CUDA SDK VolumeRender 分析 (3)

本文主要分析CUDA SDK sample如何同OpenGL相结合.

在CUDA中调用OpenGL主要有以下几个要点：

1. Interoperability with OpenGL requires that the CUDA device be specified by cudaGLSetGLDevice() before any other runtime calls.
2. Register resource to CUDA before mapping. 一个资源只需注册一次
3. After registering to CUDA, a resource should be mapped before accessing with CUDA function and unmapped after accessing it by calling cudaGraphicsMapResources() and cudaGraphicsUnmapResources().
4. A mapped resource can be read from or written to by kernels using the device memory address returned by cudaGraphicsResourceGetMappedPointer()
   for buffers and cudaGraphicsSubResourceGetMappedArray() for CUDA arrays.
5. DO NOT access a resource through OpenGL or Direct3D while it is mapped to CUDA, cause it will produce undefined results.

整体伪代码

  1:   set_OpenGL_device(); 
  2:   register_resources();
  3:   
  4:   while( is_running )
  5:   {
  6:       map_resource();
  7:       resource_pointer *pointer = get_mapped_pointetr();
  8:       process_using_cuda( pointer );
  9:       unmap_resource();
 10:       do_normal_rendering();
 11:   }
 12:   unregister_resources();

选择设备

  1: // sets device as the current device for the calling host thread.
  2: extern __host__ cudaError_t CUDARTAPI cudaGLSetGLDevice(int device);

本例中被封装在chooseCudaDevice()函数中, 自动选择性能最佳的device.

资源创建和注册

这里使用的资源是Pixel Buffer Object ： The buffer object storing pixel data is called Pixel Buffer Object (PBO). initPixelBuffer()函数负责创建并注册PBO.

  1:   // OpenGL pixel buffer object
  2:   GLuint pbo = 0;     
  3: 
  4:   // create pixel buffer object for display
  5:   glGenBuffersARB(1, &pbo);
  6:   glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
  7:   glBufferDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, width*height*sizeof(GLubyte)*4, 0, GL_STREAM_DRAW_ARB);
  8:   glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
  9: 
 10:   // register this buffer object with CUDA
 11:   cutilSafeCall(cudaGraphicsGLRegisterBuffer(&cuda_pbo_resource, pbo, cudaGraphicsMapFlagsWriteDiscard));  

使用资源

  1: // CUDA Graphics Resource (to transfer PBO)
  2: struct cudaGraphicsResource *cuda_pbo_resource; 
  3: 
  4: // render image using CUDA
  5: void render()
  6: {
  7:   // Copy inverse view matrix to const device memory
  8:   copyInvViewMatrix(invViewMatrix, sizeof(float4)*3);
  9: 
 10:   // Map graphics resources for access by CUDA
 11:   uint *d_output;
 12:   cutilSafeCall(cudaGraphicsMapResources(1, &cuda_pbo_resource, 0));
 13: 
 14:   // Get CUDA device pointer
 15:   size_t num_bytes; 
 16:   cutilSafeCall(cudaGraphicsResourceGetMappedPointer((void **)&d_output, &num_bytes,  
 17:                    cuda_pbo_resource));
 18: 
 19:   // clear image
 20:   cutilSafeCall(cudaMemset(d_output, 0, width*height*4));
 21: 
 22:   // call CUDA kernel, writing results to PBO
 23:   render_kernel(gridSize, blockSize, d_output, width, height, density, brightness, transferOffset, transferScale);
 24: 
 25:   cutilSafeCall(cudaGraphicsUnmapResources(1, &cuda_pbo_resource, 0));
 26: }

这里的cutilSafeCall()是cuda util中的函数, 负责log错误. render_kernel()是前面的d_render()函数负责写入计算出的颜色到PBO, d_output是map得到的供CUDA存取的指向PBO内存的指针.

显示

  1: // display results using OpenGL 
  2: void display()
  3: {
  4:     // use OpenGL to build view matrix
  5:     BuildViewMartix();
  6: 
  7:     // prepare pbo piexl
  8:     render();
  9: 
 10:     // display results
 11:     glClear(GL_COLOR_BUFFER_BIT);
 12: 
 13:     // draw image from PBO
 14:     glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
 15: 
 16:     // copy from pbo to texture
 17:     glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 18:     glBindTexture(GL_TEXTURE_2D, tex);
 19:     glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, 0);
 20:     glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 21: 
 22:     // draw textured quad
 23:     glEnable(GL_TEXTURE_2D);
 24:     glBegin(GL_QUADS);
 25:     glTexCoord2f(0, 0); glVertex2f(0, 0);
 26:     glTexCoord2f(1, 0); glVertex2f(1, 0);
 27:     glTexCoord2f(1, 1); glVertex2f(1, 1);
 28:     glTexCoord2f(0, 1); glVertex2f(0, 1);
 29:     glEnd();
 30: 
 31:     glDisable(GL_TEXTURE_2D);
 32:     glBindTexture(GL_TEXTURE_2D, 0);
 33: 
 34:     glutSwapBuffers();
 35:     glutReportErrors();
 36: 
 37:     cutilCheckError(cutStopTimer(timer));  
 38: 
 39:     computeFPS();
 40: }

其中render()就是上面写入PBO的函数, 这个display()函数是由glutDisplayFunc()注册的显示函数. 也就是渲染的全过程. 为了简化函数, 中间省略了一些统计和不核心的处理.

我们可以看到, 渲染的所有效果都是由CUDA通过volume render产生的, 最后OpenGL只是把结果作为一张图片贴在我们的视口上. 这里面有两个小细节glPixelStorei()函数修改数据对齐的单位, 详细介绍在这里. 第二是如何从PBO拷贝到纹理, Song Ho的OpenGL教程介绍的非常清楚, 我就不再赘述了.

看过以上几期的分析, 希望大家对Volume Render和CUDA能有一些新的理解, 欢迎大家与我讨论学习. 下一次想分析一下这个例子的一些细节技术.

参考：

CUDA C Programming Guide