Question

我正在尝试分配设备内存，复制到它，在GPU上执行计算，将结果复制回来然后释放我分配的设备内存。我想确保我没有超出限制，我想看看我是否在共享内存空间中有足够的内存来转储一些数组。

当我分配设备内存时，不会返回任何错误。当我使用cudaMemGetInfo检查分配的内存量时，看起来一个cudaMalloc没有分配任何内存。此外，当我尝试释放内存时，看起来只有一个指针被释放。

我正在使用matlab Mexfunction接口来设置GPU内存并启动内核。此时，我甚至没有调用内核，只返回结果的单位矩阵。

cudaError_t cudaErr;
size_t freeMem = 0;
size_t totalMem = 0;
size_t allocMem = 0;
cudaMemGetInfo(&freeMem, &totalMem);  
mexPrintf("Memory avaliable: Free: %lu, Total: %lu\n",freeMem, totalMem);  

/* Pointers for the device memory */
double *devicePulseDelay, *deviceTarDistance, *deviceScattDistance, *deviceScatterers;
double *deviceReceivedReal, *deviceReceivedImag;

/* Allocate memory on the device for the arrays. */
mexPrintf("Allocating memory.\n");
cudaErr = cudaMalloc( (void **) &devicePulseDelay, sizeof(double)*512);
if (cudaErr != cudaSuccess)
{
    mexPrintf("could not allocate memory to devicePulseDelay\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("devicePulseDelay: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMalloc( (void **) &deviceTarDistance, sizeof(double)*512);
if (cudaErr != cudaSuccess)
{
    mexPrintf("could not allocate memory to deviceTarDistance\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceTarDistance: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMalloc( (void **) &deviceScattDistance, sizeof(double)*999*512);
if (cudaErr != cudaSuccess)
{
    mexPrintf("could not allocate memory to deviceScattDistance\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceScattDistance: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMalloc( (void **) &deviceScatterers, sizeof(double)*999);
if (cudaErr != cudaSuccess)
{   
    mexPrintf("could not allocate memory to deviceScatterers\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}  
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceScatterers: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMalloc( (void **) &deviceReceivedReal, sizeof(double)*999*512);
if (cudaErr != cudaSuccess)
{
    mexPrintf("could not allocate memory to deviceReceivedReal\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceReceivedReal: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMalloc( (void **) &deviceReceivedImag, sizeof(double)*999*512);
if (cudaErr != cudaSuccess)
{
    mexPrintf("could not allocate memory to deviceReceivedImag\n");   
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceReceivedImag: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n", allocMem, totalMem,(freeMem - allocMem));

/* copy the input arrays across to the device */
mexPrintf("\nCopying memory.\n");
cudaErr = cudaMemcpy(devicePulseDelay, pulseDelay, sizeof(double)*512,cudaMemcpyHostToDevice);
if (cudaErr != cudaSuccess) 
{
    mexPrintf("could not copy to devicePulseDelay\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("devicePulseDelay: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMemcpy(deviceTarDistance, tarDistance, sizeof(double)*512,cudaMemcpyHostToDevice);
if (cudaErr != cudaSuccess) 
{
    mexPrintf("could not copy to deviceTarDistance\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));   
}   
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceTarDistance: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMemcpy(deviceScattDistance, scattDistance, sizeof(double)*999*512,cudaMemcpyHostToDevice);   
if (cudaErr != cudaSuccess)
{  
    mexPrintf("could not copy to deviceScattDistance\n");  
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));  
} 
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceScattDistance: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMemcpy(deviceScatterers, scatterers, sizeof(double)*999,cudaMemcpyHostToDevice); 
if (cudaErr != cudaSuccess) 
{
    mexPrintf("could not copy to deviceScatterers\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));   
}   
cudaMemGetInfo(&allocMem, &totalMem);  
mexPrintf("deviceScatterers: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));  

/* call the kernel */
// launchKernel<<<1,512>>>(........);   

/* retireve the output */  
cudaErr = cudaMemcpy(receivedReal, deviceReceivedReal, sizeof(double)*512*512,cudaMemcpyDeviceToHost);   
if (cudaErr != cudaSuccess)
{   
    mexPrintf("could not copy to receivedReal\n");  
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));  
}
cudaMemGetInfo(&allocMem, &totalMem);   
mexPrintf("receivedReal: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));
cudaErr = cudaMemcpy(receivedImag, deviceReceivedImag, sizeof(double)*512*512,cudaMemcpyDeviceToHost); 
if (cudaErr != cudaSuccess)
{ 
    mexPrintf("could not copy to receivedImag\n");   
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));   
}   
cudaMemGetInfo(&allocMem, &totalMem); 
mexPrintf("receivedImag: Memory avaliable: Free: %lu, Total: %lu, Consumed: %lu\n",allocMem, totalMem,(freeMem - allocMem));   

/* free the memory. */ 
mexPrintf("\nFree'ing memory.\n");   
cudaMemGetInfo(&freeMem, &totalMem);  
mexPrintf("Before freeing: Free %lu, Total: %lu\n", freeMem, totalMem);  
cudaErr = cudaFree(devicePulseDelay); 
if (cudaErr != cudaSuccess) 
{ 
    mexPrintf("could free devicePulseDelay\n");   
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));  
}   
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("devicePulseDelay: Memory avaliable: Free: %lu, Total: %lu, Free'd: %lu\n",allocMem, totalMem,(allocMem - freeMem));   
cudaErr = cudaFree(deviceTarDistance);   
if (cudaErr != cudaSuccess) 
{
    mexPrintf("could free deviceTarDistance\n");  
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));  
} 
cudaMemGetInfo(&allocMem, &totalMem);   
mexPrintf("deviceTarDistance: Memory avaliable: Free: %lu, Total: %lu, Free'd: %lu\n",allocMem, totalMem,(allocMem - freeMem));  
cudaErr = cudaFree(deviceScattDistance);   
if (cudaErr != cudaSuccess) 
{   
    mexPrintf("could free deviceScattDistance\n"); 
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));
}   
cudaMemGetInfo(&allocMem, &totalMem);   
mexPrintf("deviceScattDistance: Memory avaliable: Free: %lu, Total: %lu, Free'd: %lu\n",allocMem, totalMem,(allocMem - freeMem));  
cudaErr = cudaFree(deviceScatterers);  
if (cudaErr != cudaSuccess) 
{   
    mexPrintf("could free deviceScatterers\n");  
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));   
}   
cudaMemGetInfo(&allocMem, &totalMem);  
mexPrintf("deviceScatterers: Memory avaliable: Free: %lu, Total: %lu, Free'd: %lu\n",allocMem, totalMem,(allocMem - freeMem));  
cudaErr = cudaFree(deviceReceivedReal);  
if (cudaErr != cudaSuccess) 
{  
    mexPrintf("could free deviceReceivedReal\n"); 
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));  
} 
cudaMemGetInfo(&allocMem, &totalMem);  
mexPrintf("deviceReceivedReal: Memory avaliable: Free: %lu, Total: %lu, Free'd: %lu\n",allocMem, totalMem,(allocMem - freeMem));   
cudaErr = cudaFree(deviceReceivedImag);   
if (cudaErr != cudaSuccess) 
{ 
    mexPrintf("could free deviceReceivedImag\n");
    mexPrintf("Error: %s\n",cudaGetErrorString(cudaErr));  
}   
cudaMemGetInfo(&allocMem, &totalMem);
mexPrintf("deviceReceivedImag: Memory avaliable: Free: %lu, Total: %lu, Free'd: %lu\n",allocMem, totalMem,(allocMem - freeMem));

以下是此输出：

 
Memory avaliable: Free: 2523959296, Total: 2818572288
 Allocating memory.
 devicePulseDelay: Memory avaliable: Free: 2522910720, Total: 2818572288, Consumed: 1048576
 deviceTarDistance: Memory avaliable: Free: 2522910720, Total: 2818572288, Consumed: 1048576
 deviceScattDistance: Memory avaliable: Free: 2518716416, Total: 2818572288, Consumed: 5242880
 deviceScatterers: Memory avaliable: Free: 2517667840, Total: 2818572288, Consumed: 6291456
 deviceReceivedReal: Memory avaliable: Free: 2515570688, Total: 2818572288, Consumed: 8388608
 deviceReceivedImag: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760

