def test_gemm_out(self, dtype):
dtype = numpy.dtype(dtype)
tol = self._tol[dtype.char.lower()]
m, n, k = self.mnk
a = self._make_matrix(m, k, self.transa, self.ordera, dtype)
b = self._make_matrix(k, n, self.transb, self.orderb, dtype)
c = self._make_matrix(m, n, 'N', self.orderc, dtype)
alpha = 0.9
beta = 0.8
if dtype.char in 'FD':
alpha = alpha - 1j * 0.7
beta = beta - 1j * 0.6
aa = self._trans_matrix(a, self.transa)
bb = self._trans_matrix(b, self.transb)
ref = alpha * aa.dot(bb) + beta * c
if self.mode is not None:
alpha = self.mode.array(alpha)
beta = self.mode.array(beta)
cublas.gemm(self.transa,
self.transb,
a,
b,
out=c,
alpha=alpha,
beta=beta)
cupy.testing.assert_allclose(c, ref, rtol=tol, atol=tol)
def test_gemm(self, dtype):
if not (self.mode is None and self.orderc == 'C'):
pytest.skip()
dtype = numpy.dtype(dtype)
tol = self._tol[dtype.char.lower()]
m, n, k = self.mnk
a = self._make_matrix(m, k, self.transa, self.ordera, dtype)
b = self._make_matrix(k, n, self.transb, self.orderb, dtype)
aa = self._trans_matrix(a, self.transa)
bb = self._trans_matrix(b, self.transb)
ref = aa.dot(bb)
c = cublas.gemm(self.transa, self.transb, a, b)
cupy.testing.assert_allclose(c, ref, rtol=tol, atol=tol)