Merge pull request #4344 from mgermain/cudnn5

.travis.yml

@@ -37,6 +37,7 @@ install:
   - source activate pyenv
   - if [[ $TRAVIS_PYTHON_VERSION == '2.6' ]]; then pip install pydot; fi
   - pip install . --no-deps
+  - pip install nose-parameterized==0.5.0
 
 # command to run tests
 env:

doc/install.txt

@@ -49,7 +49,7 @@ instructions below for detailed installation steps):
 
 The following libraries and software are optional:
 
-    `nose <http://somethingaboutorange.com/mrl/projects/nose/>`_ >= 1.3.0
+    `nose <http://nose.readthedocs.org/en/latest/>`_ >= 1.3.0 and `nose-parameterized <https://pypi.python.org/pypi/nose-parameterized/>`_ >= 0.5.0
         Recommended, to run Theano's test-suite.
 
     `Sphinx <http://sphinx.pocoo.org/>`_ >= 0.5.1, `pygments <http://pygments.org/>`_

requirement-rtd.txt

 sphinx>=1.3.0
+pygments
 nose>=1.3.0
+nose-parameterized>=0.5.0

setup.py

@@ -163,6 +163,11 @@ def do_setup():
           packages=find_packages(),
           # 1.7.0 give too much warning related to numpy.diagonal.
           install_requires=['numpy>=1.7.1', 'scipy>=0.11', 'six>=1.9.0'],
+          # pygments is a dependency for Sphinx code highlight
+          extras_require={
+              'test': ['nose>=1.3.0', 'nose-parameterized>=0.5.0'],
+              'doc': ['Sphinx>=0.5.1', 'pygments']
+          },
           package_data={
               '': ['*.txt', '*.rst', '*.cu', '*.cuh', '*.c', '*.sh', '*.pkl',
                    '*.h', '*.cpp', 'ChangeLog'],

theano/configdefaults.py

@@ -286,6 +286,20 @@ def safe_no_dnn_algo_bwd(algo):
             '`dnn.conv.algo_bwd_filter` and `dnn.conv.algo_bwd_data` instead.')
     return True
 
+# Those are the supported algorithm by Theano,
+# The tests will reference those lists.
+SUPPORTED_DNN_CONV_ALGO_FWD = ('small', 'none', 'large', 'fft', 'fft_tiling',
+                               'winograd', 'guess_once', 'guess_on_shape_change',
+                               'time_once', 'time_on_shape_change')
+
+SUPPORTED_DNN_CONV_ALGO_BWD_DATA = ('none', 'deterministic', 'fft', 'fft_tiling',
+                                    'winograd', 'guess_once', 'guess_on_shape_change',
+                                    'time_once', 'time_on_shape_change')
+
+SUPPORTED_DNN_CONV_ALGO_BWD_FILTER = ('none', 'deterministic', 'fft', 'small',
+                                      'guess_once', 'guess_on_shape_change',
+                                      'time_once', 'time_on_shape_change')
+
 AddConfigVar('dnn.conv.algo_bwd',
              "This flag is deprecated; use dnn.conv.algo_bwd_data and "
              "dnn.conv.algo_bwd_filter.",
@@ -295,26 +309,20 @@ AddConfigVar('dnn.conv.algo_bwd',
 
 AddConfigVar('dnn.conv.algo_fwd',
              "Default implementation to use for CuDNN forward convolution.",
-             EnumStr('small', 'none', 'large', 'fft', 'fft_tiling',
-                     'guess_once', 'guess_on_shape_change',
-                     'time_once', 'time_on_shape_change'),
+             EnumStr(*SUPPORTED_DNN_CONV_ALGO_FWD),
              in_c_key=False)
 
 AddConfigVar('dnn.conv.algo_bwd_data',
              "Default implementation to use for CuDNN backward convolution to "
              "get the gradients of the convolution with regard to the inputs.",
-             EnumStr('none', 'deterministic', 'fft', 'fft_tiling',
-                     'guess_once', 'guess_on_shape_change', 'time_once',
-                     'time_on_shape_change'),
+             EnumStr(*SUPPORTED_DNN_CONV_ALGO_BWD_DATA),
              in_c_key=False)
 
