Skip to content

P
pytensor
提交 109b378c authored 作者: Arnaud Bergeron's avatar Arnaud Bergeron 提交者： Frederic Bastien
浏览文件
操作

Require cudnn v5 minimum.

上级 c251a77d
隐藏空白字符变更
内嵌 并排
正在显示 包含 11 行增加 和 58 行删除
theano/gpuarray/dnn.py
浏览文件 @ 109b378c
......@@ -59,12 +59,12 @@ def _dnn_lib():
lib_name = ctypes.util.find_library('cudnn')
if lib_name is None and sys.platform == 'win32':
# Update these names when new versions of cudnn are supported.
for name in ['cudnn64_5.dll', 'cudnn64_4.dll']:
for name in ['cudnn64_5.dll']:
lib_name = ctypes.util.find_library(name)
if lib_name:
break
if lib_name is None:
raise RuntimeError('Could not find cudnn library (looked for v4 and v5[.1])')
raise RuntimeError('Could not find cudnn library (looked for v5[.1])')
_dnn_lib.handle = ctypes.cdll.LoadLibrary(lib_name)
cudnn = _dnn_lib.handle
cudnn.cudnnCreate.argtypes = [ctypes.POINTER(ctypes.c_void_p)]
......@@ -129,7 +129,7 @@ if ((err = cudnnCreate(&_handle)) != CUDNN_STATUS_SUCCESS) {
def _dnn_check_version():
v = version()
if v < 4007:
if v < 5000:
return False, "cuDNN version is too old. Update to v5, was %d." % v
return True, None
......@@ -500,10 +500,6 @@ class GpuDnnConv(DnnBase):
if self.inplace:
self.destroy_map = {0: [2]}
if version() < 5000 and self.algo == 'winograd':
raise RuntimeError("cuDNN winograd convolution requires "
"cuDNN v5 or more recent")
assert self.algo in ['none', 'small', 'large', 'fft', 'fft_tiling',
'winograd', 'guess_once', 'guess_on_shape_change',
'time_once', 'time_on_shape_change']
......@@ -524,9 +520,9 @@ class GpuDnnConv(DnnBase):
defs.append(('CONV_INPLACE', '1'))
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM'
if self.algo == 'none': # 3d (at least in v4)
if self.algo == 'none': # 3d
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM'
elif self.algo == 'small': # 3d (at least in v4)
elif self.algo == 'small': # 3d
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM'
elif self.algo == 'large':
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_GEMM'
......@@ -534,10 +530,9 @@ class GpuDnnConv(DnnBase):
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_DIRECT'
elif self.algo == 'fft':
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_FFT'
elif self.algo == 'fft_tiling': # 3d (not in v4, in v5)
elif self.algo == 'fft_tiling': # 3d
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING'
elif self.algo == 'winograd':
# need v5
alg = 'CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD'
defs.append(('CONV_ALGO', alg))
......@@ -571,10 +566,6 @@ class GpuDnnConv(DnnBase):
if img.type.ndim == 5 and self.algo in ['large', 'fft']:
raise ValueError("convolution algo %s can't be used for "
"3d convolutions", (self.algo,))
if (img.type.ndim == 5 and
self.algo in ['fft_tiling'] and
version() < 5000):
raise ValueError("3d convolution algo fft_tiling need cudnn v5")
if (not isinstance(desc.type, CDataType) or
desc.type.ctype != 'cudnnConvolutionDescriptor_t'):
......@@ -700,13 +691,13 @@ class GpuDnnConvGradW(DnnBase):
defs.append(('CONV_INPLACE', '1'))
alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0'
if self.algo == 'none': # 3d in at least v4
if self.algo == 'none': # 3d
alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0'
if self.algo == 'deterministic':
alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1'
if self.algo == 'fft':
alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT'
if self.algo == 'small': # 3d in at least v4
if self.algo == 'small': # 3d
# non-deterministic, small workspace
alg = 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3'
if self.algo in ['guess_once', 'guess_on_shape_change',
......@@ -793,10 +784,6 @@ class GpuDnnConvGradI(DnnBase):
algo = config.dnn.conv.algo_bwd_data
self.algo = algo
if version() < 5000 and self.algo == 'winograd':
raise RuntimeError("cuDNN's winograd convolution requires cuDNN "
"v5 or more recent")
assert self.algo in ['none', 'deterministic', 'fft', 'fft_tiling',
'winograd', 'guess_once', 'guess_on_shape_change',
'time_once', 'time_on_shape_change']
......@@ -832,17 +819,16 @@ class GpuDnnConvGradI(DnnBase):
defs.append(('CONV_INPLACE', '1'))
alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_0'
if self.algo == 'none': # 3d at least v4
if self.algo == 'none': # 3d
alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_0'
elif self.algo == 'deterministic': # 3d at least v4
elif self.algo == 'deterministic': # 3d
alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_1'
elif self.algo == 'fft':
alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT'
elif self.algo == 'fft_tiling': # 3d not v4, since v5
elif self.algo == 'fft_tiling': # 3d
