BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Australia/Brisbane
X-LIC-LOCATION:Australia/Brisbane
BEGIN:STANDARD
DTSTART:19920301T030000
TZOFFSETFROM:+1100
TZOFFSETTO:+1000
TZNAME:AEST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20191203T080637Z
LOCATION:Great Hall 1&2
DTSTART;TZID=Australia/Brisbane:20191117T190400
DTEND;TZID=Australia/Brisbane:20191117T190600
UID:siggraphasia_SIGGRAPH Asia 2019_sess222_papers_171@linklings.com
SUMMARY:Taichi: A Language for High-Performance Computation on Spatially S
 parse Data Structures
DESCRIPTION:Technical Papers Fast-Forward \n\nTaichi: A Language for High-
 Performance Computation on Spatially Sparse Data Structures\n\nHu, Li, And
 erson, Ragan-Kelley, Durand\n\n3D visual computing data are often spatiall
 y sparse. To exploit such sparsity, people have developed hierarchical spa
 rse data structures, such as multi-level sparse voxel grids, particles, an
 d 3D hash tables. However, developing and using these high-performance spa
 rse data structures is challenging, due to their intrinsic complexity and 
 overhead. We propose Taichi, a new data-oriented programming language for 
 efficiently authoring, accessing, and maintaining such data structures. Th
 e language offers a high-level, data structure-agnostic interface for writ
 ing computation code. The user independently specifies the data structure.
  We provide several elementary components with different sparsity properti
 es that can be arbitrarily composed to create a wide range of multi-level 
 sparse data structures. This decoupling of data structures from computatio
 n makes it easy to experiment with different data structures without chang
 ing computation code, and allows users to write computation as if they are
  working with a dense array. Our compiler then uses the semantics of the d
 ata structure and index analysis to automatically optimize for locality, r
 emove redundant operations for coherent accesses, maintain sparsity and me
 mory allocations, and generate efficient parallel and vectorized instructi
 ons for CPUs and GPUs.\n\nOur approach yields competitive performance on c
 ommon computational kernels such as stencil applications, neighbor lookups
 , and particle scattering. We demonstrate our language by implementing sim
 ulation, rendering, and vision tasks including a material point method sim
 ulation, finite element analysis, a multigrid Poisson solver for pressure 
 projection, volumetric path tracing, and 3D convolution on sparse grids. O
 ur computation-data structure decoupling allows us to quickly experiment w
 ith different data arrangements, and to develop high-performance data stru
 ctures tailored for specific computational tasks. With 1/10 as many lines 
 of code, we achieve 4.55x higher performance on average, compared to hand-
 optimized reference implementations.\n\nRegistration Category: Platinum Pa
 ss, Full Conference Pass, Full Conference One-Day Pass, Basic Conference P
 ass, Student One-Day Pass, Exhibitor Pass
URL:https://sa2019.conference-program.com/presentation?id=papers_171&sess=
 sess222
END:VEVENT
END:VCALENDAR

