def test_zero_MN_gemm(self):
# check other border cases...
e = gulinalg.matrix_multiply(np.zeros((0,2)), np.zeros((2,2)))
assert_allclose(e, np.zeros((0,2)))
f = gulinalg.matrix_multiply(np.zeros((2,2)), np.zeros((2,0)))
assert_allclose(f, np.zeros((2,0)))
def test_zero_MN_gemm(self):
# check other border cases...
e = gulinalg.matrix_multiply(np.zeros((0, 2)), np.zeros((2, 2)))
assert_allclose(e, np.zeros((0, 2)))
f = gulinalg.matrix_multiply(np.zeros((2, 2)), np.zeros((2, 0)))
assert_allclose(f, np.zeros((2, 0)))
def test_matrix_multiply_ff_f(self):
"""matrix multiply two FORTRAN layout matrices, explicit FORTRAN array output"""
a = np.asfortranarray(np.random.randn(M, N))
b = np.asfortranarray(np.random.randn(N, K))
res = np.zeros((M, K), order='F')
gulinalg.matrix_multiply(a, b, out=res)
ref = np.dot(a, b)
assert_allclose(res, ref)
def test_uninit_gemm(self):
a = np.zeros((298, 64))
b = np.zeros((64, 298))
c = np.zeros((298, 298), order='F')
c.fill(np.nan)
gulinalg.matrix_multiply(a, b, out=c)
assert not np.isnan(c).any()
def test_matrix_multiply_fc_f(self):
"""matrix multiply FORTRAN layout by C layout matrices, explicit FORTRAN array output"""
a = np.asfortranarray(np.random.randn(M,N))
b = np.ascontiguousarray(np.random.randn(N,K))
res = np.zeros((M,K), order='F')
gulinalg.matrix_multiply(a,b, out=res)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_uninit_gemm(self):
a = np.zeros((298, 64))
b = np.zeros((64, 298))
c = np.zeros((298, 298), order='F')
c.fill(np.nan)
gulinalg.matrix_multiply(a, b, out=c)
assert not np.isnan(c).any()
def test_matrix_multiply_ff_c(self):
"""matrix multiply two FORTRAN layout matrices, explicit C array output"""
a = np.asfortranarray(np.random.randn(M,N))
b = np.asfortranarray(np.random.randn(N,K))
res = np.zeros((M,K), order='C')
gulinalg.matrix_multiply(a,b, out=res)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_matrix_multiply_fc_c(self):
"""matrix multiply FORTRAN layout by C layout matrices, explicit C array output"""
a = np.asfortranarray(np.random.randn(M, N))
b = np.ascontiguousarray(np.random.randn(N, K))
res = np.zeros((M, K), order='C')
gulinalg.matrix_multiply(a, b, out=res)
ref = np.dot(a, b)
assert_allclose(res, ref)
def test_input_non_contiguous_1(self):
"""first input not contiguous"""
a = np.ascontiguousarray(np.random.randn(M, N, 2))[:, :, 0]
b = np.ascontiguousarray(np.random.randn(N, K))
res = np.zeros((M, K), order='C')
assert not a.flags.c_contiguous and not a.flags.f_contiguous
gulinalg.matrix_multiply(a, b, out=res)
ref = np.dot(a, b)
assert_allclose(res, ref)
def test_output_non_contiguous(self):
"""output not contiguous"""
a = np.ascontiguousarray(np.random.randn(M,N))
b = np.ascontiguousarray(np.random.randn(N,K))
res = np.zeros((M,K,2), order='C')[:,:,0]
assert not res.flags.c_contiguous and not res.flags.f_contiguous
gulinalg.matrix_multiply(a, b, out=res)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_input_non_contiguous_1(self):
"""first input not contiguous"""
a = np.ascontiguousarray(np.random.randn(M,N,2))[:,:,0]
b = np.ascontiguousarray(np.random.randn(N,K))
res = np.zeros((M,K), order='C')
assert not a.flags.c_contiguous and not a.flags.f_contiguous
gulinalg.matrix_multiply(a, b, out=res)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_output_non_contiguous(self):
"""output not contiguous"""
a = np.ascontiguousarray(np.random.randn(M, N))
b = np.ascontiguousarray(np.random.randn(N, K))
res = np.zeros((M, K, 2), order='C')[:, :, 0]
assert not res.flags.c_contiguous and not res.flags.f_contiguous
gulinalg.matrix_multiply(a, b, out=res)
