Professional Documents
Culture Documents
Electronic Inverter
Electronic Inverter
SPG @ IAA-CSIC
Milne-Eddington inversion
First proposal (NSO; Harvey et al. 1972)
First code based on Unno’s equation (Auer et al. 1977)
later generalized by (Landolfi et al. 1984)
First successful application (Boulder; Skumanich & Lites
1985, Lites & Skumanich 1985, Skumanich et al. 1985, 1987)
Extensively used
Reliable (Westendorp Plaza et al. 1998, Orozco Suárez et al. 2010)
Useful for diagnostics (analytic) (Orozco Suárez & Del Toro
Iniesta, 2007)
Milne-Eddington approximation
Assumption of choice for SO/PHI
Must do inversions on board
Few observables (6 𝜆’s x 4 Stokes par.) ➜
few free parameters (9)
Analytic character ➜ easy programming
Unavailability of space-qualified processors
or DSPs ➜ specifically tailored FPGA
Architecture dependence on the device
An electronic inverter of the radiative transfer equation. Boulder, February 2017
I. Introduction
Development steps
Starting point: MILOS code in IDL
(Orozco Suárez & Del Toro Iniesta, 2007)
Optimization to C: C-MILOS
Extensive tests
Initialization
Labview graphical description
Specific SW to build the FPGA
architecture: TAPAS
Two architectures for an
embarrassingly parallel problem:
MIMD for Virtex 5
SIMD for Virtex 4
An electronic inverter of the radiative transfer equation. Boulder, February 2017
I. Introduction
Development steps
Starting point: MILOS code in IDL
(Orozco Suárez & Del Toro Iniesta, 2007)
Optimization to C: C-MILOS
Extensive tests
Initialization
Labview graphical description
Specific SW to build the FPGA
architecture: TAPAS
Two architectures for an
embarrassingly parallel problem:
MIMD for Virtex 5
SIMD for Virtex 4
An electronic inverter of the radiative transfer equation. Boulder, February 2017
I. Introduction
Development steps
Starting point: MILOS code in IDL
(Orozco Suárez & Del Toro Iniesta, 2007)
Optimization to C: C-MILOS
-2
Extensive tests
-3
Initialization
-4
Labview graphical description
log10 (𝜒2)
Specific SW to build the FPGA -5
architecture: TAPAS -6
Synthetic Yes
Model Propagation matrix Radiative transfer Stokes Final model
Convergence
atmosphere and source function equation parameters and atmosphere
RFs
Observations
Synthetic Yes
Model Propagation matrix Radiative transfer Stokes Final model
Convergence
atmosphere and source function equation parameters and atmosphere
RFs
Observations
Synthetic Yes
Model Propagation matrix Radiative transfer Stokes Final model
Convergence
atmosphere and source function equation parameters and atmosphere
RFs
Observations
Synthetic Yes
Model Propagation matrix Radiative transfer Stokes Final model
Convergence
atmosphere and source function equation parameters and atmosphere
RFs
Observations
FPGA architecture
No useful embedded processor
within FPGA
Leon 3 & NIOS II (Tong,
2006)
Leon 3 (57 WMIPS) vs Intel
Xeon (1000 WMIPS (Learn,
2011)
Virtex 4 Virtex 5
Only a few processors in
Virtex 4 or 5 (Learn, 2011)
Tailored FPGA
Our problem is an embarrass-
ingly parallel problem
An electronic inverter of the radiative transfer equation. Boulder, February 2017
II. FPGA architecture
FPGA occupation
Numerical experiments