able to run MNIST but not stable

Makefile

@@ -58,14 +58,14 @@ endif
 BIN = test/api_registry_test test/test_storage
 OBJ = narray_op_cpu.o
 # add threaded engine after it is done
-OBJCXX11 = engine.o narray.o c_api.o registry.o symbol.o fully_connected_cpu.o static_graph.o activation_cpu.o elementwise_sum_cpu.o
+OBJCXX11 = engine.o narray.o c_api.o registry.o symbol.o storage.o fully_connected_cpu.o static_graph.o activation_cpu.o elementwise_sum_cpu.o graph_executor.o pooling_cpu.o
 CUOBJ =
 SLIB = lib/libmxnet.so
 ALIB = lib/libmxnet.a
 LIB_DEP = $(DMLC_CORE)/libdmlc.a
 
 ifeq ($(USE_CUDA), 1)
-	CUOBJ += narray_op_gpu.o fully_connected_gpu.o activation_gpu.o elementwise_sum_gpu.o
+	CUOBJ += narray_op_gpu.o fully_connected_gpu.o activation_gpu.o elementwise_sum_gpu.o pooling_gpu.o
 endif
 
 .PHONY: clean all test lint doc
@@ -81,6 +81,7 @@ narray.o: src/narray/narray.cc
 narray_op_cpu.o: src/narray/narray_op_cpu.cc src/narray/narray_op-inl.h
 narray_op_gpu.o: src/narray/narray_op_gpu.cu src/narray/narray_op-inl.h
 symbol.o: src/symbol/symbol.cc
+graph_executor.o: src/symbol/graph_executor.cc
 static_graph.o : src/symbol/static_graph.cc
 registry.o: src/registry.cc
 c_api.o: src/c_api.cc
@@ -91,6 +92,9 @@ activation_cpu.o: src/operator/activation.cc
 activation_gpu.o: src/operator/activation.cu
 elementwise_sum_cpu.o: src/operator/elementwise_sum.cc
 elementwise_sum_gpu.o: src/operator/elementwise_sum.cu
+pooling_cpu.o: src/operator/pooling.cc
+pooling_gpu.o: src/operator/pooling.cu
+
 
 lib/libmxnet.a: $(OBJ) $(OBJCXX11) $(CUOBJ)
 lib/libmxnet.so: $(OBJ) $(OBJCXX11) $(CUOBJ)

