A metadata-enhanced framework for high performance visual effects
Author(s)
Cornwall, Jay L. T.
Type
Thesis
Abstract
This thesis is devoted to reducing the interactive latency of image processing computations in
visual effects. Film and television graphic artists depend upon low-latency feedback to receive
a visual response to changes in effect parameters. We tackle latency with a domain-specific optimising
compiler which leverages high-level program metadata to guide key computational and
memory hierarchy optimisations. This metadata encodes static and dynamic information about
data dependence and patterns of memory access in the algorithms constituting a visual effect –
features that are typically difficult to extract through program analysis – and presents it to the
compiler in an explicit form. By using domain-specific information as a substitute for program
analysis, our compiler is able to target a set of complex source-level optimisations that a vendor
compiler does not attempt, before passing the optimised source to the vendor compiler for
lower-level optimisation.
Three key metadata-supported optimisations are presented. The first is an adaptation of
space and schedule optimisation – based upon well-known compositions of the loop fusion and
array contraction transformations – to the dynamic working sets and schedules of a runtimeparameterised
visual effect. This adaptation sidesteps the costly solution of runtime code generation
by specialising static parameters in an offline process and exploiting dynamic metadata to
adapt the schedule and contracted working sets at runtime to user-tunable parameters. The second
optimisation comprises a set of transformations to generate SIMD ISA-augmented source code.
Our approach differs from autovectorisation by using static metadata to identify parallelism, in
place of data dependence analysis, and runtime metadata to tune the data layout to user-tunable
parameters for optimal aligned memory access. The third optimisation comprises a related set
of transformations to generate code for SIMT architectures, such as GPUs. Static dependence
metadata is exploited to guide large-scale parallelisation for tens of thousands of in-flight threads.
Optimal use of the alignment-sensitive, explicitly managed memory hierarchy is achieved by identifying
inter-thread and intra-core data sharing opportunities in memory access metadata.
A detailed performance analysis of these optimisations is presented for two industrially developed
visual effects. In our evaluation we demonstrate up to 8.1x speed-ups on Intel and AMD
multicore CPUs and up to 6.6x speed-ups on NVIDIA GPUs over our best hand-written implementations
of these two effects. Programmability is enhanced by automating the generation of
SIMD and SIMT implementations from a single programmer-managed scalar representation.
visual effects. Film and television graphic artists depend upon low-latency feedback to receive
a visual response to changes in effect parameters. We tackle latency with a domain-specific optimising
compiler which leverages high-level program metadata to guide key computational and
memory hierarchy optimisations. This metadata encodes static and dynamic information about
data dependence and patterns of memory access in the algorithms constituting a visual effect –
features that are typically difficult to extract through program analysis – and presents it to the
compiler in an explicit form. By using domain-specific information as a substitute for program
analysis, our compiler is able to target a set of complex source-level optimisations that a vendor
compiler does not attempt, before passing the optimised source to the vendor compiler for
lower-level optimisation.
Three key metadata-supported optimisations are presented. The first is an adaptation of
space and schedule optimisation – based upon well-known compositions of the loop fusion and
array contraction transformations – to the dynamic working sets and schedules of a runtimeparameterised
visual effect. This adaptation sidesteps the costly solution of runtime code generation
by specialising static parameters in an offline process and exploiting dynamic metadata to
adapt the schedule and contracted working sets at runtime to user-tunable parameters. The second
optimisation comprises a set of transformations to generate SIMD ISA-augmented source code.
Our approach differs from autovectorisation by using static metadata to identify parallelism, in
place of data dependence analysis, and runtime metadata to tune the data layout to user-tunable
parameters for optimal aligned memory access. The third optimisation comprises a related set
of transformations to generate code for SIMT architectures, such as GPUs. Static dependence
metadata is exploited to guide large-scale parallelisation for tens of thousands of in-flight threads.
Optimal use of the alignment-sensitive, explicitly managed memory hierarchy is achieved by identifying
inter-thread and intra-core data sharing opportunities in memory access metadata.
A detailed performance analysis of these optimisations is presented for two industrially developed
visual effects. In our evaluation we demonstrate up to 8.1x speed-ups on Intel and AMD
multicore CPUs and up to 6.6x speed-ups on NVIDIA GPUs over our best hand-written implementations
of these two effects. Programmability is enhanced by automating the generation of
SIMD and SIMT implementations from a single programmer-managed scalar representation.
Date Issued
2010-04
Date Awarded
2010-08
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Kelly, Paul
Sponsor
EPSRC
Creator
Cornwall, Jay L. T.
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)