Real Time Controller
Functional Requirements
Don Gavel &
Marc Reinig
UCO Lick, laboratory for
Adaptive Optics.
Keck NGAO Team Meeting
December 13, 2007
Videoconference
Outline
• Status
• Issues
• Plans
NGAO Real Time Controller FRD
2
Status
• We have an architecture that meets the performance requirements
–
–
–
–
–
–
–
9 LGS/WFS (64x64)
2 Tip/Tilt WFS
1 T/T-focus-astigmatism WFS
1 Truth WFS
5 DM’s (with T/T) (64x64)
1 Woofer
<500µSec from receipt of image to commands received at the DM
• Scalable to larger problems
• We have validated the software in the lab and through simulation
NGAO Real Time Controller FRD
3
Architecture
– Hardware and Software –
NGAO Real Time Controller FRD
4
NGAO System Context Block Diagram
(DRAFT)
Version 5, 10/31/07
Erik Johansson
Observing
Sequencer
Instrument
Sequencer
Atmospheric
Tools
Telescope
Interface
Data
Server
Laser
Sequencer
LTCS
Config Laser
Telem Telem
Data
Tel
Status
Offload
AO
Sequencer
Status
Status
Telem
RT Diags
Telem
Cmd
Cmd
Status
Status
Status
Cmd
Telem
RT Diags
PSF
Monitoring
Cameras
Motion
Control
Device
Control
Pixel
Data
dNIRI
OptoMech
Control
Relay
DM/TT
Control
Motion
Corrector
Control
Commands
Device
Control
Motion
Control
Device
Control
PSF
Cams
X?
MEM
s
DMs
x6
Acquisitio
n Control
LO WFS
Control
Device Motion
Motion
ControlControl
Control
Pixel
Sensor Corrector
Data
Control Commands
Corrector
Commands
Pixel
Data
DM/
TT
Status
Cmd
Telem
RT Diags
Acq
Cam
LO
WFS
DMs
x3
LO
WFS
x3
NGS
HOWFS
Control
RTC
Pixel
Data
Pixel
Data
Motion
Control
Sensor
Control
9 LGS
HOWFS
Control
TWFS
Control
Optics
Bench
Control
Motion
Control
Pixel
Sensor
Data
Control
Motion
Control
Sensor
Control
Motion
Control
Device
Control
Pixel
Data
LGS
WFS
x9
NGS
WFS
Uplink
TT
Legend:
TWF
S
Laser
Launch
Facility
Laser
Safety
System
Status
Motion
Control
Device
Control
Diag
Optic
s
Benc
h
Devs
Laser
LF
OptoMech
Motion
Control
Device
Control
Status
Diag
Laser
OptoMech
Laser
Servi
ce
Encl
Boresight
Cam
Spott
er
Statu
s
ESto
p
Non-Real-Time Commands and Data
Real-Time Commands and Data
Real-Time Diagnostic Data
NGAO Real Time Controller FRD
5
NGAO System Context Block Diagram
(DRAFT) cont.
Version 5, 10/31/07
Erik Johansson
Observing
Sequencer
Instrument
Sequencer
Atmospheric
Tools
Telescope
Interface
Data
Server
Laser
Sequencer
LTCS
Config Laser
Telem Telem
Data
Tel
Status
Offload
AO
Sequencer
Status
Status
Telem
RT Diags
Telem
Cmd
Cmd
Status
Status
Status
Cmd
Telem
RT Diags
PSF
Monitoring
Cameras
Motion
Control
Device
Control
Pixel
Data
PSF
Cams
X?
