Include arm_neon.h on arm platform.

cmake/cblas.cmake

@@ -14,14 +14,39 @@
 #
 
 set(CBLAS_FOUND OFF)
+set(LAPACK_FOUND OFF)
+
+macro(find_lapack)
+  ## Find clapack
+  set(CLAPACK_ROOT $ENV{CLAPACK_ROOT} CACHE PATH "Folder contain clapack")
+  find_path(CLAPACK_INC_DIR NAMES clapack.h PATHS
+    ${CLAPACK_ROOT}/include)
+  find_path(CLAPACK_F2C_INC_DIR NAMES f2c.h PATHS
+    ${CLAPACK_ROOT}/include)
+  find_library(CLAPACK_F2C_LIB NAMES f2c PATHS
+    ${CLAPACK_ROOT}/lib)
+  find_library(CLAPACK_CBLASWR_LIB NAMES cblaswr PATHS
+    ${CLAPACK_ROOT}/lib)
+  find_library(CLAPACK_CLAPACK_LIB NAMES clapack PATHS
+    ${CLAPACK_ROOT}/lib)
+  if(CLAPACK_INC_DIR AND CLAPACK_F2C_INC_DIR AND CLAPACK_F2C_LIB AND
+     CLAPACK_CBLASWR_LIB AND CLAPACK_CLAPACK_LIB AND NOT LAPACK_FOUND)
+    set(LAPACK_PROVIDER CLAPACK)
+    set(LAPACK_INC_DIR ${CLAPACK_INC_DIR})
+    set(LAPACK_LIBRARIES ${CLAPACK_F2C_LIB} ${CLAPACK_CBLASWR_LIB} ${CLAPACK_CLAPACK_LIB})
+    add_definitions(-DPADDLE_USE_CLAPACK)
+    message(STATUS "Found CLAPACK (include: ${LAPACK_INC_DIR}, library: ${LAPACK_LIBRARIES})")
+    set(LAPACK_FOUND ON)
+  endif()
+endmacro()
 
 ## Find MKL First.
 set(INTEL_ROOT "/opt/intel" CACHE PATH "Folder contains intel libs")
 set(MKL_ROOT ${INTEL_ROOT}/mkl CACHE PATH "Folder contains MKL")
 
-find_path(MKL_INCLUDE_DIR mkl.h PATHS
+find_path(MKL_INC_DIR mkl.h PATHS
   ${MKL_ROOT}/include)
-find_path(MKL_INCLUDE_DIR mkl_lapacke.h PATHS
+find_path(MKL_LAPACK_INC_DIR mkl_lapacke.h PATHS
   ${MKL_ROOT}/include)
 find_library(MKL_CORE_LIB NAMES mkl_core PATHS
   ${MKL_ROOT}/lib
@@ -34,15 +59,22 @@ find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
   ${MKL_ROOT}/lib/intel64)
 
 
-if(MKL_INCLUDE_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
+if(MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
   set(CBLAS_PROVIDER MKL)
-  set(CBLAS_INC_DIR ${MKL_INCLUDE_DIR})
+  set(CBLAS_INC_DIR ${MKL_INC_DIR})
   set(CBLAS_LIBRARIES ${MKL_INTEL_LP64}
           ${MKL_SEQUENTIAL_LIB}
           ${MKL_CORE_LIB})
   add_definitions(-DPADDLE_USE_MKL)
   message(STATUS "Found MKL (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   set(CBLAS_FOUND ON)
+  if(${MKL_LAPACK_INC_DIR})
+    set(LAPACK_INC_DIR ${MKL_LAPACK_INC_DIR})
+    message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
+    set(LAPACK_FOUND ON)
+  else()
+    find_lapack()
+  endif()
   return() # return file.
 endif()
 
@@ -68,13 +100,20 @@ find_library(ATLAS_CBLAS_LIB NAMES cblas libcblas.so.3
 find_library(ATLAS_LIB NAMES lapack_atlas liblapack_atlas.so.3
   PATHS ${ATLAS_LIB_SEARCH_PATHS})
 
-if(ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_LIB)
+if(ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_LIB AND NOT CBLAS_FOUND)
   set(CBLAS_PROVIDER ATLAS)
-  set(CBLAS_INC_DIR ${ATLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR})
+  set(CBLAS_INC_DIR ${ATLAS_INC_DIR})
   set(CBLAS_LIBRARIES ${ATLAS_LIB} ${ATLAS_CBLAS_LIB})
   add_definitions(-DPADDLE_USE_ATLAS)  
-  message(STATUS "Found Atlas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
+  message(STATUS "Found ATLAS (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   set(CBLAS_FOUND ON)
+  if(ATLAS_CLAPACK_INC_DIR)
+    set(LAPACK_INC_DIR ${ATLAS_CLAPACK_INC_DIR})
+    message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})")
+    set(LAPACK_FOUND ON)
+  else()
+    find_lapack()
+  endif()
   return()
 endif()
 
