Use CPUPinned context in ImageRecordIOParser2 (apache#13980)

* create NDArray with CPUPinned context in ImageRecordIOParser2

* update document

* use -1 device_id as an option to create CPU(0) context

* retrigger CI

* fix cpplint error

Author: yuxihu

src/io/image_iter_common.h

@@ -125,11 +125,13 @@ struct ImageRecParserParam : public dmlc::Parameter<ImageRecParserParam> {
   bool verbose;
   /*! \brief partition the data into multiple parts */
   int num_parts;
-  /*! \brief the index of the part will read*/
+  /*! \brief the index of the part will read */
   int part_index;
-  /*! \brief the size of a shuffle chunk*/
+  /*! \brief device id used to create context for internal NDArray */
+  int device_id;
+  /*! \brief the size of a shuffle chunk */
   size_t shuffle_chunk_size;
-  /*! \brief the seed for chunk shuffling*/
+  /*! \brief the seed for chunk shuffling */
   int shuffle_chunk_seed;
 
   // declare parameters
@@ -161,6 +163,11 @@ struct ImageRecParserParam : public dmlc::Parameter<ImageRecParserParam> {
         .describe("Virtually partition the data into these many parts.");
     DMLC_DECLARE_FIELD(part_index).set_default(0)
         .describe("The *i*-th virtual partition to be read.");
+    DMLC_DECLARE_FIELD(device_id).set_default(0)
+        .describe("The device id used to create context for internal NDArray. "\
+                  "Setting device_id to -1 will create Context::CPU(0). Setting "
+                  "device_id to valid positive device id will create "
+                  "Context::CPUPinned(device_id). Default is 0.");
     DMLC_DECLARE_FIELD(shuffle_chunk_size).set_default(0)
         .describe("The data shuffle buffer size in MB. Only valid if shuffle is true.");
     DMLC_DECLARE_FIELD(shuffle_chunk_seed).set_default(0)

src/io/iter_image_recordio_2.cc

@@ -285,9 +285,14 @@ inline bool ImageRecordIOParser2<DType>::ParseNext(DataBatch *out) {
     shape_vec.push_back(param_.label_width);
     TShape label_shape(shape_vec.begin(), shape_vec.end());
 
-    out->data.at(0) = NDArray(data_shape, Context::CPU(0), false,
+    auto ctx = Context::CPU(0);
+    auto dev_id = param_.device_id;
+    if (dev_id != -1) {
+      ctx = Context::CPUPinned(dev_id);
+    }
+    out->data.at(0) = NDArray(data_shape, ctx, false,
       mshadow::DataType<DType>::kFlag);
-    out->data.at(1) = NDArray(label_shape, Context::CPU(0), false,
+    out->data.at(1) = NDArray(label_shape, ctx, false,
       mshadow::DataType<real_t>::kFlag);
     unit_size_[0] = param_.data_shape.Size();
     unit_size_[1] = param_.label_width;