diff --git a/paddle/capi/Matrix.cpp b/paddle/capi/Matrix.cpp
index 85269e1885467..d898ebe2612d7 100644
--- a/paddle/capi/Matrix.cpp
+++ b/paddle/capi/Matrix.cpp
@@ -39,7 +39,7 @@ paddle_error paddle_matrix_destroy(paddle_matrix mat) {
 
 paddle_error paddle_matrix_set_row(paddle_matrix mat,
                                    uint64_t rowID,
-                                   pd_real* rowArray) {
+                                   paddle_real* rowArray) {
   if (mat == nullptr) return kPD_NULLPTR;
   auto ptr = cast(mat);
   if (ptr->mat == nullptr) return kPD_NULLPTR;
@@ -56,7 +56,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
 
 paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                    uint64_t rowID,
-                                   pd_real** rawRowBuffer) {
+                                   paddle_real** rawRowBuffer) {
   if (mat == nullptr) return kPD_NULLPTR;
   auto ptr = cast(mat);
   if (ptr->mat == nullptr) return kPD_NULLPTR;
@@ -78,3 +78,46 @@ paddle_error paddle_matrix_get_shape(paddle_matrix mat,
   return kPD_NO_ERROR;
 }
 }
+
+paddle_matrix paddle_matrix_create_sparse(
+    uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu) {
+  auto ptr = new paddle::capi::CMatrix();
+  ptr->mat = paddle::Matrix::createSparseMatrix(
+      height,
+      width,
+      nnz,
+      isBinary ? paddle::NO_VALUE : paddle::FLOAT_VALUE,
+      paddle::SPARSE_CSR,
+      false,
+      useGpu);
+  return ptr;
+}
+
+paddle_error paddle_matrix_sparse_copy_from(paddle_matrix mat,
+                                            int* rowArray,
+                                            uint64_t rowSize,
+                                            int* colArray,
+                                            uint64_t colSize,
+                                            float* valueArray,
+                                            uint64_t valueSize) {
+  if (mat == nullptr) return kPD_NULLPTR;
+  auto ptr = cast(mat);
+  if (rowArray == nullptr || colArray == nullptr ||
+      (valueSize != 0 && valueArray == nullptr) || ptr->mat == nullptr) {
+    return kPD_NULLPTR;
+  }
+  if (auto sparseMat = dynamic_cast<paddle::CpuSparseMatrix*>(ptr->mat.get())) {
+    std::vector<int> row(rowSize);
+    row.assign(rowArray, rowArray + rowSize);
+    std::vector<int> col(colSize);
+    col.assign(colArray, colArray + colSize);
+    std::vector<paddle_real> val(valueSize);
+    if (valueSize) {
+      val.assign(valueArray, valueArray + valueSize);
+    }
+    sparseMat->copyFrom(row, col, val);
+    return kPD_NO_ERROR;
+  } else {
+    return kPD_NOT_SUPPORTED;
+  }
+}
diff --git a/paddle/capi/config.h.in b/paddle/capi/config.h.in
index af4e80dea144f..d205307588eb6 100644
--- a/paddle/capi/config.h.in
+++ b/paddle/capi/config.h.in
@@ -1,7 +1,7 @@
 #ifndef __PADDLE_PADDLE_CAPI_CONFIG_H_INCLUDED__
 #define __PADDLE_PADDLE_CAPI_CONFIG_H_INCLUDED__
 
-typedef @PADDLE_FLOAT_TYPE@ pd_real;
+typedef @PADDLE_FLOAT_TYPE@ paddle_real;
 
