Add op lifter for gpu pool grad grad

theano/gpuarray/opt.py

@@ -29,6 +29,7 @@ from theano.tensor.nnet.abstract_conv import (BaseAbstractConv,
                                               AbstractConv3d,
                                               AbstractConv3d_gradWeights,
                                               AbstractConv3d_gradInputs)
+import theano.tensor.signal.pool as pool
 
 from theano.tests.breakpoint import PdbBreakpoint
 
@@ -46,7 +47,8 @@ from .blas import (gpu_dot22, GpuGemm, GpuGer, GpuGemmBatch,
                    gpugemmbatch_no_inplace,
                    gpugemv_no_inplace, gpugemv_inplace,
                    GpuCorrMM, GpuCorrMM_gradInputs, GpuCorrMM_gradWeights,
-                   GpuCorr3dMM, GpuCorr3dMM_gradInputs, GpuCorr3dMM_gradWeights)
+                   GpuCorr3dMM, GpuCorr3dMM_gradInputs, GpuCorr3dMM_gradWeights,
+                   GpuDownsampleFactorMaxGradGrad)
 from .blocksparse import (GpuSparseBlockGemv, GpuSparseBlockOuter,
                           gpu_sparse_block_outer,
                           gpu_sparse_block_outer_inplace,
@@ -62,7 +64,7 @@ from .subtensor import (GpuIncSubtensor, GpuSubtensor,
                         GpuAdvancedSubtensor1,
                         GpuAdvancedIncSubtensor1,
                         GpuAdvancedIncSubtensor1_dev20)
-from .opt_util import alpha_merge, output_merge
+from .opt_util import alpha_merge, output_merge, pad_dims, unpad_dims
 
 _logger = logging.getLogger("theano.gpuarray.opt")
 
@@ -1589,6 +1591,34 @@ def local_gpua_lift_abstractconv_graph(op, context_name, inputs, outputs):
     return [op(*inps)]
 
 
+@register_opt()
+@op_lifter([pool.DownsampleFactorMaxGradGrad])
+@register_opt2([pool.DownsampleFactorMaxGradGrad])
+def local_gpu_downsample_factor_max_grad_grad(op, ctx_name, inputs, outputs):
+    from theano.sandbox.cuda.opt import _check_constant_args_pool
+    assert op.__props__ == ('ignore_border', 'mode', 'ndim')
+    inp, out, out_grad, ws, st, pad = inputs
+    nd = op.ndim
+    ret = _check_constant_args_pool(nd, ws, st, pad, op)
+    if ret is None:
+        return
+    ws, st, pad = ret
+    if nd != 2 or max(pad) != 0 or op.mode != 'max':
+        return
+    inp = gpu_contiguous(as_gpuarray_variable(inp, ctx_name))
+    out = gpu_contiguous(as_gpuarray_variable(out, ctx_name))
+    out_grad = gpu_contiguous(as_gpuarray_variable(out_grad, ctx_name))
+    op = GpuDownsampleFactorMaxGradGrad(ws, st, op.ignore_border)
+    if inp.ndim == nd + 2:
+        return op(inp, out, out_grad)
+    else:
+        inp_4D = pad_dims(inp, 2, 2)
+        out_4D = pad_dims(out, 2, 2)
+        out_grad_4D = pad_dims(out_grad, 2, 2)
+        output_4D = op(inp_4D, out_4D, out_grad_4D)
+        return unpad_dims(output_4D, inp, 2, 2)
+
+
 @register_opt("low_memory")
 @local_optimizer([GpuCAReduceCuda])
 def local_gpu_elemwise_careduce(node):