def test_dot_diag_t_complex128(self):
d = np.asarray(np.random.rand(5), np.float64)
a = np.asarray(np.random.rand(3, 5)+1j*np.random.rand(3, 5), np.complex128)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
a = a.astype(np.complex128, order="F", copy=True)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
def test_dot_diag_t_float64(self):
d = np.asarray(np.random.rand(5), np.float64)
a = np.asarray(np.random.rand(3, 5), np.float64)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
a = a.astype(np.float64, order="F", copy=True)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
def test_dot_diag_float32(self):
d = np.asarray(np.random.rand(5), np.float32)
a = np.asarray(np.random.rand(5, 3), np.float32)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu)
assert np.allclose(np.dot(np.diag(d), a), r_gpu.get())
a = a.astype(np.float32, order="F", copy=True)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu)
assert np.allclose(np.dot(np.diag(d), a), r_gpu.get())
def test_dot_diag_complex64(self):
d = np.asarray(np.random.rand(5), np.float32)
a = np.asarray(np.random.rand(5, 3)+1j*np.random.rand(5, 3), np.complex64)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu)
assert np.allclose(np.dot(np.diag(d), a), r_gpu.get())
a = a.astype(np.complex64, order="F", copy=True)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu)
assert np.allclose(np.dot(np.diag(d), a), r_gpu.get())
def impl_test_dot_diag_t(self, dtype):
d = np.asarray(np.random.rand(5), dtype)
a = np.asarray(np.random.rand(3, 5), dtype)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
a = a.astype(dtype, order="F", copy=True)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
def ___impl_test_dot_diag_t(self, dtype):
d = np.asarray(np.random.rand(5), dtype)
a = np.asarray(np.random.rand(3, 5), dtype)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, "t")
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
a = a.astype(dtype, order="F", copy=True)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
# note: due to pycuda issue #66, this will fail when overwrite=False
r_gpu = linalg.dot_diag(d_gpu, a_gpu, "t", overwrite=True)
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
def test_dot_diag_t_float32(self):
d = np.asarray(np.random.rand(5), np.float32)
a = np.asarray(np.random.rand(3, 5), np.float32)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
def test_dot_diag_t_float32(self):
d = np.asarray(np.random.rand(5), np.float32)
a = np.asarray(np.random.rand(3, 5), np.float32)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu, 't')
assert np.allclose(np.dot(np.diag(d), a.T).T, r_gpu.get())
def test_dot_diag_float64(self):
d = np.asarray(np.random.rand(5), np.float64)
a = np.asarray(np.random.rand(5, 3), np.float64)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu)
assert np.allclose(np.dot(np.diag(d), a), r_gpu.get())
def test_dot_diag_complex128(self):
d = np.asarray(np.random.rand(5), np.float64)
a = np.asarray(np.random.rand(5, 3)+1j*np.random.rand(5, 3), np.complex128)
d_gpu = gpuarray.to_gpu(d)
a_gpu = gpuarray.to_gpu(a)
r_gpu = linalg.dot_diag(d_gpu, a_gpu)
assert np.allclose(np.dot(np.diag(d), a), r_gpu.get())
else:
nalphas = alpha
# Compute weights for each alpha
ualphas = np.unique(nalphas)
wt = np.zeros((stim.shape[1], resp.shape[1]), order='F') # Make wt column major
Vh_gpu = gpuarray.to_gpu(np.copy(Vh, order='F'))
for ua in ualphas:
selvox = np.nonzero(nalphas==ua)[0] # list of indices equal to ua
# TODO determine if this should be a GPU op
# Vh is output from SVD, i think NxN (~200x200 or 15000x15000)
# TODO determine how reduce works
Sd = S/(S**2+ua**2)
Sd_gpu = gpuarray.to_gpu(Sd)
UR_gpu = gpuarray.to_gpu(np.copy(UR[:,selvox], order='F'))
linalg.dot_diag(Sd_gpu, UR_gpu, overwrite=True)
del Sd_gpu
if selvox.shape[0] > 5000:
N=selvox.shape[0]/4
inter_gpu = linalg.dot(Vh_gpu, UR_gpu[:,0:N], transa='T')
wt[:,selvox[0:N]] = inter_gpu.get()
del inter_gpu
inter_gpu = linalg.dot(Vh_gpu, UR_gpu[:,N:2*N], transa='T')
wt[:,selvox[N:2*N]] = inter_gpu.get()
del inter_gpu
inter_gpu = linalg.dot(Vh_gpu, UR_gpu[:,2*N:3*N], transa='T')
wt[:,selvox[2*N:3*N]] = inter_gpu.get()
del inter_gpu
inter_gpu = linalg.dot(Vh_gpu, UR_gpu[:,3*N:], transa='T')
wt[:,selvox[3*N:]] = inter_gpu.get()
del inter_gpu
