def test_large(self):
t = 20_000
num_rep = 5
mat = torch.from_numpy(gen_random(t, t, np.float32, F=False, seed=123))
vec = torch.from_numpy(
gen_random(t, 1, np.float32, F=False, seed=124).reshape((-1, )))
mat_cuda = mat.cuda()
vec_cuda = vec.cuda()
cpu_times = []
for i in range(num_rep):
t_s = time.time()
out_cpu = vec_mul_triang(mat, vec, True, 1)
cpu_times.append(time.time() - t_s)
gpu_times = []
for i in range(num_rep):
t_s = time.time()
out_cuda = vec_mul_triang(mat_cuda, vec_cuda, True, 1)
torch.cuda.synchronize()
gpu_times.append(time.time() - t_s)
print("mat size %d - t_cpu: %.4fs -- t_cuda: %.4fs" %
(t, np.min(cpu_times), np.min(gpu_times)))
np.testing.assert_allclose(out_cpu, out_cuda.cpu().numpy())
def test_large(self):
t = 30_000
mat = gen_random(t, t, np.float64, F=False, seed=123)
vec = gen_random(t, 1, np.float64, F=False, seed=124).reshape((-1, ))
t_s = time.time()
vec_mul_triang(mat, vec, upper=True, side=1)
t_tri = time.time() - t_s
t_s = time.time()
mat *= vec
t_full = time.time() - t_s
print("Our took %.2fs -- Full took %.2fs" % (t_tri, t_full))
def mat(self):
return torch.from_numpy(
gen_random(TestVecMulTriang.MAT_SIZE,
TestVecMulTriang.MAT_SIZE,
'float64',
False,
seed=91))
def test_gaussian_pd():
X = gen_random(10000, 2, 'float32', F=True, seed=12)
Xt = torch.from_numpy(X)
sigma = 10.0
opt = FalkonOptions(compute_arch_speed=False, max_gpu_mem=1 * 2**30, use_cpu=False,
no_single_kernel=False)
k = GaussianKernel(sigma, opt=opt)
actual = k(Xt, Xt, opt=opt)
actual += torch.eye(Xt.shape[0]) * (1e-7 * Xt.shape[0])
# Test positive definite
np.linalg.cholesky(actual)
def vec(self):
return gen_random(self.t, self.r, np.float64, F=True, seed=124)
def test_add_diag(F):
A = torch.from_numpy(gen_random(1000, 1000, 'float64', F=F, seed=10))
diag = 10**6
falkon.preconditioner.pc_utils.inplace_add_diag_th(A, diag)
assert torch.all((A.diagonal() > 10**5) & (A.diagonal() < 20**6))
def rect(self):
return gen_random(self.t,
self.t * 2 - 1,
np.float64,
F=True,
seed=12345)
def mat(self):
return gen_random(self.t, self.t, np.float64, F=True, seed=123)
def dev_mat(request):
return torch.from_numpy(gen_random(20, 20, 'float32',
F=request.param)).cuda()
def mat(self):
return gen_random(self.t, self.t, np.float32, F=False, seed=123)
def B():
return torch.from_numpy(gen_random(m, d, 'float64', False, seed=92))
def rowmaj_arr() -> torch.Tensor:
return torch.from_numpy(gen_random(M, D, 'float64', False))
def Ac():
n, d = (4000, 10)
return gen_random(n, d, 'float64', False, seed=92)
def Bc():
m, d = (2000, 10)
return gen_random(m, d, 'float64', False, seed=92)
def A(n, d) -> torch.Tensor:
return torch.from_numpy(gen_random(n, d, 'float64', False, seed=92))
def Ac():
return gen_random(n, d, 'float64', False, seed=92)
def Bc():
return gen_random(m, d, 'float64', False, seed=92)
def vec_rhs(self, request):
return torch.from_numpy(
gen_random(self.t, request.param, 'float64', F=False, seed=9))
def data(self):
return torch.from_numpy(
gen_random(self.N, self.D, 'float64', F=False, seed=10))
def Af():
n, d = (4000, 10)
return np.asfortranarray(gen_random(n, d, 'float64', False, seed=92))
def v() -> torch.Tensor:
return torch.from_numpy(gen_random(m, t, 'float64', False, seed=92))
def Bf():
m, d = (2000, 10)
return np.asfortranarray(gen_random(m, d, 'float64', False, seed=92))
def mat():
return gen_random(M, M, 'float64', F=True, seed=10)
def w():
return torch.from_numpy(gen_random(n, t, 'float64', False, seed=92))
def arr():
return gen_random(M, T, 'float64', F=True, seed=12)
def colmaj_arr() -> torch.Tensor:
return torch.from_numpy(gen_random(M, D, 'float64', True))
def test_add_diag(F):
A = gen_random(1000, 1000, 'float64', F=F, seed=10)
diag = 10**6
falkon.preconditioner.pc_utils.inplace_add_diag(A, diag)
assert np.all(A.diagonal() > 10**5)