Convert the optimizarions to work with gpuarray.

theano/sandbox/gpuarray/dnn.py

@@ -16,7 +16,6 @@ from theano.tensor.signal.downsample import (
 
 from . import pygpu, init_dev
 from .basic_ops import (as_gpuarray_variable,
-                        host_from_gpu,
                         gpu_contiguous, HostFromGpu,
                         GpuAllocEmpty, empty_like)
 from .conv import GpuConv
@@ -24,7 +23,7 @@ from .conv import GpuConv
 # These don't exist in gpuarray
 # GpuDownsampleFactorMax, GpuDownsampleFactorMaxGrad
 from .nnet import GpuSoftmax
-from .opt import gpu_seqopt, register_opt
+from .opt import gpu_seqopt, register_opt, conv_groupopt, op_lifter
 from .opt_util import alpha_merge, output_merge
 from .comp import NVCC_compiler
 
@@ -1502,11 +1501,9 @@ def local_conv_dnn(node):
         rval = dnn_conv(img, kern,
                         border_mode=border_mode, subsample=subsample,
                         direction_hint=direction_hint)
-        if node.outputs[0].broadcastable != rval.broadcastable:
-            rval = tensor.patternbroadcast(
-                rval, node.outputs[0].type.broadcastable)
         return [rval]
 
+
 # This optimizer is registered in opt.py as part of the meta-optimizer.
 # It tries exactly the opposite code path of what local_conv_dnn() uses,
 # because for some input/kernel shape configurations, this is faster.
@@ -1540,6 +1537,11 @@ def local_conv_dnn_alternative(node):
                 rval, node.outputs[0].type.broadcastable)
         return [rval]
 
+
+conv_groupopt.register('local_conv_dnn', local_conv_dnn, 20,
+                       'conv_dnn', 'fast_compile', 'fast_run', 'cudnn')
+
+
 @local_optimizer([GpuDnnConv], inplace=True)
 def local_dnn_conv_inplace(node):
     if type(node.op) != GpuDnnConv or node.op.inplace:
@@ -1552,6 +1554,7 @@ def local_dnn_conv_inplace(node):
         inputs[2] = GpuAllocEmpty(dest.owner.op.dtype)(*dest.owner.inputs)
     return [GpuDnnConv(workmem=node.op.workmem, inplace=True)(*inputs)]
 
+
 @local_optimizer([GpuDnnConvGradW], inplace=True)
 def local_dnn_convgw_inplace(node):
     if type(node.op) != GpuDnnConvGradW or node.op.inplace:
@@ -1564,6 +1567,7 @@ def local_dnn_convgw_inplace(node):
         inputs[2] = GpuAllocEmpty(dest.owner.op.dtype)(*dest.owner.inputs)
     return [GpuDnnConvGradW(inplace=True)(*inputs)]
 
+
 @local_optimizer([GpuDnnConvGradI], inplace=True)
 def local_dnn_convgi_inplace(node):
     if type(node.op) != GpuDnnConvGradI or node.op.inplace:
@@ -1576,13 +1580,14 @@ def local_dnn_convgi_inplace(node):
         inputs[2] = GpuAllocEmpty(dest.owner.op.dtype)(*dest.owner.inputs)
     return [GpuDnnConvGradI(inplace=True)(*inputs)]
 
-optdb.register('local_dnn_conv_inplace',
+optdb.register('local_dnna_conv_inplace',
                tensor.opt.in2out(local_dnn_conv_inplace,
                                  local_dnn_convgw_inplace,
                                  local_dnn_convgi_inplace,
                                  name="local_dnn_conv_inplace"),
                70.0, 'fast_run', 'inplace', 'gpu', 'cudnn')
 
+
 @register_opt('cudnn')
 @alpha_merge(GpuDnnConv, alpha_in=4, beta_in=5, nd=4)
 def local_dnn_conv_alpha_merge(node, *inputs):
@@ -1590,6 +1595,7 @@ def local_dnn_conv_alpha_merge(node, *inputs):
         return None
     return [GpuDnnConv(workmem=node.op.workmem)(*inputs)]
 