 AddConfigVar('dnn.conv.algo_bwd_filter',
              "Default implementation to use for CuDNN backward convolution to "
              "get the gradients of the convolution with regard to the "
              "filters.",
-             EnumStr('none', 'deterministic', 'fft', 'small', 'guess_once',
-                     'guess_on_shape_change', 'time_once',
-                     'time_on_shape_change'),
+             EnumStr(*SUPPORTED_DNN_CONV_ALGO_BWD_FILTER),
              in_c_key=False)
 
 AddConfigVar('dnn.conv.precision',

theano/sandbox/cuda/dnn_base.c

@@ -54,7 +54,10 @@ c_set_filterNd(CudaNdarray *var, cudnnFilterDescriptor_t desc) {
     return -1;
   }
   int dim = CudaNdarray_NDIM(var);
-  cudnnStatus_t err = cudnnSetFilterNdDescriptor(desc, CUDNN_DATA_FLOAT, dim,
+  cudnnStatus_t err = cudnnSetFilterNdDescriptor_v4(desc,
+                                                    CUDNN_DATA_FLOAT,
+                                                    CUDNN_TENSOR_NCHW,
+                                                    dim,
                                                     CudaNdarray_HOST_DIMS(var));
   if (err != CUDNN_STATUS_SUCCESS) {
     PyErr_Format(PyExc_RuntimeError,

theano/sandbox/cuda/dnn_fwd.c

@@ -179,7 +179,7 @@ APPLY_SPECIFIC(conv_fwd)(CudaNdarray *input, CudaNdarray *kerns,
       int upscale[2];
       cudnnConvolutionMode_t mode;
       cudnnDataType_t data_type;
-      err = cudnnGetConvolutionNdDescriptor_v3(desc, 2, &nd, pad, stride,
+      err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                             upscale, &mode, &data_type);
 
       if (err != CUDNN_STATUS_SUCCESS) {
@@ -224,6 +224,19 @@ APPLY_SPECIFIC(conv_fwd)(CudaNdarray *input, CudaNdarray *kerns,
                                                   APPLY_SPECIFIC(output),
                                                   chosen_algo,
                                                   &worksize);
+    if (err == CUDNN_STATUS_NOT_SUPPORTED) {
+      // Fallback to none algo if not supported
+      // TODO: Print a warning
+      chosen_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
+
+      err = cudnnGetConvolutionForwardWorkspaceSize(_handle,
+                                                    APPLY_SPECIFIC(input),
+                                                    APPLY_SPECIFIC(kerns),
+                                                    desc,
+                                                    APPLY_SPECIFIC(output),
+                                                    chosen_algo,
+                                                    &worksize);
+    }
     if (err != CUDNN_STATUS_SUCCESS) {
       PyErr_Format(PyExc_RuntimeError,
                    "GpuDnnConv: error getting worksize: %s",

theano/sandbox/cuda/dnn_gi.c

@@ -178,7 +178,7 @@ APPLY_SPECIFIC(conv_gi)(CudaNdarray *kerns, CudaNdarray *output,
       int upscale[2];
       cudnnConvolutionMode_t mode;
       cudnnDataType_t data_type;
-      err = cudnnGetConvolutionNdDescriptor_v3(desc, 2, &nd, pad, stride,
+      err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                             upscale, &mode, &data_type);
 
       if (err != CUDNN_STATUS_SUCCESS) {
@@ -237,7 +237,7 @@ APPLY_SPECIFIC(conv_gi)(CudaNdarray *kerns, CudaNdarray *output,
       return 1;
 
     // Perform the convolution
-    err = cudnnConvolutionBackwardData_v3(
+    err = cudnnConvolutionBackwardData(
       _handle,
       (void *)&alpha,
       APPLY_SPECIFIC(kerns), CudaNdarray_DEV_DATA(kerns),

theano/sandbox/cuda/dnn_gw.c

@@ -173,7 +173,7 @@ APPLY_SPECIFIC(conv_gw)(CudaNdarray *input, CudaNdarray *output,
       int upscale[2];
       cudnnConvolutionMode_t mode;
       cudnnDataType_t data_type;
-      err = cudnnGetConvolutionNdDescriptor_v3(desc, 2, &nd, pad, stride,
+      err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                             upscale, &mode, &data_type);
 
       if (err != CUDNN_STATUS_SUCCESS) {
@@ -221,7 +221,7 @@ APPLY_SPECIFIC(conv_gw)(CudaNdarray *input, CudaNdarray *output,
       return 1;
 