 // Since we only support linux and macos in compile, always use clang or
 // gcc 4.8+. DLL_IMPORT/DLL_EXPORT is as simple as below.
diff --git a/paddle/capi/error.h b/paddle/capi/error.h
index 6a5907b869f60..44d8c2040d1aa 100644
--- a/paddle/capi/error.h
+++ b/paddle/capi/error.h
@@ -23,6 +23,7 @@ typedef enum {
   kPD_NULLPTR = 1,
   kPD_OUT_OF_RANGE = 2,
   kPD_PROTOBUF_ERROR = 3,
+  kPD_NOT_SUPPORTED = 4,
   kPD_UNDEFINED_ERROR = -1,
 } paddle_error;
 
diff --git a/paddle/capi/examples/.gitignore b/paddle/capi/examples/.gitignore
new file mode 100644
index 0000000000000..2caa0a5a298d8
--- /dev/null
+++ b/paddle/capi/examples/.gitignore
@@ -0,0 +1,2 @@
+*.bin
+build-*
diff --git a/paddle/capi/examples/common/common.h b/paddle/capi/examples/common/common.h
new file mode 100644
index 0000000000000..a78522e4a7c3c
--- /dev/null
+++ b/paddle/capi/examples/common/common.h
@@ -0,0 +1,26 @@
+#ifndef __CAPI_EXAMPLE_COMMON_H__
+#define __CAPI_EXAMPLE_COMMON_H__
+#include <stdio.h>
+#include <stdlib.h>
+
+#define CHECK(stmt)                                                \
+  do {                                                             \
+    paddle_error __err__ = stmt;                                   \
+    if (__err__ != kPD_NO_ERROR) {                                 \
+      fprintf(stderr, "Invoke paddle error %d \n" #stmt, __err__); \
+      exit(__err__);                                               \
+    }                                                              \
+  } while (0)
+
+void* read_config(const char* filename, long* size) {
+  FILE* file = fopen(filename, "r");
+  if (file == NULL) return NULL;
+  fseek(file, 0L, SEEK_END);
+  *size = ftell(file);
+  fseek(file, 0L, SEEK_SET);
+  void* buf = malloc(*size);
+  fread(buf, 1, *size, file);
+  fclose(file);
+  return buf;
+}
+#endif
diff --git a/paddle/capi/examples/dense/CMakeLists.txt b/paddle/capi/examples/dense/CMakeLists.txt
new file mode 100644
index 0000000000000..008a488fd9e6f
--- /dev/null
+++ b/paddle/capi/examples/dense/CMakeLists.txt
@@ -0,0 +1,6 @@
+project(dense)
+cmake_minimum_required(VERSION 2.8)
+aux_source_directory(. SRC_LIST)
+add_executable(${PROJECT_NAME} ${SRC_LIST})
+set_property(TARGET ${PROJECT_NAME} PROPERTY C_STANDARD 99)
+target_link_libraries(${PROJECT_NAME} -lpaddle_capi_shared)
diff --git a/paddle/capi/examples/dense/convert_protobin.sh b/paddle/capi/examples/dense/convert_protobin.sh
new file mode 100755
index 0000000000000..30ffc316ecb76
--- /dev/null
+++ b/paddle/capi/examples/dense/convert_protobin.sh
@@ -0,0 +1,2 @@
+#!/bin/bash
+python -m paddle.utils.dump_config trainer_config.py '' --binary > trainer_config.bin
diff --git a/paddle/capi/examples/dense/main.c b/paddle/capi/examples/dense/main.c
new file mode 100644
index 0000000000000..4dddd65bbfb3c
--- /dev/null
+++ b/paddle/capi/examples/dense/main.c
@@ -0,0 +1,63 @@
+#include <paddle/capi.h>
+#include <time.h>
+#include "../common/common.h"
+
+#define CONFIG_BIN "./trainer_config.bin"
+
+int main() {
+  // Initalize Paddle
+  char* argv[] = {"--use_gpu=False"};
+  CHECK(paddle_init(1, (char**)argv));
+
+  // Reading config binary file. It is generated by `convert_protobin.sh`
+  long size;
+  void* buf = read_config(CONFIG_BIN, &size);
+
+  // Create a gradient machine for inference.
+  paddle_gradient_machine machine;
+  CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
+  CHECK(paddle_gradient_machine_randomize_param(machine));
+
+  // Loading parameter. Uncomment the following line and change the directory.
+  // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
+  //                                                "./some_where_to_params"));
+  paddle_arguments in_args = paddle_arguments_create_none();
+
+  // There is only one input of this network.
+  CHECK(paddle_arguments_resize(in_args, 1));
+
+  // Create input matrix.
+  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
+                                           /* size */ 784,
+                                           /* useGPU */ false);
+  srand(time(0));
+  paddle_real* array;
+
+  // Get First row.
+  CHECK(paddle_matrix_get_row(mat, 0, &array));