+
 @register_opt('cudnn')
 @alpha_merge(GpuDnnConvGradW, alpha_in=4, beta_in=5, nd=4)
 def local_dnn_convw_alpha_merge(node, *inputs):
@@ -1597,6 +1603,7 @@ def local_dnn_convw_alpha_merge(node, *inputs):
         return None
     return [GpuDnnConvGradW()(*inputs)]
 
+
 @register_opt('cudnn')
 @alpha_merge(GpuDnnConvGradI, alpha_in=4, beta_in=5, nd=4)
 def local_dnn_convi_alpha_merge(node, *inputs):
@@ -1604,42 +1611,33 @@ def local_dnn_convi_alpha_merge(node, *inputs):
         return None
     return [GpuDnnConvGradI()(*inputs)]
 
+
 @register_opt('cudnn')
 @output_merge(GpuDnnConv, alpha_in=4, beta_in=5, out_in=2, nd=4)
 def local_dnn_conv_output_merge(node, *inputs):
     inputs = inputs[0:2] + (gpu_contiguous(inputs[2]),) + inputs[3:]
     return [GpuDnnConv(workmem=node.op.workmem)(*inputs)]
 
+
 @register_opt('cudnn')
 @output_merge(GpuDnnConvGradW, alpha_in=4, beta_in=5, out_in=2, nd=4)
 def local_dnn_convw_output_merge(node, *inputs):
     inputs = inputs[0:2] + (gpu_contiguous(inputs[2]),) + inputs[3:]
     return [GpuDnnConvGradW()(*inputs)]
 
+
 @register_opt('cudnn')
 @output_merge(GpuDnnConvGradI, alpha_in=4, beta_in=5, out_in=2, nd=4)
 def local_dnn_convi_output_merge(node, *inputs):
     inputs = inputs[0:2] + (gpu_contiguous(inputs[2]),) + inputs[3:]
     return [GpuDnnConvGradI()(*inputs)]
 
-@register_opt('cudnn')
-@local_optimizer([GpuDownsampleFactorMax])
-def local_pool_dnn(node):
-    if not dnn_available():
-        return
-    if isinstance(node.op, GpuDownsampleFactorMax):
-        if not node.op.ignore_border:
-            return
-        img, = node.inputs
-        ds = node.op.ds
-        return [dnn_pool(gpu_contiguous(img), ds, ds)]
 
 @register_opt('cudnn')
-@local_optimizer([DownsampleFactorMax])
+@op_lifter([DownsampleFactorMax])
 def local_pool_dnn_alternative(node):
     if not dnn_available():
         return
-    if isinstance(node.op, DownsampleFactorMax):
     if not node.op.ignore_border:
         return
     img, = node.inputs
@@ -1647,34 +1645,15 @@ def local_pool_dnn_alternative(node):
     stride = node.op.st
     pad = node.op.padding
     mode = node.op.mode
-        if (img.owner and isinstance(img.owner.op, HostFromGpu)):
-            ret = dnn_pool(gpu_contiguous(img.owner.inputs[0]),
+    return dnn_pool(gpu_contiguous(img.owner.inputs[0]),
                     ds, stride=stride, pad=pad, mode=mode)
-            return [host_from_gpu(ret)]
-
-@register_opt('cudnn')
-@local_optimizer([GpuDownsampleFactorMaxGrad])
-def local_pool_dnn_grad(node):
-    if not dnn_available():
-        return
-    if isinstance(node.op, GpuDownsampleFactorMaxGrad):
-        if not node.op.ignore_border:
-            return
-        inp, out, inp_grad = node.inputs
-        ds = node.op.ds
 
-        desc = GpuDnnPoolDesc(ws=ds, stride=ds, mode="max")()
-        return [GpuDnnPoolGrad()(gpu_contiguous(inp),
-                                 gpu_contiguous(out),
-                                 gpu_contiguous(inp_grad),
-                                 desc)]
 
 @register_opt('cudnn')
-@local_optimizer([DownsampleFactorMaxGrad])
+@op_lifter([DownsampleFactorMaxGrad])
 def local_pool_dnn_grad_stride(node):
     if not dnn_available():
         return
-    if isinstance(node.op, DownsampleFactorMaxGrad):
     if not node.op.ignore_border:
         return
     inp, out, inp_grad = node.inputs
@@ -1683,16 +1662,12 @@ def local_pool_dnn_grad_stride(node):
     pad = node.op.padding
     mode = node.op.mode
 
