ops)
|
||
|
|
def test_bsr_1_block(self, op):
|
||
|
|
# Check: huge bsr_matrix (1-block)
|
||
|
|
#
|
||
|
|
# The point here is that indices inside a block may overflow.
|
||
|
|
|
||
|
|
def get_matrix():
|
||
|
|
n = self.n
|
||
|
|
data = np.ones((1, n, n), dtype=np.int8)
|
||
|
|
indptr = np.array([0, 1], dtype=np.int32)
|
||
|
|
indices = np.array([0], dtype=np.int32)
|
||
|
|
m = bsr_matrix((data, indices, indptr), blocksize=(n, n), copy=False)
|
||
|
|
del data, indptr, indices
|
||
|
|
return m
|
||
|
|
|
||
|
|
gc.collect()
|
||
|
|
try:
|
||
|
|
getattr(self, "_check_bsr_" + op)(get_matrix)
|
||
|
|
finally:
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
@pytest.mark.slow
|
||
|
|
@pytest.mark.parametrize("op", _bsr_ops)
|
||
|
|
def test_bsr_n_block(self, op):
|
||
|
|
# Check: huge bsr_matrix (n-block)
|
||
|
|
#
|
||
|
|
# The point here is that while indices within a block don't
|
||
|
|
# overflow, accumulators across many block may.
|
||
|
|
|
||
|
|
def get_matrix():
|
||
|
|
n = self.n
|
||
|
|
data = np.ones((n, n, 1), dtype=np.int8)
|
||
|
|
indptr = np.array([0, n], dtype=np.int32)
|
||
|
|
indices = np.arange(n, dtype=np.int32)
|
||
|
|
m = bsr_matrix((data, indices, indptr), blocksize=(n, 1), copy=False)
|
||
|
|
del data, indptr, indices
|
||
|
|
return m
|
||
|
|
|
||
|
|
gc.collect()
|
||
|
|
try:
|
||
|
|
getattr(self, "_check_bsr_" + op)(get_matrix)
|
||
|
|
finally:
|
||
|
|
gc.collect()
|
||
|
|
|
||
|
|
def _check_bsr_matvecs(self, m): # skip name check
|
||
|
|
m = m()
|
||
|
|
n = self.n
|
||
|
|
|
||
|
|
# _matvecs
|
||
|
|
r = m.dot(np.ones((n, 2), dtype=np.int8))
|
||
|
|
assert_equal(r[0, 0], int_to_int8(n))
|
||
|
|
|
||
|
|
def _check_bsr_matvec(self, m): # skip name check
|
||
|
|
m = m()
|
||
|
|
n = self.n
|
||
|
|
|
||
|
|
# _matvec
|
||
|
|
r = m.dot(np.ones((n,), dtype=np.int8))
|
||
|
|
assert_equal(r[0], int_to_int8(n))
|
||
|
|
|
||
|
|
def _check_bsr_diagonal(self, m): # skip name check
|
||
|
|
m = m()
|
||
|
|
n = self.n
|
||
|
|
|
||
|
|
# _diagonal
|
||
|
|
r = m.diagonal()
|
||
|
|
assert_equal(r, np.ones(n))
|
||
|
|
|
||
|
|
def _check_bsr_sort_indices(self, m): # skip name check
|
||
|
|
# _sort_indices
|
||
|
|
m = m()
|
||
|
|
m.sort_indices()
|
||
|
|
|
||
|
|
def _check_bsr_transpose(self, m): # skip name check
|
||
|
|
# _transpose
|
||
|
|
m = m()
|
||
|
|
m.transpose()
|
||
|
|
|
||
|
|
def _check_bsr_matmat(self, m): # skip name check
|
||
|
|
m = m()
|
||
|
|
n = self.n
|
||
|
|
|
||
|
|
# _bsr_matmat
|
||
|
|
m2 = bsr_matrix(np.ones((n, 2), dtype=np.int8), blocksize=(m.blocksize[1], 2))
|
||
|
|
m.dot(m2) # shouldn't SIGSEGV
|
||
|
|
del m2
|
||
|
|
|
||
|
|
# _bsr_matmat
|
||
|
|
