2022
 |
|
Commodity Deep Learning Technologies Supporting Autonomy on Small, Inexpensive Platforms
Christiaan Gribble
State-of-the-Art Report, CSIAC Journal, October 2022
This report reviews state-of-the-art artificial intelligence/machine learning (AI/ML) hardware and software technologies supporting autonomy on small, inexpensive platforms. It focuses on commodity hardware components and widely available software ecosystems for deep learning, the subset of AI/ML that uses multilayered neural networks to deliver best-in-class performance and accuracy for the low-level tasks that drive higher-level applications of autonomy.
|
|
|
|
|
 |
|
Analytic Rendering and Hardware-Accelerated Simulation for Scientific Applications
Paul Navratil, Christiaan Gribble, Pascal Grosset, and John E. Stone
Computing in Science & Engineering, vol. 24, no. 2
Advances in entertainment-targeted rendering technology have been leveraged for scientific analysis. Recent progress in both hardware and software capabilities have spurred development in analytic rendering: rendering capabilities optimized for analysis, particularly 3-D spatial analysis. These efforts also leverage hardware-accelerated ray tracing for high-fidelity rendering, which unlocks the potential to use such acceleration methods directly in ray-based simulations such as radiative transfer. This special issue presents the new ANARI API standard for analytic rendering and examples of analytic rendering and hardware-accelerated simulation where such APIs could be used.
|
2021
 |
|
Physical
Simulation via Hardware-Optimized Ray Tracing Engines:
SOLAR RT Update - November 2021
Christiaan Gribble
Lightning talk, SOLAR Ray Tracing Consortium, November 2021
We provide a brief update concerning curved ray
traversal through volumes of spatially varying
refractive indices, a critical operation supporting
applications in gradient field tomography at SURVICE
Engineering. Our hardware-optimized implementation
targets RTX-enabled NVIDIA GPUs and achieves over 12
million samples per second—or nearly 7 frames per
second—with volumes up to 224×224×224
voxels and framebuffers up to 768×768 pixels in
resolution using a single NVIDIA Quadro RTX 8000 GPU.
This level of performance is sufficient to shift the
bottleneck from volume sampling to other points in the
computational pipeline that enables simulation,
reconstruction, and visualization of unknown
compressible flows. Our implementation also enables
direct visualization of these 3D refractive index
fields, as well as interactive rendering of
inhomogeneous refractive objects.
A video
recording
of the talk is also available.
|
|
|
|
|
 |
|
Curved Ray Traversal
Christiaan Gribble
Book chapter, Ray Tracing Gems II: Next Generation Real-Time Rendering with DXR, Vulkan, and OptiX, August 2021
We present an implementation of curved ray traversal
through volumes of spatially varying refractive
indices. Our work is motivated by problems in
gradient field tomography, including simulation,
reconstruction, and visualization of unknown
compressible flows. Reconstruction, in particular,
requires computing, storing, and later retrieving
sample points along each ray, which in turn
necessitates a multi-pass traversal algorithm to
overcome potentially burdensome memory requirements.
The data structures and functions implementing this
algorithm also enable direct visualization, including
rendering of objects with varying refractive indices.
We highlight a GPU implementation in OWL, the OptiX 7
Wrapper Library, and demonstrate sampling for
reconstruction and interactive rendering of refractive
objects. We also provide source code, distributed
under a permissive open source license, to enable
readers to explore, modify, or enhance our curved ray
traversal implementation.
Additional resources,
including the source code for our curved
ray traversal implementation highlighted in
this chapter, as well as the
full
text and source
code for Ray Tracing Gems II, are also available.
