Computation values turn out to be incorrect

Hi All,
I am implementing a simple bounding box algorithm on opencl and I have problem with the computations.
I have a set of vertices which I pass into the kernel and I compute the bounding box accordingly. I also pass in another array just to read back the vertices back
from the kernel to find that they are all different. I am running on WinXP with Quadro FX 570M and 197.15 (Nvidia Driver).
The code for the C part is this.


  float bbox[8], results[8];
  memset(results, 0, 8 * sizeof(float));

  bbox[0] = (float)mesh->vertex[0].pt[0]; bbox[4] = (float)mesh->vertex[0].pt[0];
  bbox[1] = (float)mesh->vertex[0].pt[1]; bbox[5] = (float)mesh->vertex[0].pt[1];
  bbox[2] = (float)mesh->vertex[0].pt[2]; bbox[6] = (float)mesh->vertex[0].pt[2];

  // GPU Calculation
  float* vertices = static_cast<float *>(malloc(4 * mesh->vertex_count * sizeof(float)));
  float* fnew_verts = static_cast<float *>(malloc(4 * mesh->vertex_count * sizeof(float)));

  memset(vertices, 0, 4 * mesh->vertex_count * sizeof(float));
  memset(fnew_verts, 0, 4 * mesh->vertex_count * sizeof(float));

  float* vertex = vertices;
  for (int i = 0; i < mesh->vertex_count; i++)
  {
    *vertex++ = (float)mesh->vertex[i].pt[0];
    *vertex++ = (float)mesh->vertex[i].pt[1];
    *vertex++ = (float)mesh->vertex[i].pt[2];
    *vertex++;
  }

  ShaderManager* shm = ShaderManager::Instance();

  cl_int err;
  cl_mem vertices_mem, bbox_mem, new_verts;
  vertices_mem = clCreateBuffer(shm->opencl_context(), CL_MEM_READ_ONLY, 4 * mesh->vertex_count * sizeof(float), NULL, NULL);
  err = clEnqueueWriteBuffer(shm->opencl_command_queue(), vertices_mem, CL_TRUE, 0, 4 * mesh->vertex_count * sizeof(float), 
                             (void *)vertices, 0, NULL, NULL);

  bbox_mem = clCreateBuffer(shm->opencl_context(), CL_MEM_READ_WRITE, 8 * sizeof(float), NULL, NULL);
  err = clEnqueueWriteBuffer(shm->opencl_command_queue(), bbox_mem, CL_TRUE, 0, 8 * sizeof(float), (void *)bbox, 0, NULL, NULL);

  new_verts = clCreateBuffer(shm->opencl_context(), CL_MEM_WRITE_ONLY, 4 * mesh->vertex_count * sizeof(float), NULL, NULL);

  clFinish(shm->opencl_command_queue());

  err = clSetKernelArg(shm->opencl_kernel(), 0, sizeof(cl_mem), &vertices_mem);
  err |= clSetKernelArg(shm->opencl_kernel(), 1, sizeof(cl_mem), &bbox_mem);
  err |= clSetKernelArg(shm->opencl_kernel(), 2, sizeof(cl_mem), &new_verts);

  size_t global_work_size = mesh->vertex_count;
  err = clEnqueueNDRangeKernel(shm->opencl_command_queue(), shm->opencl_kernel(), 1, NULL, 
                               &global_work_size, NULL, 0, NULL, NULL);

#pragma warning(disable:4189)
  const char* str = errorString(err);
  clFinish(shm->opencl_command_queue());


  err = clEnqueueReadBuffer(shm->opencl_command_queue(), bbox_mem, CL_TRUE, 0, 8 * sizeof(float), 
                            results, 0, NULL, NULL);
  err = clEnqueueReadBuffer(shm->opencl_command_queue(), new_verts, CL_TRUE, 0, 4 * mesh->vertex_count * sizeof(float),
                            fnew_verts, 0, NULL, NULL);

  clFinish(shm->opencl_command_queue());

  // Verify with input data
  for (int i = 0; i < 4 * mesh->vertex_count; i++)
    assert(fnew_verts[i] == vertices[i]);

  for (int i = 0; i < mesh->vertex_count; i++)
  {
    mesh->vertex[i].pt[0] = (double)(fnew_verts[0]);
    mesh->vertex[i].pt[1] = (double)(fnew_verts[1]);
    mesh->vertex[i].pt[2] = (double)(fnew_verts[2]);
    fnew_verts += 4;
  }

  free(vertices);


The kernel is defined like this as follows:


 __kernel void bound_mesh(__global float4* vertices, __global float4* bbox, __global float4 new_verts)
 {
  int gid = get_global_id(0);
  
  if (vertices[gid].x < bbox[0].x)
    bbox[0].x = vertices[gid].x;

  if (vertices[gid].y < bbox[0].y)
    bbox[0].y = vertices[gid].y;

  if (vertices[gid].z < bbox[0].z)
    bbox[0].z = vertices[gid].z;

  if (vertices[gid].x > bbox[1].x)
    bbox[1].x = vertices[gid].x;

  if (vertices[gid].y > bbox[1].y)
    bbox[1].y = vertices[gid].y;

  if (vertices[gid].z > bbox[2].z)
    bbox[1].z = vertices[gid].z;

  new_verts[gid = vertices[gid];

  bbox[0].w = -1.0; bbox[1].w = -1.0;
 }


So new_verts should atleast give me the correct vertices considering I am assigning them the same value for the vertices.
So neither the bounding box nor the new_verts has the correct values in them. Any suggestions as to what I might be doing wrong.
Thanks for your help.

Hi,

I didn’t tried your kernel, but it might be caused by fact the vertices parameter is defined as:

__global float4* vertices

while the new_verts:

__global float4 new_verts

(first as pointer, second not)

Maybe changing it to:

__global float4* new_verts

could fix the problem.

You might also try using:

new_verts[gid].xyzw = vertices[gid].xyzw;

If I remember correctly, it helped me to workaround similar problem long time ago with older drivers.

Petr

It looks like you are setting bbox[0] in all your work-items at the same time. This means they may overwrite each other in a non-determinisitic fashion. (E.g., work-item 302 might replace the value of bbox[0].y at the same time that work-item 303 is trying to check it, resulting in the check getting the old number and the replacement not being noticed.) Effectively this looks like a data-race between all your threads. You need to make sure that they all write to different locations or that you explicitly synchronize your writes.