include/mxnet/c_api.h

@@ -34,12 +34,11 @@ typedef void *AtomicSymbolCreator;
 typedef void *SymbolHandle;
 /*! \brief handle to a AtomicSymbol */
 typedef void *AtomicSymbolHandle;
-/*! \brief handle to a NArrayOperator */
-typedef void *OperatorHandle;
+/*! \brief handle to an Executor */
+typedef void *ExecutorHandle;
 /*! \brief handle to a DataIterator */
 typedef void *DataIterHandle;
-
-/*!
+/*
  * \brief return str message of the last error
  *  all function in this file will return 0 when success
  *  and -1 when an error occured,
@@ -325,6 +324,7 @@ MXNET_DLL int MXSymbolCompose(SymbolHandle sym,
  *  The shapes are packed into a CSR matrix represented by arg_ind_ptr and arg_shape_data
  *  The call will be treated as a kwargs call if key != nullptr or num_args==0, otherwise it is positional.
  *
+ * \param sym symbol handle
  * \param num_args numbe of input arguments.
  * \param keys the key of keyword args (optional)
  * \param arg_ind_ptr the head pointer of the rows in CSR
@@ -351,63 +351,59 @@ MXNET_DLL int MXSymbolInferShape(SymbolHandle sym,
                                  const mx_uint ***out_shape_data,
                                  int *complete);
 //--------------------------------------------
-// Part 4: operator interface on NArray
+// Part 4: Executor interface
 //--------------------------------------------
 /*!
- * \brief create operator from symbol
- * \param sym the symbol to create operator from
- * \param dev_mask device mask to indicate the device type
- * \param dev_id the device id we want to bind the symbol to
- * \param out the corresponding function handle
- * \return 0 when success, -1 when failure happens
- */
-MXNET_DLL int MXOpCreate(SymbolHandle sym,
-                         int dev_mask,
-                         int dev_id,
-                         OperatorHandle *out);
-/*!
- * \brief free the operator handle
- * \param op the handle to be freed
+ * \brief Executor forward method
+ *
+ * \param handle executor handle
  * \return 0 when success, -1 when failure happens
  */
-MXNET_DLL int MXOpFree(OperatorHandle op);
+MXNET_DLL int MXExecutorForward(ExecutorHandle handle);
 /*!
- * \brief return an array to describe the arguments
- *  of this operator
- * \param out_size the size of output array
- * \param out_array the array of parameter requirments
+ * \brief Excecutor run backward
+ *
+ * \param handle execute handle
+ * \param len lenth
+ * \param head_grads NArray handle for heads' gradient
+ *
  * \return 0 when success, -1 when failure happens
  */
-MXNET_DLL int MXOpDescribeArgs(mx_uint *out_size,
-                               int **out_array);
+MXNET_DLL int MXExecutorBackward(ExecutorHandle handle,
+                                 mx_uint len,
+                                 NArrayHandle *head_grads);
+
 /*!
- * \brief call forward on the operator
- * \param op the operator handle
- * \param in_data array of input narray to the operator
- * \param out_data array of output NArray to hold the result
+ * \brief Get executor's head NArray
+ *
+ * \param handle executor handle
+ * \param out_size output narray vector size
+ * \param out out put narray handles
  * \return 0 when success, -1 when failure happens
  */
-MXNET_DLL int MXOpForward(OperatorHandle op,
-                          NArrayHandle *in_data,
-                          NArrayHandle *out_data);
+MXNET_DLL int MXExecutorHeads(ExecutorHandle handle,
+                              mx_uint *out_size,
+                              NArrayHandle **out);
+
 /*!
- * \brief call backward on the operator
- * \param op the operator handle
- * \param grad_next array of output gradients
- * \param in_data array of input narray to the operator
- * \param out_data array of output narray to the operator
- * \param out_grad array to holds the gradient on these input
- *    can be NULL if that position request is kNullOp
- * \param reqs gradient request type
+ * \brief Generate Executor from symbol
+ *
+ * \param symbol_handle symbol handle
+ * \param len length
+ * \param in_args in args array
+ * \param arg_grad_store arg grads handle array
+ * \param grad_req_type grad req array
+ * \param out output executor handle
  * \return 0 when success, -1 when failure happens
- * \sa mxnet::Operator::GradReqType
  */
-MXNET_DLL int MXOpBackward(OperatorHandle op,
-                           NArrayHandle *grad_next,
-                           NArrayHandle *in_data,
-                           NArrayHandle *out_data,
-                           NArrayHandle *out_grad,
-                           mx_uint *reqs);
+MXNET_DLL int MXExecutorBind(SymbolHandle symbol_handle,
+                             int dev_mask,
+                             int dev_id,
+                             mx_uint len,
+                             NArrayHandle *in_args,
+                             NArrayHandle *arg_grad_store,
+                             mx_uint *grad_req_type,
+                             ExecutorHandle *out);
 
 //--------------------------------------------
 // Part 5: IO Interface

include/mxnet/context.h

@@ -33,6 +33,14 @@ struct Context {
   inline bool operator==(const Context &b) const {
     return dev_mask == b.dev_mask && dev_id == b.dev_id;
   }
+  /*!
+   * \brief check if current context not equals another one
+   * \param b another context to compare
+   * \return whether they are not the same
+   */
+  inline bool operator!=(const Context &b) const {
+    return !(*this == b);
+  }
 };
 
 /*!

