Masters' Project

Parallel volume rendering using multiple VolumePro boards

                            Student:                                                                                                 Advisor:

                            Abhijeet Ghosh                                                                                    Prof. Arie E. Kaufman

Aim: - Implementing distributed volume rendering using multiple VolumePro 500 boards over a high-speed Myrinet interconnect and parallel rendering using multiple VolumePro 1000 boards on a single PC.

Motivation: -

• Distributed Rendering:  A single VolumePro 500 board can render only 256³ voxels in real time (30 frames per second). To obtain real time frame rates for larger volumes, experiments have been carried out using multiple VP 500 boards on parallel PCI slots of a single PC. But the drivers of the board are unable to distribute the load properly to the boards and hence the obtained speedup does not scale with number of boards. The programming API for VP 500, Volume Library Interface (VLI 2.0) does not allow the programmer to access each board individually for proper load distribution. The motivation behind this project is to be able to explicitly distribute the rendering tasks to boards on 2 separate PCs in order to obtain better speedup. We use a high-speed point-to-point Myrinet interconnect (100MB/s) for this purpose to achieve high frame rates over the network.

• Parallel Rendering: Since October 2001, the VolumePro-1000 has become commercially available. This new board has better supports for multi-board rendering, with its new XY-image ordered rendering algorithm. Based on the results of the previous attempts at multi-board rendering, the new VLI 3.0 has been redesigned to allow the programmer a way to specify a particular board for rendering a sub-volume. We utilize this feature to partition the image plane and render using multiple VP 1000 boards in parallel PCI slots on 1 PC.

Resources: -

H/W: - VolumePro 500 & 1000, Pentium III PC with 512 MB RAM, Myrinet LANi interconnect.

S/W: - Volume Library Interface (VLI 2.0, 3.0), Visual C++ 6.0, OpenGL 1.2, Myrinet GM API 1.3.

VolumePro 500

                                                                                                                                                                                  

                     VolumePro 500 PCI board                                                                      4 VP500 boards in parallel PCI slots of 1 machine

Distributed Rendering Algorithm

We have one PC set up as the master node and another as the slave. The basic procedure is as follows: - 

• Each rendering node (master & slave) has VolumePro board(s) in its PCI slots.

• The volume is divided into 2 equal halves at the master node.

• One half is transferred to the remote node (slave).

• Rendering context (model view matrix, lighting information, color LUTs) are transferred to the slave once per frame.

• Post-warped image of the half volume is transferred to the master once per frame.

• Hexagonal textures of the baseplanes (silhouette of a cubic volume is a hexagon) representing the corresponding halves are composited at the master node using alpha blending by the texture mapping graphics hardware.

 Important - The application (volume) has to be large enough to justify rendering over this networked set-up. The intuition is that for certain applications distributed rendering including the network transmission times and synchronization delays will be faster than rendering on a single machine.

• 1000 million tri-linearly interpolated, Phong-shaded samples/second.

• Ability to embed opaque and translucent surfaces.

• 8-, 16-, and 32-bit voxels with up to four customizable fields.

• XY image order rendering.

• Multi-pass volume rendering.

• Classification and interpolation in either order.

• Up to 8192 (8K) voxels in any dimension in one pass.

• Support for super volumes and multi-board rendering

• Space-leaping and early-ray termination.

High Quality Visualization with XY image ordered Rendering Algorithm

                                                                                                                         

                                                                                           CT_CAROTID_MGH 12 bit dataset rendered using VP 1000

Parallel Volume Rendering with multiple VP 1000 boards

• First create separate volume objects (of the same volume) and use these to load the volume on to the available board (one for every board).

• Create image & depth buffers for every volume object.

• Partition the image buffer in application memory (to be displayed by OpenGL) into equal ranges, one range for every board.

• For every frame generated, make calls to Render function with the appropriate image & depth buffers for every volume object.

• For every frame generated, make calls to Unload function to copy pixel values from the corresponding board into that range of the image buffer.

• Make call to Drawpixels() for drawing the image with OpenGL.

                                                            F-15 (213x300x82): Top half (in red) rendered by board 1 and bottom half (in black) by board 2

Bibliography: -

•  H. Pfister, J. Hardenbergh, J. Knittel, H. Lauer, and L. Seiler, The VolumePro Real-Time Ray-Casting System, Proceedings of SIGGRAPH 99.

• Volume Library Interface User’s Guide – Copyright RTViz.

•  VolumePro 1000 Programmer’s Guide – Copyright RTViz.

• The GM Message Passing System – Copyright Myricom, Inc.

• OpenGL Programming Guide – OpenGL ARB.