Copying memory.
 devicePulseDelay: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760
 deviceTarDistance: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760
 deviceScattDistance: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760
 deviceScatterers: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760
 receivedReal: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760
 receivedImag: Memory avaliable: Free: 2513473536, Total: 2818572288, Consumed: 10485760

Free'ing memory.
 Before freeing: Free 2513473536, Total: 2818572288
 devicePulseDelay: Memory avaliable: Free: 2513473536, Total: 2818572288, Free'd: 0
 deviceTarDistance: Memory avaliable: Free: 2513473536, Total: 2818572288, Free'd: 0
 deviceScattDistance: Memory avaliable: Free: 2513473536, Total: 2818572288, Free'd: 0
 deviceScatterers: Memory avaliable: Free: 2514522112, Total: 2818572288, Free'd: 1048576
 deviceReceivedReal: Memory avaliable: Free: 2514522112, Total: 2818572288, Free'd: 1048576
 deviceReceivedImag: Memory avaliable: Free: 2514522112, Total: 2818572288, Free'd: 1048576

我觉得有一些明显的东西我不知道了。任何人都可以帮助解释发生了什么？

编辑：平台是带有Tesla C2050 GPu卡的Windows 7。

Answer 1

一种常见的误解是malloc在被调用时直接从主机操作系统获取内存分配，free在调用时直接将它们释放回主机操作系统。但它们几乎总是不那样工作，而是标准库维护一个自由和malloc内存的循环列表，通过与主机操作系统的交互进行机会扩展和收缩（参见{{3的一些答案）如果您有兴趣了解更多细节）。无论它是如何工作的，这都会导致许多非直观的结果，包括通常不可能分配尽可能多的内存，因为操作系统说是免费的，分配有时似乎不会改变可用内存的数量，并且free有时对OS所说的内存量没有影响。

虽然我只有经验证据支持这一点，但我相信CUDA的工作方式完全相同。上下文维护自己的malloc'd和free'd内存列表，并将扩展和收缩该列表中保存的内存作为主机驱动程序/窗口管理器，GPU本身允许。所有硬件都具有特定的MMU页面大小，并且有证据表明NVIDIA GPU上的页面大小相当大。这意味着cudaMalloc调用中的粒度相当粗糙，并且有时意味着malloc似乎不会影响可用内存量或消耗比请求的内存更多的内存，有时free调用似乎没有任何效果（如果您感兴趣，可以找到一个有助于说明CUDA驱动程序How do malloc() and free() work?的页面大小行为的小工具，尽管它是为早期版本的CUDA API编写的，可能需要用现代版本编译的几个变化）。我相信这是你正在观察的行为的最可能的解释。

顺便提一下，如果我使用GT200系列设备运行您在MacOS 10.6上发布的简化版代码：

#include <cstdio>

#define mexPrintf printf

inline void gpuAssert(cudaError_t code, char *file, int line, 
                 bool abort=true)
{
   if (code != cudaSuccess) 
   {
      mexPrintf("GPUassert: %s %s %d\n", cudaGetErrorString(code),
          file, line);
      if (abort) exit(code);
   }
}

#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }

inline void gpuMemReport(size_t * avail, size_t * total, 
        const char * title = 0, const size_t * free = 0, const bool sense = true) 
{
    char tstring[32] = { '\0' };
    gpuErrchk( cudaMemGetInfo(avail, total) );  

    if (free) {
        if (title) {
            strncpy(tstring, title, 31);
        }
        mexPrintf("%s Memory avaliable: Free: %zu, Total: %zu, %s: %zu\n",
                tstring, *avail, *total, (sense) ? "Allocated\0" : "Freed\0", 
                (sense) ? (*free - *avail) : (*avail - *free));
    } else {
        mexPrintf("Memory avaliable: Free: %zu, Total: %zu\n", *avail, *total);  
    }
}

int main()
{
    size_t freeMem = 0;
    size_t totalMem = 0;
    size_t allocMem = 0;

    gpuErrchk( cudaFree(0) );
    gpuMemReport(&freeMem, &totalMem);