dNIRI
OptoMech
Control
Relay
DM/TT
Control
Motion
Corrector
Control
Commands
Device
Control
Motion
Control
Device
Control
MEM
s
DMs
x6
Acquisitio
n Control
LO WFS
Control
Device Motion
Motion
ControlControl
Control
Pixel
Corrector
Sensor
Data
Control Commands
Corrector
Commands
Pixel
Data
DM/
TT
Status
Cmd
Telem
RT Diags
Acq
Cam
LO
WFS
DMs
x3
LO
WFS
x3
NGS
HOWFS
Control
RTC
Pixel
Data
Pixel
Data
Motion
Control
Sensor
Control
9 LGS
HOWFS
Control
Motion
Control
Pixel
Sensor
Data
Control
Motion
Control
Sensor
Control
Pixel
Data
LGS
WFS
x9
NGS
WFS
Uplink
TT
Legend:
TWFS
Control
TWF
S
Optics
Bench
Control
Motion
Control
Device
Control
Laser
Launch
Facility
Laser
Safety
System
Status
Motion
Control
Device
Control
Diag
Optic
s
Benc
h
Devs
Laser
LF
OptoMech
Motion
Control
Device
Control
Status
Diag
Laser
OptoMech
Laser
Servi
ce
Encl
Boresight
Cam
Spott
er
Statu
s
ESto
p
Non-Real-Time Commands and Data
Real-Time Commands and Data
Real-Time Diagnostic Data
NGAO Real Time Controller FRD
6
Interface to SRT System
• SRT controls all aspects of RTC - Ethernet
– There will be a non-RT Linux PC controlling the RTC
– Interface to the SRT will be Ethernet between the SRT and the RTC
control PC
– Load and read the RTC code for all operations
– Load and read the RTC HW configuration
• Telemetry streams to monitoring, PSF, etc - Ethernet
• Interface between the RTC and the high-speed disk arrays for
diagnostic data - Fiber Link or equivalent
NGAO Real Time Controller FRD
7
NGAO RTC System Context Block Diagram
(DRAFT)
Observing Sequencer
Data
Server
Telescope Interface
AO
Sequencer
Telem
RT Diags
Offload
High Speed Data
Server
Version 2, Nov 11, 2007
Marc Reinig
Telem
Config
Data
High Speed Diagnostic Data
Cmd
Status
Status
Cmd
Telem
RT Diags
LO WFS
Control
Device
Control Motion
Control
Corrector
Commands
Corrector
Commands
RTC
Uplink
TT
Corrector
Commands
DM/TT
Legend:
LO WFS
DMs
x3
LO WFS
x3
9 LGS
HOWFS
Control
TWFS
Control
Motion
Control
Sensor
Control
Motion
Control
Sensor
Control
Motion
Control
Sensor
Control
Pixel
Data
Pixel
Data
Pixel
Data
MEMs
DMs
x6
Pixel
Data
NGS
HOWFS
Control
LF
OptoMech
NGS
WFS
Pixel
Data
LGS WFS
x9
TWFS
Non-Real-Time Commands and Data
Real-Time Commands and Data
Real-Time Diagnostic Data
NGAO Real Time Controller FRD
8
Tomography Is Part of a Larger System
– With characteristics that lend themselves to a similar solution –
Kolmogorov filter
Guide star
Layer heights
height
Wavefront
Wavefront
Sensors
Wavefront
Sensors
Wavefront
Sensors
Sensors
Image
Image
Processors
Image
Processors
Image
Processors
Processors
Tomography
Unit
Centroid algorithm
r0, guidstar brightness,
Guidestar position
Precondition
Cn2 profile
FT-1
FT
X
Wavefront
Wavefront
Sensors
Wavefront
Sensors
DM
Sensors
Projection
DM conjugate
altitude
Backpropagate
Independent parallel
streams of parallelizable
processes
+
FT
X
FT-1
Aperture
Image
Image
Processors
Image
Processors
Deformable
Processors
Mirrors
Actuator
influence
function
Postcondition
Volumetric
dOPD
estimates
Aperture
WFS
data
Image
Image
Processors
Image
Processors
DM
Processors
Fit
Forwardpropagate
More independent parallel
streams!