     // Perform the convolution
-    err = cudnnConvolutionBackwardFilter_v3(
+    err = cudnnConvolutionBackwardFilter(
       _handle,
       (void *)&alpha,
       APPLY_SPECIFIC(input), CudaNdarray_DEV_DATA(input),

theano/sandbox/cuda/tests/test_dnn.py

@@ -392,8 +392,9 @@ def test_pooling_with_tensor_vars():
 
 
 def test_old_pool_interface():
-    if not cuda.dnn.dnn_available():
+    if not cuda.dnn.dnn_available() or cuda.dnn.version() > (5000, 5000):
         raise SkipTest(cuda.dnn.dnn_available.msg)
+
     testfile_dir = os.path.dirname(os.path.realpath(__file__))
     fname = 'old_pool_interface.pkl'
     with open(os.path.join(testfile_dir, fname), 'rb') as fp:

theano/sandbox/gpuarray/conv_desc.c

@@ -35,7 +35,7 @@ int APPLY_SPECIFIC(conv_desc)(PyArrayObject *filt_shp,
     return -1;
   }
 
-  err = cudnnSetConvolutionNdDescriptor_v3(*desc, NB_DIMS, pad, strides,
+  err = cudnnSetConvolutionNdDescriptor(*desc, NB_DIMS, pad, strides,
                                         upscale, CONV_MODE, PRECISION);
   return 0;
 }

theano/sandbox/gpuarray/dnn.py

@@ -33,6 +33,8 @@ from .nnet import GpuSoftmax
 from .opt import gpu_seqopt, register_opt, conv_groupopt, op_lifter
 from .opt_util import alpha_merge, output_merge, inplace_allocempty
 
+from theano.configdefaults import SUPPORTED_DNN_CONV_ALGO_BWD_FILTER
+
 
 def raise_no_cudnn(msg="CuDNN is required for convolution and pooling"):
     raise RuntimeError(msg)
@@ -232,6 +234,7 @@ def version(raises=True):
 
     :raises: If True, raise an exception if CuDNN is not present or badly installed.
         Otherwise, return -1.
+
     """
     if not dnn_present():
         if raises:
@@ -397,9 +400,9 @@ class GpuDnnConv(DnnBase):
     ----------
     image
     kernel
-    descr
+    descr :
         The convolution descriptor.
-    algo : {'small', 'none', 'large', 'fft', 'fft_tiling', 'guess_once',
+    algo : {'small', 'none', 'large', 'fft', 'fft_tiling', 'winograd', 'guess_once',
             'guess_on_shape_change', 'time_once', 'time_on_shape_change'}
         Default is the value of :attr:`config.dnn.conv.algo_fwd`.
 
@@ -435,8 +438,12 @@ class GpuDnnConv(DnnBase):
                 raise RuntimeError("CuDNN tiled-FFT convolution requires "
                                    "CuDNN v4 or more recent")
 
+        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',
-                             'guess_once', 'guess_on_shape_change',
+                             'winograd', 'guess_once', 'guess_on_shape_change',
                              'time_once', 'time_on_shape_change']
 
     def __setstate__(self, d):
@@ -468,6 +475,9 @@ class GpuDnnConv(DnnBase):
         elif self.algo == 'fft_tiling':
             # need v4
             alg = 'CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING'
+        elif self.algo == 'winograd':
+            # need v5
+            alg = 'CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD'
         defs.append(('CONV_ALGO', alg))
 
         if self.algo in ['guess_once', 'guess_on_shape_change',
@@ -565,8 +575,11 @@ class GpuDnnConvGradW(DnnBase):
     ----------
     image
     kernel
-    descr
+    descr :
         The convolution descriptor.
+    algo : {'none', 'deterministic', 'fft', 'small', 'guess_once',
+            'guess_on_shape_change', 'time_once', 'time_on_shape_change'}
+        Default is the value of :attr:`config.dnn.conv.algo_bwd_filter`.
 
