cuda runtime库不是线程安全的？

因为应用需要cpu侧需要多个线程并发调用一个包含cuda计算的库，发现在运行中有些线程会挂死的现象。后来经过简化发现就算是单独的一个cudaMalloc也会出现线程挂死的问题。测试平台为windows xp，cuda 2.3。
是用法不对还是cuda运行库本身不是线程安全的？

测试代码如下：

#include <stdio.h>

#include <cuda_runtime.h>

#include <cutil_inline.h>

#include <process.h>

#include <windows.h>



#define GRD 45

#define BLK 16

#define SIZE (GRD * GRD * BLK * BLK)



DWORD test(void *tid)

{

    char *src_g;

    char *src_c;



    int id = (int)tid;



    printf("before cudaSetDevice(%d)\n", id);

    cutilSafeCall(cudaSetDevice(0));



    printf("before cudaMalloc(%d)\n", id);

    cutilSafeCall(cudaMalloc((void**) &src_g, SIZE));

    cutilSafeCall(cudaFree(src_g));



    printf("thread %d ok.\n", id);



    return 1;

}



int main(int argc, char *argv[])

{

    HANDLE h1 = CreateThread(0, 0, (LPTHREAD_START_ROUTINE)test, (void*)1, 0, 0);

    HANDLE h2 = CreateThread(0, 0, (LPTHREAD_START_ROUTINE)test, (void*)2, 0, 0);

    test(0);



    // wait other threads

    WaitForSingleObject(h1, INFINITE);

    WaitForSingleObject(h2, INFINITE);

    CloseHandle(h1);

    CloseHandle(h2);



    return 0;

}

上面测试程序经常出现如下打印然后就无法退出了：
before cudaSetDevice(1)
before cudaSetDevice(0)
before cudaSetDevice(2)
before cudaMalloc(1)
before cudaMalloc(2)
before cudaMalloc(0)
thread 1 ok.