NGAO Real Time Controller FRD
9
A Systolic Array
Voxel
Boundary
Chip
Boundary
Board
Boundary
NGAO Real Time Controller FRD
10
Data Rate and Location
Nasmyth
analog cable
length
constraints
WFS
Cam
Local
frame
grab
..
.
Hartmann sensor
optics
digital cable
(prefer DMA
transfer)
HV cable length
constraints
Woofer
Deformable
Mirror
Tweeter
Deformable
Mirror
..
.
Nasmyth
Computer
Room
Local
dedicated
centroider
Fiber Link or
equivalent
Telemetry
Data
(>160MB/Sec each)
Tomography
Engine
High
Voltage
D/A
converter
Woofer
fits
High
Voltage
D/A
converter
Tweeter
fits
Diagnostic Data
(To Disk) (>1 GB/Sec)
Nasmyth
NGAO Real Time Controller FRD
11
Telemetry and Diagnostics Streams
Telemetry (Low bandwidth, on-line monitoring) (>1Hz)
• Set and read the telemetry rate for each data stream
• Set and read the data streams that are enabled for telemetry
–
–
–
–
Centroids
Residual tomography error
DM commands, including uplink
Tip/Tilt commands
Diagnostics (High bandwidth, stored to disk) (full frame rate)
(Could be > 1GB/Sec)
• Set and read the diagnostic rate for each data stream
• Set and read the data streams that are enabled for diagnostics
–
–
–
–
–
–
Centroids
Residual tomography error
Raw layer information
DM commands, including uplink
Collapsed layer information
Tip/Tilt commands
NGAO Real Time Controller FRD
12
RTC Software Flow
ATM +
+
Tomography
+
-
Woofer
+
Woofer
Fit
Woofer
+
Woofer
go-to
error
Tweeter

Collapse
Average
Tweeter
..
.
Tweeter
Fit
+
..
.
NGAO Real Time Controller FRD
13
Tomography Basic Loop
Fourier Domain
Guide Star = 1
(G = 1)
Spatial Domain
Dynamic data
Start
Static Data
Forward Prop(G)
current estimated values for
voxels passed through
layers
IDFT(FP)
forward propagated values
passed by rows (east/west)
then columns (north/south)
IDFT coefficients
apply aperture
Spatial fwd prop value
1 or 0 depending on location
measure error
Spatial adj fwd prop value
Measured value
multiplied by
shift(layer,guide star)
Calculate
New
Estimate
G=G+1
error values passed by rows
(east/west) then columns
(north/south)
DFT(error)
The error vectors for this guide star are
stored for back propagation
Store Error(G)
Sum of Squares
and Store(G)
no
DFT coefficients
Back
Propagate
An error magnitude for this guide star:
|E| = (real)2 + (imag)2 is summed across all
columns and rows. This value is stored for
global error.
G == Total #
Guide Stars?
Error > Limit
yes
Calculate
Error
Sum the Sum of Squares
error values across all guide
stars
Global Error
Error <
Cutoff?
Forward
Propagate
yes
Done
no
Measured Value
From WFS
G=1
Back propagation
G=G+1
no
Accumulate error values for
each guide star in each
voxel in each layer
Error <= Limit
CN2 value(Layer)
G == Total #
Guide Stars?
yes
Kolmogorov Filter
Coeff error
K_filter value
Error
Old estimate
Done
Error Limit
New Estimate
NGAO Real Time Controller FRD
14
Forward Propagation
New Voxel
Estimate,
Layer(0)
Adjust for
Shift for
Layer(0)
New Voxel
Estimate,
Layer(N)
Adjust for
Shift for
Layer(N)
Start
Calculate New Estimate
Back Propagate
Forward Propagate
Error > Limit
Calculate Error
Measured Value From
WFS
Error Limit
Sum All
Adjusted
Values in the
Ray Path
Error <= Limit
For each sub aperture, sum all the
estimated Voxel values, along the
path of a ray to each Guide Star.
