Merge pull request #486 from lyq998/support-lite-fork-debug-mode

feat(lite): support base lite ipc fork debug

lite/example/c_example/main.c

@@ -14,6 +14,7 @@
 #include "lite-c/tensor_c.h"
 
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
 
 #define LITE_CAPI_CHECK(_expr)                                       \
@@ -42,7 +43,7 @@ int basic_c_interface(const char* mode_path) {
             LITE_get_tensor_total_size_in_byte(c_input_tensor, &length_in_byte));
     LITE_CAPI_CHECK(LITE_get_tensor_memory(c_input_tensor, &dst_ptr));
     //! copy or forward data to network
-    memset(dst_ptr, 5, length_in_byte);
+    LITE_memset(dst_ptr, 5, length_in_byte);
 
     //! forward
     LITE_CAPI_CHECK(LITE_forward(c_network));
@@ -66,23 +67,41 @@ int basic_c_interface(const char* mode_path) {
 
     float max = -1.0f;
     float sum = 0.0f;
+    int is_enable_ipc_debug = LITE_is_enable_ipc_debug_mode();
+    float* copy_ptr = NULL;
+    float* final_dst_ptr = (float*)output_ptr;
+    if (is_enable_ipc_debug) {
+        copy_ptr = (float*)(malloc(length_output_in_byte));
+        LITE_CAPI_CHECK(LITE_copy_server_tensor_memory(
+                output_ptr, copy_ptr, length_output_in_byte));
+        final_dst_ptr = (float*)copy_ptr;
+    }
+
     for (size_t i = 0; i < out_length; i++) {
-        float data = ((float*)(output_ptr))[i];
+        float data = final_dst_ptr[i];
         sum += data;
         if (max < data)
             max = data;
     }
     printf("max=%e, sum=%e\n", max, sum);
+    LITE_destroy_network(c_network);
+    if (is_enable_ipc_debug) {
+        free(copy_ptr);
+    }
     return 0;
 }
 
 int main(int argc, char** argv) {
-    if (argc < 2) {
-        printf("usage: lite_c_examples <model file> , just test C interface "
+    if (argc < 3) {
+        printf("usage: lite_c_examples is_enable_fork_debug_model <model file> , just "
+               "test C interface "
                "build.\n");
         return -1;
     }
-    return basic_c_interface(argv[1]);
+    if (atoi(argv[1])) {
+        LITE_enable_lite_ipc_debug();
+    }
+    return basic_c_interface(argv[2]);
 }
 
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

lite/lite-c/include/lite-c/global_c.h

@@ -26,6 +26,15 @@ LITE_API void LITE_clear_last_error();
  */
 LITE_API const char* LITE_get_last_error();
 
+/*! \brief LITE is enable ipc debug mode or not.
+ * \return if enable ipc debug mode, will return 1, if not will return 0
+ */
+LITE_API int LITE_is_enable_ipc_debug_mode();
+
+/*! \brief LITE enable ipc debug mode.
+ */
+LITE_API void LITE_enable_lite_ipc_debug();
+
 /*! \brief Get device count
  * \param[in] device_type device type
  * \return the device count

lite/lite-c/include/lite-c/tensor_c.h

@@ -158,6 +158,15 @@ LITE_API int LITE_tensor_share_memory_with(
  */
 LITE_API int LITE_get_tensor_memory(const LiteTensor tensor, void** data);
 
+/**
+ * \brief copy tensor memory if fork debug mode.
+ * \param[server_ptr] the ptr valid in lite server
+ * \param[client_ptr] the ptr only valid in lite client
+ * \param[size_in_byte] copy size
+ */
+LITE_API int LITE_copy_server_tensor_memory(
+        void* server_ptr, void* client_ptr, size_t size_in_byte);
+
 /**
  * \brief get the memory pointer of a Tensor object.
  * \param[in] tensor The input Tensor
@@ -230,6 +239,11 @@ LITE_API int LITE_tensor_concat(
         LiteTensor* tensors, int nr_tensor, int dim, LiteDeviceType dst_device,
         int device_id, LiteTensor* result_tensor);
 
+/**
+ * \brief fuction like memset
+ */
+LITE_API void* LITE_memset(void* s, int c, size_t n);
+
 #ifdef __cplusplus
 }
 #endif

lite/lite-c/src/global.cpp

@@ -1,5 +1,6 @@
 #include "lite/global.h"
 #include "common.h"
+#include "ipc_helper.h"
 #include "lite-c/global_c.h"
 
 namespace {
@@ -47,7 +48,24 @@ void LITE_clear_last_error() {
 
 const char* LITE_get_last_error() {
     LITE_LOCK_GUARD(mtx_error);
-    return get_global_error().get_error_msg().c_str();
+    if (ipc_imp::is_server()) {
+        return get_global_error().get_error_msg().c_str();
+    } else {
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_GET_LAST_ERROR);
+
+        char* ret_ptr = static_cast<char*>(raw_shm_ptr);
+        return ret_ptr;
+    }
+}
+
+int LITE_is_enable_ipc_debug_mode() {
+    return ipc::IpcHelper::is_enable_fork_debug_mode();
+}
+
+void LITE_enable_lite_ipc_debug() {
+    ipc_imp::enable_lite_ipc_debug();
 }
 