# big workspace but less than fft
alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING'
elif self.algo == 'winograd':
# need v5
alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD'
if self.algo in ['guess_once', 'guess_on_shape_change',
......@@ -877,10 +863,6 @@ class GpuDnnConvGradI(DnnBase):
if kern.type.ndim == 5 and self.algo in ['fft']:
raise ValueError("convolution algo %s can't be used for "
"3d convolutions", (self.algo,))
if (kern.type.ndim == 5 and
self.algo == 'fft_tiling' and
version() < 5000):
raise ValueError("3d convolution algo fft_tiling need cudnn v5")
if (not isinstance(desc.type, CDataType) or
desc.type.ctype != 'cudnnConvolutionDescriptor_t'):
......@@ -1316,11 +1298,7 @@ class GpuDnnPoolDesc(Op):
static const int pad[%(nd)d] = {%(pad)s};
static const int str[%(nd)d] = {%(str)s};
#if CUDNN_VERSION >= 5000
err = cudnnSetPoolingNdDescriptor(%(desc)s, %(mode_flag)s, CUDNN_PROPAGATE_NAN, %(nd)d, win, pad, str);
#else
err = cudnnSetPoolingNdDescriptor(%(desc)s, %(mode_flag)s, %(nd)d, win, pad, str);
#endif
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError, "could not set op descriptor: %%s",
......@@ -1664,8 +1642,6 @@ class GpuDnnBatchNorm(DnnBase):
DnnBase.__init__(self, ['dnn_batchnorm_base.c', 'dnn_batchnorm.c'],
'dnn_batchnorm_op')
if version() < 5000:
raise RuntimeError("cuDNN Batch Normalization requires cuDNN v5 or later")
assert (mode in ('per-activation', 'spatial'))
self.mode = mode
......@@ -1724,8 +1700,6 @@ class GpuDnnBatchNormInference(DnnBase):
DnnBase.__init__(self, ['dnn_batchnorm_base.c', 'dnn_batchnorm_inf.c'],
'dnn_batchnorm_op')
if version() < 5000:
raise RuntimeError("cuDNN Batch Normalization requires cuDNN v5 or later")
assert (mode in ('per-activation', 'spatial'))
self.mode = mode
......@@ -1788,8 +1762,6 @@ class GpuDnnBatchNormGrad(DnnBase):
DnnBase.__init__(self, ['dnn_batchnorm_base.c', 'dnn_batchnorm_grad.c'],
'dnn_batchnorm_grad')
if version() < 5000:
raise RuntimeError("cuDNN Batch Normalization requires cuDNN v5 or later")
assert (mode in ('per-activation', 'spatial'))
self.mode = mode
......
theano/gpuarray/dnn_base.c
浏览文件 @ 109b378c
......@@ -115,11 +115,7 @@ c_set_filter(PyGpuArrayObject *var, cudnnFilterDescriptor_t desc) {
if (nd < 3)
nd = 3;
#if CUDNN_VERSION >= 5000
err = cudnnSetFilterNdDescriptor(desc, dt, CUDNN_TENSOR_NCHW, nd, dims);
#else
err = cudnnSetFilterNdDescriptor(desc, dt, nd, dims);
#endif
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError,
......
theano/gpuarray/dnn_pool.c
浏览文件 @ 109b378c
......@@ -71,11 +71,7 @@ int APPLY_SPECIFIC(dnn_pool)(PyGpuArrayObject *img,
s[i] = *((npy_intp*)PyArray_GETPTR1(stride, i));
}
#if CUDNN_VERSION >= 5000
err = cudnnSetPoolingNdDescriptor(APPLY_SPECIFIC(pool), MODE_FLAG, CUDNN_PROPAGATE_NAN, ndims, w, p, s);
#else
err = cudnnSetPoolingNdDescriptor(APPLY_SPECIFIC(pool), MODE_FLAG, ndims, w, p, s);
#endif
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError, "could not set op descriptor %s", cudnnGetErrorString(err));
......
theano/gpuarray/dnn_pool_grad.c
浏览文件 @ 109b378c
......@@ -111,11 +111,7 @@ int APPLY_SPECIFIC(dnn_pool_grad)(PyGpuArrayObject *inp,
s[i] = *((npy_intp*)PyArray_GETPTR1(stride, i));
}
#if CUDNN_VERSION >= 5000
err = cudnnSetPoolingNdDescriptor(APPLY_SPECIFIC(pool), MODE_FLAG, CUDNN_PROPAGATE_NAN, ndims, w, p, s);
#else
err = cudnnSetPoolingNdDescriptor(APPLY_SPECIFIC(pool), MODE_FLAG, ndims, w, p, s);
#endif
if (err != CUDNN_STATUS_SUCCESS) {
PyErr_Format(PyExc_RuntimeError, "could not set op descriptor %s", cudnnGetErrorString(err));
......
theano/gpuarray/tests/test_dnn.py
浏览文件 @ 109b378c
......@@ -604,9 +604,6 @@ class TestDnnInferShapes(utt.InferShapeTester):
[conv_modes[0]])),
testcase_func_name=utt.custom_name_func)
def test_conv(self, algo, border_mode, conv_mode):
if algo == 'winograd' and dnn.version(raises=False) < 5000:
raise SkipTest(dnn.dnn_available.msg)
self._test_conv(T.tensor4('img'),
T.tensor4('kerns'),
T.tensor4('out'),
......@@ -1361,8 +1358,6 @@ class test_SoftMax(test_nnet.test_SoftMax):
def test_dnn_batchnorm_train():
if not dnn.dnn_available(test_ctx_name):
raise SkipTest(dnn.dnn_available.msg)
if dnn.version(raises=False) < 5000:
raise SkipTest("batch normalization requires cudnn v5+")
utt.seed_rng()
for mode in ('per-activation', 'spatial'):
......@@ -1416,8 +1411,6 @@ def test_dnn_batchnorm_train():
def test_batchnorm_inference():
if not dnn.dnn_available(test_ctx_name):
raise SkipTest(dnn.dnn_available.msg)
if dnn.version(raises=False) < 5000:
raise SkipTest("batch normalization requires cudnn v5+")
utt.seed_rng()
for mode in ('per-activation', 'spatial'):
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论