教材Programming Massively Parallel Processors 2nd Edition

331

Programming Massively Parallel Processors. DOI:

Copyright © David B. Kirk/NVIDIA Corporation and Wen-mei W. Hwu. Published by Elsevier Inc. All rights reserved

2017

http://dx.doi.org/10.1016/B978-0-12-811986-0.00015-7

Application case study—

molecular visualization

and analysis

15

CHAPTER

15.5 Summary .........................................................................................................342

15.6 Exercises .........................................................................................................343

References ...............................................................................................................344

The previous case study used a statistical estimation application to illustrate the pro-

cess of selecting an appropriate level of a loop nest for parallel execution, transform-

ing, and fast boundary condition checking. This application case study shows that

the effective use of these practical techniques can signicantly improve the execution

throughput of the application.

332 CHAPTER 15 Application case study—molecular visualization

This case study is based on VMD (Visual Molecular Dynamics) [HDS 1996], a popular

software system designed for displaying, animating, and analyzing bio-molecular sys-

tems. VMD has more than 200,000 registered users. It is an important foundation for

cation for interactive 3D visualization and analysis. For computation that runs too

long for interactive use, VMD can also be used in a batch mode to render movies

for later use. A motivation for accelerating VMD is to make batch mode jobs fast

enough for interactive use. This can drastically improve the productivity of scientic

investigations. With CUDA devices widely available in PCs, such acceleration can

have broad impact on the VMD user community. To date, multiple aspects of VMD

have been accelerated with CUDA, including electrostatic potential map calculation,

2009), molecular surfaces (KSES 2012), radial distribution histograms (LSK 2011),

and electron density map quality-of-t (SMIS 2014), and high delity ray tracing of

In this application, the electrostatic potential map is used to identify spatial locations where

simulation process as well as the visualization/analysis of simulation results.

potential value of each grid point as the sum of contributions from all atoms in the

system. This is illustrated in Fig. 15.2. The contribution of atom i to a lattice point

j is the charge of atom i divided by the distance from lattice point j to atom i. Since

33315.2 A simple kernel implementation

Fig. 15.3 shows the base C code of the DCS code. The function is written to process a

two-dimensional (2D) slice of a three-dimensional (3D) grid. The function will be called

334 CHAPTER 15 Application case study—molecular visualization

structure. The outer loops then iterate and take the execution to the next grid point.

point on the y by multiplying the grid point index values by the spacing between

grid points. This is a uniform grid method where all grid points are spaced at the

same distance in all three dimensions. The function does take advantage of the fact

precalculated by the caller of the function and passed in as a function parameter (z).

Based on what we learned from the MRI case study, two attributes of the DCS

atomic memory operations to coordinate the updates done by different threads to each

33515.2 A simple kernel implementation

can either perform a loop ssion (as we did in the previous case study), or we move

to create a new array to hold all y values and result in two kernels communicating

data through global memory. The latter increases the number of times that the y coor-

only a small amount of calculation that can be easily accommodated into the inner

loop without signicant increase in execution time of the inner loop. The amount of

work to be absorbed into the inner loop is much smaller than that in the previous case

study. The former would have added a kernel launch overhead for a kernel where

the atomic electrical charges are not modied during the computation. This means

that the atomic charge values can be efciently stored in the constant memory (in the

GPU box in Fig. 15.4).