In the Fourier domain, you must
account for the spatial shift that
occurs in the spatial domain. This
will be the complex conjugate of the
shift value used in the back
propagation.
Done
Calculate
Error Guide
Star (N)
Measured Value
From WFS(0)
NGAO Real Time Controller FRD
15
Start
Calculate Error
Calculate New Estimate
Back Propagate
Forward Propagate
Back
Propagate
Error > Limit
Forward
Propagate
Calculate Error
For each sub aperture, for each
ray from a Guide Star, calculate
the error between the forward
propagated value and the
measured value from the WFS for
that Guide Star.
Error <= Limit
Measured Value From WFS
Error > Limit
Error Limit
Done
Compare
Error to
Limit
Subtract
Forward
Propagated
Value from
Measured Value
for each WFS
If you have reached the limit stop,
else continue with another loop
Error <= Limit
Done
Error Limit
Measured
Value From
WFS(0)
NGAO Real Time Controller FRD
16
Back Propagate
Calculate
New
Estimate
for this
Voxel
Start
Calculate New
Estimate
Back Propagate
Adjust
Error(0) for
Shift for this
Layer
(Multiply)
Forward Propagate
Error > Limit
Calculate Error
Measured Value From
WFS
Error Limit
Adjust
Error(N) for
Shift
(Multiply)
For each guide star, for each sub aperture,
back propagate the error for a guide star to all
the Voxels along the ray back to the Guide
Star (These should be the same Voxels used
in calculating the Forward Propagated value.)
Adjust
Error(N) For
Gain and Cn2
(Multiply)
Adjust the error for each Voxel for the Cn2
value for its layer
Error <= Limit
Adjust
Error(0) For
Gain and Cn2
(Multiply)
Done
In the Fourier domain, you must account for
the spatial shift that occurs in the spatial
domain. This will be the complex conjugate of
the shift value used in the Forward
propagation.
Calculate
Error Guide
Star (N)
Calculate
Error Guide
Star (0)
Forward
Propagate
Guide Star
(N)
Forward
Propagate
Guide Star
(0)
Measured Value
From WFS(0)
Measured Value
From WFS(N)
Error Limit
NGAO Real Time Controller FRD
17
Calculate New Estimate
Start
Calculate New
Estimate
Forward Propagate
Back Propagate
Error > Limit
Average
Adjusted
Errors
Calculate Error
Measured Value From
WFS
Error Limit
Error <= Limit
Add to
Current
Estimated
Value
For each Voxel, calculate the
average of the adjusted errors that
were back propagated along the
rays to various Guide Stars and add
this to the estimated value for the
Voxel to get a new estimated value.
Done
Back
Propagate
d GS
Error(0)
Back
Propagate
d GS
Error(N)
Forward
Propagate
GS(0)
Forward
Propagate
GS(N)
NGAO Real Time Controller FRD
18
WFS Control Commands
Camera
• Set and read the camera frame rate
• Set and read the camera pixel rate
• Set and read the camera gain
Centroiding
• Load the dark pattern
• Load pixel gain pattern
• Set and read the threshold level
• Set and read the centroiding algorithm
• Load and read the centroid weights
• Load and read the reference centroids
• Load and read the centroid offsets
• Set and read pixel offsets for the camera image
• Set and read the guidestar mode (NGS or LGS) for any WFS
• Load and read centroid non linearity tables
NGAO Real Time Controller FRD
19
Tomography Control Commands
• Load and read the cone-effect scaling array for any guide star or
target
• Load and read the sky position for any guidestar or target
• Load and read the Kolmogorov filter array
• Load and read the pre-conditioning arrays
• Load and read the tomography loop gain arrays
• Load and read the tomography bailout value
• Load and read the Cn2 values for all layers
• Load and read the wind array for all layers
NGAO Real Time Controller FRD
20