+
+  for (int i = 0; i < 784; ++i) {
+    array[i] = rand() / ((float)RAND_MAX);
+  }
+
+  CHECK(paddle_arguments_set_value(in_args, 0, mat));
+
+  paddle_arguments out_args = paddle_arguments_create_none();
+  CHECK(paddle_gradient_machine_forward(machine,
+                                        in_args,
+                                        out_args,
+                                        /* isTrain */ false));
+  paddle_matrix prob = paddle_matrix_create_none();
+
+  CHECK(paddle_arguments_value(out_args, 0, prob));
+
+  CHECK(paddle_matrix_get_row(prob, 0, &array));
+
+  printf("Prob: ");
+  for (int i = 0; i < 10; ++i) {
+    printf("%.2f ", array[i]);
+  }
+  printf("\n");
+
+  return 0;
+}
diff --git a/paddle/capi/examples/dense/trainer_config.py b/paddle/capi/examples/dense/trainer_config.py
new file mode 100644
index 0000000000000..873ec119e7a3d
--- /dev/null
+++ b/paddle/capi/examples/dense/trainer_config.py
@@ -0,0 +1,18 @@
+from paddle.trainer_config_helpers import *
+
+img = data_layer(name='pixel', size=784)
+
+hidden = fc_layer(
+    input=img,
+    size=200,
+    param_attr=ParamAttr(name='hidden.w'),
+    bias_attr=ParamAttr(name='hidden.b'))
+
+prob = fc_layer(
+    input=hidden,
+    size=10,
+    act=SoftmaxActivation(),
+    param_attr=ParamAttr(name='prob.w'),
+    bias_attr=ParamAttr(name='prob.b'))
+
+outputs(prob)
diff --git a/paddle/capi/examples/multi_thread/.gitignore b/paddle/capi/examples/multi_thread/.gitignore
new file mode 100644
index 0000000000000..fab7372d796ea
--- /dev/null
+++ b/paddle/capi/examples/multi_thread/.gitignore
@@ -0,0 +1,73 @@
+# This file is used to ignore files which are generated
+# ----------------------------------------------------------------------------
+
+*~
+*.autosave
+*.a
+*.core
+*.moc
+*.o
+*.obj
+*.orig
+*.rej
+*.so
+*.so.*
+*_pch.h.cpp
+*_resource.rc
+*.qm
+.#*
+*.*#
+core
+!core/
+tags
+.DS_Store
+.directory
+*.debug
+Makefile*
+*.prl
+*.app
+moc_*.cpp
+ui_*.h
+qrc_*.cpp
+Thumbs.db
+*.res
+*.rc
+/.qmake.cache
+/.qmake.stash
+
+# qtcreator generated files
+*.pro.user*
+
+# xemacs temporary files
+*.flc
+
+# Vim temporary files
+.*.swp
+
+# Visual Studio generated files
+*.ib_pdb_index
+*.idb
+*.ilk
+*.pdb
+*.sln
+*.suo
+*.vcproj
+*vcproj.*.*.user
+*.ncb
+*.sdf
+*.opensdf
+*.vcxproj
+*vcxproj.*
+
+# MinGW generated files
+*.Debug
+*.Release
+
+# Python byte code
+*.pyc
+
+# Binaries
+# --------
+*.dll
+*.exe
+
diff --git a/paddle/capi/examples/multi_thread/CMakeLists.txt b/paddle/capi/examples/multi_thread/CMakeLists.txt
new file mode 100644
index 0000000000000..98e411ddc02a4
--- /dev/null
+++ b/paddle/capi/examples/multi_thread/CMakeLists.txt
@@ -0,0 +1,8 @@
+project(multi_thread)
+cmake_minimum_required(VERSION 2.8)
+aux_source_directory(. SRC_LIST)
+add_executable(${PROJECT_NAME} ${SRC_LIST})
+find_package (Threads)
+set_property(TARGET ${PROJECT_NAME} PROPERTY C_STANDARD 99)
+target_link_libraries(${PROJECT_NAME} -lpaddle_capi_shared
+  ${CMAKE_THREAD_LIBS_INIT})
diff --git a/paddle/capi/examples/multi_thread/convert_protobin.sh b/paddle/capi/examples/multi_thread/convert_protobin.sh
new file mode 120000
index 0000000000000..3c1b3533523cf
--- /dev/null
+++ b/paddle/capi/examples/multi_thread/convert_protobin.sh
@@ -0,0 +1 @@
+../dense/convert_protobin.sh
\ No newline at end of file
diff --git a/paddle/capi/examples/multi_thread/main.c b/paddle/capi/examples/multi_thread/main.c
new file mode 100644
index 0000000000000..23f8629765d8a
--- /dev/null
+++ b/paddle/capi/examples/multi_thread/main.c
@@ -0,0 +1,96 @@
+#include <paddle/capi.h>
+#include <pthread.h>
+#include <time.h>
+#include "../common/common.h"
+
+#define CONFIG_BIN "./trainer_config.bin"
+#define NUM_THREAD 1000
+#define NUM_ITER 1000
+
+pthread_mutex_t mutex;
+
+void* thread_main(void* gm_ptr) {
+  paddle_gradient_machine machine = (paddle_gradient_machine)(gm_ptr);
+
+  for (int iter = 0; iter < NUM_ITER; ++iter) {
+    paddle_arguments in_args = paddle_arguments_create_none();
+    // There is only one input of this network.
+    CHECK(paddle_arguments_resize(in_args, 1));
+
+    // Create input matrix.
