cudaFree -无效的设备指针错误

Question

我正在尝试释放我在CUDA + OpenGL互操作代码中分配的设备内存OpenGL。在进行错误检查时，我收到了Invalid Device Pointer错误，程序在cudaFree(dev_inp);函数结束时停止了执行。一切都很好，但我担心的是内存泄漏。

问题：

为什么我不能释放我分配的本地设备内存？我将cuda_resource从像素缓冲区对象中取消映射，并取消资源注册。

摘自数据自动化系统C方案编制指南B.17节：

Memory allocated via malloc() cannot be freed using the runtime (i.e. by calling any of the free memory functions from Sections 3.2.2).

这就引出了另外两个问题：

我在内核中没有malloced内存，因为我没有。因此，利用cudaFree函数应该(技术上？)在这里工作对吗？是由程序员来释放提供给本地定义的指针的内存，还是nvcc编译器在程序退出时或在程序超出本地范围时处理解除分配？我不想在代码中出现内存泄漏，所以通过处理以前分配的内存，我感到更安全。

c.在cudaDeviceReset()函数的末尾调用renderScene()是否谨慎，从而破坏了主CUDA上下文(以及它的变量和指针，如CUDA C编程指南所示)？我已经看到NVidia Visual也提到了这一点:当我调用它时，呈现速度似乎比平常慢。如果我可以简单地在这里cudaFree内存，那就太好了，但我似乎无法让它工作起来。

完整代码：

#define GET_PROC_ADDRESS( str ) wglGetProcAddress( str )

GLuint tex; 
GLuint pbo;
struct cudaGraphicsResource *cuda_resource;    

PFNGLBINDBUFFERARBPROC    glBindBuffer     = NULL;
PFNGLDELETEBUFFERSARBPROC glDeleteBuffers  = NULL;
PFNGLGENBUFFERSARBPROC    glGenBuffers     = NULL;
PFNGLBUFFERDATAARBPROC    glBufferData     = NULL;

// ==========================================================================================
// CUDA ERROR CHECKING CODE
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
   if (code != cudaSuccess) 
   {
      fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
      if (abort) getchar();
   }
}

// ==========================================================================================

void initCUDADevice() { 

    gpuErrchk(cudaGLSetGLDevice( cutGetMaxGflopsDeviceId() ));    

}

// ==========================================================================================

void changeSize(int w, int h) {

    //cudaDeviceReset();
    //initCUDADevice();

    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
    glLoadIdentity();

    // Prevent a divide by zero, when window is too short
    // (you cant make a window of zero width).
    if (h == 0)
        h = 1;

    float ratio =  w * 1.0 / h;

    // Use the Projection Matrix
    glMatrixMode(GL_PROJECTION);

    // Reset Matrix
    //glLoadIdentity();

    //// Set the viewport to be the entire window
    glViewport(0, 0, w, h);

    //// Get Back to the Modelview
    glMatrixMode(GL_MODELVIEW);
}

// ==========================================================================================

void renderScene(void) {

    // Clear Color and Depth Buffers
    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
    // Reset transformations
    glLoadIdentity();

    // ====================================================================================
    // initiate GPU by setting it correctly 
    //initCUDADevice(); 

    // ====================================================================================
    // read the image that needs to be textured 

    Mat image, flipped;
    image = imread("K:/Ultrasound experiment images/PA_175.png", CV_LOAD_IMAGE_GRAYSCALE);   // Read the file from disk

    if(!image.data)                              // Check for invalid input
    {
        cout <<  "Could not open or find the image" << std::endl ;


    }

    cv::flip(image, flipped, 0);

    imshow("OpenCV - image", image);    // displays output

    // ====================================================================================
    // allocate the PBO, texture, and CUDA resource

    glBindBuffer    = (PFNGLBINDBUFFERARBPROC)GET_PROC_ADDRESS("glBindBuffer");
    glDeleteBuffers = (PFNGLDELETEBUFFERSARBPROC)GET_PROC_ADDRESS("glDeleteBuffers");
    glGenBuffers    = (PFNGLGENBUFFERSARBPROC)GET_PROC_ADDRESS("glGenBuffers");
    glBufferData    = (PFNGLBUFFERDATAARBPROC)GET_PROC_ADDRESS("glBufferData");

    // ====================================================================================
    // generate the pixel buffer object (PBO)

    // Generate a buffer ID called a PBO (Pixel Buffer Object)
    glGenBuffers(1, &pbo);

    // Make this the current UNPACK buffer (OpenGL is state-based)
    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);

    // Allocate data for the buffer. 4-channel 8-bit image
    glBufferData(GL_PIXEL_UNPACK_BUFFER, sizeof(unsigned char) * flipped.rows * flipped.cols, NULL, GL_STREAM_DRAW);
    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);

    gpuErrchk(cudaGraphicsGLRegisterBuffer(&cuda_resource, pbo, cudaGraphicsMapFlagsNone)); 

    // ====================================================================================
    // create the texture object 

    // enable 2D texturing
    glEnable(GL_TEXTURE_2D);

