def test_svd_ss_float64(self):
a = np.asarray(np.random.randn(9, 6), np.float64)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = linalg.svd(a_gpu, 's', 's')
assert np.allclose(a, np.dot(u_gpu.get(),
np.dot(np.diag(s_gpu.get()),
vh_gpu.get())),
atol=atol_float64)
def test_svd_so_complex128(self):
a = np.asarray(np.random.randn(6, 6) + 1j*np.random.randn(6, 6), np.complex128)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = linalg.svd(a_gpu, 's', 'o')
assert np.allclose(a, np.dot(u_gpu.get(),
np.dot(np.diag(s_gpu.get()),
vh_gpu.get())),
atol=atol_float64)
def test_svd_aa_cusolver_float32(self):
a = np.asarray(np.random.randn(6, 6), np.float32)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = linalg.svd(a_gpu, lib='cusolver')
assert np.allclose(a, np.dot(u_gpu.get(),
np.dot(np.diag(s_gpu.get()),
vh_gpu.get())),
atol=atol_float32)
def test_svd_ss_float64(self):
a = np.asarray(np.random.randn(9, 6), np.float64)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = linalg.svd(a_gpu, 's', 's')
assert np.allclose(a,
np.dot(u_gpu.get(),
np.dot(np.diag(s_gpu.get()), vh_gpu.get())),
atol=atol_float64)
def test_svd_so_complex128(self):
a = np.asarray(
np.random.randn(6, 6) + 1j * np.random.randn(6, 6), np.complex128)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = linalg.svd(a_gpu, 's', 'o')
assert np.allclose(a,
np.dot(u_gpu.get(),
np.dot(np.diag(s_gpu.get()), vh_gpu.get())),
atol=atol_float64)
def svd_skcuda(X, typ):
"""
Singular Value Decomposition with scikit-cuda and CULA library
Usage: U_gpu, S, Vt_gpu = svd_skcuda(X, typ)
Input: X matrix to decompose (array)
typ data type (string)
CULA free only support typ = float32 and complex64
Output: U_gpu left unitary matrix on GPU (array)
S singular values on CPU (vector)
Vt_gpu transposed right unitary matrix on GPU (array)
"""
X = X.astype(typ)
X_gpu = gpuarray.to_gpu(X) # send data to gpu
U_gpu, S_gpu, Vt_gpu = cu_linalg.svd(
X_gpu, jobu='S', jobvt='S', lib='cula') # small U and Vt matrices
S = S_gpu.get()
X_gpu.gpudata.free() # release gpu memory
S_gpu.gpudata.free()
return U_gpu, S, Vt_gpu
"""
import pycuda.autoinit
import pycuda.driver as drv
import pycuda.gpuarray as gpuarray
import numpy as np
import skcuda.linalg as culinalg
import skcuda.misc as cumisc
culinalg.init()
# Double precision is only supported by devices with compute
# capability >= 1.3:
import string
import scikits.cuda.cula as cula
demo_types = [np.float32, np.complex64]
if cula._libcula_toolkit == 'premium' and \
cumisc.get_compute_capability(pycuda.autoinit.device) >= 1.3:
demo_types.extend([np.float64, np.complex128])
for t in demo_types:
print 'Testing svd for type ' + str(np.dtype(t))
a = np.asarray((np.random.rand(50, 50) - 0.5) / 10, t)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = culinalg.svd(a_gpu)
a_rec = np.dot(u_gpu.get(), np.dot(np.diag(s_gpu.get()), vh_gpu.get()))
print 'Success status: ', np.allclose(a, a_rec, atol=1e-3)
print 'Maximum error: ', np.max(np.abs(a - a_rec))
print ''
def gpu_skcuda_linalg_svd(input):
input_gpu = gpuarray.to_gpu(input)
linalg.init()
u_gpu, s_gpu, vh_gpu = linalg.svd(input_gpu, 'A', 'A', 'cusolver')
exit()
"""
Using svd to factorize the copa_matrix into
cause/effect vector representations
"""
epsilon = 1e-3
data = json.loads(open('copa_matrix.json').read())
arr = np.asarray(np.array(data), np.float32)
arr[np.diag(np.diag(arr) < epsilon)] = 0.0
#print arr.shape[0]
#print np.identity(arr.shape[0])
#print gpuarray.to_gpu(arr)
a_gpu = gpuarray.to_gpu(arr + epsilon*np.identity(arr.shape[0]))
#u_gpu, s_gpu, vh_gpu = culinalg.svd(a_gpu)
u_gpu, s_gpu, vh_gpu = culinalg.svd(a_gpu, jobu='S', jobvt='S')
print u_gpu.shape
print vh_gpu.shape
print type(u_gpu.get())
with open('u.json','wb') as outf:
outf.write(json.dumps(u_gpu.get().tolist()))
with open('v.json','wb') as outf:
outf.write(json.dumps(vh_gpu.get().tolist()))
# This example demonstrates simple SVD-computation using CUDA-Solver library.
import pycuda.autoinit
from pycuda import gpuarray
import numpy as np
from skcuda import linalg
a = np.random.rand(1000, 5000).astype(np.float32)
a_gpu = gpuarray.to_gpu(a)
U_d, s_d, V_d = linalg.svd(a_gpu, lib='cusolver')
U = U_d.get()
s = s_d.get()
V = V_d.get()
S = np.zeros((1000, 5000))
S[:1000, :1000] = np.diag(s)
print('Can we reconstruct a from its SVD decomposition? : {}'.format(
np.allclose(a, np.dot(U, np.dot(S, V), atol=1e-5))))
from __future__ import print_function
import pycuda.autoinit
import pycuda.driver as drv
import pycuda.gpuarray as gpuarray
import numpy as np
import skcuda.linalg as culinalg
import skcuda.misc as cumisc
culinalg.init()
# Double precision is only supported by devices with compute
# capability >= 1.3:
import string
import scikits.cuda.cula as cula
demo_types = [np.float32, np.complex64]
if cula._libcula_toolkit == 'premium' and \
cumisc.get_compute_capability(pycuda.autoinit.device) >= 1.3:
demo_types.extend([np.float64, np.complex128])
for t in demo_types:
print('Testing svd for type ' + str(np.dtype(t)))
a = np.asarray((np.random.rand(50, 50) - 0.5) / 10, t)
a_gpu = gpuarray.to_gpu(a)
u_gpu, s_gpu, vh_gpu = culinalg.svd(a_gpu)
a_rec = np.dot(u_gpu.get(), np.dot(np.diag(s_gpu.get()), vh_gpu.get()))
print('Success status: ', np.allclose(a, a_rec, atol=1e-3))
print('Maximum error: ', np.max(np.abs(a - a_rec)))
print('')