+    paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
+                                             /* size */ 784,
+                                             /* useGPU */ false);
+
+    paddle_real* array;
+
+    // Get First row.
+    CHECK(paddle_matrix_get_row(mat, 0, &array));
+
+    for (int i = 0; i < 784; ++i) {
+      array[i] = rand() / ((float)RAND_MAX);
+    }
+
+    CHECK(paddle_arguments_set_value(in_args, 0, mat));
+
+    paddle_arguments out_args = paddle_arguments_create_none();
+    CHECK(paddle_gradient_machine_forward(machine,
+                                          in_args,
+                                          out_args,
+                                          /* isTrain */ false));
+    paddle_matrix prob = paddle_matrix_create_none();
+
+    CHECK(paddle_arguments_value(out_args, 0, prob));
+
+    CHECK(paddle_matrix_get_row(prob, 0, &array));
+
+    pthread_mutex_lock(&mutex);
+    printf("Prob: ");
+    for (int i = 0; i < 10; ++i) {
+      printf("%.2f ", array[i]);
+    }
+    printf("\n");
+    pthread_mutex_unlock(&mutex);
+  }
+
+  CHECK(paddle_gradient_machine_destroy(machine));
+  return NULL;
+}
+
+int main() {
+  // Initalize Paddle
+  char* argv[] = {"--use_gpu=False"};
+  CHECK(paddle_init(1, (char**)argv));
+
+  // Reading config binary file. It is generated by `convert_protobin.sh`
+  long size;
+  void* buf = read_config(CONFIG_BIN, &size);
+
+  // Create a gradient machine for inference.
+  paddle_gradient_machine machine;
+  CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
+  CHECK(paddle_gradient_machine_randomize_param(machine));
+
+  // Loading parameter. Uncomment the following line and change the directory.
+  // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
+  //                                                "./some_where_to_params"));
+  srand(time(0));
+  pthread_mutex_init(&mutex, NULL);
+
+  pthread_t threads[NUM_THREAD];
+
+  for (int i = 0; i < NUM_THREAD; ++i) {
+    paddle_gradient_machine thread_local_machine;
+    CHECK(paddle_gradient_machine_create_shared_param(
+        machine, buf, size, &thread_local_machine));
+    pthread_create(&threads[i], NULL, thread_main, thread_local_machine);
+  }
+
+  for (int i = 0; i < NUM_THREAD; ++i) {
+    pthread_join(threads[i], NULL);
+  }
+
+  pthread_mutex_destroy(&mutex);
+
+  return 0;
+}
diff --git a/paddle/capi/examples/multi_thread/trainer_config.py b/paddle/capi/examples/multi_thread/trainer_config.py
new file mode 120000
index 0000000000000..70cfb1f7f4cfe
--- /dev/null
+++ b/paddle/capi/examples/multi_thread/trainer_config.py
@@ -0,0 +1 @@
+../dense/trainer_config.py
\ No newline at end of file
diff --git a/paddle/capi/examples/sparse_binary/.gitignore b/paddle/capi/examples/sparse_binary/.gitignore
new file mode 100644
index 0000000000000..fab7372d796ea
--- /dev/null
+++ b/paddle/capi/examples/sparse_binary/.gitignore
@@ -0,0 +1,73 @@
+# This file is used to ignore files which are generated
+# ----------------------------------------------------------------------------
+
+*~
+*.autosave
+*.a
+*.core
+*.moc
+*.o
+*.obj
+*.orig
+*.rej
+*.so
+*.so.*
+*_pch.h.cpp
+*_resource.rc
+*.qm
+.#*
+*.*#
+core
+!core/
+tags
+.DS_Store
+.directory
+*.debug
+Makefile*
+*.prl
+*.app
+moc_*.cpp
+ui_*.h
+qrc_*.cpp
+Thumbs.db
+*.res
+*.rc
+/.qmake.cache
+/.qmake.stash
+
+# qtcreator generated files
+*.pro.user*
+
+# xemacs temporary files
+*.flc
+
+# Vim temporary files
+.*.swp
+
+# Visual Studio generated files
+*.ib_pdb_index
+*.idb
+*.ilk
+*.pdb
+*.sln
+*.suo
+*.vcproj
+*vcproj.*.*.user
+*.ncb
+*.sdf
+*.opensdf
+*.vcxproj
+*vcxproj.*
+
+# MinGW generated files
+*.Debug
+*.Release
+
+# Python byte code
+*.pyc
+
+# Binaries
+# --------
+*.dll
+*.exe
+
diff --git a/paddle/capi/examples/sparse_binary/CMakeLists.txt b/paddle/capi/examples/sparse_binary/CMakeLists.txt
new file mode 100644
index 0000000000000..c82195688902a
--- /dev/null
+++ b/paddle/capi/examples/sparse_binary/CMakeLists.txt
@@ -0,0 +1,7 @@
+project(sparse_binary)
+cmake_minimum_required(VERSION 2.8)
+aux_source_directory(. SRC_LIST)
+add_executable(${PROJECT_NAME} ${SRC_LIST})
+find_package (Threads)
