Cuda + OpenGL 互操作性，glDrawArrays() 访问冲突 [关闭]答案

【问题标题】：Cuda + OpenGL interoperability , glDrawArrays() Access violation [closed]Cuda + OpenGL 互操作性，glDrawArrays() 访问冲突 [关闭]
【发布时间】：2016-12-04 09:40:07
【问题描述】：

首先，我想为我在下面提供的代码量道歉，我不确定，但我可能违反了发布规则。

我正在尝试修改 nvidia 的 cuda 示例流体GL（可以在此处查看示例 https://www.youtube.com/watch?v=jU9wgSu4_zw），以便我可以加载自定义图像（例如 PNG），而不是修改随机像素，然后在其上应用流体求解器，以流体般的方式有效地扭曲它。

在尝试了许多不同的实现之后，我发现使用两个 vbo 可能是可能的，一个用于将发送到流体求解器内核的 float2 粒子，另一个用于我填写的颜色图像中的 RGBA 信息。

虽然我在 glDrawArrays 中遇到了访问冲突，但我仍然没有弄清楚为什么，在之前的尝试中，我尝试使用一个带有名为 vertex_data 的结构的单个 vbo，该结构包含一个 float2 和一个 uchar4，通过 glbufferdata 发送它并让内核计算它的float2部分，但也发生了访问冲突异常。

如果有人愿意帮助我，我将不胜感激

typedef unsigned char ubyte;

#define DEVICE __device__
#define GLOBAL __global__

#define MAX(a,b) ((a > b) ? a : b)
#define DIM 512
#define DS DIM*DIM

glm::mat4 m_mat;

// CUFFT plan handle
cufftHandle planr2c;
cufftHandle planc2r;
static float2 *vxfield = NULL;
static float2 *vyfield = NULL;

float2 *hvfield = NULL;
float2 *dvfield = NULL;
static int wWidth = MAX(512, DIM);
static int wHeight = MAX(512, DIM);

/*-----CUSTOM STRUCT-----------------------------------------------------*/

struct GLTexture
{
    GLuint id;
    int width;
    int height;
};


vertex_data data[DS];

//c linkage
/*--------------------------------------------------------------------------------------------------------------------------------*/
extern "C" void addForces(float2 *v, int dx, int dy, int spx, int spy, float fx, float fy, int r);
extern "C" void advectVelocity(float2 *v, float *vx, float *vy, int dx, int pdx, int dy, float dt);
extern "C" void diffuseProject(float2 *vx, float2 *vy, int dx, int dy, float dt, float visc);
extern "C" void updateVelocity(float2 *v, float *vx, float *vy, int dx, int pdx, int dy);
extern "C" void advectParticles(GLuint vbo, float2 *v, int dx, int dy, float dt);
/*--------------------------------------------------------------------------------------------------------------------------------*/

GLSLProgram prog;
IOManager m_manager;
GLTexture m_tex;
std::vector<ubyte> in_img;
std::vector<ubyte> out_img;
vertex_data vData[6];


GLuint positionsVBO;
GLuint colorsVBO;
cudaGraphicsResource* positionsVBO_CUDA;
float2 *particles = NULL;

float2 *part_cuda = NULL;

int lastx = 0, lasty = 0;
int clicked = 0;
size_t tPitch = 0;

float myrand(void)
{
    return rand() / (float)RAND_MAX;
}

void initParticles(float2 *p, int dx, int dy)
{
    int i, j;

    for (i = 0; i < dy; i++)
    {
        for (j = 0; j < dx; j++)
        {
            p[i*dx + j].x = (j + 0.5f + (myrand() - 0.5f)) / dx;
            p[i*dx + j].y = (i + 0.5f + (myrand() - 0.5f)) / dy;
        }
    }

}




void keyboard(unsigned char key, int x, int y)
{
    switch (key)
    {
    case 27:

        glutDestroyWindow(glutGetWindow());
        exit(0);
        return;


    default:
        break;
    }
}

void click(int button, int updown, int x, int y)
{
    lastx = x;
    lasty = y;
    clicked = !clicked;
}

void motion(int x, int y)
{
    // Convert motion coordinates to domain
    float fx = (lastx / (float)wWidth);
    float fy = (lasty / (float)wHeight);
    int nx = (int)(fx * DIM);
    int ny = (int)(fy * DIM);

    if (clicked && nx < DIM - FR && nx > FR - 1 && ny < DIM - FR && ny > FR - 1)
    {
        int ddx = x - lastx;
        int ddy = y - lasty;
        fx = ddx / (float)wWidth;
        fy = ddy / (float)wHeight;
        int spy = ny - FR;
        int spx = nx - FR;
        addForces(dvfield, DIM, DIM, spx, spy, FORCE * DT * fx, FORCE * DT * fy, FR);
        lastx = x;
        lasty = y;
    }

    glutPostRedisplay();
}

void cleanup(void)
{
    cudaGraphicsUnregisterResource(positionsVBO_CUDA);

    unbindTexture();
    deleteTexture();