@@ -103,8 +142,15 @@ if(OPENBLAS_INC_DIR AND OPENBLAS_LIB)
   set(CBLAS_PROVIDER OPENBLAS)
   set(CBLAS_INC_DIR ${OPENBLAS_INC_DIR})
   set(CBLAS_LIBRARIES ${OPENBLAS_LIB})
-  message(STATUS "Found OpenBlas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
+  message(STATUS "Found OpenBLAS (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   set(CBLAS_FOUND ON)
+  if(OPENBLAS_LAPACKE_INC_DIR)
+    set(LAPACK_INC_DIR ${OPENBLAS_LAPACKE_INC_DIR})
+    message(STATUS "Found lapack in OpenBLAS (include: ${OPENBLAS_LAPACKE_INC_DIR})")
+    set(LAPACK_FOUND ON)
+  else()
+    find_lapack()
+  endif()
   return()
 endif()
 

cmake/external/openblas.cmake

@@ -20,6 +20,7 @@ IF(NOT ${CBLAS_FOUND})
     SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
     SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
     SET(CBLAS_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
+    SET(LAPACK_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "lapack (in openblas) include directory." FORCE)
 
     IF(WIN32)
         SET(CBLAS_LIBRARIES "${CBLAS_INSTALL_DIR}/lib/openblas.lib" CACHE FILEPATH "openblas library." FORCE)
@@ -64,3 +65,4 @@ IF(NOT ${CBLAS_FOUND})
 ENDIF(NOT ${CBLAS_FOUND})
 
 INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
+INCLUDE_DIRECTORIES(${LAPACK_INC_DIR})

paddle/cuda/include/hl_matrix_type.cuh

@@ -17,13 +17,20 @@ limitations under the License. */
 
 #include "hl_base.h"
 
-#ifdef __CUDA_ARCH__
+#if defined(__CUDA_ARCH__)
 #include <vector_types.h>
 #ifndef PADDLE_TYPE_DOUBLE
 typedef float4 vecType;
 #else
 typedef double2 vecType;
 #endif
+#elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
+#include <arm_neon.h>
+#ifndef PADDLE_TYPE_DOUBLE
+typedef float32x4_t  vecType;
+#else
+#error NEON instructions does not support double precision
+#endif
 #else
 #include <mmintrin.h>
 #include <xmmintrin.h>

paddle/math/MathFunctions.h

@@ -17,7 +17,11 @@ limitations under the License. */
 
 #ifdef PADDLE_USE_MKL
 #include <mkl.h>
+#ifdef PADDLE_USE_CLAPACK
+#include <clapack.h>
+#else
 #include <mkl_lapacke.h>
+#endif
 #else
 extern "C" {
 #include <cblas.h>
@@ -27,9 +31,13 @@ extern "C" {
 #include <clapack.h>
 }
 #else
+#ifdef PADDLE_USE_CLAPACK
+#include <clapack.h>
+#else
 #include <lapacke.h>
 #endif
 #endif
+#endif
 
 #include <cmath>
 

paddle/utils/CpuId.cpp

@@ -19,7 +19,7 @@ limitations under the License. */
 /// for MSVC
 #define CPUID(info, x) __cpuidex(info, x, 0)
 
-#else
+#elif !defined(__ANDROID__)
 
 #include <cpuid.h>
 
@@ -31,6 +31,7 @@ limitations under the License. */
 namespace paddle {
 
 SIMDFlags::SIMDFlags() {
+#if !defined(__ANDROID__)
   unsigned int cpuInfo[4];
   // CPUID: https://en.wikipedia.org/wiki/CPUID
   // clang-format off
@@ -51,6 +52,9 @@ SIMDFlags::SIMDFlags() {
   CPUID(cpuInfo, 0x80000001);
   simd_flags_ |= cpuInfo[2] & (1 << 16) ? SIMD_FMA4  : SIMD_NONE;
   // clang-fotmat on
+#else
+  simd_flags_ = SIMD_NEON;
+#endif
 }
 
 SIMDFlags const* SIMDFlags::instance() {

paddle/utils/CpuId.h

@@ -18,6 +18,7 @@ namespace paddle {
 // clang-format off
 enum simd_t {
   SIMD_NONE   = 0,          ///< None
+#if !defined(__ANDROID__)
   SIMD_SSE    = 1 << 0,     ///< SSE
   SIMD_SSE2   = 1 << 1,     ///< SSE 2
   SIMD_SSE3   = 1 << 2,     ///< SSE 3
@@ -29,6 +30,9 @@ enum simd_t {
   SIMD_AVX    = 1 << 8,     ///< AVX
   SIMD_AVX2   = 1 << 9,     ///< AVX 2
   SIMD_AVX512 = 1 << 10,    ///< AVX 512
+#else
+  SIMD_NEON   = 1 << 0,     ///< NEON
+#endif
 };
 // clang-format on
 

paddle/utils/Logging.cpp

@@ -18,6 +18,7 @@ limitations under the License. */
  */
 
 #include "Logging.h"
+#include <cstdlib>
 
 namespace paddle {
 

paddle/utils/Util.cpp

@@ -15,11 +15,16 @@ limitations under the License. */
 #include "Util.h"
 
 #include <dirent.h>
-#include <pmmintrin.h>
 #include <signal.h>
 #include <sys/stat.h>
 #include <sys/types.h>
+
+#ifdef __SSE__
 #include <xmmintrin.h>
+#endif
+#ifdef __SSE3__
+#include <pmmintrin.h>
+#endif
 
 #include <fstream>
 #include <mutex>
@@ -162,8 +167,12 @@ void initMain(int argc, char** argv) {
 
   installProfilerSwitch();
 
+#ifdef __SSE__
   _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+#endif
+#ifdef __SSE3__
   _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#endif
 
   if (FLAGS_seed == 0) {
     unsigned int t = time(NULL);