def test_cublas_bug():
'''
The SGEMM call would cause all calls after it to fail for some unknown
reason. Likely this is caused swaprows causing memory corruption.
NOTE: this was confirmed by nvidia to be a bug within CUDA, and should be
fixed in CUDA 6.5
'''
from pycuda.driver import Stream
from skcuda.cublas import cublasSgemm
from skcuda.misc import _global_cublas_handle as handle
n = 131
s = slice(128, n)
X = gpuarray.to_gpu(np.random.randn(n, 2483).astype(np.float32))
a = gpuarray.empty((X.shape[1], 3), dtype=np.float32)
c = gpuarray.empty((a.shape[0], X.shape[1]), dtype=np.float32)
b = gpuarray.empty_like(X)
m, n = a.shape[0], b[s].shape[1]
k = a.shape[1]
lda = m
ldb = k
ldc = m
#cublasSgemm(handle, 0, 0, m, n, k, 0.0, b.gpudata, lda, a.gpudata, ldb, 0.0, c.gpudata, ldc)
cublasSgemm(handle, 'n', 'n', m, n, k, 1.0, b[s].gpudata, lda, a.gpudata,
ldb, 0.0, c.gpudata, ldc)
#print handle, 'n', 'n', m, n, k, 1.0, b[s].gpudata, lda, a.gpudata, ldb, 0.0, c.gpudata, ldc
#gpuarray.dot(d, Xoutd[s])
#op.sgemm(a, b[s], c)
stream = Stream()
stream.synchronize()
def test_todense_stream():
''' Test GPUCSRArray.todense()'''
X = np.random.laplace(size=(600, 300)).astype(np.float32)
X[X < 0.1] = 0
X = csr_matrix(X, dtype=np.float32)
Xd = GPUCSRArray(X)
stream = Stream()
Yd = Xd.todense(stream=stream)
stream.synchronize()
assert_allclose(Yd.get(), X.A, rtol=1e-3, err_msg="todense")
def test_cusparseSetStream():
A = np.random.laplace(size=(3, 5)).astype(np.float32)
A[A<0.1] = 0
A = sparse.csr_matrix(A, dtype=np.float32)
A.sort_indices()
a_data = gpu.to_gpu(A.data)
a_indptr = gpu.to_gpu(A.indptr)
a_indices = gpu.to_gpu(A.indices)
out = gpu.empty((A.shape[0], A.shape[1]), dtype=A.dtype, order="F")
h = cusparse.cusparseCreate()
descrA = cusparse.cusparseCreateMatDescr()
stream = Stream()
cusparse.cusparseSetStream(h, stream.handle)
cusparse.cusparseScsr2dense(h, A.shape[0], A.shape[1],
descrA, a_data.gpudata, a_indptr.gpudata, a_indices.gpudata,
out.gpudata, out.shape[0])
cusparse.cusparseSetStream(h, 0)
stream.synchronize()
assert_allclose(out.get(), A.A, rtol=1e-4)
def test_unary_func_kwargs(self):
"""tests if the kwargs to the unary functions work"""
from pycuda.driver import Stream
name, a, b, threshold = ("exp", -3, 3, 1e-5)
gpu_func = getattr(cumath, name)
cpu_func = getattr(np, numpy_func_names.get(name, name))
for s in sizes:
for dtype in dtypes:
np.random.seed(1)
A = (np.random.random(s) * (b - a) + a).astype(dtype)
if complex:
A = A + (np.random.random(s) * (b - a) + a) * 1j
np.random.seed(1)
A = (np.random.random(s) * (b - a) + a).astype(dtype)
args = gpuarray.to_gpu(A)
# 'out' kw
gpu_results = gpuarray.empty_like(args)
gpu_results = gpu_func(args, out=gpu_results).get()
cpu_results = cpu_func(A)
max_err = np.max(np.abs(cpu_results - gpu_results))
assert (max_err <= threshold).all(), (max_err, name, dtype)
# 'out' position
gpu_results = gpuarray.empty_like(args)
gpu_results = gpu_func(args, gpu_results).get()
cpu_results = cpu_func(A)
max_err = np.max(np.abs(cpu_results - gpu_results))
assert (max_err <= threshold).all(), (max_err, name, dtype)
# 'stream' kw
mystream = Stream()
np.random.seed(1)
A = (np.random.random(s) * (b - a) + a).astype(dtype)
args = gpuarray.to_gpu(A)
gpu_results = gpuarray.empty_like(args)
gpu_results = gpu_func(args, stream=mystream).get()
cpu_results = cpu_func(A)
max_err = np.max(np.abs(cpu_results - gpu_results))
assert (max_err <= threshold).all(), (max_err, name, dtype)
# 'stream' position
mystream = Stream()
np.random.seed(1)
A = (np.random.random(s) * (b - a) + a).astype(dtype)
args = gpuarray.to_gpu(A)
gpu_results = gpuarray.empty_like(args)
gpu_results = gpu_func(args, mystream).get()
cpu_results = cpu_func(A)
max_err = np.max(np.abs(cpu_results - gpu_results))
assert (max_err <= threshold).all(), (max_err, name, dtype)
# 'out' and 'stream' kw
mystream = Stream()
np.random.seed(1)
A = (np.random.random(s) * (b - a) + a).astype(dtype)
args = gpuarray.to_gpu(A)
gpu_results = gpuarray.empty_like(args)
gpu_results = gpu_func(args, stream=mystream,
out=gpu_results).get()
cpu_results = cpu_func(A)
max_err = np.max(np.abs(cpu_results - gpu_results))
assert (max_err <= threshold).all(), (max_err, name, dtype)