ref = np.dot(a, b)
assert_allclose(res, ref)
def test_matrix_multiply_ff_f_batch(self):
"""matrix multiply two C layout matrices"""
n_outer = 4
for workers in [1, -1]:
a = np.random.randn(M, N)
b = np.random.randn(N, K)
ref = np.dot(a, b)
a = np.asfortranarray(np.stack((a, ) * n_outer, axis=0))
b = np.asfortranarray(np.stack((b, ) * n_outer, axis=0))
ref = np.stack((ref, ) * n_outer, axis=0)
res = np.zeros((n_outer, M, K), order='F')
gulinalg.matrix_multiply(a, b, workers=workers, out=res)
assert_allclose(res, ref)
def test_output_non_contiguous_batch(self):
"""output not contiguous"""
n_outer = 4
for workers in [1, -1]:
a = np.random.randn(M, N)
b = np.random.randn(N, K)
ref = np.dot(a, b)
a = np.ascontiguousarray(np.stack((a, ) * n_outer, axis=0))
b = np.ascontiguousarray(np.stack((b, ) * n_outer, axis=0))
ref = np.stack((ref, ) * n_outer, axis=0)
res = np.zeros((n_outer, M, K, 2), order='C')[..., 0]
assert not res.flags.c_contiguous and not res.flags.f_contiguous
gulinalg.matrix_multiply(a, b, workers=workers, out=res)
assert_allclose(res, ref)
def test_input_non_contiguous_3_batch(self):
"""neither input contiguous"""
n_outer = 4
for workers in [1, -1]:
a = np.random.randn(M, N, 2)
b = np.random.randn(N, K, 2)
ref = np.dot(a[..., 0], b[..., 0])
a = np.ascontiguousarray(np.stack((a, ) * n_outer, axis=0))[..., 0]
b = np.ascontiguousarray(np.stack((b, ) * n_outer, axis=0))[..., 0]
assert not a.flags.c_contiguous and not a.flags.f_contiguous
assert not b.flags.c_contiguous and not b.flags.f_contiguous
ref = np.stack((ref, ) * n_outer, axis=0)
res = np.zeros((n_outer, M, K), order='C')
gulinalg.matrix_multiply(a, b, workers=workers, out=res)
assert_allclose(res, ref)
def test_matrix_multiply_cc(self):
"""matrix multiply two C layout matrices"""
a = np.ascontiguousarray(np.random.randn(M,N))
b = np.ascontiguousarray(np.random.randn(N,K))
res = gulinalg.matrix_multiply(a,b)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_real_vector_f(self):
m = 10
rstate = np.random.RandomState(5)
a = rstate.randn(n_batch, m, m)
a = gulinalg.matrix_multiply(a, a.swapaxes(-2, -1)) # make symmetric
a = np.asfortranarray(a)
self._run_vector(a)
def test_matrix_multiply_ff(self):
"""matrix multiply two FORTRAN layout matrices"""
a = np.asfortranarray(np.random.randn(M,N))
b = np.asfortranarray(np.random.randn(N,K))
res = gulinalg.matrix_multiply(a,b)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_matrix_multiply_ff(self):
"""matrix multiply two FORTRAN layout matrices"""
a = np.asfortranarray(np.random.randn(M, N))
b = np.asfortranarray(np.random.randn(N, K))
res = gulinalg.matrix_multiply(a, b)
ref = np.dot(a, b)
assert_allclose(res, ref)
def test_matrix_multiply_cc(self):
"""matrix multiply two C layout matrices"""
a = np.ascontiguousarray(np.random.randn(M, N))
b = np.ascontiguousarray(np.random.randn(N, K))
res = gulinalg.matrix_multiply(a, b)
ref = np.dot(a, b)
assert_allclose(res, ref)
def test_complex_vector(self):
m = 10
rstate = np.random.RandomState(5)
a = rstate.randn(n_batch, m, m)
a = a + 1j * rstate.randn(n_batch, m, m)
a = gulinalg.matrix_multiply(
a, np.conj(a).swapaxes(-2, -1) # make symmetric
)
self._run_vector(a)
def test_zero_K_gemm(self):
a = np.zeros((2, 0))
b = np.zeros((0, 2))
c = np.empty((2, 2), order="C")
c.fill(np.nan)
# make sure that the result is set to zero
gulinalg.matrix_multiply(a, b, out=c)
assert_allclose(c, np.zeros((2, 2)))
# make sure that this works also when order is FORTRAN
c = np.empty((2, 2), order="F")
c.fill(np.nan)
gulinalg.matrix_multiply(a, b, out=c)
assert_allclose(c, np.zeros((2, 2)))