+set_property(TARGET ${PROJECT_NAME} PROPERTY C_STANDARD 99)
+target_link_libraries(${PROJECT_NAME} -lpaddle_capi_shared)
diff --git a/paddle/capi/examples/sparse_binary/convert_protobin.sh b/paddle/capi/examples/sparse_binary/convert_protobin.sh
new file mode 120000
index 0000000000000..3c1b3533523cf
--- /dev/null
+++ b/paddle/capi/examples/sparse_binary/convert_protobin.sh
@@ -0,0 +1 @@
+../dense/convert_protobin.sh
\ No newline at end of file
diff --git a/paddle/capi/examples/sparse_binary/main.c b/paddle/capi/examples/sparse_binary/main.c
new file mode 100644
index 0000000000000..c5e653dbc2876
--- /dev/null
+++ b/paddle/capi/examples/sparse_binary/main.c
@@ -0,0 +1,64 @@
+#include <paddle/capi.h>
+#include <time.h>
+#include "../common/common.h"
+
+#define CONFIG_BIN "./trainer_config.bin"
+
+int main() {
+  // Initalize Paddle
+  char* argv[] = {"--use_gpu=False"};
+  CHECK(paddle_init(1, (char**)argv));
+
+  // Reading config binary file. It is generated by `convert_protobin.sh`
+  long size;
+  void* buf = read_config(CONFIG_BIN, &size);
+
+  // Create a gradient machine for inference.
+  paddle_gradient_machine machine;
+  CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
+  CHECK(paddle_gradient_machine_randomize_param(machine));
+
+  // Loading parameter. Uncomment the following line and change the directory.
+  // CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
+  //                                                "./some_where_to_params"));
+  paddle_arguments in_args = paddle_arguments_create_none();
+
+  // There is only one input of this network.
+  CHECK(paddle_arguments_resize(in_args, 1));
+
+  // Create input matrix.
+  paddle_matrix mat = paddle_matrix_create_sparse(1, 784, 3, true, false);
+  srand(time(0));
+  paddle_real* array;
+  int colBuf[] = {9, 93, 109};
+  int rowBuf[] = {0, sizeof(colBuf) / sizeof(int)};
+
+  CHECK(paddle_matrix_sparse_copy_from(mat,
+                                       rowBuf,
+                                       sizeof(rowBuf) / sizeof(int),
+                                       colBuf,
+                                       sizeof(colBuf) / sizeof(int),
+                                       NULL,
+                                       0));
+
+  CHECK(paddle_arguments_set_value(in_args, 0, mat));
+
+  paddle_arguments out_args = paddle_arguments_create_none();
+  CHECK(paddle_gradient_machine_forward(machine,
+                                        in_args,
+                                        out_args,
+                                        /* isTrain */ false));
+  paddle_matrix prob = paddle_matrix_create_none();
+
+  CHECK(paddle_arguments_value(out_args, 0, prob));
+
+  CHECK(paddle_matrix_get_row(prob, 0, &array));
+
+  printf("Prob: ");
+  for (int i = 0; i < 10; ++i) {
+    printf("%.2f ", array[i]);
+  }
+  printf("\n");
+
+  return 0;
+}
diff --git a/paddle/capi/examples/sparse_binary/trainer_config.py b/paddle/capi/examples/sparse_binary/trainer_config.py
new file mode 120000
index 0000000000000..70cfb1f7f4cfe
--- /dev/null
+++ b/paddle/capi/examples/sparse_binary/trainer_config.py
@@ -0,0 +1 @@
+../dense/trainer_config.py
\ No newline at end of file
diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp
index 6e7740a455b5d..00f76e0152366 100644
--- a/paddle/capi/gradient_machine.cpp
+++ b/paddle/capi/gradient_machine.cpp
@@ -113,3 +113,11 @@ paddle_error paddle_gradient_machine_create_shared_param(
   return kPD_NO_ERROR;
 }
 }
+
+paddle_error paddle_gradient_machine_randomize_param(
+    paddle_gradient_machine machine) {
+  auto m = cast(machine);
+  if (m == nullptr || m->machine == nullptr) return kPD_NULLPTR;
+  m->machine->randParameters();
+  return kPD_NO_ERROR;
+}
diff --git a/paddle/capi/gradient_machine.h b/paddle/capi/gradient_machine.h
index 36c1a2b1b4b91..d7e2dd9bf8037 100644
--- a/paddle/capi/gradient_machine.h
+++ b/paddle/capi/gradient_machine.h
@@ -74,6 +74,9 @@ paddle_gradient_machine_create_shared_param(paddle_gradient_machine origin,
                                             int size,
                                             paddle_gradient_machine* slave);
 