 int LITE_get_version(int* major, int* minor, int* patch) {

lite/lite-c/src/ipc_helper.cpp

@@ -0,0 +1,260 @@
+#include "ipc_helper.h"
+
+#include "lite-c/global_c.h"
+#include "lite-c/network_c.h"
+
+#include "misc.h"
+
+using namespace ipc_imp;
+namespace ipc {
+
+static void api_remote_call_cb(struct MsgBody* msg) {
+    LITE_DEBUG(
+            "into %s: %d remote_func_id: %zu", __func__, __LINE__, msg->remote_func_id);
+    switch (static_cast<RemoteFuncId>(msg->remote_func_id)) {
+        case RemoteFuncId::LITE_MAKE_NETWORK: {
+            LiteNetwork network;
+            //! second args is const LiteConfig config
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteConfig config;
+            memcpy(&config, shm_ptr_c, sizeof(LiteConfig));
+            //! third args is network_io
+            LiteNetworkIO network_io;
+            memcpy(&network_io, shm_ptr_c + sizeof(LiteConfig), sizeof(LiteNetworkIO));
+
+            int ret = LITE_make_network(&network, config, network_io);
+
+            //! API is block, put ret to shm_ptr
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+            ret_ptr++;
+            void* network_p = static_cast<void*>(ret_ptr);
+            memcpy(network_p, &network, sizeof(LiteNetwork));
+        }; break;
+        case RemoteFuncId::LITE_LOAD_MODEL_FROM_PATH: {
+            LiteNetwork network;
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            memcpy(&network, shm_ptr_c, sizeof(LiteNetwork));
+
+            int ret =
+                    LITE_load_model_from_path(network, shm_ptr_c + sizeof(LiteNetwork));
+
+            //! API is block, put ret to shm_ptr
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+        }; break;
+        case RemoteFuncId::LITE_GET_LAST_ERROR: {
+            auto shm_size = base_get_shm_size();
+
+            const char* ret = LITE_get_last_error();
+            char* ret_ptr = static_cast<char*>(msg->shm_ptr);
+
+            auto last_error_str_len = strlen(ret) + 1;
+            ASSERT_SHM_SIZE(shm_size, last_error_str_len);
+            strcpy(ret_ptr, ret);
+        }; break;
+        case RemoteFuncId::LITE_GET_IO_TENSOR: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteNetwork network;
+            LiteTensorPhase phase;
+            LiteTensor tensor;
+            memcpy(&network, shm_ptr_c, sizeof(LiteNetwork));
+            memcpy(&phase, shm_ptr_c + sizeof(LiteNetwork), sizeof(LiteTensorPhase));
+
+            int ret = LITE_get_io_tensor(
+                    network, shm_ptr_c + sizeof(LiteNetwork) + sizeof(LiteTensorPhase),
+                    phase, &tensor);
+
+            //! API is block, put ret to shm_ptr
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+            ret_ptr++;
+            void* lite_tensor_p = static_cast<void*>(ret_ptr);
+            memcpy(lite_tensor_p, &tensor, sizeof(LiteTensor));
+        }; break;
+        case RemoteFuncId::LITE_GET_TENSOR_TOTAL_SIZE_IN_BYTE: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteTensor tensor;
+            size_t size;
+            memcpy(&tensor, shm_ptr_c, sizeof(LiteTensor));
+
+            int ret = LITE_get_tensor_total_size_in_byte(tensor, &size);
+
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+            ret_ptr++;
+            memcpy(ret_ptr, &size, sizeof(size_t));
+        }; break;
+        case RemoteFuncId::LITE_GET_TENSOR_MEMORY: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteTensor tensor;
+            void* data;
+            memcpy(&tensor, shm_ptr_c, sizeof(LiteTensor));
+
+            int ret = LITE_get_tensor_memory(tensor, &data);
+
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+            ret_ptr++;
+            memcpy(ret_ptr, &data, sizeof(void*));
+        }; break;
+        case RemoteFuncId::LITE_MEMSET: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            void* s;
+            int c;
+            size_t n;
+            memcpy(&s, shm_ptr_c, sizeof(void*));
+            memcpy(&c, shm_ptr_c + sizeof(void*), sizeof(int));
+            memcpy(&n, shm_ptr_c + sizeof(void*) + sizeof(int), sizeof(size_t));
+            LITE_memset(s, c, n);
+        }; break;
+        case RemoteFuncId::LITE_FORWARD: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteNetwork network;
+            memcpy(&network, shm_ptr_c, sizeof(LiteNetwork));
+
+            int ret = LITE_forward(network);
+
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+        }; break;
+        case RemoteFuncId::LITE_WAIT: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteNetwork network;
+            memcpy(&network, shm_ptr_c, sizeof(LiteNetwork));
+
+            int ret = LITE_wait(network);
+
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+        }; break;
+        case RemoteFuncId::LITE_GET_OUTPUT_NAME: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteNetwork network;
+            size_t index;
+            const char* name;
+            memcpy(&network, shm_ptr_c, sizeof(LiteNetwork));
+            memcpy(&index, shm_ptr_c + sizeof(LiteNetwork), sizeof(size_t));
+
+            int ret = LITE_get_output_name(network, index, &name);
+            auto output_name_len = strlen(name) + 1;
+            auto shm_size = base_get_shm_size();
+            ASSERT_SHM_SIZE(shm_size, output_name_len);
+
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+            ret_ptr++;
+            void* p = static_cast<void*>(ret_ptr);
+            char* p_c = static_cast<char*>(p);
+            strcpy(p_c, name);
+        }; break;
+        case RemoteFuncId::LITE_COPY_SERVER_TENSOR_MEMORY: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            void* server_ptr;
+            size_t size_in_byte;
+            memcpy(&server_ptr, shm_ptr_c, sizeof(void*));
+            memcpy(&size_in_byte, shm_ptr_c + sizeof(void*), sizeof(size_t));
+            memcpy(shm_ptr_c, server_ptr, size_in_byte);
+        }; break;
+        case RemoteFuncId::LITE_DESTROY_NETWORK: {
+            char* shm_ptr_c = static_cast<char*>(msg->shm_ptr);
+            LiteNetwork network;
+            memcpy(&network, shm_ptr_c, sizeof(LiteNetwork));
+
+            int ret = LITE_destroy_network(network);
+
+            int* ret_ptr = static_cast<int*>(msg->shm_ptr);
+            *ret_ptr = ret;
+        }; break;
+        default:
+            LITE_THROW("code issue happened: do not handle RemoteFuncId");
+    }
+};
+bool IpcHelper::sm_is_enable_fork_ipc = false;
+IpcHelper::IpcHelper() {
+    if (!is_server()) {
+        sm_is_enable_fork_ipc = is_enable_ipc();
+        auto shm_mem_conut = base_get_shm_count();
+        shm_mem_for_null_consumer_ptr = base_get_shm_ptr(0);
+        LITE_ASSERT(
+                shm_mem_for_null_consumer_ptr, "invalid shm_ptr: %p",
+                shm_mem_for_null_consumer_ptr);
+
+        for (size_t i = 1; i < shm_mem_conut; i++) {
+            void* shm_ptr = base_get_shm_ptr(i);
+            LITE_ASSERT(shm_ptr, "invalid shm_ptr: %p", shm_ptr);
+            //! init consumer ptr as nullptr
+            m_shm_ptr2consumer_ptr[shm_ptr] = nullptr;
+        }
+
+        shm_size = base_get_shm_size();
+        register_remote_call_cb(&api_remote_call_cb);
+    } else {
+        LITE_THROW("code issue happened: can not use for client");
+    }
+};
+
+IpcHelper::~IpcHelper() {
+    struct MsgBody msg;
+    msg.type = IPC_SERVER_EXIT;
+    send_ipc_msg(&msg);
+    join_server();
+};
+
+void* IpcHelper::get_shm_ptr(void* consumer_ptr) {
+    LITE_LOCK_GUARD(m_mtx);
+
+    if (!consumer_ptr) {
+        return shm_mem_for_null_consumer_ptr;
+    }
+
+    void* ret = nullptr;
+    //! try find old one
+    for (auto&& i : m_shm_ptr2consumer_ptr) {
+        if (consumer_ptr == i.second) {
+            ret = i.first;
+            break;
+        }
+    }
+
+    //! if not find, try alloc a new one
+    if (!ret) {
+        for (auto&& i : m_shm_ptr2consumer_ptr) {
+            if (nullptr == i.second) {
+                i.second = consumer_ptr;
+                ret = i.first;
+                break;
+            }
+        }
+    }
+
+    if (!ret) {
+        LITE_ERROR(
+                "can not find any usable shm_mem, may config LITE_DEBUG_IPC_SHM_COUNT "
+                "big "
+                "than :%zu",
+                m_shm_ptr2consumer_ptr.size() + 1);
+        LITE_ASSERT(ret);
+    }
+
+    return ret;
+};
+
+void IpcHelper::release_shm_ptr(void* consumer_ptr) {
+    LITE_LOCK_GUARD(m_mtx);
+
+    LITE_ASSERT(consumer_ptr, "invalid consumer_ptr ptr");
+    for (auto&& i : m_shm_ptr2consumer_ptr) {
+        if (consumer_ptr == i.second) {
+            //! release use of shm_mem, then other consumer can use it
+            i.second = nullptr;
+            return;
+        }
+    }
+
+    LITE_THROW(
+            "error happened!!, can not find any consumer_ptr in "
+            "m_shm_ptr2consumer_ptr");
+};
+
+}  // namespace ipc