     """
 
@@ -582,9 +595,7 @@ class GpuDnnConvGradW(DnnBase):
             algo = config.dnn.conv.algo_bwd_filter
         self.algo = algo
 
-        assert self.algo in ['none', 'deterministic', 'fft', 'small',
-                             'guess_once', 'guess_on_shape_change',
-                             'time_once', 'time_on_shape_change']
+        assert self.algo in SUPPORTED_DNN_CONV_ALGO_BWD_FILTER
 
     def __setstate__(self, d):
         self.__dict__.update(d)
@@ -688,6 +699,9 @@ class GpuDnnConvGradI(DnnBase):
     kernel
     descr
         The convolution descriptor.
+    algo : {'none', 'deterministic', 'fft', 'fft_tiling', 'winograd', 'guess_once',
+            'guess_on_shape_change', 'time_once', 'time_on_shape_change'}
+        Default is the value of :attr:`config.dnn.conv.algo_bwd_data`.
 
     """
 
@@ -708,9 +722,12 @@ class GpuDnnConvGradI(DnnBase):
         if version() < 4000 and self.algo == 'fft_tiling':
             raise RuntimeError("CuDNN's tiled-FFT convolution requires CuDNN "
                                "v4 or more recent")
+        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',
-                             'guess_once', 'guess_on_shape_change',
+                             'winograd', 'guess_once', 'guess_on_shape_change',
                              'time_once', 'time_on_shape_change']
 
     def __setstate__(self, d):
@@ -749,13 +766,16 @@ class GpuDnnConvGradI(DnnBase):
             alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_0'
             if self.algo == 'none':
                 alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_0'
-            if self.algo == 'deterministic':
+            elif self.algo == 'deterministic':
                 alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_1'
-            if self.algo == 'fft':
+            elif self.algo == 'fft':
                 alg = 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT'
-            if self.algo == 'fft_tiling':
+            elif self.algo == 'fft_tiling':
                 # 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',
                              'time_once', 'time_on_shape_change']:
@@ -1047,9 +1067,13 @@ class GpuDnnPoolDesc(Op):
   static const int win[%(nd)d] = {%(win)s};
   static const int pad[%(nd)d] = {%(pad)s};
   static const int str[%(nd)d] = {%(str)s};
-  err = cudnnSetPoolingNdDescriptor(
-    %(desc)s, %(mode_flag)s, %(nd)d,
-    win, pad, str);
+
+#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",
                  cudnnGetErrorString(err));
@@ -1062,7 +1086,7 @@ class GpuDnnPoolDesc(Op):
            str=', '.join(map(str, self.stride)))
 
     def c_code_cache_version(self):
-        return (3, version())
+        return (4, version())
 
 
 class GpuDnnPool(DnnBase):
@@ -1070,18 +1094,17 @@ class GpuDnnPool(DnnBase):
     """
     Parameters
     ----------
-    img
+    img : tensor
         The image 4d or 5d tensor.
-    Parameters
-    ----------
-    ws : tensor variable
+    ws : tensor
         Window size.
-    stride : tensor variable
+    stride : tensor
         (dx, dy) or (dx, dy, dz).
     mode : {'max', 'average_inc_pad', 'average_exc_pad'}
         The old deprecated name 'average' corresponds to 'average_inc_pad'.
     pad : tensor
         (padX, padY) or (padX, padY, padZ)
+
     """
 
     __props__ = ('mode',)
@@ -1255,14 +1278,12 @@ class GpuDnnSoftmaxBase(DnnBase):
 
     Parameters
     ----------
-    algo
-        'fast', 'accurate' or 'log' indicating whether, respectively,
-        computations should be optimized for speed, for accuracy, or if CuDNN
-        should rather compute the log-softmax instead.
-    mode
-        'instance' or 'channel' indicating whether the softmax should be
-        computed per image across 'c01' or per spatial location '01' per
-        image across 'c'.
+    algo : {'fast', 'accurate', 'log'}
+        Indicating whether, respectively, computations should be optimized for
+        speed, for accuracy, or if CuDNN should rather compute the log-softmax instead.
+    mode : {'instance', 'channel'}
+        Indicating whether the softmax should be computed per image across 'c01'
+        or per spatial location '01' per image across 'c'.
 
     """
 
@@ -1306,14 +1327,12 @@ class GpuDnnSoftmax(GpuDnnSoftmaxBase):
     """
     Op for the cuDNN Softmax.
 
