cudaFree - 设备指针错误无效

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dev_inp。检查错误后，我收到Invalid Device Pointer错误，程序在cudaFree(dev_inp);函数末尾的renderScene()调用处停止执行。一切都很好，但我担心内存泄漏。

问题：

一个。为什么我无法释放我分配的本地设备内存？我从像素缓冲区对象中取消映射cuda_resource，并取消注册资源。

CUDA C编程指南中的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).

所以，这引出了另外两个问题：

湾我在内核中没有malloc内存，因为我没有内存。那么，利用cudaFree函数应该（技术上？）在这里工作吗？是由程序员解除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者当它超出本地范围时负责解除分配？我不想在我的代码中发生内存泄漏，所以通过解除我之前分配的内存来解决问题，我觉得更安全。

℃。在cudaDeviceReset()函数的末尾调用renderScene()是否明智，以便根据CUDA C编程指南销毁主CUDA上下文（及其变量和指针）？我已经看到NVidia Visual Profiler文档也提到了这一点：cudaDeviceReset() 当我调用它时，渲染似乎比平常慢。如果我能简单地cudaFree这里的记忆会很棒，但我似乎无法让它发挥作用。

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

此行不是必需的，应从代码中删除：

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

此行创建设备分配，并将该分配的指针分配给dev_inp。

问题出现在这里：

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

此行获取从cuda_resource对象派生的 new 指针到另一个不同的分配，并将该指针放入dev_inp，覆盖您之前分配的指针（来自cudaMalloc）。在此行中获取的新指针已经具有底层设备分配。此时您无需单独/另外分配。

此时，如果您尝试释放dev_inp：

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

您正在尝试释放您的程序未明确分配的数据（通过cudaMalloc），并且还是持久性（此时）cuda_resource对象的必要组件。你不想那样做。不幸的是，放在dev_inp中的原始指针现在丢失（覆盖），因此无法在程序中“释放”它，并且只要程序是执行。

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) );

这意味着也应该删除相应的cudaFree操作：

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

在程序实际退出之前，我不会在CUDA代码中的任何地方使用cudaDeviceReset，尤其是CUDA / OpenGL代码。还有一些非常特殊的情况，您可能希望在实际意图退出程序之前使用cudaDeviceReset，但它们不适用于此处。