lite/lite-c/src/ipc_helper.h

@@ -0,0 +1,101 @@
+#pragma once
+
+#include <cstdio>
+#include <cstring>
+#include <map>
+#include <string>
+#include <unordered_map>
+
+#include "ipc_imp.h"
+
+#define IPC_INSTACE() ipc::IpcHelper::Instance()
+
+#define IPC_HELP_REMOTE_CALL(SHM_PTR, REMOTEFUNCID)         \
+    struct ipc_imp::MsgBody msg;                            \
+    msg.type = ipc_imp::IPC_CALL_REMOTE_API;                \
+    msg.shm_ptr = static_cast<void*>(SHM_PTR);              \
+    msg.remote_func_id = static_cast<size_t>(REMOTEFUNCID); \
+    IPC_INSTACE().send_ipc_msg(&msg);
+
+#define ASSERT_SHM_SIZE(SHM_SIZE, NEED_SIZE)                                        \
+    do {                                                                            \
+        if (SHM_SIZE < NEED_SIZE) {                                                 \
+            LITE_ERROR(                                                             \
+                    "shm_size not enough to run this api need vs config: (%fMB "    \
+                    "%fMB), please config it by env: LITE_DEBUG_IPC_SHM_SIZE, for " \
+                    "example config to 20MB by: export LITE_DEBUG_IPC_SHM_SIZE=20", \
+                    NEED_SIZE / 1024.0 / 1024.0, SHM_SIZE / 1024.0 / 1024.0);       \
+            __builtin_trap();                                                       \
+        }                                                                           \
+    } while (0)
+
+namespace ipc {
+
+template <class T>
+class Singleton {
+public:
+    Singleton() {}
+    static T& Instance() {
+        static T _;
+        return _;
+    }
+};
+
+class IpcHelper : public Singleton<IpcHelper> {
+public:
+    IpcHelper(const IpcHelper&) = delete;
+    IpcHelper& operator=(const IpcHelper&) = delete;
+    IpcHelper();
+
+    ~IpcHelper();
+
+    //! send msg with default timeout
+    struct ipc_imp::MsgBody send_ipc_msg(struct ipc_imp::MsgBody* msg) {
+        return send_msg(msg, &tv);
+    }
+
+    //! get shm ptr
+    void* get_shm_ptr(void* consumer_ptr);
+
+    //! release shm_ptr, NOT free shm_ptr
+    void release_shm_ptr(void* consumer_ptr);
+
+    //! check shm size
+    void check_shm_size(size_t need_size) { ASSERT_SHM_SIZE(shm_size, need_size); }
+
+    //! is enable ipc fork debug mode
+    static bool is_enable_fork_debug_mode() { return sm_is_enable_fork_ipc; }
+
+    static bool sm_is_enable_fork_ipc;
+
+private:
+    //! 5 minutes
+    struct timeval tv = {300, 0};
+
+    //! map of <shm_ptr, consumer_ptr>,
+    std::map<void*, void*> m_shm_ptr2consumer_ptr;
+
+    size_t shm_size = 0;
+
+    //! shm_mem for consumer_ptr == nullptr
+    void* shm_mem_for_null_consumer_ptr;
+
+    LITE_MUTEX m_mtx;
+};
+
+enum class RemoteFuncId : size_t {
+    LITE_MAKE_NETWORK = 1,
+    LITE_LOAD_MODEL_FROM_PATH = 2,
+    LITE_GET_LAST_ERROR = 3,
+    LITE_GET_IO_TENSOR = 4,
+    LITE_GET_TENSOR_TOTAL_SIZE_IN_BYTE = 5,
+    LITE_GET_TENSOR_MEMORY = 6,
+    LITE_MEMSET = 7,
+    LITE_FORWARD = 8,
+    LITE_WAIT = 9,
+    LITE_GET_OUTPUT_NAME = 10,
+    LITE_COPY_SERVER_TENSOR_MEMORY = 11,
+    LITE_DESTROY_NETWORK = 12,
+};
+
+}  // namespace ipc