    // Free all host and device resources
    free(hvfield);
    free(particles);
    cudaFree(dvfield);
    cudaFree(vxfield);
    cudaFree(vyfield);
    cufftDestroy(planr2c);
    cufftDestroy(planc2r);

    glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);
    glDeleteBuffersARB(1, &positionsVBO);
}

void run_simulation(void){

    //Execute kernel

    advectVelocity(dvfield, (float *)vxfield, (float *)vyfield, DIM, RPADW, DIM, DT);
    diffuseProject(vxfield, vyfield, CPADW, DIM, DT, VIS);
    updateVelocity(dvfield, (float *)vxfield, (float *)vyfield, DIM, RPADW, DIM);
    advectParticles(positionsVBO, dvfield, DIM, DIM, DT);



}


void initShaders(){
    prog.compileShaders("vShader.vertex", "fShader.frag");
    prog.addAttribute("vertexPos");
    prog.addAttribute("vertexColor");

    prog.linkShaders();
}



void pre_display()
{

    glViewport(0, 0, 512, 512);
    glutPostRedisplay();

}

void display()
{
    pre_display();

    // render points from vertex buffer
    glClear(GL_COLOR_BUFFER_BIT);

    initShaders();

    run_simulation();

    prog.use();

    //GLint textureUniform = prog.getUniformLocation("mySampler");
    //glUniform1i(textureUniform, 0);
    //glActiveTexture(GL_TEXTURE0);

    GLint pUniform = prog.getUniformLocation("P");
    glUniformMatrix4fv(pUniform, 1, GL_FALSE, &m_mat[0][0]);

    glBindBufferARB(GL_ARRAY_BUFFER_ARB, positionsVBO);

    glPointSize(1);

    glEnable(GL_POINT_SMOOTH);
    glEnable(GL_BLEND);
    glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);

    //glEnableVertexAttribArray(0); tried to manually enable the arrays
    //glEnableVertexAttribArray(1);


    glDisable(GL_DEPTH_TEST);
    glDisable(GL_CULL_FACE);

    glVertexAttribPointer(0, 2, GL_FLOAT, GL_TRUE, sizeof(float2), 0);

    glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0); //dont need this but access violaton persists without it anyway

    glBindBufferARB(GL_ARRAY_BUFFER_ARB, colorsVBO);

    glVertexAttribPointer(1, 4, GL_UNSIGNED_BYTE, GL_TRUE, sizeof(vertex_data) ,(void*)(offsetof(vertex_data, col)));


    glDrawArrays(GL_POINTS, 0, DS);

    glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);

    prog.unuse();


    glDisable(GL_TEXTURE_2D); //from nvidia's probably linked to the cudaarray_t


    // Swap buffers
    glutSwapBuffers();

}

void initGL()
{
    int foo = 1;
    char *bar = "bar";
    glutInit(&foo, &bar);
    glutInitDisplayMode(GLUT_DEPTH | GLUT_DOUBLE | GLUT_RGBA);
    glutInitWindowSize(DIM, DIM);
    glutCreateWindow("mate21");

    glClearColor(0.0, 0.0, 0.0, 1.0);


    glutKeyboardFunc(keyboard);
    glutMouseFunc(click);
    glutMotionFunc(motion);



    glutDisplayFunc(display);

    glewInit();
}

void setGLDevice(){
    cudaDeviceProp prop;
    int dev;

    memset(&prop, 0, sizeof(cudaDeviceProp));
    prop.major = 1;
    prop.minor = 0;
    cudaChooseDevice(&dev, &prop);
    cudaGLSetGLDevice(dev);
}



void createVBO(){

    //reading rgba information from image to out_img
    unsigned long width, height;
    IOManager::readFileToBuffer("jojo_test.png", in_img);
    decodePNG(out_img, width, height, &(in_img[0]), in_img.size());


    //data.resize(DS); data used to be a vector, gave up on that
    for (int i = 0; i < DS; ++i){

        //data[i].pos = particles[i]; edited vertex_data struct for rgba only
        data[i].col.x = out_img[i * 4 + 0];
        data[i].col.y = out_img[i * 4 + 1];
        data[i].col.z = out_img[i * 4 + 2];
        data[i].col.w = out_img[i * 4 + 3];

    }


    glGenBuffers(1, &positionsVBO);
    glBindBufferARB(GL_ARRAY_BUFFER_ARB, positionsVBO);

    glBufferDataARB(GL_ARRAY_BUFFER_ARB, sizeof(float2)*DS, particles , GL_DYNAMIC_DRAW_ARB);
    glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);

    cudaGraphicsGLRegisterBuffer(&positionsVBO_CUDA, positionsVBO, cudaGraphicsMapFlagsNone);

    glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);

    glGenBuffers(1, &colorsVBO);
    glBindBuffer(GL_ARRAY_BUFFER_ARB, colorsVBO);
    glBufferDataARB(GL_ARRAY_BUFFER_ARB, sizeof(vertex_data)*DS, data, GL_DYNAMIC_DRAW_ARB);
    glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);



}

int main()
{
    setGLDevice();

    initGL();

