Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[backport] Do not call cudnnGetVersion on import #5446

Merged
merged 1 commit into from
Jul 6, 2021
Merged

[backport] Do not call cudnnGetVersion on import #5446

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 28 additions & 19 deletions cupy/cudnn.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,12 +19,20 @@ from cupy._core._ufuncs import elementwise_copy as _elementwise_copy
from cupy import _util
from cupy_backends.cuda.libs import cudnn as _py_cudnn

cdef int _cudnn_version = cudnn.getVersion()

cdef int _cudnn_version = -1
cdef _thread_local = _threading.local()

cdef vector.vector[size_t] _handles


cdef inline int cudnn_version():
global _cudnn_version
if _cudnn_version == -1:
_cudnn_version = cudnn.getVersion()
return _cudnn_version


cpdef size_t get_handle() except? 0:
cdef int dev
dev = device.get_device_id()
Expand Down Expand Up @@ -216,7 +224,7 @@ cpdef _create_convolution_descriptor(
c_dilation.assign(ndim, 1)
else:
c_dilation = dilation
if _cudnn_version < 6000:
if cudnn_version() < 6000:
for i in c_dilation:
if i != 1:
raise ValueError(
Expand All @@ -229,13 +237,13 @@ cpdef _create_convolution_descriptor(
d0 = d1 = 1
else:
d0, d1 = dilation
if _cudnn_version < 6000 and (d0 != 1 or d1 != 1):
if cudnn_version() < 6000 and (d0 != 1 or d1 != 1):
raise ValueError('dilation must be one when cuDNN < 6.0')
p0, p1 = pad
s0, s1 = stride
cudnn.setConvolution2dDescriptor_v5(
desc, p0, p1, s0, s1, d0, d1, mode, compute_type)
if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
if use_tensor_core:
math_type = cudnn.CUDNN_TENSOR_OP_MATH
cudnn.setConvolutionMathType(desc, math_type)
Expand Down Expand Up @@ -812,7 +820,7 @@ def create_rnn_descriptor(hidden_size, num_layers, dropout_desc,
input_mode, direction, mode, data_type, algo=None):
desc = Descriptor(cudnn.createRNNDescriptor(),
_py_cudnn.destroyRNNDescriptor)
if _cudnn_version >= 6000:
if cudnn_version() >= 6000:
_handle = get_handle()
if algo is None:
algo = cudnn.CUDNN_RNN_ALGO_STANDARD
Expand Down Expand Up @@ -1230,7 +1238,7 @@ def create_reduce_tensor_descriptor(reduce_type, dtype):


cpdef bint is_tensor_core_available(dtype) except *:
return (_cudnn_version >= 7000 and
return (cudnn_version() >= 7000 and
(<str>dtype.char) == 'e' and
int(device.get_compute_capability()) == 70)