    // generate and bind the texture    
    glGenTextures(1, &tex);
    glBindTexture(GL_TEXTURE_2D, tex);

    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);

    // put flipped.data at the end for cpu rendering 
    glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE,  image.cols, image.rows,  0, GL_LUMINANCE, GL_UNSIGNED_BYTE, 0 );

    // put tex at the end for cpu rendering 
    glBindTexture(GL_TEXTURE_2D, 0);

    // ====================================================================================
    // copy OpenCV flipped image data into the device pointer

    glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);

    unsigned char *dev_inp; 

    gpuErrchk( cudaMalloc((void**)&dev_inp, sizeof(unsigned char)*flipped.rows*flipped.cols) );

    gpuErrchk( cudaGraphicsMapResources(1, &cuda_resource, 0) );

    size_t size; 
    gpuErrchk( cudaGraphicsResourceGetMappedPointer((void **)&dev_inp, &size, cuda_resource) );

    gpuErrchk( cudaMemcpy(dev_inp, flipped.data, sizeof(unsigned char)*flipped.rows*flipped.cols, cudaMemcpyHostToDevice) );

    gpuErrchk( cudaGraphicsUnmapResources(1, &cuda_resource, 0) ); 

    // ====================================================================================
    // bind pbo and texture to render data now 

    glBindBuffer( GL_PIXEL_UNPACK_BUFFER, pbo);
    //
    glBindTexture(GL_TEXTURE_2D, tex);

    glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, flipped.cols, flipped.rows, GL_LUMINANCE, GL_UNSIGNED_BYTE, NULL);

    gpuErrchk( cudaGraphicsUnregisterResource(cuda_resource));
    gpuErrchk( cudaThreadSynchronize());

    //gpuErrchk(cudaFree(dev_inp));

    // ====================================================================================
    // map the texture coords to the vertex coords 

    glBegin(GL_QUADS);
    // Front Face
    glTexCoord2f(0.0f, 0.0f); glVertex3f(-1.0f, -1.0f,  1.0f);  // Bottom Left Of The Texture and Quad
    glTexCoord2f(1.0f, 0.0f); glVertex3f( 1.0f, -1.0f,  1.0f);  // Bottom Right Of The Texture and Quad
    glTexCoord2f(1.0f, 1.0f); glVertex3f( 1.0f,  1.0f,  1.0f);  // Top Right Of The Texture and Quad
    glTexCoord2f(0.0f, 1.0f); glVertex3f(-1.0f,  1.0f,  1.0f);  // Top Left Of The Texture and Quad

    glEnd();

    glFlush();  // force rendering

    glDisable(GL_TEXTURE_2D);

    //glutSwapBuffers();
    gpuErrchk(cudaFree(dev_inp));        // <--- Error here
    //cudaGraphicsUnregisterResource(cuda_resource);

}


// ==========================================================================================


int main(int argc, char **argv) {


    // init GLUT and create window
    glutInit(&argc, argv);
    glutInitDisplayMode(GLUT_DEPTH | GLUT_RGB );
    glutInitWindowPosition(100,100);
    glutInitWindowSize(1024,256);
    glutCreateWindow("CUDA + OpenGL interop");


    // register callbacks
    glutDisplayFunc(renderScene);
    glutReshapeFunc(changeSize);
    //glutIdleFunc(renderScene);

    // enter GLUT event processing cycle
    glutMainLoop();

    return 1;
}

Answer 1

这一行不是必需的，应该从代码中删除：

gpuErrchk( cudaMalloc((void**)&dev_inp, sizeof(unsigned char)*flipped.rows*flipped.cols) );

这一行创建一个设备分配，并将该分配的指针分配给dev_inp。

这里出现的问题是：

gpuErrchk( cudaGraphicsResourceGetMappedPointer((void **)&dev_inp, &size, cuda_resource) );

这一行获取一个从cuda_resource对象派生到另一个不同分配的新指针，并将该指针放置到dev_inp，覆盖以前分配的指针的(从cudaMalloc)。在此行中获得的新指针已经具有基础设备分配。在这一点上，您不需要单独/另外分配它。

此时，如果您试图释放dev_inp

gpuErrchk(cudaFree(dev_inp));        // <--- Error here

您正在尝试释放程序未显式分配的数据(通过cudaMalloc)，并且是持久性(此时) cuda_resource对象的必要组件。你不会想那样做的。不幸的是，放置在dev_inp中的原始指针现在丢失了(覆盖)，因此在您的程序中无法“释放”它，而且只要程序正在执行，您就会有内存泄漏。

解决方案不是执行额外的、不需要的分配：

gpuErrchk( cudaMalloc((void**)&dev_inp, sizeof(unsigned char)*flipped.rows*flipped.cols) );

这就意味着应该删除相应的cudaFree操作：

gpuErrchk(cudaFree(dev_inp));        // <--- Error here

在程序实际退出之前，我不会在任何地方使用cudaDeviceReset代码，特别是CUDA/OpenGL代码。在其他一些非常特殊的情况下，您可能希望在实际退出程序之前使用cudaDeviceReset，但在这里不适用。