-        if ((inp.owner and isinstance(inp.owner.op, HostFromGpu)) or
-            (out.owner and isinstance(out.owner.op, HostFromGpu)) or
-            (inp_grad.owner and isinstance(inp_grad.owner.op,
-                                           HostFromGpu))):
     desc = GpuDnnPoolDesc(ws=ds, stride=st, mode=mode, pad=pad)()
-            ret = GpuDnnPoolGrad()(gpu_contiguous(inp),
+    return GpuDnnPoolGrad()(gpu_contiguous(inp),
                             gpu_contiguous(out),
                             gpu_contiguous(inp_grad),
                             desc)
-            return [host_from_gpu(ret)]
+
 
 @register_opt('cudnn')
 @local_optimizer([GpuSoftmax])
@@ -1706,6 +1681,7 @@ def local_softmax_dnn(node):
         out = as_gpuarray_variable(out.dimshuffle(0, 1))
         return [out]
 
+
 class NoCuDNNRaise(Optimizer):
     def apply(self, fgraph):
         """ Raise a RuntimeError if cudnn can't be used"""
@@ -1716,16 +1692,13 @@ class NoCuDNNRaise(Optimizer):
                 "cuDNN optimization was enabled, but Theano was not able"
                 " to use it. We got this error: \n" +
                 dnn_available.msg)
+
 gpu_seqopt.register("NoCuDNNRaise", NoCuDNNRaise(), 0, 'cudnn')
 
+
 @register_opt('cudnn')
-@local_optimizer([SoftmaxGrad])
+@op_lifter([SoftmaxGrad])
 def local_softmax_dnn_grad(node):
-    if (isinstance(node.op, SoftmaxGrad) and
-        ((node.inputs[0].owner and
-          isinstance(node.inputs[0].owner.op, HostFromGpu)) or
-         (node.inputs[1].owner and
-          isinstance(node.inputs[1].owner.op, HostFromGpu)))):
     if not dnn_available():
         return
     ins = []
@@ -1736,8 +1709,6 @@ def local_softmax_dnn_grad(node):
             return
         ins.append(n.dimshuffle(0, 1, 'x', 'x'))
 
-        out = GpuDnnSoftmaxGrad('bc01',
-                                'accurate',
-                                'channel')(gpu_contiguous(ins[0]),
-                                           gpu_contiguous(ins[1]))
+    out = GpuDnnSoftmaxGrad('bc01', 'accurate', 'channel')(
+        gpu_contiguous(ins[0]), gpu_contiguous(ins[1]))
     return [out.dimshuffle(0, 1)]

theano/sandbox/gpuarray/opt.py

@@ -12,11 +12,13 @@ from theano import tensor, scalar, gof
 from theano.compile import optdb
 from theano.gof import (local_optimizer, EquilibriumDB,
                         SequenceDB, Optimizer, toolbox)
+from theano.gof.optdb import LocalGroupDB
 
 from theano.scan_module import scan_utils, scan_op, scan_opt
 
 from theano.tensor.nnet.conv import ConvOp
 from theano.tests.breakpoint import PdbBreakpoint
+
 from .type import GpuArrayType, GpuArrayConstant
 from .basic_ops import (host_from_gpu, gpu_from_host,
                         HostFromGpu, GpuFromHost,
@@ -39,6 +41,10 @@ gpu_cut_copies = EquilibriumDB()
 
 gpu_seqopt = SequenceDB()
 
+# Don't register this right now
+conv_groupopt = LocalGroupDB()
+conv_groupopt.__name__ = "gpua_conv_opts"
+
 gpu_seqopt.register('gpuarray_local_optimiziations', gpu_optimizer, 1,
                     'fast_compile', 'fast_run', 'inplace', 'gpuarray')
 gpu_seqopt.register('gpuarray_cut_transfers', gpu_cut_copies, 2,
@@ -689,6 +695,9 @@ def local_gpu_conv(node):
     out.values_eq_approx = values_eq_approx
     return [out]
 
+# Register this here so that it goes after 'local_gpu_conv'
+register_opt()(conv_groupopt)
+
 
 @register_opt("low_memory")
 @local_optimizer([GpuCAReduceCuda])