lite/lite-c/src/ipc_imp.cpp

@@ -0,0 +1,370 @@
+#include "ipc_imp.h"
+
+#if __linux__
+#include <sys/prctl.h>
+#include <sys/wait.h>
+#endif
+
+#ifdef __ANDROID__
+#include <android/log.h>
+#include <sys/system_properties.h>
+#endif
+
+namespace ipc_imp {
+
+namespace {
+
+//! max count if shm
+#define MAX_SHM_COUNT 15
+struct ServerConfig {
+#ifdef _LITE_SUPPORT_IPC
+    pid_t server_id;
+#else
+    size_t server_id;
+#endif
+    void* cb;
+    int fd_s[2];
+    int fd_c[2];
+    fd_set select_s;
+    fd_set select_c;
+    void* shm_ptr[MAX_SHM_COUNT];
+    size_t shm_mem_conut;
+    //! all shm use the same shm_size
+    size_t shm_size;
+};
+
+static LITE_MUTEX m_mtx;
+
+static ServerConfig server_config;
+
+#ifdef _LITE_SUPPORT_IPC
+static size_t config_shm_memory() {
+    //! default config to 10MB
+    size_t shm_size = 10 * 1024 * 1024;
+    //! env to config LITE_DEBUG_IPC_SHM_SIZE
+    //! for example , export LITE_DEBUG_IPC_SHM_SIZE=20, means config SHM size 20MB
+    if (auto env = ::std::getenv("LITE_DEBUG_IPC_SHM_SIZE"))
+        shm_size = std::stoi(env) * 1024 * 1024;
+
+#ifdef __ANDROID__
+    //! special for Android prop, attention: getprop may need permission
+    char buf[PROP_VALUE_MAX];
+    if (__system_property_get("LITE_DEBUG_IPC_SHM_SIZE", buf) > 0) {
+        shm_size = std::stoi(buf) * 1024 * 1024;
+    }
+#endif
+
+    return shm_size;
+}
+
+//! FIXME: detail at create_server(), at this stage, we only support enable lite fork
+//! debug by: LITE_enable_lite_ipc_debug, after issue fix, may support env:
+//! LITE_ENABLE_C_API_WITH_FORK_MODE
+// bool config_enable_debug_fork() {
+//    //! debug off, we only support enable fork debug mode with env
+//    //! LITE_ENABLE_C_API_WITH_FORK_MODE, not support api to config
+//    //! as we will fork as soon as possible by __attribute__((constructor)),
+//    //! user may not have to chance to call any normal api before it
+//    bool ret = false;
+//    //! env to config LITE_ENABLE_C_API_WITH_FORK_MODE
+//    //! for example , export LITE_ENABLE_C_API_WITH_FORK_MODE=1, means enable LITE c
+//    api
+//    //! with fork mode
+//    if (auto env = ::std::getenv("LITE_ENABLE_C_API_WITH_FORK_MODE")) {
+//        if (std::stoi(env) > 0) {
+//            ret = true;
+//        }
+//    }
+//
+//#ifdef __ANDROID__
+//    //! special for Android prop, attention: getprop may need permission
+//    char buf[PROP_VALUE_MAX];
+//    if (__system_property_get("LITE_ENABLE_C_API_WITH_FORK_MODE", buf) > 0) {
+//        ret = std::stoi(buf);
+//        if (std::stoi(buf) > 0) {
+//            ret = true;
+//        }
+//    }
+//#endif
+//
+//    return ret;
+//}
+#endif
+
+static bool is_enable_debug_fork = false;
+
+//! internal recycle server when IPC_ASSERT happen
+static void recycle_server() {
+    static struct timeval tv = {1, 0};
+    struct MsgBody msg;
+    msg.type = IPC_SERVER_EXIT;
+    if (server_config.server_id > 0) {
+        send_msg(&msg, &tv);
+    }
+}
+
+#define ipc_unlikely(v) __builtin_expect((v), 0)
+#define IPC_ASSERT(expr, msg)                                                          \
+    do {                                                                               \
+        if (ipc_unlikely(!(expr))) {                                                   \
+            LITE_ERROR("ipc fatal error: assert failed: %s with msg: %s", #expr, msg); \
+            recycle_server();                                                          \
+            __builtin_trap();                                                          \
+        }                                                                              \
+    } while (0)
+
+#ifdef _LITE_SUPPORT_IPC
+static size_t config_shm_memory_count() {
+    //! default config to 2
+    size_t shm_cnt = 2;
+    //! env to config LITE_DEBUG_IPC_SHM_COUNT
+    //! for example , export LITE_DEBUG_IPC_SHM_COUNT=8, means config SHM count 8
+    if (auto env = ::std::getenv("LITE_DEBUG_IPC_SHM_COUNT"))
+        shm_cnt = std::stoi(env);
+
+#ifdef __ANDROID__
+    //! special for Android prop, attention: getprop may need permission
+    char buf[PROP_VALUE_MAX];
+    if (__system_property_get("LITE_DEBUG_IPC_SHM_COUNT", buf) > 0) {
+        shm_cnt = std::stoi(buf);
+    }
+#endif
+    IPC_ASSERT(
+            shm_cnt >= 2 && shm_cnt <= MAX_SHM_COUNT,
+            "error config LITE_DEBUG_IPC_SHM_COUNT, should be range of [2, "
+            "MAX_SHM_COUNT]");
+
+    return shm_cnt;
+}
+#endif
+
+#ifdef _LITE_SUPPORT_IPC
+static void handle_remote_call(struct MsgBody* msg) {
+    LITE_DEBUG("into %s: %d", __func__, __LINE__);
+    IPC_ASSERT(
+            server_config.cb,
+            "handle_remote_call failed: can not find valid "
+            "remote_call_cb, please call "
+            "register_remote_call_cb firstly!!");
+    remote_call_cb cb = (remote_call_cb)server_config.cb;
+    cb(msg);
+}
+
+static void* ipc_mmap(
+        void* addr, size_t length, int prot, int flags, int fd, off_t offset) {
+    void* ret = mmap(addr, length, prot, flags, fd, offset);
+    IPC_ASSERT(ret != MAP_FAILED, "call mmap failed");
+    return ret;
+}
+
+static int ipc_munmap(void* addr, size_t length) {
+    int ret = munmap(addr, length);
+    IPC_ASSERT(0 == ret, "call munmap failed");
+    return ret;
+}
+#endif
+
+//! start server as soon as possible
+//! FIXME: when use __attribute__((constructor)) on clang, will init before all
+//! static class object, which will lead to flatbuffer runtime issue, env config
+//! with init_priority, do not take effect on diff cpp src file on clang
+// static __attribute__((constructor)) void create_server() {
+void create_server() {
+#ifdef _LITE_SUPPORT_IPC
+    LITE_LOCK_GUARD(m_mtx);
+    LITE_LOG("try to config lite fork debug model");
+    if (is_enable_debug_fork)
+        return;
+
+    is_enable_debug_fork = true;
+    //! is_enable_debug_fork = config_enable_debug_fork();
+    //! init default server_config
+    server_config.server_id = 0;
+
+    if (!is_enable_debug_fork)
+        return;
+
+    server_config.cb = nullptr;
+    server_config.shm_size = config_shm_memory();
