Eu estou tentando alocar memória do dispositivo, copiar para ele, executar os cálculos na GPU, copiar os resultados de volta e, em seguida, liberar a memória do dispositivo que eu aloquei. Eu queria ter certeza de que não ultrapassaria o limite e queria ver se teria memória suficiente no espaço de memória compartilhada para despejar alguns arrays.

Quando aloco a memória do dispositivo, não há erros sendo retornados. Quando eu usocudaMemGetInfo para verificar a quantidade de memória alocada, parece umcudaMalloc não alocou memória alguma. Além disso, quando tento liberar a memória, parece que apenas um ponteiro é liberado.

Estou usando o matlabMexfunction interface para configurar a memória da GPU e iniciar o kernel. Neste ponto, eu nem estou chamando para o kernel e apenas retornando uma matriz de unidade para os resultados.

<code>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));
</code>

Aqui está a saída disso:

 
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

Eu sinto que há algo óbvio que estou perdendo. Alguém pode ajudar a explicar o que está acontecendo?

EDIT: plataforma é o Windows 7 com uma placa GPU Tesla C2050.