Refactor for windows CI 'out of heap space' errors

src/operator/tensor/broadcast_reduce_minmax_value.cc

@@ -0,0 +1,59 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2016 by Contributors
+ * \file broadcast_reduce_minmax_value.cc
+ * \brief CPU Implementation of broadcast and reduce min and max functions based on value.
+ */
+#include "./broadcast_reduce_op.h"
+
+namespace mxnet {
+namespace op {
+
+MXNET_OPERATOR_REGISTER_MINMAX_REDUCE(max)
+.add_alias("max_axis")
+.describe(get_reduce_axes_description("max", __LINE__))
+.set_attr<FCompute>("FCompute<cpu>", ReduceAxesCompute<cpu, mshadow::red::maximum>)
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<nnvm::FGradient>("FGradient", ReduceGrad{"_backward_max"});
+
+MXNET_OPERATOR_REGISTER_REDUCE_BACKWARD(_backward_max)
+.set_num_inputs(3)
+.set_attr<FCompute>("FCompute<cpu>", ReduceAxesBackwardUseInOut<cpu, mshadow_op::eq>);
+
+MXNET_OPERATOR_REGISTER_MINMAX_REDUCE(min)
+.add_alias("min_axis")
+.describe(get_reduce_axes_description("min", __LINE__))
+.set_attr<FCompute>("FCompute<cpu>", ReduceAxesCompute<cpu, mshadow::red::minimum>)
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<nnvm::FGradient>("FGradient", ReduceGrad{"_backward_min"});
+
+MXNET_OPERATOR_REGISTER_REDUCE_BACKWARD(_backward_min)
+.set_num_inputs(3)
+.set_attr<FCompute>("FCompute<cpu>", ReduceAxesBackwardUseInOut<cpu, mshadow_op::eq>);
+
+}  // namespace op
+}  // namespace mxnet

src/operator/tensor/broadcast_reduce_minmax_value.cu

@@ -0,0 +1,43 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2016 by Contributors
+ * \file broadcast_reduce_minmax_value.cu
+ * \brief GPU Implementation of broadcast and reduce min and max functions based on value.
+ */
+#include "./broadcast_reduce_op.h"
+
+namespace mxnet {
+namespace op {
+
+NNVM_REGISTER_OP(max)
+.set_attr<FCompute>("FCompute<gpu>", ReduceAxesCompute<gpu, mshadow::red::maximum>);
+
+NNVM_REGISTER_OP(_backward_max)
+.set_attr<FCompute>("FCompute<gpu>", ReduceAxesBackwardUseInOut<gpu, mshadow_op::eq>);
+
+NNVM_REGISTER_OP(min)
+.set_attr<FCompute>("FCompute<gpu>", ReduceAxesCompute<gpu, mshadow::red::minimum>);
+
+NNVM_REGISTER_OP(_backward_min)
+.set_attr<FCompute>("FCompute<gpu>", ReduceAxesBackwardUseInOut<gpu, mshadow_op::eq>);
+
+}  // namespace op
+}  // namespace mxnet

src/operator/tensor/broadcast_reduce_norm_value.cc

@@ -0,0 +1,118 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2016 by Contributors
+ * \file broadcast_reduce_norm_value.cc
+ * \brief CPU Implementation of broadcast and reduce norm functions based on value.
+ */
+#include "./broadcast_reduce_op.h"
+
+namespace mxnet {
+namespace op {
+DMLC_REGISTER_PARAMETER(NormParam);
+
+template<>
+void L2NormComputeEx<cpu>(const nnvm::NodeAttrs& attrs,
+                          const OpContext& ctx,
+                          const std::vector<NDArray>& inputs,
+                          const std::vector<OpReqType>& req,
+                          const std::vector<NDArray>& outputs) {
+  CHECK_EQ(inputs.size(), 1U);
+  CHECK_EQ(outputs.size(), 1U);
+  CHECK_EQ(req.size(), 1U);
+  const NormParam& param = nnvm::get<NormParam>(attrs.parsed);
+  mshadow::Stream<cpu>* s = ctx.get_stream<cpu>();
+  const NDArrayStorageType istype = inputs[0].storage_type();
+  const mxnet::TShape axis = param.axis.has_value() ? param.axis.value() : mxnet::TShape();
+  if ((istype == kRowSparseStorage || istype == kCSRStorage) && axis.ndim() == 0 &&
+       param.ord == 2) {
+    // l2 norm on the entire array
+    L2NormComputeSparseImpl<cpu>(s, inputs[0], req[0], outputs[0].data());
+  } else if (istype == kCSRStorage && axis.ndim() == 1 && (axis[0] == 0 || axis[0] == 1) &&
+             !param.keepdims && param.ord == 2) {
+    // l2 norm on a particular axis
+    NDArray output = outputs[0];
+    ReduceCsrImpl<cpu, sq_sum, false>(s, ctx, inputs[0], req[0], &output, axis);
+    CHECK_EQ(outputs[0].storage_type(), kDefaultStorage);
+    SqRootForL2<cpu>(ctx, req[0], outputs[0].data());
+  } else {
+    LogUnimplementedOp(attrs, ctx, inputs, req, outputs);
+  }
+}
+
+NNVM_REGISTER_OP(norm)
+MXNET_ADD_SPARSE_OP_ALIAS(norm)
+.describe(R"code(Computes the norm on an NDArray.
+
+This operator computes the norm on an NDArray with the specified axis, depending
+on the value of the ord parameter. By default, it computes the L2 norm on the entire
+array. Currently only ord=2 supports sparse ndarrays.
+
+Examples::
+
+  x = [[[1, 2],
+        [3, 4]],
+       [[2, 2],
+        [5, 6]]]
+
+  norm(x, ord=2, axis=1) = [[3.1622777 4.472136 ]
+                            [5.3851647 6.3245554]]
+
+  norm(x, ord=1, axis=1) = [[4., 6.],
+                            [7., 8.]]
+
+  rsp = x.cast_storage('row_sparse')
+
+  norm(rsp) = [5.47722578]
+
+  csr = x.cast_storage('csr')
+
+  norm(csr) = [5.47722578]
+
+)code" ADD_FILELINE)
+.set_num_inputs(1)
+.set_num_outputs(1)
+.set_attr_parser(ParamParser<NormParam>)
+.set_attr<mxnet::FInferShape>("FInferShape", NormShape)
+.set_attr<nnvm::FInferType>("FInferType", NormType)
+.set_attr<FInferStorageType>("FInferStorageType", LpNormStorageType)
+.set_attr<nnvm::FGradient>("FGradient", ReduceGrad{ "_backward_norm" })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", LpNormCompute<cpu>)
+.set_attr<FComputeEx>("FComputeEx<cpu>", L2NormComputeEx<cpu>)
+.add_argument("data", "NDArray-or-Symbol", "The input")
+.add_arguments(NormParam::__FIELDS__());
+
+NNVM_REGISTER_OP(_backward_norm)
+.set_num_outputs(1)
+.set_attr_parser(ParamParser<NormParam>)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", LpNormGradCompute<cpu>);
+
+
+}  // namespace op
+}  // namespace mxnet

