[cherry-pick] NV JETSON support and auto_growth strategy for inference. (#21500)

* ADD NV JETSON SUPPORT
test=release/1.6

* CHERRY_PICK: specify the auto growth allocator for inference.
test=release/1.6

CMakeLists.txt

@@ -69,6 +69,7 @@ option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
 option(ON_INFER         "Turn on inference optimization."               OFF)
 ################################ Internal Configurations #######################################
 option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
+option(WITH_NV_JETSON     "Compile PaddlePaddle with NV JETSON"             OFF)
 option(WITH_NGRAPH      "Compile PaddlePaddle with nGraph support."     OFF)
 option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler and gperftools"        OFF)
 option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)

cmake/cuda.cmake

@@ -2,11 +2,20 @@ if(NOT WITH_GPU)
     return()
 endif()
 
-set(paddle_known_gpu_archs "30 35 50 52 60 61 70")
-set(paddle_known_gpu_archs7 "30 35 50 52")
-set(paddle_known_gpu_archs8 "30 35 50 52 60 61")
-set(paddle_known_gpu_archs9 "30 35 50 52 60 61 70")
-set(paddle_known_gpu_archs10 "30 35 50 52 60 61 70 75")
+
+if (WITH_NV_JETSON)
+  set(paddle_known_gpu_archs "53 62 72")
+  set(paddle_known_gpu_archs7 "53")
+  set(paddle_known_gpu_archs8 "53 62")
+  set(paddle_known_gpu_archs9 "53 62")
+  set(paddle_known_gpu_archs10 "53 62 72")
+else()
+  set(paddle_known_gpu_archs "30 35 50 52 60 61 70")
+  set(paddle_known_gpu_archs7 "30 35 50 52")
+  set(paddle_known_gpu_archs8 "30 35 50 52 60 61")
+  set(paddle_known_gpu_archs9 "30 35 50 52 60 61 70")
+  set(paddle_known_gpu_archs10 "30 35 50 52 60 61 70 75")
+endif()
 
 ######################################################################################
 # A function for automatic detection of GPUs installed  (if autodetection is enabled)

cmake/external/openblas.cmake

@@ -33,7 +33,7 @@ IF(NOT ${CBLAS_FOUND})
 
     IF (NOT WIN32)
     SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
-    SET(OPENBLAS_COMMIT "v0.2.20")
+    SET(OPENBLAS_COMMIT "v0.3.7")
 
     IF(APPLE)
         SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
@@ -54,7 +54,6 @@ IF(NOT ${CBLAS_FOUND})
         BUILD_IN_SOURCE     1
         BUILD_COMMAND       ${CMAKE_MAKE_PROGRAM} ${COMMON_ARGS} ${OPTIONAL_ARGS}
         INSTALL_COMMAND     ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 NO_LAPACK=1 PREFIX=<INSTALL_DIR> 
-                            && rm -r ${CBLAS_INSTALL_DIR}/lib/cmake ${CBLAS_INSTALL_DIR}/lib/pkgconfig
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )

cmake/flags.cmake

@@ -187,7 +187,9 @@ set(GPU_COMMON_FLAGS
     -Wno-error=unused-function  # Warnings in Numpy Header.
     -Wno-error=array-bounds # Warnings in Eigen::array
 )
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
+if (NOT WITH_NV_JETSON) 
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
+endif()
 endif(NOT WIN32)
 
 if (APPLE)

paddle/fluid/framework/CMakeLists.txt

@@ -140,26 +140,28 @@ cc_library(op_call_stack SRCS op_call_stack.cc DEPS op_proto_maker enforce)
 
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
-py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
-py_proto_compile(trainer_py_proto SRCS trainer_desc.proto data_feed.proto)
+if(WITH_PYTHON)
+  py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
+  py_proto_compile(trainer_py_proto SRCS trainer_desc.proto data_feed.proto)
 #Generate an empty \
     #__init__.py to make framework_py_proto as a valid python module.
-add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
-add_dependencies(framework_py_proto framework_py_proto_init)
-if (NOT WIN32)
+  add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
+  add_dependencies(framework_py_proto framework_py_proto_init)
+  if (NOT WIN32)
     add_custom_command(TARGET framework_py_proto POST_BUILD
       COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
       COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/
       COMMENT "Copy generated python proto into directory paddle/fluid/proto."
       WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
-else(NOT WIN32)
+  else(NOT WIN32)
     string(REPLACE "/" "\\" proto_dstpath "${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/")
     add_custom_command(TARGET framework_py_proto POST_BUILD
           COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto
           COMMAND copy /Y *.py ${proto_dstpath}
           COMMENT "Copy generated python proto into directory paddle/fluid/proto."
           WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
-endif(NOT WIN32)
+  endif(NOT WIN32)
+endif()
 
 cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
 

paddle/fluid/framework/io/shell.cc

@@ -19,7 +19,7 @@ namespace framework {
 
 std::shared_ptr<FILE> shell_fopen(const std::string& path,
                                   const std::string& mode) {
-#if defined _WIN32 || defined __APPLE__
+#if defined _WIN32 || defined __APPLE__ || defined PADDLE_ARM
   return nullptr;
 #else
   if (shell_verbose()) {
@@ -44,7 +44,7 @@ std::shared_ptr<FILE> shell_fopen(const std::string& path,
 // The implementation is async signal safe
 // Mostly copy from CPython code
 static int close_open_fds_internal() {
-#if defined _WIN32 || defined __APPLE__
+#if defined _WIN32 || defined __APPLE__ || defined PADDLE_ARM
   return 0;
 #else
   struct linux_dirent {

paddle/fluid/framework/io/shell.h

@@ -31,6 +31,11 @@
 #include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/string/string_helper.h"
 
+#if defined(__arm__) || defined(__aarch64__) || defined(__ARM_NEON) || \
+    defined(__ARM_NEON__)
+#define PADDLE_ARM
+#endif
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/inference/api/analysis_predictor.cc

@@ -500,6 +500,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
       std::string flag = "--fraction_of_gpu_memory_to_use=" +
                          std::to_string(fraction_of_gpu_memory);
       flags.push_back(flag);
+      // use auto growth strategy here.
+      flags.push_back("--allocator_strategy=auto_growth");
       flags.push_back("--cudnn_deterministic=True");
       VLOG(3) << "set flag: " << flag;
       framework::InitGflags(flags);

paddle/fluid/platform/CMakeLists.txt

 proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto simple_threadpool)
-py_proto_compile(profiler_py_proto SRCS profiler.proto)
 proto_library(error_codes_proto SRCS error_codes.proto)
 
-add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
-add_dependencies(profiler_py_proto profiler_py_proto_init)
 
-if (NOT WIN32)
-add_custom_command(TARGET profiler_py_proto POST_BUILD
+if (WITH_PYTHON)
+  py_proto_compile(profiler_py_proto SRCS profiler.proto)
+  add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
+  add_dependencies(profiler_py_proto profiler_py_proto_init)
+
+  if (NOT WIN32)
+    add_custom_command(TARGET profiler_py_proto POST_BUILD
         COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMENT "Copy generated python proto into directory paddle/fluid/proto/profiler."
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
-else(NOT WIN32)
-string(REPLACE "/" "\\" proto_dstpath "${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler/")
-add_custom_command(TARGET profiler_py_proto POST_BUILD
+  else(NOT WIN32)
+    string(REPLACE "/" "\\" proto_dstpath "${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler/")
+    add_custom_command(TARGET profiler_py_proto POST_BUILD
         COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMAND copy /Y *.py ${proto_dstpath}
         COMMENT "Copy generated python proto into directory paddle/fluid/proto/profiler."
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
-endif(NOT WIN32)
+  endif(NOT WIN32)
+endif()
 
 cc_library(flags SRCS flags.cc DEPS gflags) 
 

paddle/fluid/platform/float16.h

@@ -38,24 +38,6 @@ limitations under the License. */
 #include <cuda_fp16.h>
 #endif
 
-#if defined(__arm__) || defined(__aarch64__)
-#define PADDLE_ARM
-#endif
-
-#if defined(__ARM_NEON) || defined(__ARM_NEON__)
-#define PADDLE_NEON
-#include <arm_neon.h>
-#endif
-
-#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
-    (PADDLE_GNUC_VER >= 62 || PADDLE_CLANG_VER >= 37)
-#define PADDLE_WITH_NATIVE_FP16
-#endif
-
-#ifndef PADDLE_ARM
-#include <immintrin.h>
-#endif  // PADDLE_ARM
-
 #if !defined(_WIN32)
 #define PADDLE_ALIGN(x) __attribute__((aligned(x)))
 #else