    double *devicePulseDelay, *deviceTarDistance, *deviceScattDistance, *deviceScatterers;
    double *deviceReceivedReal, *deviceReceivedImag;

    mexPrintf("Allocating memory.\n");
    gpuErrchk( cudaMalloc( (void **) &devicePulseDelay, sizeof(double)*512) );
    gpuMemReport(&allocMem, &totalMem, "devicePulseDelay:", &freeMem);

    gpuErrchk( cudaMalloc( (void **) &deviceTarDistance, sizeof(double)*512) );
    gpuMemReport(&allocMem, &totalMem, "deviceTarDistance:", &freeMem);

    gpuErrchk( cudaMalloc( (void **) &deviceScattDistance, sizeof(double)*999*512) );
    gpuMemReport(&allocMem, &totalMem, "deviceScattDistance:", &freeMem);

    gpuErrchk( cudaMalloc( (void **) &deviceScatterers, sizeof(double)*999) );
    gpuMemReport(&allocMem, &totalMem, "deviceScatterers:", &freeMem);

    gpuErrchk( cudaMalloc( (void **) &deviceReceivedReal, sizeof(double)*999*512) );
    gpuMemReport(&allocMem, &totalMem, "deviceReceivedReal:", &freeMem);

    gpuErrchk( cudaMalloc( (void **) &deviceReceivedImag, sizeof(double)*999*512) );
    gpuMemReport(&allocMem, &totalMem, "deviceReceivedImag:", &freeMem);

    mexPrintf("\nFree'ing memory.\n");   
    gpuMemReport(&freeMem, &totalMem);

    gpuErrchk( cudaFree(devicePulseDelay) ); 
    gpuMemReport(&allocMem, &totalMem, "devicePulseDelay:", &freeMem, false);

    gpuErrchk( cudaFree(deviceTarDistance) ); 
    gpuMemReport(&allocMem, &totalMem, "deviceTarDistance:", &freeMem, false);

    gpuErrchk( cudaFree(deviceScattDistance) ); 
    gpuMemReport(&allocMem, &totalMem, "deviceScattDistance:", &freeMem, false);

    gpuErrchk( cudaFree(deviceScatterers) ); 
    gpuMemReport(&allocMem, &totalMem, "deviceScatterers:", &freeMem, false);

    gpuErrchk( cudaFree(deviceReceivedReal) ); 
    gpuMemReport(&allocMem, &totalMem, "deviceReceivedReal:", &freeMem, false);

    gpuErrchk( cudaFree(deviceReceivedImag) ); 
    gpuMemReport(&allocMem, &totalMem, "deviceReceivedImag:", &freeMem, false);

    return 0;
}

我得到了不同的结果，但也有一个显示相同的现象：

Allocating memory.
devicePulseDelay: Memory avaliable: Free: 202870784, Total: 265027584, Allocated: 1048576
deviceTarDistance: Memory avaliable: Free: 202870784, Total: 265027584, Allocated: 1048576
deviceScattDistance: Memory avaliable: Free: 198778880, Total: 265027584, Allocated: 5140480
deviceScatterers: Memory avaliable: Free: 197730304, Total: 265027584, Allocated: 6189056
deviceReceivedReal: Memory avaliable: Free: 193638400, Total: 265027584, Allocated: 10280960
deviceReceivedImag: Memory avaliable: Free: 189546496, Total: 265027584, Allocated: 14372864

Free'ing memory.
Memory avaliable: Free: 189546496, Total: 265027584
devicePulseDelay: Memory avaliable: Free: 189546496, Total: 265027584, Freed: 0
deviceTarDistance: Memory avaliable: Free: 190595072, Total: 265027584, Freed: 1048576
deviceScattDistance: Memory avaliable: Free: 194686976, Total: 265027584, Freed: 5140480
deviceScatterers: Memory avaliable: Free: 195735552, Total: 265027584, Freed: 6189056
deviceReceivedReal: Memory avaliable: Free: 199827456, Total: 265027584, Freed: 10280960
deviceReceivedImag: Memory avaliable: Free: 203919360, Total: 265027584, Freed: 14372864

这表明该行为也依赖于硬件/主机操作系统。

为什么CudaFree似乎没有释放内存？

1 个答案: