diff --git a/cmake/cross_compiling/android.cmake b/cmake/cross_compiling/android.cmake
index 5e3e437a8da9624df35a5c754fe77be73f20361d..84219cfa5587f5b765b2e8f35180797d7053169f 100644
--- a/cmake/cross_compiling/android.cmake
+++ b/cmake/cross_compiling/android.cmake
@@ -20,6 +20,7 @@
 # The supported variables are listed belows:
 # 
 # ANDROID_STANDALONE_TOOLCHAIN
+# ANDROID_TOOLCHAIN
 # ANDROID_ABI
 # ANDROID_NATIVE_API_LEVEL
 # ANDROID_ARM_MODE
@@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL)
     ENDIF()
 ENDIF()
 
+IF(NOT DEFINED ANDROID_TOOLCHAIN)
+    SET(ANDROID_TOOLCHAIN clang)
+ENDIF()
+
 IF(NOT DEFINED ANDROID_ABI)
     SET(ANDROID_ABI "armeabi-v7a")
 ENDIF()
@@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
             "${CMAKE_VERSION}), when cross-compiling for Android.")
 
     IF(ANDROID_STANDALONE_TOOLCHAIN)
+        # Use standalone toolchain
         SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot")
 
         IF(NOT CMAKE_SYSTEM_VERSION)
@@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ENDIF()
 
         # Toolchain
-        SET(ANDROID_TOOLCHAIN "gcc")
         SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN})
-        IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-            SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
-            IF(ANDROID_ABI STREQUAL "armeabi")
-                SET(CMAKE_SYSTEM_PROCESSOR armv5te)
-            ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
-                SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
-            ENDIF()
-        ENDIF()
-        IF(ANDROID_ABI STREQUAL "arm64-v8a")
-            SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
-            SET(CMAKE_SYSTEM_PROCESSOR aarch64)
+    ELSE(ANDROID_NDK)
+        # TODO: use android ndk
+    ENDIF()
+
+    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
+        SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
+        IF(ANDROID_ABI STREQUAL "armeabi")
+            SET(CMAKE_SYSTEM_PROCESSOR armv5te)
+            SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi)
+        ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
+            SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
+            SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi)
         ENDIF()
-        SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
+    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
+        SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
+        SET(CMAKE_SYSTEM_PROCESSOR aarch64)
+        SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android)
+    ELSE()
+        MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.")
+    ENDIF()
+    SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
+
+    IF(ANDROID_TOOLCHAIN STREQUAL clang)
+        SET(ANDROID_C_COMPILER_NAME clang)
+        SET(ANDROID_CXX_COMPILER_NAME clang++)
+        SET(CMAKE_C_COMPILER_TARGET   ${ANDROID_CLANG_TRIPLE})
+        SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
+    ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc)
+        SET(ANDROID_C_COMPILER_NAME gcc)
+        SET(ANDROID_CXX_COMPILER_NAME g++)
+    ELSE()
+        MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}")
     ENDIF()
 
     # C compiler
     IF(NOT CMAKE_C_COMPILER)
-        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}gcc")
+        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}")
     ELSE()
         GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
     ENDIF()
@@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
 
     # CXX compiler
     IF(NOT CMAKE_CXX_COMPILER)
-        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}g++")
+        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}")
     ELSE()
         GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
     ENDIF()
@@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
     SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
 
     # Toolchain and ABI specific flags.
-    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections -finline-limit=64")
+    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections")
     SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections")
 
     IF(ANDROID_ABI STREQUAL "armeabi")
@@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
              -march=armv5te
              -mtune=xscale
              -msoft-float)
-    ENDIF()
-    IF(ANDROID_ABI STREQUAL "armeabi-v7a")
+    ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
         LIST(APPEND ANDROID_COMPILER_FLAGS
              -march=armv7-a
              -mfloat-abi=softfp)
@@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
             LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16)
         ENDIF()
         LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8)
+    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
+        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
     ENDIF()
 
     IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
@@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ELSE()
             LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb)
         ENDIF()
+        IF(ANDROID_TOOLCHAIN STREQUAL clang)
+            # Disable integrated-as for better compatibility.
+            LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as)
+        ENDIF()
     ENDIF()
 
-    IF(ANDROID_ABI STREQUAL "arm64-v8a")
-        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
+    IF(ANDROID_TOOLCHAIN STREQUAL clang)
+        # CMake automatically forwards all compiler flags to the linker,
+        # and clang doesn't like having -Wa flags being used for linking.
+        # To prevent CMake from doing this would require meddling with
+        # the CMAKE_<LANG>_COMPILE_OBJECT rules, which would get quite messy.
+        LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments)
     ENDIF()
 
     STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")
diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake
index 3cc652bed5ec6e745c38294d2f17f6e80650cfa2..2d7daed9bcd5b8d854ffae6dc1ea191d154c16fe 100644
--- a/cmake/external/warpctc.cmake
+++ b/cmake/external/warpctc.cmake
@@ -41,7 +41,6 @@ IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "App
 ELSE()
     SET(USE_OMP ON)
 ENDIF()
-SET(USE_OMP OFF FORCE)
 
 ExternalProject_Add(
     extern_warpctc
diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index 82d598d885bd698f5a36c7b45f1f56c21c2dc915..4a2132c8d1bfa329ced575f9b78052bdbfe3e4d5 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -2774,28 +2774,23 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
   CHECK(!isTransposed()) << "Not supported";
 
   size_t a_col, b_col, a_row, b_row;
-  // CBLAS_TRANSPOSE a_trans, b_trans;
   bool a_trans, b_trans;
   if (!a->isTransposed()) {
     a_col = a->getWidth();
     a_row = a->getHeight();
-    // a_trans = CblasNoTrans;
     a_trans = false;
   } else {
     a_col = a->getHeight();
     a_row = a->getWidth();
-    // a_trans = CblasTrans;
     a_trans = true;
   }
   if (!b->isTransposed()) {
     b_col = b->getWidth();
     b_row = b->getHeight();
-    // b_trans = CblasNoTrans;
     b_trans = false;
   } else {
     b_col = b->getHeight();
     b_row = b->getWidth();
-    // b_trans = CblasTrans;
     b_trans = true;
   }
 
diff --git a/paddle/scripts/docker/build_android.sh b/paddle/scripts/docker/build_android.sh
index 98f66fa6f1ab2df87987c381728dc2164eb217e4..512a37166c7ade8cb0b7ca8d18d73080193c9c27 100644
--- a/paddle/scripts/docker/build_android.sh
+++ b/paddle/scripts/docker/build_android.sh
@@ -9,20 +9,15 @@ rm -rf $BUILD_ROOT 2>/dev/null || true
 mkdir -p $BUILD_ROOT
 cd $BUILD_ROOT
 
-THIRD_PARTY_PATH=/paddle/third_party_android$SUFFIX/$ANDROID_ABI
-
 if [ $ANDROID_ABI == "armeabi-v7a" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
         -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_NEON=ON \
         -DANDROID_ARM_MODE=ON \
-        -DCMAKE_C_COMPILER=$ANDROID_ARM_STANDALONE_TOOLCHAIN/bin/arm-linux-androideabi-clang \
-        -DCMAKE_CXX_COMPILER=$ANDROID_ARM_STANDALONE_TOOLCHAIN/bin/arm-linux-androideabi-clang++ \
         -DHOST_C_COMPILER=/usr/bin/gcc \
         -DHOST_CXX_COMPILER=/usr/bin/g++ \
         -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
-        -DTHIRD_PARTY_PATH=$THIRD_PARTY_PATH \
         -DCMAKE_BUILD_TYPE=Release \
         -DUSE_EIGEN_FOR_BLAS=ON \
         -DWITH_C_API=ON \
@@ -34,12 +29,9 @@ elif [ $ANDROID_ABI == "arm64-v8a" ]; then
         -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_MODE=ON \
-        -DCMAKE_C_COMPILER=$ANDROID_ARM64_STANDALONE_TOOLCHAIN/bin/aarch64-linux-android-clang \
-        -DCMAKE_CXX_COMPILER=$ANDROID_ARM64_STANDALONE_TOOLCHAIN/bin/aarch64-linux-android-clang++ \
         -DHOST_C_COMPILER=/usr/bin/gcc \
         -DHOST_CXX_COMPILER=/usr/bin/g++ \
         -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
-        -DTHIRD_PARTY_PATH=$THIRD_PARTY_PATH \
         -DCMAKE_BUILD_TYPE=Release \
         -DUSE_EIGEN_FOR_BLAS=OFF \
         -DWITH_C_API=ON \
@@ -53,7 +45,6 @@ elif [ $ANDROID_ABI == "armeabi" ]; then
         -DHOST_C_COMPILER=/usr/bin/gcc \
         -DHOST_CXX_COMPILER=/usr/bin/g++ \
         -DCMAKE_INSTALL_PREFIX=/paddle/install \
-        -DTHIRD_PARTY_PATH=$THIRD_PARTY_PATH \
         -DCMAKE_BUILD_TYPE=Release \
         -DWITH_C_API=ON \
         -DWITH_SWIG_PY=OFF \