def test_cublasSgetriBatched(self):
l,m = 11,7
np.random.seed(1)
A = np.random.rand(l,m, m).astype(np.float32)
a_gpu = gpuarray.to_gpu(A)
a_arr = bptrs(a_gpu)
c_gpu = gpuarray.empty((l,m,m), np.float32)
c_arr = bptrs(c_gpu)
p_gpu = gpuarray.empty((l,m), np.int32)
i_gpu = gpuarray.zeros(l, np.int32)
cublas.cublasSgetrfBatched(self.cublas_handle,
m, a_arr.gpudata, m, p_gpu.gpudata,
i_gpu.gpudata, l)
cublas.cublasSgetriBatched(self.cublas_handle,
m, a_arr.gpudata, m, p_gpu.gpudata, c_arr.gpudata,m,
i_gpu.gpudata, l)
X = np.array(map(np.linalg.inv,A))
X_ = c_gpu.get()
assert np.allclose(X,X_,6)
def thunk():
input_shape = inputs[0][0].shape
size = input_shape[1] # matrices to invert are (size x size)
batch_size = input_shape[0]
z = outputs[0]
# only allocate if there is no previous allocation of the right size.
if z[0] is None or z[0].shape != input_shape:
z[0] = cuda.CudaNdarray.zeros(input_shape)
pivot_alloc[0] = pycuda.gpuarray.empty((batch_size, size), np.int32)
info_alloc[0] = pycuda.gpuarray.zeros(batch_size, np.int32)
input_pycuda = to_gpuarray(inputs[0][0])
output_pycuda = to_gpuarray(z[0])
pivot = pivot_alloc[0]
info = info_alloc[0]
# construct pointer arrays for batched operations
input_arr = bptrs(input_pycuda)
output_arr = bptrs(output_pycuda)
if not self.destructive:
input_pycuda = input_pycuda.copy() # to prevent destruction of the input
handle = scikits.cuda.misc._global_cublas_handle
# perform LU factorization
cublas.cublasSgetrfBatched(handle, size, input_arr.gpudata, size, pivot.gpudata, info.gpudata, batch_size)
# the LU factorization is now in input_pycuda (destructive operation!)
# use factorization to perform inversion
cublas.cublasSgetriBatched(handle, size, input_arr.gpudata, size, pivot.gpudata, output_arr.gpudata, size, info.gpudata, batch_size)
def solve_gpu(As, Bs):
batch_size, num_factors = As.shape
if allocated_shape[0] == As.shape: # reuse previous allocations
As_gpu, Bs_gpu, P_gpu, info_gpu, Cs_gpu, Rs_gpu, A_arr, B_arr, C_arr, R_arr = allocations[0]
As_gpu.set(As)
Bs_gpu.set(Bs)
else: # allocate
# transfer As and Bs to GPU
As_gpu = pycuda.gpuarray.to_gpu(As.astype('float32'))
Bs_gpu = pycuda.gpuarray.to_gpu(Bs.astype('float32'))
# allocate arrays
P_gpu = pycuda.gpuarray.empty((batch_size, num_factors), np.int32)
info_gpu = pycuda.gpuarray.zeros(batch_size, np.int32)
Cs_gpu = pycuda.gpuarray.empty_like(Bs_gpu) # inverted Bs.
Rs_gpu = pycuda.gpuarray.empty_like(As_gpu) # final output, As * inverted Bs.
# get pointer arrays
A_arr = bptrs(As_gpu)
B_arr = bptrs(Bs_gpu)
C_arr = bptrs(Cs_gpu)
R_arr = bptrs(Rs_gpu)
allocated_shape[0] = As.shape
allocations[0] = As_gpu, Bs_gpu, P_gpu, info_gpu, Cs_gpu, Rs_gpu, A_arr, B_arr, C_arr, R_arr
handle = scikits.cuda.misc._global_cublas_handle
# perform LU factorization
cublas.cublasSgetrfBatched(handle, num_factors, B_arr.gpudata, num_factors, P_gpu.gpudata, info_gpu.gpudata, batch_size)
# the LU factorization is now in Bs_gpu!
# use factorization to perform inversion
cublas.cublasSgetriBatched(handle, num_factors, B_arr.gpudata, num_factors, P_gpu.gpudata, C_arr.gpudata, num_factors, info_gpu.gpudata, batch_size)
# the inverted matrices are now in Cs_gpu!
# compute dot products dot(A, C) = dot(A, Binv). Note that the As are actually vectors!
transb = 'n'
transa = 'n'
N, k, m = Cs_gpu.shape
N2, l = As_gpu.shape
n = 1 # As_gpu is a batch of vectors, not matrices, but we treat it as a batch of matrices with leading dimension 1.