+    server_config.shm_mem_conut = config_shm_memory_count();
+
+    for (size_t i = 0; i < server_config.shm_mem_conut; i++) {
+        server_config.shm_ptr[i] = ipc_mmap(
+                NULL, server_config.shm_size, PROT_READ | PROT_WRITE,
+                MAP_SHARED | MAP_ANON, -1, 0);
+    }
+
+    IPC_ASSERT(-1 != pipe(server_config.fd_s), "create server pipe failed");
+    IPC_ASSERT(-1 != pipe(server_config.fd_c), "create client pipe failed");
+
+    FD_ZERO(&server_config.select_s);
+    FD_ZERO(&server_config.select_c);
+    //! config server and client
+    FD_SET(server_config.fd_s[0], &server_config.select_s);
+    FD_SET(server_config.fd_c[0], &server_config.select_c);
+
+    server_config.server_id = fork();
+
+    IPC_ASSERT(server_config.server_id >= 0, "call fork failed");
+
+    if (server_config.server_id > 0) {
+        LITE_LOG("create lite_ipc_server success pid is: %d", server_config.server_id);
+    } else {
+        std::string server_name = "lite_ipc_server";
+        // TODO: __APPLE__ do not support PR_SET_NAME and PR_SET_PDEATHSIG
+        // if caller crash, no have chance to send IPC_SERVER_EXIT
+#if __linux__
+        LITE_LOG("start server with name: %s....", server_name.c_str());
+        prctl(PR_SET_NAME, (unsigned long)server_name.c_str(), 0, 0, 0);
+        //! auto die if father crash
+        prctl(PR_SET_PDEATHSIG, SIGKILL);
+#else
+        LITE_LOG("start server....");
+#endif
+
+        while (1) {
+            LITE_DEBUG("lite_ipc_server wait msg now.....");
+            int res =
+                    select(server_config.fd_s[0] + 1, &server_config.select_s, NULL,
+                           NULL, NULL);
+
+            IPC_ASSERT(
+                    res > 0,
+                    "select issue happened or timeout(but we do not support timeout)");
+
+            struct MsgBody msg;
+            size_t r_size = read(server_config.fd_s[0], &msg, sizeof(msg));
+            IPC_ASSERT(r_size == sizeof(msg), "broken pipe msg");
+
+            struct MsgBody response;
+            response.type = IPC_SERVER_RESPONSE;
+            switch (msg.type) {
+                case IPC_CALL_REMOTE_API:
+                    LITE_DEBUG("handle remote call");
+                    handle_remote_call(&msg);
+                    break;
+                case IPC_CONFIG_REMOTE_HANDLE_API:
+                    LITE_DEBUG("handle register remote cb");
+                    server_config.cb = msg.cb;
+                    break;
+                default:
+                    IPC_ASSERT(IPC_SERVER_EXIT == msg.type, "code issue happened!!");
+            }
+
+            size_t w_size = write(server_config.fd_c[1], &response, sizeof(response));
+            IPC_ASSERT(w_size == sizeof(response), "write pip failed");
+
+            if (IPC_SERVER_EXIT == msg.type) {
+                LITE_DEBUG("handle exit now");
+                for (size_t i = 0; i < server_config.shm_mem_conut; i++) {
+                    ipc_munmap(server_config.shm_ptr[i], server_config.shm_size);
+                }
+                exit(0);
+            }
+        }
+    }
+#else
+    //! TODO: imp for Windows with CreateProcess
+    server_config.server_id = 0;
+    LITE_ERROR("lite do not support fork debug ipc on this PLATFORM");
+    __builtin_trap();
+
+    return;
+#endif
+}
+
+}  // namespace
+//////////////////////////////////////////////// api imp /////////////////////////
+void register_remote_call_cb(remote_call_cb cb) {
+    IPC_ASSERT(!server_config.cb, "already register remote_call_cb");
+    IPC_ASSERT(cb, "invalid remote_call_cb");
+    IPC_ASSERT(server_config.server_id, "register cb need server already up");
+
+    server_config.cb = (void*)cb;
+    static struct timeval tv = {5, 0};
+    struct MsgBody msg;
+    msg.type = IPC_CONFIG_REMOTE_HANDLE_API;
+    msg.cb = (void*)cb;
+    send_msg(&msg, &tv);
+}
+
+struct MsgBody send_msg(struct MsgBody* msg, struct timeval* timeout) {
+#ifdef _LITE_SUPPORT_IPC
+    IPC_ASSERT(server_config.server_id > 0, "server not ready");
+    if (IPC_CALL_REMOTE_API == msg->type) {
+        IPC_ASSERT(
+                server_config.cb,
+                "can not find valid remote_call_cb, please "
+                "call register_remote_call_cb firstly!!");
+    }
+
+    //! send msg to server
+    size_t w_size = write(server_config.fd_s[1], msg, sizeof(struct MsgBody));
+    IPC_ASSERT(w_size == sizeof(struct MsgBody), "write pipe failed");
+
+    //! now wait server response
+    struct MsgBody response;
+    LITE_DEBUG("wait server response");
+
+    int res = select(
+            server_config.fd_c[0] + 1, &server_config.select_c, NULL, NULL, timeout);
+    if (0 == res) {
+        LITE_ERROR("wait server timeout");
+    }
+    IPC_ASSERT(res > 0, "select issue happened or timeout");
+
+    size_t r_size = read(server_config.fd_c[0], &response, sizeof(response));
+    IPC_ASSERT(sizeof(response) == r_size, "broken pipe msg");
+    IPC_ASSERT(IPC_SERVER_RESPONSE == response.type, "error server response type");
+
+    return response;
+#else
+    struct MsgBody response;
+    LITE_ERROR("This code should not be called");
+    __builtin_trap();
+
+    return response;
+#endif
+}
+
+bool is_server() {
+    return !server_config.server_id;
+}
+
+bool is_enable_ipc() {
+    return is_enable_debug_fork;
+}
+
+void join_server() {
+#ifdef _LITE_SUPPORT_IPC
+    if (!is_enable_debug_fork)
+        return;
+
+    int ret;
+    waitpid(server_config.server_id, &ret, 0);
+    if (ret) {
+        //! when server crash, we mark server_config.server_id to 0
+        //! to prevent handle more msg, for example recycle_server
+        server_config.server_id = 0;
+    }
+    IPC_ASSERT(
+            ret == 0, "child process exit !zero, please check server log for detail");
+#endif
+}
+
+void* base_get_shm_ptr(size_t index) {
+    return server_config.shm_ptr[index];
+}
+
+size_t base_get_shm_count() {
+    return server_config.shm_mem_conut;
+}
+
+size_t base_get_shm_size() {
+    return server_config.shm_size;
+}
+
+void enable_lite_ipc_debug() {
+    create_server();
+}
+}  // namespace ipc_imp

