Split local_abstractconv_cudnn into 3 opt to have better stats.

theano/gpuarray/dnn.py

@@ -32,7 +32,7 @@ from .elemwise import GpuElemwise
 # These don't exist in gpuarray
 # GpuDownsampleFactorMax, GpuDownsampleFactorMaxGrad
 from .nnet import GpuSoftmax
-from .opt import (gpu_seqopt, register_opt, conv_groupopt,
+from .opt import (gpu_seqopt, register_opt,
                   op_lifter, register_opt2)
 
 from .opt_util import alpha_merge, output_merge, inplace_allocempty
@@ -1472,18 +1472,31 @@ def local_abstractconv_cudnn_graph(op, context_name, inputs, outputs):
     return [rval]
 
 
-@local_optimizer([AbstractConv2d, AbstractConv2d_gradWeights,
-                  AbstractConv2d_gradInputs])
+@register_opt('fast_compile',  'conv_dnn', 'cudnn')
+@local_optimizer([AbstractConv2d])
 def local_abstractconv_cudnn(node):
     ctx = infer_context_name(*node.inputs)
     if not isinstance(node.inputs[0].type, GpuArrayType):
         return
     return local_abstractconv_cudnn_graph(node.op, ctx, node.inputs, node.outputs)
 
-conv_groupopt.register('local_abstractconv_cudnn',
-                       local_abstractconv_cudnn, 20,
-                       'fast_compile', 'fast_run',
-                       'gpuarray', 'conv_dnn', 'cudnn')
+
+@register_opt('fast_compile',  'conv_dnn', 'cudnn')
+@local_optimizer([AbstractConv2d_gradWeights])
+def local_abstractconv_gw_cudnn(node):
+    ctx = infer_context_name(*node.inputs)
+    if not isinstance(node.inputs[0].type, GpuArrayType):
+        return
+    return local_abstractconv_cudnn_graph(node.op, ctx, node.inputs, node.outputs)
+
+
+@register_opt('fast_compile',  'conv_dnn', 'cudnn')
+@local_optimizer([AbstractConv2d_gradInputs])
+def local_abstractconv_gi_cudnn(node):
+    ctx = infer_context_name(*node.inputs)
+    if not isinstance(node.inputs[0].type, GpuArrayType):
+        return
+    return local_abstractconv_cudnn_graph(node.op, ctx, node.inputs, node.outputs)
 
 
 @inplace_allocempty(GpuDnnConv, 2)

theano/gpuarray/opt.py

@@ -79,10 +79,6 @@ class GraphToGPUDB(DB):
 
 gpu_seqopt = SequenceDB()
 
-# Don't register this right now
-conv_groupopt = LocalGroupDB()
-conv_groupopt.__name__ = "gpua_conv_opts"
-
 gpu_seqopt.register('gpuarray_graph_optimization', GraphToGPUDB(), -0.5,
                     'fast_compile', 'fast_run', 'gpuarray')
 
@@ -1297,9 +1293,6 @@ def local_gpua_lift_abstractconv2d_graph(op, context_name, inputs, outputs):
                                    context_name=context_name)
     return [op(*inps)]
 
-# Register this here so that it goes after the abstract lifting
-register_opt('fast_compile')(conv_groupopt)
-
 
 @register_opt("low_memory")
 @local_optimizer([GpuCAReduceCuda])