DM and Tip/Tilt Control Commands
• Load and read the layer-to-DM collapse arrays
• Load and read the DM command matrixes and non linearity lookup
tables for open and closed loop control
• Set and read DM open/closed loop mode
• Set and read the Tip/Tilt command matrixes
NGAO Real Time Controller FRD
21
Calibration Capabilities
•
•
•
•
•
•
•
•
•
Create a system matrix for any DM
Set the number of frames to use when creating the system matrix
Clear the DM of any manually set values
Calibrate science target DMs in MOAO
Woofer
Tweeter
Force an arbitrary camera image for any WFS
Force an arbitrary centroid output for any tomography WFS input
Force an arbitrary layer pattern to be used by the DM command
module
• Send an arbitrary pattern to any DM
• Save the current centroid positions as an reference centroid file
(averaged over a settable number of frames)
NGAO Real Time Controller FRD
22
System Diagnostics Capabilities
•
•
•
•
Read all chip temperatures
Read system temperatures at various predetermined locations
Monitor chip and system fan operation
Set and read temperature alarm trip levels
NGAO Real Time Controller FRD
23
Reliability and Reparability
•
•
•
•
Graceful degradation
Spares requirement
MTBF
Self Healing (SEU)
NGAO Real Time Controller FRD
24
Physical Characteristics
• Size
– <20 m3
• Weight
– TBD
• Location
– WFS and DM Drive are on the Nasmyth
– RTC and High Speed Disk array are in the computer room
• Power
– <20KW
NGAO Real Time Controller FRD
25
Issues Remaining
• Determine Single Event Upset (SEU) rate from Gamma/Cosmic rays
• Verify detection and recovery scheme
NGAO Real Time Controller FRD
26
Plans
• Complete FRD
• Finish writing report
• Validate SEU issues
NGAO Real Time Controller FRD
27
Backups
NGAO Real Time Controller FRD
28
Data Rate Work Sheet
Sub apertures across primary
Sub aperature at altitued
Number of WFSs
Number of Pixels per WFS Camera
Camera transfer time to WFS (µSec)
Number of Layers
Number of MCAO DM's
Number of observing DMs (only relevant for MOAO)
Word size (Bytes)
Portion of a frame used for transfer time (% frame)
(Transfers of WFS data-in and DM data-out are overlapped)
Frame time (µSec)
Input and output control data rates (MB/Sec)
Per Camera
Total Camera
Per WFS
Total WFS
Per DM
Total DM
Diagnostic data rates (MB/Sec)
Per Camera
Total Camera
Per WFS
Total WFS
Per Layer
Total Layer
Per DM
Total DM
Diagnostic Storage Rate w/o layer info (GB/Sec)
Maximum diagnostic Storage Rate (GB/Sec)
Giga Bytes of Mass Storage per minute
Radial
MCAO MOAO Voxels
64
64
64
3
3
3
10
10
10
32,000 32,000 32,000
400
400
400
7
7
7
2
2
2
6
5
5
2
2
2
10%
500
10%
500
10%
500
160
1,600
164
1,638
164
328
160
1,600
164
1,638
164
983
160
1,600
164
1,638
164
983
128
1,280
16
164
1,475
10,322
16
33
1.5
12
708
128
1,280
16
164
1,475
10,322
16
98
1.5
12
712
128
1,280
16
164
164
1,147
16
98
1.5
3
161
NGAO Real Time Controller FRD
29
Sample Code to Configure the Hardware
– Simple ALU –
module alu4 ( input_a, input_b, output_c, ctrl_0, ctrl_1,
ctrl_2,ctrl_3, clr, clk );
input ctrl_0, ctrl_1, ctrl_2, ctrl_3;
input [15:0] input_a, input_b;
output [31:0] output_c;
input clr, clk;
wire [3:0] ctrl;
reg [31:0] output_c;
assign ctrl =
parameter ADD
parameter MUL
parameter NOP
{ctrl_3, ctrl_2, ctrl_1, ctrl_0};
= 4'b1001;
= 4'b0001;
= 4'b1000;
always @(posedge clk or posedge clr)
begin
if (clr)
output_c <= 0;
else
case (ctrl)
ADD:
output_c <= input_a + input_b;
MUL:
output_c <= input_a * input_b;
default:;
endcase
end
endmodule
NGAO Real Time Controller FRD
30
Processor Array Illustration
Global System
State Information
Forward Propagation
Path
GS1
...