+PD_API paddle_error
+paddle_gradient_machine_randomize_param(paddle_gradient_machine machine);
+
 /**
  * @brief Destroy a gradient machine
  * @param machine that need to destroy
diff --git a/paddle/capi/matrix.h b/paddle/capi/matrix.h
index 2f6488f38386a..f15f7f3bbbd14 100644
--- a/paddle/capi/matrix.h
+++ b/paddle/capi/matrix.h
@@ -15,6 +15,7 @@ limitations under the License. */
 #ifndef __PADDLE_CAPI_MATRIX_H__
 #define __PADDLE_CAPI_MATRIX_H__
 
+#include <stdbool.h>
 #include <stdint.h>
 #include "config.h"
 #include "error.h"
@@ -39,6 +40,18 @@ PD_API paddle_matrix paddle_matrix_create(uint64_t height,
                                           uint64_t width,
                                           bool useGpu);
 
+/**
+ * @brief paddle_matrix_create_sparse Create a sparse matrix.
+ * @param height the matrix height.
+ * @param width the matrix width.
+ * @param nnz the number of non-zero elements.
+ * @param isBinary is binary (either 1 or 0 in matrix) or not.
+ * @param useGpu is using GPU or not.
+ * @return paddle_matrix.
+ */
+PD_API paddle_matrix paddle_matrix_create_sparse(
+    uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu);
+
 /**
  * @brief paddle_matrix_destroy Destroy a matrix.
  * @param mat
@@ -55,7 +68,7 @@ PD_API paddle_error paddle_matrix_destroy(paddle_matrix mat);
  */
 PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
                                           uint64_t rowID,
-                                          pd_real* rowArray);
+                                          paddle_real* rowArray);
 
