Change Interface of NDArray & TBlob for DLPack Compatible (apache#6345)

* Change Interface of NDArray & TBlob for DLPack Compatible

Fix for cudnn operator

Fix cpp tests

* Update nnvm

* Fix for MKL mem

* Fix for windows macro

* Bump up version number to 0.10.1

* Update NDArray Save&Load

* trigger update

* Add test for legacy data load

* Use LegacyTShapeLoad

* trigger update

* Update tensor_blob.h

.gitmodules

@@ -10,6 +10,9 @@
 [submodule "nnvm"]
 	path = nnvm
 	url = https://github.com/dmlc/nnvm
+[submodule "dlpack"]
+	path = dlpack
+	url = https://github.com/dmlc/dlpack
 [submodule "cub"]
 	path = cub
 	url = https://github.com/NVlabs/cub

CMakeLists.txt

@@ -142,6 +142,7 @@ include_directories("mshadow")
 include_directories("cub")
 include_directories("nnvm/include")
 include_directories("dmlc-core/include")
+include_directories("dlpack/include")
 
 if(NOT MSVC)
   set(BEGIN_WHOLE_ARCHIVE -Wl,--whole-archive)

Makefile

@@ -18,6 +18,10 @@ ifndef NNVM_PATH
 	NNVM_PATH = $(ROOTDIR)/nnvm
 endif
 
+ifndef DLPACK_PATH
+  	DLPACK_PATH = $(ROOTDIR)/dlpack
+endif
+
 ifneq ($(USE_OPENMP), 1)
 	export NO_OPENMP = 1
 endif
@@ -49,7 +53,7 @@ ifeq ($(DEBUG), 1)
 else
 	CFLAGS += -O3 -DNDEBUG=1
 endif
-CFLAGS += -I$(ROOTDIR)/mshadow/ -I$(ROOTDIR)/dmlc-core/include -fPIC -I$(NNVM_PATH)/include -Iinclude $(MSHADOW_CFLAGS)
+CFLAGS += -I$(ROOTDIR)/mshadow/ -I$(ROOTDIR)/dmlc-core/include -fPIC -I$(NNVM_PATH)/include -I$(DLPACK_PATH)/include -Iinclude $(MSHADOW_CFLAGS)
 LDFLAGS = -pthread $(MSHADOW_LDFLAGS) $(DMLC_LDFLAGS)
 ifeq ($(DEBUG), 1)
 	NVCCFLAGS = -std=c++11 -Xcompiler -D_FORCE_INLINES -g -G -O0 -ccbin $(CXX) $(MSHADOW_NVCCFLAGS)

R-package/DESCRIPTION

@@ -1,7 +1,7 @@
 Package: mxnet
 Type: Package
 Title: MXNet
-Version: 0.10.0
+Version: 0.10.1
 Date: 2015-12-23
 Author: Tianqi Chen, Qiang Kou, Tong He
 Maintainer: Qiang Kou <qkou@umail.iu.edu>

include/mxnet/base.h

@@ -87,7 +87,7 @@
 /*! \brief minor version */
 #define MXNET_MINOR 10
 /*! \brief patch version */
-#define MXNET_PATCH 0
+#define MXNET_PATCH 1
 /*! \brief mxnet version */
 #define MXNET_VERSION (MXNET_MAJOR*10000 + MXNET_MINOR*100 + MXNET_PATCH)
 /*! \brief helper for making version number */

include/mxnet/c_api.h

@@ -390,7 +390,7 @@ MXNET_DLL int MXNDArrayGetShape(NDArrayHandle handle,
                                 const mx_uint **out_pdata);
 /*!
  * \brief get the content of the data in NDArray
- * \param handle the handle to the narray
+ * \param handle the handle to the ndarray
  * \param out_pdata pointer holder to get pointer of data
  * \return 0 when success, -1 when failure happens
  */