m2 = bsr_matrix(np.ones((2, n), dtype=np.int8), blocksize=(2, m.blocksize[0]))
|
||
|
|
m2.dot(m) # shouldn't SIGSEGV
|
||
|
|
|
||
|
|
|
||
|
|
@pytest.mark.skip(reason="64-bit indices in sparse matrices not available")
|
||
|
|
def test_csr_matmat_int64_overflow():
|
||
|
|
n = 3037000500
|
||
|
|
assert n**2 > np.iinfo(np.int64).max
|
||
|
|
|
||
|
|
# the test would take crazy amounts of memory
|
||
|
|
check_free_memory(n * (8*2 + 1) * 3 / 1e6)
|
||
|
|
|
||
|
|
# int64 overflow
|
||
|
|
data = np.ones((n,), dtype=np.int8)
|
||
|
|
indptr = np.arange(n+1, dtype=np.int64)
|
||
|
|
indices = np.zeros(n, dtype=np.int64)
|
||
|
|
a = csr_matrix((data, indices, indptr))
|
||
|
|
b = a.T
|
||
|
|
|
||
|
|
assert_raises(RuntimeError, a.dot, b)
|
||
|
|
|
||
|
|
|
||
|
|
def test_upcast():
|
||
|
|
a0 = csr_matrix([[np.pi, np.pi*1j], [3, 4]], dtype=complex)
|
||
|
|
b0 = np.array([256+1j, 2**32], dtype=complex)
|
||
|
|
|
||
|
|
for a_dtype in supported_dtypes:
|
||
|
|
for b_dtype in supported_dtypes:
|
||
|
|
msg = f"({a_dtype!r}, {b_dtype!r})"
|
||
|
|
|
||
|
|
if np.issubdtype(a_dtype, np.complexfloating):
|
||
|
|
a = a0.copy().astype(a_dtype)
|
||
|
|
else:
|
||
|
|
a = a0.real.copy().astype(a_dtype)
|
||
|
|
|
||
|
|
if np.issubdtype(b_dtype, np.complexfloating):
|
||
|
|
b = b0.copy().astype(b_dtype)
|
||
|
|
else:
|
||
|
|
with np.errstate(invalid="ignore"):
|
||
|
|
# Casting a large value (2**32) to int8 causes a warning in
|
||
|
|
# numpy >1.23
|
||
|
|
b = b0.real.copy().astype(b_dtype)
|
||
|
|
|
||
|
|
if not (a_dtype == np.bool_ and b_dtype == np.bool_):
|
||
|
|
c = np.zeros((2,), dtype=np.bool_)
|
||
|
|
assert_raises(ValueError, _sparsetools.csr_matvec,
|
||
|
|
2, 2, a.indptr, a.indices, a.data, b, c)
|
||
|
|
|
||
|
|
if ((np.issubdtype(a_dtype, np.complexfloating) and
|
||
|
|
not np.issubdtype(b_dtype, np.complexfloating)) or
|
||
|
|
(not np.issubdtype(a_dtype, np.complexfloating) and
|
||
|
|
np.issubdtype(b_dtype, np.complexfloating))):
|
||
|
|
c = np.zeros((2,), dtype=np.float64)
|
||
|
|
assert_raises(ValueError, _sparsetools.csr_matvec,
|
||
|
|
2, 2, a.indptr, a.indices, a.data, b, c)
|
||
|
|
|
||
|
|
c = np.zeros((2,), dtype=np.result_type(a_dtype, b_dtype))
|
||
|
|
_sparsetools.csr_matvec(2, 2, a.indptr, a.indices, a.data, b, c)
|
||
|
|
assert_allclose(c, np.dot(a.toarray(), b), err_msg=msg)
|
||
|
|
|
||
|
|
|
||
|
|
def test_endianness():
|
||
|
|
d = np.ones((3,4))
|
||
|
|
offsets = [-1,0,1]
|
||
|
|
|
||
|
|
a = dia_matrix((d.astype('<f8'), offsets), (4, 4))
|
||
|
|
b = dia_matrix((d.astype('>f8'), offsets), (4, 4))
|
||
|
|
v = np.arange(4)
|
||
|
|
|
||
|
|
assert_allclose(a.dot(v), [1, 3, 6, 5])
|
||
|
|
assert_allclose(b.dot(v), [1, 3, 6, 5])
|