lite/lite-c/src/ipc_imp.h

@@ -0,0 +1,63 @@
+#pragma once
+
+#undef _LITE_SUPPORT_IPC
+#if __linux__ || __unix__ || __APPLE__
+#define _LITE_SUPPORT_IPC
+#endif
+
+#ifdef _LITE_SUPPORT_IPC
+#include <sys/mman.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#endif
+
+#include "misc.h"
+
+namespace ipc_imp {
+//! server call api ret
+enum MsgType {
+    IPC_SERVER_RESPONSE = 1,
+    IPC_CALL_REMOTE_API = 2,
+    IPC_SERVER_EXIT = 3,
+    IPC_CONFIG_REMOTE_HANDLE_API = 4,
+};
+
+struct MsgBody {
+    enum MsgType type;
+    //! remote call handle callback
+    void* cb;
+    //! remote call function emum, define by user
+    size_t remote_func_id;
+    //! mmap region ptr
+    void* shm_ptr;
+};
+
+//! block wait server return response msg
+struct MsgBody send_msg(struct MsgBody* msg, struct timeval* timeout);
+
+//! wait server exit
+void join_server();
+
+typedef void (*remote_call_cb)(struct MsgBody* msg);
+
+//! register remote call
+void register_remote_call_cb(remote_call_cb cb);
+
+//! is server or not, server or do not enable ipc mode will return true
+bool is_server();
+
+//! is enable ipc or not
+bool is_enable_ipc();
+
+//! get shm ptr
+void* base_get_shm_ptr(size_t index);
+
+//! get shm count
+size_t base_get_shm_count();
+
+// get shm size
+size_t base_get_shm_size();
+
+// enable fork ipc debug
+void enable_lite_ipc_debug();
+}  // namespace ipc_imp

lite/lite-c/src/network.cpp

@@ -1,5 +1,6 @@
 #include "lite/network.h"
 #include "common.h"
+#include "ipc_helper.h"
 #include "lite-c/network_c.h"
 