include/mxnet/ndarray.h

@@ -57,10 +57,10 @@ class AutogradRuntime;
  */
 class NDArray {
  public:
-  /*! \brief default cosntructor */
+  /*! \brief default constructor */
   NDArray() {
 #if MKL_EXPERIMENTAL == 1
-      Mkl_mem_ = MKLMemHolder::create();
+    Mkl_mem_ = MKLMemHolder::create();
 #endif
   }
   /*!
@@ -75,7 +75,7 @@ class NDArray {
       : ptr_(std::make_shared<Chunk>(shape.Size(), ctx, delay_alloc, dtype)),
         shape_(shape), offset_(0), dtype_(dtype), entry_({nullptr, 0, 0}) {
 #if MKL_EXPERIMENTAL == 1
-      Mkl_mem_ = std::make_shared<MKLMemHolder>();
+    Mkl_mem_ = std::make_shared<MKLMemHolder>();
 #endif
   }
   /*!
@@ -89,29 +89,32 @@ class NDArray {
       : ptr_(std::make_shared<Chunk>(data, dev_id)), shape_(data.shape_), offset_(0),
         dtype_(data.type_flag_), entry_({nullptr, 0, 0}) {
 #if MKL_EXPERIMENTAL == 1
-      Mkl_mem_ = std::make_shared<MKLMemHolder>();
+    Mkl_mem_ = std::make_shared<MKLMemHolder>();
 #endif
   }
   /*!
    * \return the shape of current NDArray
    */
-  inline const TShape &shape() const {
+  inline const TShape& shape() const {
     return shape_;
   }
   /*!
    * \return the data TBlob
    */
-  inline TBlob data() const {
+  inline const TBlob& data() const {
     CheckAndAlloc();
-    TBlob res;
+#if MKL_EXPERIMENTAL == 1
     MSHADOW_TYPE_SWITCH(dtype_, DType, {
-      res = TBlob(static_cast<DType*>(ptr_->shandle.dptr)
-        + offset_, shape_, ptr_->shandle.ctx.dev_mask());
+      tblob_ = TBlob(static_cast<DType*>(ptr_->shandle.dptr) + offset_,
+        shape_, ptr_->shandle.ctx.dev_mask(), ptr_->shandle.ctx.dev_id, Mkl_mem_);
+    });
+#else
+    MSHADOW_TYPE_SWITCH(dtype_, DType, {
+      tblob_ = TBlob(static_cast<DType*>(ptr_->shandle.dptr) + offset_,
+        shape_, ptr_->shandle.ctx.dev_mask(), ptr_->shandle.ctx.dev_id);
     });
-#if MKL_EXPERIMENTAL == 1
-    res.Mkl_mem_ = Mkl_mem_;
 #endif
-    return res;
+    return tblob_;
   }
   /*!
    * \return a chunk of raw data in TBlob
@@ -122,8 +125,8 @@ class NDArray {
     TShape raw_shape(1);
     raw_shape[0] = length;
     MSHADOW_TYPE_SWITCH(dtype_, DType, {
-      res = TBlob(static_cast<DType*>(ptr_->shandle.dptr)
-        + offset_ + offset, raw_shape, ptr_->shandle.ctx.dev_mask());
+      res = TBlob(static_cast<DType*>(ptr_->shandle.dptr) + offset_ + offset,
+        raw_shape, ptr_->shandle.ctx.dev_mask(), ptr_->shandle.ctx.dev_id);
     });
 #if MKL_EXPERIMENTAL == 1
     res.Mkl_mem_ = Mkl_mem_;
@@ -326,7 +329,7 @@ class NDArray {
     ptr_->CheckAndAlloc();
   }
   /*!
-   * \brief Save list of narray into the Stream.x
+   * \brief Save list of ndarray into the Stream.x
    * \param fo The stream of output.
    * \param data the NDArrays to be saved.
    * \param names the name of the NDArray, optional, can be zero length.
@@ -335,7 +338,7 @@ class NDArray {
                    const std::vector<NDArray>& data,
                    const std::vector<std::string>& names);
   /*!
-   * \brief Load list of narray into from the stream.
+   * \brief Load list of ndarray into from the stream.
    * \param fi The stream of the input file.
    * \param data the NDArrays to be loaded
    * \param keys the name of the NDArray, if saved in the file.
@@ -368,10 +371,10 @@ class NDArray {
         : static_data(true),
           delay_alloc(false) {
       var = Engine::Get()->NewVariable();
-      if (data.dev_mask_ == cpu::kDevMask) {
+      if (data.dev_mask() == cpu::kDevMask) {
         shandle.ctx = Context::CPU();
       } else {
-        CHECK_EQ(data.dev_mask_, gpu::kDevMask);
+        CHECK_EQ(data.dev_mask(), gpu::kDevMask);
         shandle.ctx = Context::GPU(dev_id);
       }
       shandle.dptr = data.dptr_;
@@ -418,6 +421,14 @@ class NDArray {
   int dtype_ = -1;
   /*! \brief node entry for autograd */
   autograd::AGNodeEntry entry_;
+  /*!
+   * \brief internal TBlob
+   * \note When user access tblob_ by some const methods like
+   *     NDArray::data(), the dptr in tblob_ still need to be updated
+   *     in case that allocation happens. So we make it mutable for
+   *     this situation.
+   */
+  mutable TBlob tblob_;
 };
 
 /*!