|
|
|
|
|
 |
|
Render-Accelerated
Physical Simulation
Christiaan Gribble
Sponsored session,
GPU Technology Conference, March 2021
We discuss two recent applications of
hardware-accelerated ray tracing to problems in
physical simulation—or so-called non-optical
rendering—at SURVICE Engineering. In the first, we
leverage multi-hit ray tracing to compute ray/surface
intervals and thus simulate phenomena governed by
equations similar to the Beer-Lambert Law. In the
second, we leverage fast ray marching to relate
three-dimensional (3D) ray deflections to 3D
refractive index gradients via line integrals and thus
recover 3D refractive index distributions in gradient
field tomography. We implement the requisite
functionality for RTX-enabled NVIDIA GPUs to exploit
hardware-accelerated ray tracing in both problem
domains. We also suggest that the need for a
standardized ray-based simulation API is stronger than
ever, as such an API will facilitate the expanding
role of ray tracing throughout scientific workflows.
|
|
|
|
|
 |
|
Curved
Ray Traversal
Christiaan Gribble
Technical session,
GPU Technology Conference, March 2021
We present an implementation of curved ray traversal
through volumes of spatially varying refractive
indices. Our work is motivated by problems in gradient
field tomography, including simulation,
reconstruction, and visualization of unknown
compressible flows. Reconstruction, in particular,
requires computing, storing, and later retrieving
sample points along each ray, which in turn
necessitates a multi-pass traversal algorithm to
overcome potentially burdensome memory
requirements. The data structures and functions
implementing this algorithm also enable direct
visualization, including rendering of inhomogeneous
refractive objects. We discuss a GPU implementation in
OWL, the OptiX 7 Wrapper Library, and we highlight our
application source code, which is distributed under a
permissive open source license, to enable interested
attendees to explore, modify, or enhance our curved
ray traversal implementation.
|
2020
 |
|
Physical
Simulation via Hardware-Optimized Ray Tracing Engines:
SOLAR RT Update - November 2020
Christiaan Gribble
Lightning talk, SOLAR Ray Tracing Consortium, November 2020
We provide a brief update concerning two applications
of hardware-optimized ray tracing to problems in
physical simulation at SURVICE Engineering: multi-hit
ray traversal for interval computation and fast ray
marching for gradient field tomography. We implement
the requisite functionality for RTX-enabled NVIDIA
GPUs to exploit hardware-accelerated ray tracing in
both problem domains. We also suggest that the need
for a standardized ray-based simulation API is
stronger than ever, particularly as such an API will
potentially enable effective, efficient solutions to a
wider range of applications in physical simulation
across the diverse collection of anticipated hardware
ray tracing platforms.
|
|
|
|
|
 |
|
Implementing a Prototype System for 3D Reconstruction of Compressible Flow
Christiaan Gribble, Victor Eijkhout, and Paul Navratil
Practice and Experience in Advanced Research Computing, July 2020
PowerFlow3D is a prototype system for
acquiring, reconstructing, and visualizing
three-dimensional structure of complex flows around
objects in wind tunnel test procedures.
PowerFlow3D combines modern
high-performance computing (HPC) with existing
acquisition, reconstruction, and visualization methods
to provide a foundational capability that helps to
reveal critical information about the underlying
structure of unknown flows. We describe the
implementation of our system, focusing on tomographic
reconstruction, in particular, and highlight the
practical challenges encountered throughout our
initial research and development (R&D) process. The
resulting prototype achieves both reasonable
performance and fidelity and provides opportunities
for enhanced performance, fidelity, and scale. The
results of this initial R&D effort thus enable
continued progress toward a scalable HPC-accelerated
system for guiding real-time decisions during wind
tunnel tests.
|
|
|
|
|
 |
|
Effective
Parallelization Strategies for Scalable, High-Performance Iterative Reconstruction
Christiaan Gribble
Eurographics Symposium on Parallel Graphics and Visualization, May 2020
Iterative reconstruction techniques in X-ray computed
tomography converge to a result by successively
refining increasingly accurate estimates. Compared to
alternative approaches, iterative reconstruction
imposes significant computational demand but generally
leads to higher reconstruction quality and is more
robust to inherently imperfect scan data. We explore
several strategies for exploiting parallelism in
iterative reconstruction and evaluate their
scalability and performance on modern
workstation-class systems. Results show that
scalable, high performance iterative reconstruction is
possible with careful attention to the expression of
parallelism in both the projection and backprojection
phases of computation.