# kind of tricky, but it seems to work. The same goes for the output array Rs_gpu.
lda = max(1, m)
ldb = max(1, k)
ldc = max(1, m)
alpha = np.float32(1.0)
beta = np.float32(0.0)
cublas.cublasSgemmBatched(handle, transb, transa, m, n, k, alpha, C_arr.gpudata,
lda, A_arr.gpudata, ldb, beta, R_arr.gpudata, ldc, N)
# the resulting batch of vectors is now in Rs_gpu.
return Rs_gpu.get()
def test_cublasSgetrfBatched(self):
from scipy.linalg import lu_factor
l, m = 11, 7
A = np.random.rand(l, m, m).astype(np.float32)
A = np.array([np.matrix(a) * np.matrix(a).T for a in A])
a_gpu = gpuarray.to_gpu(A)
a_arr = bptrs(a_gpu)
p_gpu = gpuarray.empty((l, m), np.int32)
i_gpu = gpuarray.zeros(1, np.int32)
X = np.array([lu_factor(a)[0] for a in A])
cublas.cublasSgetrfBatched(self.cublas_handle, m, a_arr.gpudata, m,
p_gpu.gpudata, i_gpu.gpudata, l)
X_ = np.array([a.T for a in a_gpu.get()])
assert np.allclose(X, X_)
def test_cublasSgetrfBatched(self):
from scipy.linalg import lu_factor
l,m = 11,7
A = np.random.rand(l,m, m).astype(np.float32)
A = np.array([np.matrix(a)*np.matrix(a).T for a in A])
a_gpu = gpuarray.to_gpu(A)
a_arr = bptrs(a_gpu)
p_gpu = gpuarray.empty((l,m), np.int32)
i_gpu = gpuarray.zeros(l, np.int32)
X = np.array([ lu_factor(a)[0] for a in A])
cublas.cublasSgetrfBatched(self.cublas_handle,
m, a_arr.gpudata, m, p_gpu.gpudata, i_gpu.gpudata, l)
X_ = np.array([a.T for a in a_gpu.get()])
assert np.allclose(X,X_,6)
def solve_gpu(As, Bs):
batch_size, num_factors = As.shape
if allocated_shape[0] == As.shape: # reuse previous allocations
As_gpu, Bs_gpu, P_gpu, info_gpu, Cs_gpu, Rs_gpu, A_arr, B_arr, C_arr, R_arr = allocations[
0]
As_gpu.set(As)
Bs_gpu.set(Bs)
else: # allocate
# transfer As and Bs to GPU
As_gpu = pycuda.gpuarray.to_gpu(As.astype('float32'))
Bs_gpu = pycuda.gpuarray.to_gpu(Bs.astype('float32'))
# allocate arrays
P_gpu = pycuda.gpuarray.empty((batch_size, num_factors), np.int32)
info_gpu = pycuda.gpuarray.zeros(batch_size, np.int32)
Cs_gpu = pycuda.gpuarray.empty_like(Bs_gpu) # inverted Bs.
Rs_gpu = pycuda.gpuarray.empty_like(
As_gpu) # final output, As * inverted Bs.
# get pointer arrays
A_arr = bptrs(As_gpu)
B_arr = bptrs(Bs_gpu)
C_arr = bptrs(Cs_gpu)
R_arr = bptrs(Rs_gpu)
allocated_shape[0] = As.shape
allocations[
0] = As_gpu, Bs_gpu, P_gpu, info_gpu, Cs_gpu, Rs_gpu, A_arr, B_arr, C_arr, R_arr
handle = scikits.cuda.misc._global_cublas_handle
# perform LU factorization
cublas.cublasSgetrfBatched(handle, num_factors, B_arr.gpudata, num_factors,
P_gpu.gpudata, info_gpu.gpudata, batch_size)
# the LU factorization is now in Bs_gpu!
# use factorization to perform inversion
cublas.cublasSgetriBatched(handle, num_factors, B_arr.gpudata, num_factors,
P_gpu.gpudata, C_arr.gpudata, num_factors,
info_gpu.gpudata, batch_size)
# the inverted matrices are now in Cs_gpu!
# compute dot products dot(A, C) = dot(A, Binv). Note that the As are actually vectors!
transb = 'n'
transa = 'n'
N, k, m = Cs_gpu.shape
N2, l = As_gpu.shape
n = 1 # As_gpu is a batch of vectors, not matrices, but we treat it as a batch of matrices with leading dimension 1.
# kind of tricky, but it seems to work. The same goes for the output array Rs_gpu.
lda = max(1, m)
ldb = max(1, k)
ldc = max(1, m)
alpha = np.float32(1.0)
beta = np.float32(0.0)
cublas.cublasSgemmBatched(handle, transb, transa, m, n, k, alpha,
C_arr.gpudata, lda, A_arr.gpudata, ldb, beta,
R_arr.gpudata, ldc, N)
# the resulting batch of vectors is now in Rs_gpu.
return Rs_gpu.get()