Back Propagation
Path
GS1
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
...
Global System
State Information
Forward Propagation
Path
GSn
GS1
...
GS1 Error
GS2 Error
GS1
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
Voxel
Local Registers
Back Propagation
Path
Cn2
Cumulative Value
GSn
...
GS1
GSn
GS1
...
GS1
...
GSn
GS1
...
GSn
GSn
GS1
...
GS1 Error
GS2 Error
GS1
GS1
GS1
...
GSn
Path
Back Propagation
Path
GS1
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
...
Back Propagation
Path
Forward Propagation
Path
...
...
GS2 Error
GSn
GS1
...
Global System
State Information
Control Logic
Forward Propagation
Path
GS1
...
GSn
GS1
...
Back Propagation
Path
Global System
State Information
GSn
...
An Array of Voxel Processors
GSn
GSN Error
Current
Estimated
Value
Cn2
GS1
GS1
...
...
Control Logic
GSn
ALU
(Word Size + NGS) wide
Path
Back Propagation
Path
GS1
GSn
...
GSn
Voxel
Local Registers
GS1 Error
GS2 Error
GS3 Error
...
GSN Error
Control Logic
Cumulative Value
GSn
ALU
(Word Size + NGS) wide
...
GSn
Back Propagation
Path
GS1
Global System
State Information
GS1
Cumulative Value
GSn
Cumulative Value
GS1
GS1
Cumulative Value
GSn
Cumulative Value
GS1
Cn2
...
GSn
Forward Propagation
Path
GS1
...
...
GSn
GSn
Back Propagation
Path
Global System
State Information
•
...
...
GSn
Back Propagation
Path
Forward Propagation
Path
Current
Estimated
Value
ALU
(Word Size + NGS) wide
Forward Propagation
Path
GS1
GS1 Error
GS2 Error
Cumulative Value
GSn
...
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
Cn2
GSn
Voxel
Local Registers
Cn2
Forward Propagation
Global System
State
Information
Global
System
State Information
GSN Error
...
GSn
GSN Error
GS1
GSn
GS2 Error
GS1
...
ALU
(Word Size + NGS) wide
GSn
GS1 Error
Cumulative Value
GS1
GS1
GS3 Error
GS1 Error
...
Cumulative Value
GSn
GSn
GSn
GS2 Error
GS3 Error
ALU
(Word Size + NGS) wide
GSn
...
Voxel
Local Registers
Current
Estimated
Value
Cn2
...
GS1
GSn
Cumulative Value
GS1
Back Propagation
GSn
...
...
Path
Back Propagation
Path
...
GSN Error
GS1
GS1
GS3 Error
GS1
Path
Forward Propagation
Path
...
Back Propagation
Path
Back Propagation
Path
Voxel
Local Registers
Control Logic
Forward Propagation
Global System
State
Information
Global
System
State Information
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
GS1 Error
Cumulative Value
GS1
...
Current
Estimated
Value
Cumulative Value
GSn
GS3 Error
Control Logic
GS1
Cn2
GS1
GSn
Voxel
Local Registers
Current
Estimated
Value
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
GSN Error
Cumulative Value
GS1
GS1
Forward Propagation
Global System
State
Information
Global
System
State Information
GSn
ALU
(Word Size + NGS) wide
Control Logic
Cumulative Value
GSn
GSn
...
GS1
Forward Propagation
Path
Forward Propagation
Path
Global System
State Information
Global System
State Information
...