Expand Down Expand Up @@ -1354,7 +1362,7 @@ cpdef _Algorithm _find_algorithm_fwd(
if algo is not None:
return algo
workspace = memory.alloc(max_workspace_size)
if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
perf = cudnn.findConvolutionForwardAlgorithmEx_v7(
handle, x_desc, x.data.ptr, filter_desc, W.data.ptr, conv_desc,
y_desc, y.data.ptr, 1, workspace.ptr, max_workspace_size)[0]
Expand Down Expand Up @@ -1383,7 +1391,8 @@ cpdef _Algorithm _get_algorithm_fwd(
return algo
cdef list ret
cdef bint skip
if (use_tensor_core and _cudnn_version >= 7000) or _cudnn_version >= 8000:
cdef int cudnn_ver = cudnn_version()
if (use_tensor_core and cudnn_ver >= 7000) or cudnn_ver >= 8000:
ret = cudnn.getConvolutionForwardAlgorithm_v7(
handle, x_desc, filter_desc, conv_desc, y_desc, 10)
skip = False
Expand Down Expand Up @@ -1441,7 +1450,7 @@ cpdef _Algorithm _find_algorithm_bwd_filter(
if algo is not None:
return algo
workspace = memory.alloc(max_workspace_size)
if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
if deterministic:
ret = cudnn.findConvolutionBackwardFilterAlgorithmEx_v7(
handle, x_desc, x.data.ptr, dy_desc, dy.data.ptr, conv_desc,
Expand Down Expand Up @@ -1485,7 +1494,7 @@ cpdef _Algorithm _get_algorithm_bwd_filter(
return algo
cdef list ret
cdef bint skip
if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
ret = cudnn.getConvolutionBackwardFilterAlgorithm_v7(
handle, x_desc, gy_desc, conv_desc, filter_desc, 10)
skip = False
Expand Down Expand Up @@ -1546,7 +1555,7 @@ cpdef _Algorithm _find_algorithm_bwd_data(
if algo is not None:
return algo
workspace = memory.alloc(max_workspace_size)
if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
if deterministic:
ret = cudnn.findConvolutionBackwardDataAlgorithmEx_v7(
handle, filter_desc, W.data.ptr, x_desc, x.data.ptr, conv_desc,
Expand Down Expand Up @@ -1588,7 +1597,7 @@ cpdef _Algorithm _get_algorithm_bwd_data(
return algo
cdef list ret
cdef bint skip
if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
ret = cudnn.getConvolutionBackwardDataAlgorithm_v7(
handle, filter_desc, x_desc, conv_desc, y_desc, 10)
skip = False
Expand Down Expand Up @@ -1708,7 +1717,7 @@ def convolution_forward(
x, W, y, conv_param, handle, x_desc, filter_desc,
conv_desc, y_desc, max_workspace_size, use_tensor_core)

if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
cudnn.setConvolutionMathType(conv_desc, perf.mathType)

workspace = memory.alloc(perf.memory)
Expand Down Expand Up @@ -1802,7 +1811,7 @@ def convolution_backward_filter(
conv_desc, pad, stride, dilation, groups, x.dtype,
cudnn.CUDNN_CROSS_CORRELATION, use_tensor_core)

if deterministic and _cudnn_version < 7000:
if deterministic and cudnn_version() < 7000:
# TODO(imanishi): Support Tensor Core in deterministic mode.
algo = cudnn.CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1
workspace_size = cudnn.getConvolutionBackwardFilterWorkspaceSize(
Expand All @@ -1827,7 +1836,7 @@ def convolution_backward_filter(
workspace_size = perf.memory
math_type = perf.mathType

if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
cudnn.setConvolutionMathType(conv_desc, math_type)

workspace = memory.alloc(workspace_size)
Expand Down Expand Up @@ -1902,7 +1911,7 @@ def convolution_backward_data(
conv_desc, pad, stride, dilation, groups, x.dtype,
cudnn.CUDNN_CROSS_CORRELATION, use_tensor_core)

if deterministic and _cudnn_version < 7000:
if deterministic and cudnn_version() < 7000:
# TODO(imanishi): Support Tensor Core in deterministic mode.
algo = cudnn.CUDNN_CONVOLUTION_BWD_DATA_ALGO_1
workspace_size = cudnn.getConvolutionBackwardDataWorkspaceSize(
Expand All @@ -1927,7 +1936,7 @@ def convolution_backward_data(
workspace_size = perf.memory
math_type = perf.mathType

if _cudnn_version >= 7000:
if cudnn_version() >= 7000:
cudnn.setConvolutionMathType(conv_desc, math_type)

workspace = memory.alloc(workspace_size)
Expand Down Expand Up @@ -2172,7 +2181,7 @@ cdef _batch_normalization_forward_training(
# current implementation of our BN depends on this behavior so that
# we can reduce the number of reduction kernels.

if _cudnn_version >= 7401:
if cudnn_version() >= 7401:

bn_ops = cudnn.CUDNN_BATCHNORM_OPS_BN

Expand Down Expand Up @@ -2349,7 +2358,7 @@ def batch_normalization_backward(
ggamma = core.ndarray(gamma._shape, dtype_param)
gbeta = core.ndarray(gamma._shape, dtype_param)

if _cudnn_version >= 7401:
if cudnn_version() >= 7401:
bn_ops = cudnn.CUDNN_BATCHNORM_OPS_BN

workspace_size = (
Expand Down