 #include "../../src/network_impl_base.h"
@@ -204,11 +205,32 @@ int LITE_make_network(
         LiteNetwork* network, const LiteConfig config, const LiteNetworkIO network_io) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
-    auto lite_network = std::make_shared<lite::Network>(
-            convert_to_lite_config(config), convert_to_lite_io(network_io));
-    LITE_LOCK_GUARD(mtx_network);
-    get_gloabl_network_holder()[lite_network.get()] = lite_network;
-    *network = lite_network.get();
+    if (ipc_imp::is_server()) {
+        auto lite_network = std::make_shared<lite::Network>(
+                convert_to_lite_config(config), convert_to_lite_io(network_io));
+        LITE_LOCK_GUARD(mtx_network);
+        get_gloabl_network_holder()[lite_network.get()] = lite_network;
+        *network = lite_network.get();
+    } else {
+        size_t need_size = sizeof(LiteConfig) + sizeof(LiteNetworkIO);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        //! second args is const LiteConfig config
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &config, sizeof(LiteConfig));
+        //! third args is network_io
+        memcpy(shm_ptr_c + sizeof(LiteNetworkIO), &network_io, sizeof(LiteNetworkIO));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_MAKE_NETWORK);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        memcpy(network, ret_ptr, sizeof(LiteNetwork));
+        return ret;
+    }
     LITE_CAPI_END();
 }
 
@@ -234,17 +256,53 @@ int LITE_load_model_from_path(LiteNetwork network, const char* model_path) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
     LITE_ASSERT(model_path, "The model path pass to LITE api is null");
-    static_cast<lite::Network*>(network)->load_model(model_path);
+    if (ipc_imp::is_server()) {
+        static_cast<lite::Network*>(network)->load_model(model_path);
+    } else {
+        size_t need_size = sizeof(LiteNetwork) + strlen(model_path) + 1;
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(network);
+
+        //! first args is LiteNetwork network
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &network, sizeof(LiteNetwork));
+        //! second args is const char* model_path
+        strcpy(shm_ptr_c + sizeof(LiteNetwork), model_path);
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_LOAD_MODEL_FROM_PATH);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        return ret;
+    }
     LITE_CAPI_END();
 }
 
 int LITE_destroy_network(LiteNetwork network) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
-    LITE_LOCK_GUARD(mtx_network);
-    auto& global_holder = get_gloabl_network_holder();
-    if (global_holder.find(network) != global_holder.end()) {
-        global_holder.erase(network);
+    if (ipc_imp::is_server()) {
+        LITE_LOCK_GUARD(mtx_network);
+        auto& global_holder = get_gloabl_network_holder();
+        if (global_holder.find(network) != global_holder.end()) {
+            global_holder.erase(network);
+        }
+    } else {
+        size_t need_size = sizeof(LiteNetwork);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(network);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &network, sizeof(LiteNetwork));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_DESTROY_NETWORK);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        return ret;
     }
     LITE_CAPI_END();
 }
@@ -252,14 +310,48 @@ int LITE_destroy_network(LiteNetwork network) {
 int LITE_forward(const LiteNetwork network) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
-    static_cast<lite::Network*>(network)->forward();
+    if (ipc_imp::is_server()) {
+        static_cast<lite::Network*>(network)->forward();
+    } else {
+        size_t need_size = sizeof(LiteNetwork);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(network);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &network, sizeof(LiteNetwork));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_FORWARD);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        return ret;
+    }
     LITE_CAPI_END();
 }
 
 int LITE_wait(const LiteNetwork network) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
-    static_cast<lite::Network*>(network)->wait();
+    if (ipc_imp::is_server()) {
+        static_cast<lite::Network*>(network)->wait();
+    } else {
+        size_t need_size = sizeof(LiteNetwork);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(network);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &network, sizeof(LiteNetwork));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_WAIT);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        return ret;
+    }
     LITE_CAPI_END();
 }
 
@@ -268,9 +360,31 @@ int LITE_get_io_tensor(
         LiteTensor* tensor) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
-    auto io_tensor =
-            static_cast<lite::Network*>(network)->get_io_tensor(io_name, phase);
-    *tensor = io_tensor.get();
+    if (ipc_imp::is_server()) {
+        auto io_tensor =
+                static_cast<lite::Network*>(network)->get_io_tensor(io_name, phase);
+        *tensor = io_tensor.get();
+    } else {
+        size_t need_size =
+                sizeof(LiteNetwork) + strlen(io_name) + 1 + sizeof(LiteTensorPhase);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(network);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &network, sizeof(LiteNetwork));
+        memcpy(shm_ptr_c + sizeof(LiteNetwork), &phase, sizeof(LiteTensorPhase));
+        strcpy(shm_ptr_c + sizeof(LiteNetwork) + sizeof(LiteTensorPhase), io_name);
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_GET_IO_TENSOR);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        memcpy(tensor, ret_ptr, sizeof(LiteTensor));
+        IPC_INSTACE().release_shm_ptr(network);
+        return ret;
+    }
     LITE_CAPI_END();
 }
 
@@ -286,8 +400,31 @@ int LITE_get_output_name(const LiteNetwork network, size_t index, const char** n
     LITE_CAPI_BEGIN();
     LITE_ASSERT(network, "The network pass to LITE api is null");
     LITE_ASSERT(name, "The name ptr pass to LITE api is null");
-    *name = lite::NetworkHelper::implement(static_cast<lite::Network*>(network))
-                    ->get_output_name(index);
+    if (ipc_imp::is_server()) {
+        *name = lite::NetworkHelper::implement(static_cast<lite::Network*>(network))
+                        ->get_output_name(index);
+    } else {
+        size_t need_size = sizeof(LiteNetwork) + sizeof(index);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        //! warn: we use get_shm_ptr with nullptr, not with network
+        //! so caller need consume this ptr as soon as possible
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &network, sizeof(LiteNetwork));
+        memcpy(shm_ptr_c + sizeof(LiteNetwork), &index, sizeof(size_t));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_GET_OUTPUT_NAME);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        void* p = static_cast<void*>(ret_ptr);
+        char* p_c = static_cast<char*>(p);
+        *name = p_c;
+        return ret;
+    }
     LITE_CAPI_END();
 }
 