ALU
(Word Size + NGS) wide
...
...
GS1 Error
Current
Estimated
Value
Cn2
GS1
Cumulative Value
GSn
GS3 Error
GSN Error
Cumulative Value
GS1
Cn2
ALU
(Word Size + NGS) wide
GSn
GS2 Error
...
Control Logic
...
Voxel
Local Registers
GS3 Error
Current
Estimated
Value
GS1
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
GSN Error
Cumulative Value
GS1
Back Propagation
Path
Back Propagation
Path
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
Voxel
Local Registers
Back Propagation
Path
GSn
GS1 Error
GS2 Error
...
Cumulative Value
GSn
...
GS1
Forward Propagation
Path
Voxel
Local Registers
Control Logic
Forward Propagation
Path
Forward Propagation
Path
Global System
State Information
Global System
State Information
...
Global System
State Information
GS3 Error
Cn2
Cumulative Value
GS1
Back Propagation
Path
Back Propagation
Path
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
GS1
GSn
Current
Estimated
Value
ALU
(Word Size + NGS) wide
Control Logic
Forward Propagation
Path
Forward Propagation
Path
Back Propagation
Path
...
GSN Error
Current
Estimated
Value
ALU
(Word Size + NGS) wide
Global System
State Information
Global System
State Information
...
GS3 Error
GSN Error
Cumulative Value
GS1
GS1
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
GS1 Error
GS2 Error
...
Control Logic
Forward Propagation
Path
GSn
Voxel
Local Registers
GS3 Error
Current
Estimated
Value
...
Global System
State Information
•
Forward Propagation
Path
Back Propagation
Path
A Single Voxel Processor
NGAO Real Time Controller FRD
31
The Tomography Engine
- A scalable, programmable, configurable processor •
•
•
Scalable: System can be easily put together (by replication) to work for any
sized problem
Programmable: Changes in the algorithm can be loaded at run time
Configurable: Parameters are easily changed (sub aps, guide stars, layers,
Cn2; mixes and positions of natural and laser guidestars; etc.
… scales to this
This …
Multiple Voxels in each FPGA
(One column for each sub ap.)
Multiple Chips on a Board(s)
NGAO Real Time Controller FRD
32
A Single FPGA Can Hold
Hundreds of Voxels
A Circuit Board Can Hold
A Dozen FPGA’s
NGAO Real Time Controller FRD
33
Basing Our Architecture on
Our Knowledge of Structure
We divide the atmosphere into
regions called “voxels”
And use that model for our
structure for the iterative solution
NGAO Real Time Controller FRD
34
Using Our Knowledge of Structure
•
For a single Voxel
•
•
•
Examining the problem, we find:
All calculations for all voxels are identical
Most operations on voxels are independent of
other voxels
Operations are simple
Memory requirements are modest
This is an “Embarrassingly” parallel problem
So …
•
•
•
•
We give each voxel its own simple processor
Global System
State Information
Forward Propagation
Path
GS1
...
Back Propagation
Path
GSn
Note:
Because the Forward propagation
and Back Progagation paths are
parallel, but are used at different
times, they will actually be a single
bus in the physical implementation.
GS1
...
GSn
GS1
...
GSn
Voxel
Local Registers
GS1 Error
GS2 Error
GS3 Error
...
GSN Error
Current
Estimated
Value
Control Logic
Cn2
ALU
(Word Size + NGS) wide
Cumulative Value
GSn
Cumulative Value
GS1
These Voxel processors are
interconnected, modeled after the
structure of our problem
GS1
Global System
State Information
...
GSn
Forward Propagation
Path
Back Propagation
Path
This …
… scales to this
Each column of voxels is a complete sub aperture
containing all layers in our model of the atmosphere.
Multiple sub apertures are arrayed in each FPGA chip
Multiple Chips on a Board, and multiple boards to
satisfy the size of the problem
NGAO Real Time Controller FRD
35