-    algo
-        'fast', 'accurate' or 'log' indicating whether, respectively,
-        computations should be optimized for speed, for accuracy, or if CuDNN
-        should rather compute the log-softmax instead.
-    mode
-        'instance' or 'channel' indicating whether the softmax should be
-        computed per image across 'c01' or per spatial location '01' per
-        image across 'c'.
+    algo : {'fast', 'accurate', 'log'}
+        Indicating whether, respectively, computations should be optimized for
+        speed, for accuracy, or if CuDNN should rather compute the log-softmax instead.
+    mode : {'instance', 'channel'}
+        Indicating whether the softmax should be computed per image across 'c01'
+        or per spatial location '01' per image across 'c'.
 
     """
     direction = "forward"

theano/sandbox/gpuarray/dnn_base.c

@@ -51,6 +51,8 @@ c_set_tensorNd(PyGpuArrayObject *var, cudnnTensorDescriptor_t desc) {
 static int
 c_set_filter(PyGpuArrayObject *var, cudnnFilterDescriptor_t desc) {
   cudnnDataType_t dt;
+  cudnnStatus_t err;
+
   if (!GpuArray_IS_C_CONTIGUOUS(&var->ga)) {
     PyErr_SetString(PyExc_ValueError,
 		    "Only contiguous filters (kernels) are supported.");
@@ -86,7 +88,12 @@ c_set_filter(PyGpuArrayObject *var, cudnnFilterDescriptor_t desc) {
     dims[i] = PyGpuArray_DIM(var, i);
   }
 
-  cudnnStatus_t err = cudnnSetFilterNdDescriptor(desc, dt, nd, dims);
+#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,
 		 "Could not set filter descriptor: %s.",

theano/sandbox/gpuarray/dnn_fwd.c

@@ -92,12 +92,12 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
     }
     algo = choice.algo;
 #else
-    size_t free = 0, total = 0;
-    cudaError_t err2 = cudaMemGetInfo(&free, &total);
-    if (err2 != cudaSuccess) {
+    size_t free;
+    int err2 = c->ops->property(c->ctx, NULL, NULL, GA_CTX_PROP_FREE_GMEM, &free);
+
+    if (err2 != GA_NO_ERROR) {
       PyErr_Format(PyExc_RuntimeError, "Error when trying to find the "
-                   "memory information on the GPU: %s\n",
-                   cudaGetErrorString(err2));
+                   "memory information on the GPU");
       cuda_exit(c->ctx);
       return 1;
     }
@@ -154,7 +154,7 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
     int upscale[2];
     cudnnConvolutionMode_t mode;
     cudnnDataType_t data_type;
-    err = cudnnGetConvolutionNdDescriptor_v3(desc, 2, &nd, pad, stride,
+    err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                              upscale, &mode, &data_type);
     if (err != CUDNN_STATUS_SUCCESS) {
       PyErr_Format(PyExc_RuntimeError,
@@ -193,6 +193,21 @@ APPLY_SPECIFIC(conv_fwd)(PyGpuArrayObject *input, PyGpuArrayObject *kerns,
                                                   APPLY_SPECIFIC(output),
                                                   algo,
                                                   &worksize);
+
+    if (err == CUDNN_STATUS_NOT_SUPPORTED) {
+      // Fallback to none algo if not supported
+      // TODO: Print a warning
+      algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
+
+      err = cudnnGetConvolutionForwardWorkspaceSize(APPLY_SPECIFIC(_handle),
+                                                    APPLY_SPECIFIC(input),
+                                                    APPLY_SPECIFIC(kerns),
+                                                    desc,
+                                                    APPLY_SPECIFIC(output),
+                                                    algo,
+                                                    &worksize);
+    }
+
     if (err != CUDNN_STATUS_SUCCESS) {
       PyErr_Format(PyExc_RuntimeError,
                    "error getting worksize: %s",