lite/lite-c/src/tensor.cpp

@@ -4,6 +4,7 @@
 #include <unordered_map>
 #include "../../src/tensor_impl_base.h"
 #include "common.h"
+#include "ipc_helper.h"
 #include "lite-c/tensor_c.h"
 
 const LiteLayout default_layout = {
@@ -166,7 +167,70 @@ int LITE_get_tensor_memory(const LiteTensor tensor, void** data) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(tensor, "The tensor pass to LITE c_api is null");
     LITE_ASSERT(data, "The data ptr pass to LITE c_api is null");
-    *data = static_cast<lite::Tensor*>(tensor)->get_memory_ptr();
+    if (ipc_imp::is_server()) {
+        *data = static_cast<lite::Tensor*>(tensor)->get_memory_ptr();
+    } else {
+        size_t need_size = sizeof(LiteTensor);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &tensor, sizeof(LiteTensor));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_GET_TENSOR_MEMORY);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        memcpy(data, ret_ptr, sizeof(void*));
+        return ret;
+    }
+    LITE_CAPI_END();
+}
+
+void* LITE_memset(void* s, int c, size_t n) {
+    if (ipc_imp::is_server()) {
+        return memset(s, c, n);
+    } else {
+        size_t need_size = sizeof(void*) + sizeof(int) + sizeof(size_t);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &s, sizeof(void*));
+        memcpy(shm_ptr_c + sizeof(void*), &c, sizeof(int));
+        memcpy(shm_ptr_c + sizeof(void*) + sizeof(int), &n, sizeof(size_t));
+
+        IPC_HELP_REMOTE_CALL(raw_shm_ptr, ipc::RemoteFuncId::LITE_MEMSET);
+
+        return s;
+    }
+}
+
+int LITE_copy_server_tensor_memory(
+        void* server_ptr, void* client_ptr, size_t size_in_byte) {
+    LITE_CAPI_BEGIN();
+    if (ipc_imp::is_server()) {
+        LITE_ASSERT(
+                false, "lite not in fork debug mode, please do not call this function");
+    } else {
+        size_t need_size = sizeof(void*) + sizeof(size_t);
+        IPC_INSTACE().check_shm_size(need_size);
+        IPC_INSTACE().check_shm_size(size_in_byte);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &server_ptr, sizeof(void*));
+        memcpy(shm_ptr_c + sizeof(void*), &size_in_byte, sizeof(size_t));
+
+        IPC_HELP_REMOTE_CALL(
+                raw_shm_ptr, ipc::RemoteFuncId::LITE_COPY_SERVER_TENSOR_MEMORY);
+        memcpy(client_ptr, raw_shm_ptr, size_in_byte);
+        return 0;
+    }
     LITE_CAPI_END();
 }
 
@@ -186,7 +250,26 @@ int LITE_get_tensor_total_size_in_byte(const LiteTensor tensor, size_t* size) {
     LITE_CAPI_BEGIN();
     LITE_ASSERT(tensor, "The tensor pass to LITE c_api is null");
     LITE_ASSERT(size, "The size ptr pass to LITE c_api is null");
-    *size = static_cast<lite::Tensor*>(tensor)->get_tensor_total_size_in_byte();
+    if (ipc_imp::is_server()) {
+        *size = static_cast<lite::Tensor*>(tensor)->get_tensor_total_size_in_byte();
+    } else {
+        size_t need_size = sizeof(LiteTensor);
+        IPC_INSTACE().check_shm_size(need_size);
+
+        void* raw_shm_ptr = IPC_INSTACE().get_shm_ptr(nullptr);
+
+        char* shm_ptr_c = static_cast<char*>(raw_shm_ptr);
+        memcpy(shm_ptr_c, &tensor, sizeof(LiteTensor));
+
+        IPC_HELP_REMOTE_CALL(
+                raw_shm_ptr, ipc::RemoteFuncId::LITE_GET_TENSOR_TOTAL_SIZE_IN_BYTE);
+
+        int* ret_ptr = static_cast<int*>(raw_shm_ptr);
+        auto ret = *ret_ptr;
+        ret_ptr++;
+        memcpy(size, ret_ptr, sizeof(size_t));
+        return ret;
+    }
     LITE_CAPI_END();
 }
 

lite/src/misc.cpp

@@ -11,6 +11,7 @@
 
 #ifdef __ANDROID__
 #include <android/log.h>
+#include <sys/system_properties.h>
 #endif
 
 using namespace lite;
@@ -18,7 +19,26 @@ using namespace lite;
 namespace lite {
 namespace log_detail {
 
-LiteLogLevel current_log_level = LiteLogLevel::ERROR;
+LiteLogLevel config_default_log_level() {
+    auto default_level = LiteLogLevel::ERROR;
+    //! env to config LogLevel
+    //!  DEBUG = 0, INFO = 1, WARN = 2, ERROR = 3, NO_LOG = 4
+    //! for example , export RUNTIME_OVERRIDE_LOG_LEVEL=0, means set LogLevel to
+    //! DEBUG
+    if (auto env = ::std::getenv("RUNTIME_OVERRIDE_LOG_LEVEL"))
+        default_level = static_cast<LiteLogLevel>(std::stoi(env));
+
+#ifdef __ANDROID__
+    //! special for Android prop, attention: getprop may need permission
+    char buf[PROP_VALUE_MAX];
+    if (__system_property_get("RUNTIME_OVERRIDE_LOG_LEVEL", buf) > 0) {
+        default_level = static_cast<LiteLogLevel>(atoi(buf));
+    }
+#endif
+
+    return default_level;
+}
+LiteLogLevel current_log_level = config_default_log_level();
 
 template <class T, size_t N>
 constexpr size_t countof(T (&)[N]) {