# check that the correct shape is created as well...
d = gulinalg.matrix_multiply(a, b)
assert_allclose(d, np.zeros((2, 2)))
def test_broadcast(self):
"""test broadcast matrix multiply"""
a = np.ascontiguousarray(np.random.randn(M, N))
b = np.ascontiguousarray(np.random.randn(10, N, K))
res = gulinalg.matrix_multiply(a,b)
assert res.shape == (10, M, K)
ref = np.stack([np.dot(a, b[i]) for i in range(len(b))])
assert_allclose(res, ref)
def test_broadcast(self):
"""test broadcast matrix multiply"""
a = np.ascontiguousarray(np.random.randn(M, N))
b = np.ascontiguousarray(np.random.randn(10, N, K))
res = gulinalg.matrix_multiply(a, b)
assert res.shape == (10, M, K)
ref = np.stack([np.dot(a, b[i]) for i in range(len(b))])
assert_allclose(res, ref)
def test_zero_K_gemm(self):
a = np.zeros((2,0))
b = np.zeros((0,2))
c = np.empty((2,2), order="C")
c.fill(np.nan)
# make sure that the result is set to zero
gulinalg.matrix_multiply(a, b, out=c)
assert_allclose(c, np.zeros((2,2)))
# make sure that this works also when order is FORTRAN
c = np.empty((2,2), order="F")
c.fill(np.nan)
gulinalg.matrix_multiply(a, b, out=c)
assert_allclose(c, np.zeros((2,2)))
# check that the correct shape is created as well...
d = gulinalg.matrix_multiply(a,b)
assert_allclose(d, np.zeros((2,2)))
def test_stride_tricks(self):
"""test that matrices that are contiguous but have their dimension
overlapped *copy*, as BLAS does not support them"""
a = np.ascontiguousarray(np.random.randn(M + N))
a = np.lib.stride_tricks.as_strided(a, shape=(M,N),
strides=(a.itemsize, a.itemsize))
b = np.ascontiguousarray(np.random.randn(N,K))
res = gulinalg.matrix_multiply(a,b)
ref = np.dot(a,b)
assert_allclose(res, ref)
def test_real_broadcast_L(self):
m = 5
nbatch = 16
a = np.asfortranarray(np.random.randn(m, m))
a = np.dot(a, a.T) # make Hermetian symmetric
a = np.stack((a, ) * nbatch, axis=0)
for workers in [1, -1]:
L = gulinalg.poinv(np.tril(a), UPLO='L')
assert_allclose(gulinalg.matrix_multiply(a, L),
np.stack((np.eye(m), ) * nbatch, axis=0),
atol=1e-10)
def test_matrix_multiply_fc_batch(self):
"""matrix multiply two C layout matrices"""
n_outer = 4
for workers in [1, -1]:
a = np.random.randn(M, N)
b = np.random.randn(N, K)
ref = np.dot(a, b)
a = np.asfortranarray(np.stack((a, ) * n_outer, axis=0))
b = np.ascontiguousarray(np.stack((b, ) * n_outer, axis=0))
ref = np.stack((ref, ) * n_outer, axis=0)
res = gulinalg.matrix_multiply(a, b, workers=workers)
assert_allclose(res, ref)
def test_real_vector_noncontig(self):
m = 10
rstate = np.random.RandomState(5)
a = rstate.randn(n_batch, m, m, 2)[..., 0]
a = gulinalg.matrix_multiply(a, a.swapaxes(-2, -1)) # make symmetric
self._run_vector(a)
def _check_eigen(self, A, w, v):
'''vectorial check of Mv==wv'''
lhs = gulinalg.matrix_multiply(A, v)
rhs = w*v
assert_allclose(lhs, rhs, rtol=1e-6)