Supplemental
materials, including information about the
source
code for our prototype XCT reconstruction
system and scaling results for TP1
and TP2, are also available.
|
|
|
|
|
 |
|
Using Robust Networks to Inform Lightweight Models in Semi-Supervised Learning for Object Detection
Jonathan Worobey, Shawn Recker, and Christiaan Gribble
Poster, GPU Technology Conference, March 2020
We propose a semi-supervised model training approach
that temporarily utilizes the capacity of robust
networks to efficiently train low latency models with
limited hand-labeled data and a larger pool of
unlabeled data. This approach results in more accurate
lightweight models with minimal cost from hand-labeled
data while also providing an efficient way of curating
ground-truth datasets.
|
2019
 |
|
Using Robust Networks to Inform Lightweight Models in Semi-Supervised Learning for Object Detection
Jonathan Worobey, Shawn Recker, and Christiaan Gribble
Applied Imagery Pattern Recognition Workshop, October 2019
A common trade-off among object detection algorithms
is accuracy-for-speed (or vice versa). To meet our
application’s real-time requirement, we use a Single
Shot MultiBox Detector (SSD) model. This architecture
meets our latency requirements; however, a large
amount of training data is required to achieve an
acceptable accuracy level. While unusable for our end
application, more robust network architectures, such
as Regions with CNN features (R-CNN), provide an
important advantage over SSD models—they can
be more reliably trained on small datasets. By
fine-tuning R-CNN models on a small number of
hand-labeled examples, we create new, larger training
datasets by running inference on the remaining
unlabeled data. We show that these new, inferenced
labels are beneficial to the training of lightweight
models. These inferenced datasets are imperfect, and
we explore various methods of dealing with the errors,
including hand-labeling mislabeled data, discarding
poor examples, and simply ignoring errors. Further,
we explore the total cost, measured in human and
computer time, required to execute this workflow
compared to a hand-labeling baseline.
An addendum
with additional results for ATO-Action, a video
dataset capturing human subjects performing aircraft
handling signals,
including forward, back, left, right, stop, wave
off, and land, is also available. The
ATO-Action dataset supports our exploration of gesture
recognition techniques for ground-based control of
unmanned aerial vehicles, as outlined in the main
text.
|
|
|
|
|
 |
|
Physical Simulation via Hardware-Optimized Ray Tracing Engines
Christiaan Gribble
Keynote presentation, SOLAR Ray Tracing Consortium, May 2019
We discuss two recent applications of
hardware-optimized ray tracing engines to problems in
physical simulation—or so-called non-optical
rendering—at SURVICE Engineering. In the
first, we leverage multi-hit ray tracing to compute
ray/surface intervals and thus simulate phenomena
governed by equations similar to the Beer-Lambert Law.
In the second, we leverage fast ray marching to relate
three-dimensional (3D) ray deflections to 3D
refractive index gradients via line integrals and thus
recover 3D refractive index distributions in gradient
field tomography. In each case, modern
high-performance ray tracing engines are used to not
only visualize simulation results, but to effectively
and efficiently solve the underlying problem at hand.
|
|
|
|
|
 |
|
Multi-Hit Ray Tracing in DXR
Christiaan Gribble
Book chapter, Ray Tracing Gems: High-Quality and Real-Time Rendering with DXR and Other APIs,
February 2019
Multi-hit ray traversal is a class of ray traversal
algorithm that finds one or more, and possibly all,
primitives intersected by a ray ordered by point of
intersection. Multi-hit traversal generalizes
traditional first-hit ray traversal and is useful in
computer graphics and physics-based simulation. We
present several possible multi-hit implementations
using Microsoft DirectX Raytracing and explore the
performance of these implementations in an example GPU
ray tracer.
Additional resources,
including source
and binary distributions of the example
multi-hit GPU ray tracing application highlighted in
this chapter, as well as the
full
text and source
code for Ray Tracing Gems, are also available.
|
[ 2001-2006 ]
[ 2007-2012 ]
[ 2013-2018 ]
[ 2019-present ]
|
| | |
