Added register_opt2 tags for other opts

theano/gpuarray/dnn.py

@@ -31,7 +31,7 @@ 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, op_lifter
+from .opt import gpu_seqopt, register_opt, conv_groupopt, op_lifter, register_opt2
 from .opt_util import alpha_merge, output_merge, inplace_allocempty
 
 from theano.configdefaults import SUPPORTED_DNN_CONV_ALGO_BWD_FILTER
@@ -1498,6 +1498,7 @@ def local_dnn_convi_output_merge(node, *inputs):
 
 @register_opt('cudnn', 'fast_compile')
 @op_lifter([Pool])
+@register_opt2([Pool], 'fast_compile')
 def local_pool_dnn_alternative(op, ctx_name, inputs):
     if not dnn_available(ctx_name):
         raise_no_cudnn()
@@ -1514,10 +1515,11 @@ def local_pool_dnn_alternative(op, ctx_name, inputs):
 
 @register_opt('cudnn', 'fast_compile')
 @op_lifter([MaxPoolGrad])
+@register_opt2([MaxPoolGrad], 'fast_compile')
 def local_pool_dnn_grad_stride(op, ctx_name, inputs):
     if not dnn_available(ctx_name):
         raise_no_cudnn()
-    if not node.op.ignore_border:
+    if not op.ignore_border:
         return
     inp, out, out_grad = inputs
     inp = as_gpuarray_variable(inp, ctx_name)
@@ -1538,6 +1540,7 @@ def local_pool_dnn_grad_stride(op, ctx_name, inputs):
 
 @register_opt('cudnn', 'fast_compile')
 @op_lifter([AveragePoolGrad])
+@register_opt2([AveragePoolGrad], 'fast_compile')
 def local_avg_pool_dnn_grad_stride(op, ctx_name, inputs):
     if not dnn_available(ctx_name):
         raise_no_cudnn()
@@ -1591,6 +1594,7 @@ def local_log_softmax_dnn(node):
 
 @register_opt('cudnn', 'fast_compile')
 @op_lifter([LogSoftmax])
+@register_opt2([LogSoftmax], 'fast_compile')
 def local_logsoftmax_to_dnn(op, ctx_name, inputs):
     # Transform the input in the format expected by GpuDnnSoftmax
     inp = inputs[0]
@@ -1629,6 +1633,7 @@ gpu_seqopt.register("NoCuDNNRaise", NoCuDNNRaise(), 0, 'cudnn')
 
 @register_opt('cudnn', 'fast_compile')
 @op_lifter([SoftmaxGrad])
+#@register_opt2([SoftmaxGrad], 'fast_compile')
 def local_softmax_dnn_grad(op, ctx_name, inputs):
     if not dnn_available(ctx_name):
         raise_no_cudnn("cuDNN needed for SoftmaxGrad")

theano/gpuarray/extra_ops.py

@@ -10,7 +10,7 @@ except ImportError:
 
 from .basic_ops import (as_gpuarray_variable, GpuKernelBase, Kernel,
                         infer_context_name, GpuFromHost)
-from .opt import register_opt as register_gpu_opt, op_lifter
+from .opt import register_opt, op_lifter, register_opt2
 
 
 class GpuCumsum(GpuKernelBase, Op):
@@ -450,8 +450,9 @@ class GpuCumsum(GpuKernelBase, Op):
         """ % locals()
         return super(GpuCumsum, self).c_support_code_struct(node, nodename) + code
 
-
+@register_opt('fast_compile')
 @op_lifter([CumsumOp])
+@register_opt2([CumsumOp], 'fast_compile')
 def use_gpu_cumsumop(op, ctx_name, inputs):
     if inputs[0].dtype == 'float32':
         axis = op.axis
@@ -471,4 +472,5 @@ def use_gpu_cumsumop(op, ctx_name, inputs):
 
         return GpuCumsum(axis)(x)
 
-register_gpu_opt()(use_gpu_cumsumop)
+#register_opt('fast_compile')(use_gpu_cumsumop)
+#register_opt2([CumsumOp], 'fast_compile')(use_gpu_cumsumop)
\ No newline at end of file

theano/gpuarray/multinomial.py

@@ -14,7 +14,7 @@ from theano.gof import Op
 from theano.tensor import NotScalarConstantError, get_scalar_constant_value
 from theano import gpuarray
 from .basic_ops import as_gpuarray_variable, infer_context_name
-from .opt import register_opt, op_lifter
+from .opt import register_opt, op_lifter, register_opt2
 from .type import GpuArrayType
 
 
@@ -227,8 +227,9 @@ KERNEL void k_multi_warp_multinomial(
         return (1,)
 
 
-@register_opt()
+@register_opt('fast_compile')
 @op_lifter([theano.sandbox.multinomial.MultinomialFromUniform])
+@register_opt2([theano.sandbox.multinomial.MultinomialFromUniform], 'fast_compile')
 def local_gpua_multinomial(op, context_name, inputs):
     # TODO : need description for function
 
@@ -242,6 +243,9 @@ def local_gpua_multinomial(op, context_name, inputs):
             return None
     except NotScalarConstantError:
         return None
+
+    node = op.make_node(*inputs)
+    outputs = node.outputs
     m, = outputs
     if (p.dtype == u.dtype == m.dtype == 'float32'):
         gpu_op = GPUAMultinomialFromUniform(op.odtype)

theano/gpuarray/nerv.py

@@ -147,8 +147,9 @@ if (GpuKernel_init(&k_%(name)s, c->ctx, 1, &bcode, &sz,
         return '\n'.join(codel)
 
 
-@opt.register_opt()
+@opt.register_opt('fast_compile')
 @opt.op_lifter([tensor.Dot])
+@opt.register_opt2([tensor.Dot], 'fast_compile')
 def local_dot_to_gemm16(op, ctx_name, inputs):
     if nerv is None:
         return