    //orthogonal view matrix with glm
    m_mat = glm::ortho(0, 1, 1, 0, 0, 1);





    hvfield = (float2 *)malloc(sizeof(float2) * DS);
    memset(hvfield, 0, sizeof(float2) * DS);

    // Allocate and initialize device data
    cudaMallocPitch((void **)&dvfield, &tPitch, sizeof(float2)*DIM, DIM);

    cudaMemcpy(dvfield, hvfield, sizeof(float2) * DS,
        cudaMemcpyHostToDevice);
    // Temporary complex velocity field data
    cudaMalloc((void **)&vxfield, sizeof(float2) * PDS);
    cudaMalloc((void **)&vyfield, sizeof(float2) * PDS);

    setupTexture(DIM, DIM);
    bindTexture();

    // Create particle array
    particles = (float2 *)malloc(sizeof(float2) * DS);
    memset(particles, 0, sizeof(float2) * DS);

    initParticles(particles, DIM, DIM);

    // Create CUFFT transform plan configuration
    cufftPlan2d(&planr2c, DIM, DIM, CUFFT_R2C);
    cufftPlan2d(&planc2r, DIM, DIM, CUFFT_C2R);

    cufftSetCompatibilityMode(planr2c, CUFFT_COMPATIBILITY_FFTW_PADDING);
    cufftSetCompatibilityMode(planc2r, CUFFT_COMPATIBILITY_FFTW_PADDING);


    createVBO();

    //cleanup
    glutCloseFunc(cleanup);
    //Launch rendering loop
    glutMainLoop();
}

这是在 float2 粒子上有效模拟流体的相关内核

extern "C"
void advectParticles(GLuint vbo, float2 *v, int dx, int dy, float dt)
{
    dim3 grid((dx/TILEX)+(!(dx%TILEX)?0:1), (dy/TILEY)+(!(dy%TILEY)?0:1));
    dim3 tids(TIDSX, TIDSY);

    float2 *p;
    cudaGraphicsMapResources(1, &positionsVBO_CUDA, 0);


   size_t num_bytes;
   cudaGraphicsResourceGetMappedPointer((void **)&p, &num_bytes,positionsVBO_CUDA);


    advectParticles_k<<<grid, tids>>>(p, v, dx, dy, dt, TILEY/TIDSY, tPitch);


    cudaGraphicsUnmapResources(1, &positionsVBO_CUDA, 0);

}

还有 advectParticles_k：

__global__ void
advectParticles_k(float2 *part, float2 *v, int dx, int dy,
                  float dt, int lb, size_t pitch)
{

    int gtidx = blockIdx.x * blockDim.x + threadIdx.x;
    int gtidy = blockIdx.y * (lb * blockDim.y) + threadIdx.y * lb;
    int p;

    // gtidx is the domain location in x for this thread
    float2 pterm, vterm;

    if (gtidx < dx)
    {
        for (p = 0; p < lb; p++)
        {
            // fi is the domain location in y for this thread
            int fi = gtidy + p;

            if (fi < dy)
            {
                int fj = fi * dx + gtidx;
                pterm = part[fj];

                int xvi = ((int)(pterm.x * dx));
                int yvi = ((int)(pterm.y * dy));
                vterm = *((float2 *)((char *)v + yvi * pitch) + xvi);

                pterm.x += dt * vterm.x;
                pterm.x = pterm.x - (int)pterm.x;
                pterm.x += 1.f;
                pterm.x = pterm.x - (int)pterm.x;
                pterm.y += dt * vterm.y;
                pterm.y = pterm.y - (int)pterm.y;
                pterm.y += 1.f;
                pterm.y = pterm.y - (int)pterm.y;

                part[fj] = pterm;
            }
        } // If this thread is inside the domain in Y
    } // If this thread is inside the domain in X
}

谢谢

【问题讨论】：

标签： c++ c opengl cuda interop

【解决方案1】：

我认为您的缓冲区中没有为粒子分配足够的内存。

让我们假设：

sizeof(float2) = 4
DS = 10

调用时：

 glBufferDataARB(GL_ARRAY_BUFFER_ARB, sizeof(float2)*DS, particles , GL_DYNAMIC_DRAW_ARB);

您已分配 4 * 10 = 40 字节的内存

那么你正在调用这个函数：

glVertexAttribPointer(0, 2, GL_FLOAT, GL_TRUE, sizeof(float2), 0);

这表示单个粒子需要 2 * 4 = 8 字节的内存。

然后你正在渲染：

glDrawArrays(GL_POINTS, 0, DS);

通过此调用，您尝试渲染 10 个消耗 10 * 8 = 80 字节内存的粒子

通过此调用，您会遇到访问冲突，因为您只分配了 40 字节的内存。

【讨论】：

嗯，我看到了，乍一看，float2 cuda 原生类型是包含两个浮点值 x 和 y 的结构。对于 opengl 兼容性，它应该可以与 glVertexAttribPointer 调用一起正常工作，我会考虑您的评论并尝试进一步检查。如果我设法取得任何积极成果，我会再次发布。谢谢 Amadeusz。
你为什么会假设 sizeof(float2) = 4 ？它是 8，一个简单的测试就可以证明这一点。