theano/sandbox/gpuarray/dnn_gi.c

@@ -91,12 +91,12 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
 
     algo = choice.algo;
 #else
-    size_t free = 0, total = 0;
-    cudaError_t err2 = cudaMemGetInfo(&free, &total);
-    if (err2 != cudaSuccess){
-      cudaGetLastError();
-      PyErr_Format(PyExc_RuntimeError, "Error when trying to find the memory "
-                   "information on the GPU: %s\n", cudaGetErrorString(err2));
+    size_t free;
+    int err2 = c->ops->property(c->ctx, NULL, NULL, GA_CTX_PROP_FREE_GMEM, &free);
+
+    if (err2 != GA_NO_ERROR) {
+      PyErr_Format(PyExc_RuntimeError, "Error when trying to find the "
+                   "memory information on the GPU");
       cuda_exit(c->ctx);
       return 1;
     }
@@ -146,7 +146,7 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
     int upscale[2];
     cudnnConvolutionMode_t mode;
     cudnnDataType_t data_type;
-    err = cudnnGetConvolutionNdDescriptor_v3(desc, 2, &nd, pad, stride,
+    err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                              upscale, &mode, &data_type);
     if (err != CUDNN_STATUS_SUCCESS) {
       PyErr_Format(PyExc_RuntimeError,
@@ -203,7 +203,7 @@ APPLY_SPECIFIC(conv_gi)(PyGpuArrayObject *kerns, PyGpuArrayObject *output,
   cuda_wait(output->ga.data, GPUARRAY_CUDA_WAIT_READ);
   cuda_wait((*input)->ga.data, GPUARRAY_CUDA_WAIT_WRITE);
 
-  err = cudnnConvolutionBackwardData_v3(
+  err = cudnnConvolutionBackwardData(
     APPLY_SPECIFIC(_handle),
     alpha_p,
     APPLY_SPECIFIC(kerns), PyGpuArray_DEV_DATA(kerns),

theano/sandbox/gpuarray/dnn_gw.c

@@ -92,12 +92,12 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
 
     algo = choice.algo;
 #else
-    size_t free = 0, total = 0;
-    cudaError_t err2 = cudaMemGetInfo(&free, &total);
-    if (err2 != cudaSuccess){
-      cudaGetLastError();
-      PyErr_Format(PyExc_RuntimeError, "Error when trying to find the memory "
-                   "information on the GPU: %s\n", cudaGetErrorString(err2));
+    size_t free;
+    int err2 = c->ops->property(c->ctx, NULL, NULL, GA_CTX_PROP_FREE_GMEM, &free);
+
+    if (err2 != GA_NO_ERROR) {
+      PyErr_Format(PyExc_RuntimeError, "Error when trying to find the "
+                   "memory information on the GPU");
       cuda_exit(c->ctx);
       return 1;
     }
@@ -146,7 +146,7 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
     int upscale[2];
     cudnnConvolutionMode_t mode;
     cudnnDataType_t data_type;
-    err = cudnnGetConvolutionNdDescriptor_v3(desc, 2, &nd, pad, stride,
+    err = cudnnGetConvolutionNdDescriptor(desc, 2, &nd, pad, stride,
                                              upscale, &mode, &data_type);
     if (err != CUDNN_STATUS_SUCCESS) {
       PyErr_Format(PyExc_RuntimeError,
@@ -190,7 +190,7 @@ APPLY_SPECIFIC(conv_gw)(PyGpuArrayObject *input, PyGpuArrayObject *output,
   cuda_wait(output->ga.data, GPUARRAY_CUDA_WAIT_READ);
   cuda_wait((*kerns)->ga.data, GPUARRAY_CUDA_WAIT_WRITE);
 
-  err = cudnnConvolutionBackwardFilter_v3(
+  err = cudnnConvolutionBackwardFilter(
     APPLY_SPECIFIC(_handle),
     alpha_p,
     APPLY_SPECIFIC(input), PyGpuArray_DEV_DATA(input),

theano/sandbox/gpuarray/dnn_pool.c

@@ -69,7 +69,12 @@ int APPLY_SPECIFIC(dnn_pool)(PyGpuArrayObject *img,
   for(int i = 0; i < ndims; i++) {
      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/sandbox/gpuarray/dnn_pool_grad.c

