【发布时间】:2017-02-09 10:38:17
【问题描述】:
根据这个答案(https://groups.google.com/forum/#!msg/torch7/rmIcBYCiFG8/IC68Xzd3DgAJ),torch 似乎自动释放了不再使用的张量。我已经遇到了这个问题,张量指针被 lua 自动释放,导致 c 程序出现段错误。但是,从我的实验来看,似乎没有在 C 代码上手动调用 free 会导致内存泄漏,即程序内存不断增加。
这是我的测试代码:
#include <iostream>
#include <chrono>
#include <thread>
#include <lua.hpp>
extern "C" {
#include <TH.h>
#include <luaT.h>
}
using namespace std;
int main(int argc, char** argv)
{
lua_State* L = luaL_newstate();
luaL_openlibs(L);
lua_getglobal(L, "require");
lua_pushstring(L, "torch");
lua_pcall(L,1,0,0); // require "torch"
int h=224, w=224, nb_channel=3;
int len = h * w * nb_channel;
long stride_1 = h * w;
long stride_2 = w;
long stride_3 = 1;
for (size_t i = 0 ; i < 1000000 ; ++i)
{
float *tensorData = (float*)malloc(sizeof(float)*len); // Should be when the storage is freed
THFloatStorage *storage = THFloatStorage_newWithData(tensorData, len); // Memory released when tensor input is freed by lua garbadge collector
THFloatTensor* input = THFloatTensor_newWithStorage3d(storage,
0, // Offset
nb_channel, stride_1, // Channels
h, stride_2, // Height
w, stride_3); // Width
// Do some initialisation of the tensor...
luaT_pushudata(L, (void*) input, "torch.FloatTensor"); // Send the tensor to lua
// Do some stuff with the input... (call lua torch scripts...)
lua_pop(L, 1);
// If those two lines is not set, we get a memory leak
THFloatTensor_free(input); // Will sometimes create a segfault as the tensor seems to be deleted by the lua garbadge collector
THFloatStorage_free(storage);
if (i%10000 == 0)
std::cout << i << std::endl;
std::this_thread::sleep_for(std::chrono::microseconds(1));
}
return 0;
}
如何用c和lua正确管理内存,避免segfault又不造成内存泄漏?
【问题讨论】:
-
那不是 C 代码。
-
我知道,但问题与 c 库(luaT.h、lua.h)有关。示例的 cpp 部分与问题无关。