for i in range(batchSize):
x = np.random.randn(n, n) + 1j * np.random.randn(n, n)
x = x + x.conj().T
x = x.astype(np.complex64)
A[i * n:(i + 1) * n, :] = x
# Need to reverse dimensions because CUSOLVER expects column-major matrices:
B[i * n:(i + 1) * n, :] = x.T.copy()
x_gpu = gpuarray.to_gpu(B)
# Set up output buffers:
w_gpu = gpuarray.empty((batchSize, n), dtype=np.float32)
# Set up parameters
params = solver.cusolverDnCreateSyevjInfo()
solver.cusolverDnXsyevjSetTolerance(params, 1e-7)
solver.cusolverDnXsyevjSetMaxSweeps(params, 15)
# Set up work buffers:
lwork = solver.cusolverDnCheevjBatched_bufferSize(handle,
'CUSOLVER_EIG_MODE_VECTOR',
'u', n, x_gpu.gpudata, n,
w_gpu.gpudata, params,
batchSize)
workspace_gpu = gpuarray.zeros(lwork, dtype=A.dtype)
info = gpuarray.zeros(batchSize, dtype=np.int32)
# Compute:
solver.cusolverDnCheevjBatched(handle, 'CUSOLVER_EIG_MODE_VECTOR', 'u', n,
x_gpu.gpudata, n, w_gpu.gpudata,
A = np.empty((n*batchSize, n), dtype = np.double)
for i in range(batchSize):
x = np.random.randn(n, n)
x = x+x.T
A[i*n:(i+1)*n, :] = x
x_gpu = gpuarray.to_gpu(A)
# Set up output buffers:
w_gpu = gpuarray.empty((batchSize, n), dtype = A.dtype)
# Set up parameters
params = solver.cusolverDnCreateSyevjInfo()
solver.cusolverDnXsyevjSetTolerance(params, 1e-7)
solver.cusolverDnXsyevjSetMaxSweeps(params, 15)
# Set up work buffers:
lwork = solver.cusolverDnDsyevjBatched_bufferSize(handle, 'CUSOLVER_EIG_MODE_VECTOR',
'u', n, x_gpu.gpudata, n,
w_gpu.gpudata, params, batchSize)
workspace_gpu = gpuarray.zeros(lwork, dtype = A.dtype)
info = gpuarray.zeros(batchSize, dtype = np.int32)
# Compute:
solver.cusolverDnDsyevjBatched(handle, 'CUSOLVER_EIG_MODE_VECTOR',
'u', n, x_gpu.gpudata, n,
w_gpu.gpudata, workspace_gpu.gpudata,
lwork, info.gpudata, params, batchSize)