src/operator/tensor/broadcast_reduce_norm_value.cu

@@ -0,0 +1,60 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2016 by Contributors
+ * \file broadcast_reduce_norm_value.cu
+ * \brief GPU Implementation of broadcast and reduce norm functions based on value.
+ */
+#include "./broadcast_reduce_op.h"
+
+namespace mxnet {
+namespace op {
+
+template<>
+void L2NormComputeEx<gpu>(const nnvm::NodeAttrs& attrs,
+                          const OpContext& ctx,
+                          const std::vector<NDArray>& inputs,
+                          const std::vector<OpReqType>& req,
+                          const std::vector<NDArray>& outputs) {
+  CHECK_EQ(inputs.size(), 1U);
+  CHECK_EQ(outputs.size(), 1U);
+  CHECK_EQ(req.size(), 1U);
+  const NormParam& param = nnvm::get<NormParam>(attrs.parsed);
+  mshadow::Stream<gpu>* s = ctx.get_stream<gpu>();
+  const NDArrayStorageType istype = inputs[0].storage_type();
+  const mxnet::TShape axis = param.axis.has_value() ? param.axis.value() : mxnet::TShape();
+  if ((istype == kRowSparseStorage || istype == kCSRStorage) && axis.ndim() == 0 &&
+       param.ord == 2) {
+    // l2 norm on the entire array
+    L2NormComputeSparseImpl<gpu>(s, inputs[0], req[0], outputs[0].data());
+  } else {
+    LogUnimplementedOp(attrs, ctx, inputs, req, outputs);
+  }
+}
+
+NNVM_REGISTER_OP(norm)
+.set_attr<FCompute>("FCompute<gpu>", LpNormCompute<gpu>)
+.set_attr<FComputeEx>("FComputeEx<gpu>", L2NormComputeEx<gpu>);
+
+NNVM_REGISTER_OP(_backward_norm)
+.set_attr<FCompute>("FCompute<gpu>", LpNormGradCompute<gpu>);
+
+}  // namespace op
+}  // namespace mxnet

src/operator/tensor/broadcast_reduce_op.h

@@ -26,6 +26,7 @@
 #define MXNET_OPERATOR_TENSOR_BROADCAST_REDUCE_OP_H_
 
 #include <mxnet/operator_util.h>
+#include <string>
 #include <vector>
 #include <utility>
 #include <algorithm>
@@ -1580,6 +1581,20 @@ void PickOpBackward(const nnvm::NodeAttrs& attrs,
   });
 }
 
+inline std::string get_reduce_axes_description(const std::string& op_name, int line) {
+  std::string doc = R"code(Computes the __op__ of array elements over given axes.
+
+Defined in )code";
+  doc += std::string(__FILE__) + std::string(":L") + std::to_string(line);
+  size_t pos = 0;
+  std::string holder("__op__");
+  while ((pos = doc.find(holder, pos)) != std::string::npos) {
+    doc.replace(pos, holder.length(), op_name);
+    pos += op_name.length();
+  }
+  return doc;
+}
+
 #define MXNET_OPERATOR_REGISTER_REDUCE_AXIS(name)               \
   NNVM_REGISTER_OP(name)                                        \
   .set_num_inputs(1)                                            \