@@ -109,7 +109,12 @@ int APPLY_SPECIFIC(dnn_pool_grad)(PyGpuArrayObject *inp,
   for(int i = 0; i < ndims; i++) {
      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/sandbox/gpuarray/tests/test_dnn.py

 import logging
 
 from nose.plugins.skip import SkipTest
+from nose_parameterized import parameterized
 import numpy
-from itertools import product
+from itertools import product, chain
 
 import theano
 from six import StringIO
@@ -18,6 +19,8 @@ from ..basic_ops import GpuAllocEmpty
 from .config import mode_with_gpu, mode_without_gpu, test_ctx_name
 from . import test_nnet
 
+from theano.configdefaults import SUPPORTED_DNN_CONV_ALGO_FWD
+
 
 def test_dnn_conv_desc_merge():
     if not dnn.dnn_available(test_ctx_name):
@@ -393,6 +396,9 @@ def test_dnn_tag():
 
 class TestDnnInferShapes(utt.InferShapeTester):
 
+    border_modes = ['valid', 'full', 'half']
+    conv_modes = ['conv', 'cross']
+
     def setUp(self):
         super(TestDnnInferShapes, self).setUp()
         self.mode = mode_with_gpu
@@ -427,37 +433,25 @@ class TestDnnInferShapes(utt.InferShapeTester):
             dnn.GpuDnnSoftmaxGrad
         )
 
-    def test_conv(self):
+    def _test_conv(self, img, kerns, out, img_val, kern_vals, border_mode, conv_mode, subsamples, algo):
         if not dnn.dnn_available(test_ctx_name):
             raise SkipTest(dnn.dnn_available.msg)
-        img = T.ftensor4('img')
-        kerns = T.ftensor4('kerns')
-        out = T.ftensor4('out')
-        img_val = numpy.asarray(
-            numpy.random.rand(7, 2, 6, 4),
-            dtype='float32'
-        )
-        kern_vals = numpy.asarray(
-            numpy.random.rand(8, 2, 4, 3),
-            dtype='float32'
-        )
 
-        for params in product(
-            ['valid', 'full', 'half'],
-            [(1, 1), (2, 2)],
-            ['conv', 'cross']
-        ):
+        img_val = numpy.asarray(img_val, dtype='float32')
+        kern_vals = numpy.asarray(kern_vals, dtype='float32')
+
+        for subsample in subsamples:
             out_vals = numpy.zeros(
                 dnn.GpuDnnConv.get_out_shape(img_val.shape, kern_vals.shape,
-                                             border_mode=params[0],
-                                             subsample=params[1]),
+                                             border_mode=border_mode,
+                                             subsample=subsample),
                 dtype='float32')
             desc = dnn.GpuDnnConvDesc(
-                border_mode=params[0],
-                subsample=params[1],
-                conv_mode=params[2]
+                border_mode=border_mode,
+                subsample=subsample,
+                conv_mode=conv_mode
             )(kerns.shape)
-            conv = dnn.GpuDnnConv()(img, kerns, out, desc)
+            conv = dnn.GpuDnnConv(algo=algo)(img, kerns, out, desc)
             self._compile_and_check(
                 [img, kerns, out],
                 [conv],
@@ -465,29 +459,56 @@ class TestDnnInferShapes(utt.InferShapeTester):
                 dnn.GpuDnnConv
             )
 
-    def test_conv_gradw(self):
+    @parameterized.expand(chain(product([SUPPORTED_DNN_CONV_ALGO_FWD[0]],
+                                        border_modes,
+                                        conv_modes),
+                                product(SUPPORTED_DNN_CONV_ALGO_FWD[1:],
+                                        [border_modes[0]],
+                                        [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() < 5000:
+            raise SkipTest(dnn.dnn_available.msg)
+
+        self._test_conv(T.ftensor4('img'),
+                        T.ftensor4('kerns'),
+                        T.ftensor4('out'),
+                        numpy.random.rand(7, 2, 8, 4),
+                        numpy.random.rand(8, 2, 4, 3),
+                        border_mode,
+                        conv_mode,
+                        [(1, 1), (2, 2)],
+                        algo)
+
+    @parameterized.expand(product(border_modes, conv_modes), utt.custom_name_func)
+    def test_conv3d_none(self, border_mode, conv_mode):
+        ftensor5 = T.TensorType(dtype="float32", broadcastable=(False,) * 5)
+        self._test_conv(ftensor5('img'),
+                        ftensor5('kerns'),
+                        ftensor5('out'),
+                        numpy.random.rand(10, 2, 6, 4, 11),
+                        numpy.random.rand(8, 2, 4, 3, 1),
+                        border_mode,
+                        conv_mode,
+                        [(1, 1, 1), (2, 2, 2)],
+                        'none')
+
+    def _test_conv_gradw(self, img, kerns, out, img_val, kern_vals, border_mode, conv_mode, subsample):
         if not dnn.dnn_available(test_ctx_name):
             raise SkipTest(dnn.dnn_available.msg)
-        img = T.ftensor4('img')
-        kerns = T.ftensor4('kerns')
-        out = T.ftensor4('out')
+
         img_val = numpy.asarray(
-            numpy.random.rand(2, 5, 6, 8),
+            img_val,
             dtype='float32'
         )
         kern_vals = numpy.asarray(
-            numpy.random.rand(2, 1, 5, 6),
+            kern_vals,
             dtype='float32'
         )
 
