Require cudnn v5 minimum.

theano/gpuarray/dnn.py

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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'):