 /**
  * @brief PDMatGetRow Get raw row buffer from matrix
@@ -66,7 +79,7 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
  */
 PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                           uint64_t rowID,
-                                          pd_real** rawRowBuffer);
+                                          paddle_real** rawRowBuffer);
 
 /**
  * @brief PDMatCreateNone Create None Matrix
@@ -85,6 +98,27 @@ PD_API paddle_error paddle_matrix_get_shape(paddle_matrix mat,
                                             uint64_t* height,
                                             uint64_t* width);
 
+/**
+ * @brief paddle_matrix_sparse_copy_from Copy from a CSR format matrix
+ * @param [out] mat output matrix
+ * @param [in] rowArray row array. The array slices in column array.
+ * @param [in] rowSize length of row array.
+ * @param [in] colArray the column array. It means the non-zero element indices
+ * in each row.
+ * @param [in] colSize length of column array.
+ * @param [in] valueArray the value array. It means the non-zero elemnt values.
+ * NULL if the matrix is binary.
+ * @param [in] valueSize length of value array. Zero if the matrix is binary.
+ * @return paddle_error
+ */
+PD_API paddle_error paddle_matrix_sparse_copy_from(paddle_matrix mat,
+                                                   int* rowArray,
+                                                   uint64_t rowSize,
+                                                   int* colArray,
+                                                   uint64_t colSize,
+                                                   float* valueArray,
+                                                   uint64_t valueSize);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/paddle/capi/tests/test_Arguments.cpp b/paddle/capi/tests/test_Arguments.cpp
index e6e4ac9937e5e..f56391d51e32a 100644
--- a/paddle/capi/tests/test_Arguments.cpp
+++ b/paddle/capi/tests/test_Arguments.cpp
@@ -17,10 +17,10 @@ limitations under the License. */
 #include "gtest/gtest.h"
 #include "paddle/utils/ThreadLocal.h"
 
