Merge pull request #3 from dmlc/master

merge dmlc/master

.travis.yml

@@ -11,6 +11,7 @@ env:
     - TASK=python CXX=g++
     - TASK=python3 CXX=g++
     - TASK=python_naive CXX=g++
+    - TASK=python_perdev CXX=g++
     - TASK=cpp_unittest CXX=g++
 
 # dependent apt packages

include/mxnet/base.h

@@ -1,12 +1,13 @@
 /*!
  *  Copyright (c) 2015 by Contributors
  * \file base.h
- * \brief configuation of mxnet
+ * \brief configuation of mxnet as well as basic data structure.
  */
 #ifndef MXNET_BASE_H_
 #define MXNET_BASE_H_
 
 #include <dmlc/base.h>
+#include <dmlc/io.h>
 #include <dmlc/type_traits.h>
 #include <dmlc/parameter.h>
 #include <mshadow/tensor.h>
@@ -62,6 +63,84 @@ typedef mshadow::default_real_t real_t;
 typedef mshadow::TShape TShape;
 /*! \brief storage container type */
 typedef mshadow::TBlob TBlob;
+
+/*! \brief Context information about the execution enviroment */
+struct Context {
+  /*! \brief the device type we run the op can be cpu::kDevMask or gpu::kDevMask */
+  int32_t dev_mask;
+  /*! \brief device id we are going to run it on */
+  int32_t dev_id;
+  /*! \brief constructor */
+  Context() : dev_mask(cpu::kDevMask), dev_id(0) {}
+  /*!
+   * \brief constructor of context
+   * \param dev_mask the device mask
+   * \param dev_id the device id
+   */
+  Context(int dev_mask, int dev_id)
+      : dev_mask(dev_mask), dev_id(dev_id) {}
+  /*!
+   * \brief check if current context equals another one
+   * \param b another context to compare
+   * \return whether dev mask and id are same
+   */
+  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);
+  }
+  /*!
+   * \brief save the content into binary stream
+   * \param strm the output stream
+   */
+  void Save(dmlc::Stream *strm) const {
+    strm->Write(&dev_mask, sizeof(dev_mask));
+    strm->Write(&dev_id, sizeof(dev_id));
+  }
+  /*!
+   * \brief load the content from binary stream
+   * \param strm the output stream
+   * \return whether the load is successful
+   */
+  bool Load(dmlc::Stream *strm) {
+    if (strm->Read(&dev_mask, sizeof(int32_t)) != sizeof(int32_t)) return false;
+    if (strm->Read(&dev_id, sizeof(int32_t)) != sizeof(int32_t)) return false;
+    return true;
+  }
+  /*! \brief the maximal device mask, cpu = 1, gpu = 2 */
+  static const int32_t kMaxDevMask = 2;
+  /*!
+   * \brief A dedicate ID for pinned cpu memory.
+   *  Any normal CPU ID should be less than this number.
+   */
+  static const int32_t kPinnedMemoryID = 16;
+};
+
+/*!
+ * \brief execution time context.
+ *  The information needed in runtime for actual execution.
+ */
+struct RunContext {
+  /*!
+   * \brief the stream of the device, can be NULL or Stream<gpu>* in GPU mode
+   */
+  void *stream;
+  /*!
+   * \brief get mshadow stream from Context
+   * \return the mshadow stream
+   * \tparam xpu the device type of the stream
+   */
+  template<typename xpu>
+  inline mshadow::Stream<xpu>* get_stream() const {
+    return static_cast<mshadow::Stream<xpu>*>(stream);
+  }
+};
 }  // namespace mxnet
 
 //! \cond Doxygen_Suppress

include/mxnet/engine.h

@@ -12,7 +12,6 @@
 #endif
 #include <vector>
 #include "./base.h"
-#include "./context.h"
 
 namespace mxnet {
 /*! \brief namespace of engine internal types. */
@@ -28,13 +27,14 @@ typedef Opr* OprHandle;
 }  // namespace engine
 
 #if DMLC_USE_CXX11
-
 /*! \brief Function property, used to hint what action is pushed to engine. */
 enum class FnProperty {
   /*! \brief Normal operation */
   kNormal,
-  /*! \brief Copy operation between CPU and GPU */
-  kCopy,
+  /*! \brief Copy operation from GPU to other devices */
+  kCopyFromGPU,
+  /*! \brief Copy operation from CPU to other devices */
+  kCopyToGPU,
   /*! \brief Asynchronous function call */
   kAsync
 };  // enum class FnProperty

include/mxnet/kvstore.h

@@ -10,7 +10,7 @@
 #if DMLC_USE_CXX11
 #include <functional>
 #endif  // DMLC_USE_CXX11
-#include "ndarray.h"
+#include "./ndarray.h"
 
 namespace mxnet {
 

include/mxnet/ndarray.h

@@ -13,10 +13,9 @@
 #include <dmlc/registry.h>
 #include <memory>
 #include "./base.h"
-#include "./context.h"
 #include "./storage.h"
-#include "./context.h"
 #include "./engine.h"
+
 // check c++11
 #if DMLC_USE_CXX11 == 0
 #error "cxx11 was required for ndarray module"

include/mxnet/operator.h

@@ -14,7 +14,7 @@
 #include <string>
 #include <utility>
 #include "./base.h"
-#include "./context.h"
+#include "./resource.h"
 
 namespace mxnet {
 /*! \brief operation request type to Forward and Backward */
@@ -230,18 +230,22 @@ class OperatorProperty {
    * \brief Declare additional resource required in forward pass.
    *  These additional resources will be presented in OpContext.requested
    *  in the same order of the returned Resource.
+   * \param in_shape The input shape to the operator, corresponds to shapes of in_data.
    * \return Additional resource request
    */
-  virtual std::vector<ResourceRequest> ForwardResource() const {
+  virtual std::vector<ResourceRequest> ForwardResource(
+      const std::vector<TShape> &in_shape) const {
     return std::vector<ResourceRequest>();
   }
   /*!
    * \brief Decalre additional resource required in backward pass.
    *  These additional resources will be presented in OpContext.requested
    *  in the same order of the returned Resource.
+   * \param in_shape The input shape to the operator, corresponds to shapes of in_data.
    * \return Additional resource request
    */
-  virtual std::vector<ResourceRequest> BackwardResource() const {
+  virtual std::vector<ResourceRequest> BackwardResource(
+      const std::vector<TShape> &in_shape) const {
     return std::vector<ResourceRequest>();
   }
   /*!