-        for params in product(
-            ['valid', 'full', 'half'],
-            [(1, 1)],  # strides besides (1, 1)
-            ['conv', 'cross']
-        ):
         temp_img = img.dimshuffle(1, 0, 2, 3)
         temp_kerns = kerns
-            if params[2] == 'conv':
+        if conv_mode == 'conv':
             temp_kerns = temp_kerns[:, :, ::-1, ::-1]
         temp_kerns = temp_kerns.dimshuffle(1, 0, 2, 3)
         shape = (
@@ -497,9 +518,9 @@ class TestDnnInferShapes(utt.InferShapeTester):
         )
         out_vals = numpy.zeros(shape, dtype='float32')
         desc = dnn.GpuDnnConvDesc(
-                border_mode=params[0],
-                subsample=params[1],
-                conv_mode=params[2]
+            border_mode=border_mode,
+            subsample=subsample,
+            conv_mode=conv_mode
         )(out.shape)
         conv_grad_w = dnn.GpuDnnConvGradW()(
             temp_img,
@@ -514,6 +535,17 @@ class TestDnnInferShapes(utt.InferShapeTester):
             dnn.GpuDnnConvGradW
         )
 
+    @parameterized.expand(product(border_modes, conv_modes), utt.custom_name_func)
+    def test_conv_gradw(self, border_mode, conv_mode):
+        self._test_conv_gradw(T.ftensor4('img'),
+                              T.ftensor4('kerns'),
+                              T.ftensor4('out'),
+                              numpy.random.rand(2, 5, 6, 8),
+                              numpy.random.rand(2, 1, 5, 6),
+                              border_mode,
+                              conv_mode,
+                              (1, 1))
+
     def test_conv_gradi(self):
         if not dnn.dnn_available(test_ctx_name):
             raise SkipTest(dnn.dnn_available.msg)

theano/tests/unittest_tools.py

@@ -4,6 +4,7 @@ from functools import wraps
 import logging
 import sys
 import unittest
+from nose_parameterized import parameterized
 
 from six import integer_types
 from six.moves import StringIO
@@ -31,6 +32,13 @@ except ImportError:
 _logger = logging.getLogger("theano.tests.unittest_tools")
 
 
+def custom_name_func(testcase_func, param_num, param):
+    return "%s_%s" % (
+        testcase_func.__name__,
+        parameterized.to_safe_name("_".join(str(x) for x in param.args)),
+    )
+
+
 def fetch_seed(pseed=None):
     """
     Returns the seed to use for running the unit tests.
@@ -96,6 +104,7 @@ verify_grad.E_grad = T.verify_grad.E_grad
 
 
 class TestOptimizationMixin(object):
+
     def assertFunctionContains(self, f, op, min=1, max=sys.maxsize):
         toposort = f.maker.fgraph.toposort()
         matches = [node for node in toposort if node.op == op]
@@ -172,6 +181,7 @@ class T_OpContractMixin(object):
 
 
 class InferShapeTester(unittest.TestCase):
+
     def setUp(self):
         seed_rng()
         # Take into account any mode that may be defined in a child class
@@ -311,6 +321,7 @@ def str_diagnostic(expected, value, rtol, atol):
 
 
 class WrongValue(Exception):
+
     def __init__(self, expected_val, val, rtol, atol):
         Exception.__init__(self)  # to be compatible with python2.4
         self.val1 = expected_val