-static std::vector<pd_real> randomBuffer(size_t bufSize) {
+static std::vector<paddle_real> randomBuffer(size_t bufSize) {
   auto& eng = paddle::ThreadLocalRandomEngine::get();
-  std::uniform_real_distribution<pd_real> dist(-1.0, 1.0);
-  std::vector<pd_real> retv;
+  std::uniform_real_distribution<paddle_real> dist(-1.0, 1.0);
+  std::vector<paddle_real> retv;
   retv.reserve(bufSize);
   for (size_t i = 0; i < bufSize; ++i) {
     retv.push_back(dist(eng));
@@ -42,7 +42,7 @@ TEST(CAPIArguments, value) {
 
   paddle_matrix mat = paddle_matrix_create(128, 64, false);
   for (size_t i = 0; i < 128; ++i) {
-    std::vector<pd_real> sampleBuf = randomBuffer(64);
+    std::vector<paddle_real> sampleBuf = randomBuffer(64);
     paddle_matrix_set_row(mat, i, sampleBuf.data());
   }
   ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(args, 0, mat));
@@ -52,8 +52,8 @@ TEST(CAPIArguments, value) {
   ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_value(args, 0, val));
 
   for (size_t i = 0; i < 128; ++i) {
-    pd_real* row1;
-    pd_real* row2;
+    paddle_real* row1;
+    paddle_real* row2;
 
     ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, i, &row1));
     ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(val, i, &row2));
diff --git a/paddle/capi/tests/test_GradientMachine.cpp b/paddle/capi/tests/test_GradientMachine.cpp
index b37fe83a30bd1..be7dfadddc35e 100644
--- a/paddle/capi/tests/test_GradientMachine.cpp
+++ b/paddle/capi/tests/test_GradientMachine.cpp
@@ -21,10 +21,10 @@ limitations under the License. */
 #include "capi.h"
 #include "paddle/utils/ThreadLocal.h"
 
-static std::vector<pd_real> randomBuffer(size_t bufSize) {
+static std::vector<paddle_real> randomBuffer(size_t bufSize) {
   auto& eng = paddle::ThreadLocalRandomEngine::get();
-  std::uniform_real_distribution<pd_real> dist(-1.0, 1.0);
-  std::vector<pd_real> retv;
+  std::uniform_real_distribution<paddle_real> dist(-1.0, 1.0);
+  std::vector<paddle_real> retv;
   retv.reserve(bufSize);
   for (size_t i = 0; i < bufSize; ++i) {
     retv.push_back(dist(eng));
@@ -60,12 +60,12 @@ TEST(GradientMachine, testPredict) {
   paddle_arguments inArgs = paddle_arguments_create_none();
   ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(inArgs, 1));
   paddle_matrix mat = paddle_matrix_create(1, 100, false);
-  static_assert(std::is_same<pd_real, paddle::real>::value, "");
+  static_assert(std::is_same<paddle_real, paddle::real>::value, "");
 
   auto data = randomBuffer(100);
-  pd_real* rowPtr;
+  paddle_real* rowPtr;
   ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr));
-  memcpy(rowPtr, data.data(), data.size() * sizeof(pd_real));
+  memcpy(rowPtr, data.data(), data.size() * sizeof(paddle_real));
 
   ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(inArgs, 0, mat));
   ASSERT_EQ(kPD_NO_ERROR,
diff --git a/paddle/capi/tests/test_Matrix.cpp b/paddle/capi/tests/test_Matrix.cpp
index 162df448d2b27..71dc2064dd02d 100644
--- a/paddle/capi/tests/test_Matrix.cpp
+++ b/paddle/capi/tests/test_Matrix.cpp
@@ -17,7 +17,7 @@ limitations under the License. */
 
 TEST(CAPIMatrix, create) {
   paddle_matrix mat = paddle_matrix_create(128, 32, false);
-  std::vector<pd_real> sampleRow;
+  std::vector<paddle_real> sampleRow;
   sampleRow.resize(32);
   for (size_t i = 0; i < sampleRow.size(); ++i) {
     sampleRow[i] = 1.0 / (i + 1.0);
@@ -26,7 +26,7 @@ TEST(CAPIMatrix, create) {
   ASSERT_EQ(kPD_OUT_OF_RANGE,
             paddle_matrix_set_row(mat, 128, sampleRow.data()));
 
-  pd_real* arrayPtr;
+  paddle_real* arrayPtr;
 
   ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &arrayPtr));
   for (size_t i = 0; i < sampleRow.size(); ++i) {