include/mxnet/narray.h

@@ -35,7 +35,7 @@ class NArray {
    */
   NArray(const TShape &shape, Context ctx,
          bool delay_alloc = false)
-      : ptr_(new Chunk(shape, ctx, delay_alloc)) {
+      : ptr_(new Chunk(shape.Size(), ctx, delay_alloc)), shape_(shape), offset_(0) {
   }
   /*!
    * \brief constructing a static NArray that shares data with TBlob
@@ -45,19 +45,20 @@ class NArray {
    * \param dev_id the device id this tensor sits at
    */
   NArray(const TBlob &data, int dev_id)
-      : ptr_(new Chunk(data, dev_id)) {
+      : ptr_(new Chunk(data, dev_id)), shape_(data.shape_), offset_(0) {
   }
   /*!
    * \return the shape of current NArray
    */
   inline const TShape &shape() const {
-    return ptr_->data.shape_;
+    return shape_;
   }
   /*!
    * \return the data TBlob
    */
-  inline const TBlob &data() const {
-    return ptr_->data;
+  inline TBlob data() const {
+    return TBlob(static_cast<real_t*>(ptr_->shandle.dptr) + offset_, \
+                                      shape_, ptr_->shandle.ctx.dev_mask);
   }
   /*!
    * \return the context of NArray, this function is only valid when the NArray is not empty
@@ -123,6 +124,43 @@ class NArray {
    * \return the new copy
    */
   NArray Copy(Context ctx) const;
+  /*!
+   * \brief Slice a NArray
+   *
+   * \param begin begin index in first dim
+   * \param end end index in first dim
+   *
+   * \return sliced NArray
+   */
+  inline NArray Slice(index_t begin, index_t end) const {
+    NArray ret = *this;
+    CHECK_GE(shape_.ndim(), 0) << "NArray not initialized";
+    CHECK_GE(shape_[0], end) << "Chunk is smaller than required";
+    size_t length = 1;
+    if (shape_.ndim() == 1) {
+      ret.offset_= begin;
+    } else {
+      for (index_t i = 1; i < shape_.ndim(); ++i) {
+        length *= shape_[i];
+      }
+      ret.offset_ = begin * length;
+    }
+    ret.shape_[0] = end - begin;
+    return ret;
+  }
+  /*!
+   * \brief Reshape current NArray
+   *
+   * \param shape new shape
+   * \return NArray in new shape
+   */
+  inline NArray Reshape(const TShape &shape) const {
+    CHECK_GE(shape_.Size(), shape.Size()) \
+      << "required shape is larger than chunk";
+    NArray ret = *this;
+    ret.shape_ = shape;
+    return ret;
+  }
 
  private:
   /*! \brief the real data chunk that backs NArray */
@@ -131,8 +169,6 @@ class NArray {
     Storage::Handle shandle;
     /*! \brief variable from DAG engine */
     DAGEngine::Variable var;
-    /*! \brief holds the data content */
-    TBlob data;
     /*!
      * \brief if this is true, this means the data do not come
      * from Storage, and do not need to be freed
@@ -146,25 +182,25 @@ class NArray {
     }
     /*! \brief construct from static data */
     Chunk(const TBlob &data, int dev_id)
-        : data(data),
-          static_data(true),
+        : static_data(true),
           delay_alloc(false) {
       var = DAGEngine::Get()->NewVar();
       shandle.ctx = Context(data.dev_mask_, dev_id);
+      shandle.dptr = data.dptr_;
+      shandle.size = data.shape_.Size() * sizeof(real_t);
     }
     /*! \brief construct a new chunk */
-    Chunk(const TShape &shape, Context ctx, bool delay_alloc_)
+    Chunk(uint64_t size, Context ctx, bool delay_alloc_)
         : static_data(false), delay_alloc(true) {
       var = DAGEngine::Get()->NewVar();
-      data.shape_ = shape;
+      shandle.size = size * sizeof(real_t);
       shandle.ctx = ctx;
       if (!delay_alloc_) this->CheckAndAlloc();
     }
     /*! \brief check if delay alloc is on, do alloc if not yet done */
     inline void CheckAndAlloc(void) {
       if (delay_alloc) {
-        shandle = Storage::Get()->Alloc(data.shape_.Size() * sizeof(real_t), shandle.ctx);
-        data = TBlob(static_cast<real_t*>(shandle.dptr), data.shape_, shandle.ctx.dev_mask);
+        shandle = Storage::Get()->Alloc(shandle.size, shandle.ctx);
         delay_alloc = false;
       }
     }
@@ -183,6 +219,11 @@ class NArray {
   };
   /*! \brief internal data of NArray */
   std::shared_ptr<Chunk> ptr_;
+  /*! \brief shape of current NArray */
+  TShape shape_;
+  /*! \brief offset in chunk */
+  size_t offset_;
+
   // add friend to helper functions
   friend void CopyFromTo(const NArray &from, NArray *to);
   template<typename OP>