Question

我有一个简单的基于粒子的刚体动力学代码，每个刚体由许多具有质量，位置，速度等的小粒子组成......现在我想把这个cpu代码移植到gpu。

对于结构，我选择使用指针数组int** d_rigid_particle_indices将其每个元素点指向其子粒子索引的数组（因为我使用基数排序来对每个帧排序粒子，索引将在每帧中更改所以我将使用双缓冲方案 - ＆gt; int** d_rigid_particle_indices[2];）

这是数据结构＆amp;将数据从主机复制到设备的功能：

typedef struct{
// Particle based rigid body dynamics
float4*       d_rb_pos;                     
float4*       d_rb_vel;
float4*       d_rb_angular_velocity;
float4*       d_rb_linear_momentum;
float4*       d_rb_angular_momentum;
float4*       d_rb_force;
float4*       d_rb_torque;
float4*       d_rb_quaternion;
float*        d_rb_rotation_matrix;         
float*        d_rb_inv_inertia_local;       
float*        d_rb_inv_inertia_world;   
float*        d_rb_mass;
int**         d_rigid_particle_indices[2];

} d_RigidBody;

void CopyRigidBodyDataHostToDevice(d_RigidBody& d_rb,std::vector<RigidBody*>& rigidbodies_h){
const int num_rigid_bodies = rigidbodies_h.size();

// allocate host mem for top level pointer array
int** h_rigid_bodies = (int**)malloc(num_rigid_bodies*sizeof(int*));
for (int i = 0; i < num_rigid_bodies; ++i)
    h_rigid_bodies[i] = NULL;

// allocate host&device mem for child particle index array
for (int i = 0; i < num_rigid_bodies; ++i)
{
    RigidBody* rb = rigidbodies_h[i];
    if (rb)
    {
        std::vector<int>& rp_indices = rb->getRigidBodyParticleIndicesArray();
        const int num_particles = rp_indices.size();
        int* temp_rp_indices = (int*)malloc(num_particles*sizeof(int));
        for (int j = 0; j < num_particles; ++j)
        {
            temp_rp_indices[j] = rp_indices[j];
        }           
        int * temp_ptr;
        CUDA_SAFE_CALL( cudaMalloc( ( (void**)&temp_ptr, num_particles*sizeof(int) ) ) );       // allocate device mem for child pointer
        h_rigid_bodies[i] = temp_ptr;
        temp_ptr = NULL;
        for (int k = 0; k < num_particles; ++k)
        {
            // copy RigidParticleIndices per rigid body from host to device
            CUDA_SAFE_CALL( cudaMemcpy(h_rigid_bodies[i], temp_rp_indices, num_particles*sizeof(int), cudaMemcpyHostToDevice) );    //h_rigid_bodies[i] stores the device mem pointer
        }
        if (temp_rp_indices)
            free(temp_rp_indices);              
    }
}

// allocate device mem for top level pointer array
CUDA_SAFE_CALL( cudaMalloc( (void**)&(d_rb.d_rigid_particle_indices[0]), num_rigid_bodies*sizeof(int*) ) ); // pointer array
CUDA_SAFE_CALL( cudaMalloc( (void**)&(d_rb.d_rigid_particle_indices[1]), num_rigid_bodies*sizeof(int*) ) ); // pointer array

// now we're ready to copy top level pointer array from host to device 
CUDA_SAFE_CALL( cudaMemcpy(d_rb.d_rigid_particle_indices[0], h_rigid_bodies, num_rigid_bodies*sizeof(int*), cudaMemcpyHostToDevice) );

// clean up
if (h_rigid_bodies)
{
    for (int i = 0; i < num_rigid_bodies; ++i)
    {
        if (h_rigid_bodies[i])
        {
            free(h_rigid_bodies[i]);
        }
    }
    free(h_rigid_bodies);
}

}

编译之后，我得到一个错误“没有重载函数的实例”cudaMalloc“匹配参数列表参数类型是:( unsigned int）”这段代码：

CUDA_SAFE_CALL( cudaMalloc( ( (void**)&temp_ptr, num_particles*sizeof(int) ) ) );       // allocate device mem for child pointer.

有什么建议吗？

感谢。

Answer 1

你过分使用了一对括号。因此编译器只会看到一个参数传递给cudaMalloc调用。将行更改为

CUDA_SAFE_CALL( cudaMalloc( (void**)&temp_ptr, num_particles*sizeof(int) ) );

它有效。

我尝试将指针数组复制到设备时出现cudaMalloc错误

1 个答案: