
2010.08.24

  Remaining work to be done:

  * test and turn on CR masking
  * double-check source size analysis:
    * are we doing the right thing with sources flagged as bad in 'rough'?
    * are we doing the right thing for SAT, CR, EXT, PSF after SourceSize?
  * convert CR_SIGMA and EXT_SIGMA to probabilities
  * example results of model fits vs fake inputs
  * what is the right solution for the stack PSF photometry?
    * Nigel sees 4-6%
    * we see 1.9% at the worst
    * non-poisson errors are clearly better -- check on the chisq values
    * why is QGAUSS so good for the one example (better for poisson, if trendy)

2010.08.12

  speeding up the PCM fitting.  

  * first, I need to convert the convolutions to FFTs (double check that this will actually be faster!)
  * second, I need to thread the PCM fitting process

  I have three options:

  1) use the FFT threading in fftw -- each convolution would run threaded, but separate ones would be run in series
     * con: the images are generally small, so the threading overhead may be large compared to the operation (test?)
  2) each pass of the minimization has 1 + N convolutions to perform: run these in separate threads
     * can only use 1+N threads, where N is the number of free parameters (Io, Sxx, Syy, Sxy, but maybe also Xo, Yo)
  3) thread by source, using the system defined for the other fitting. 
     * since these source are large, this is a bit trickier to guarantee non-collision. 
     + infrastructure is easy -- copy psphotBlendFit.

2010.07.15

various psphot issues that need to be addressed:

* update extNsigma based on kron mags 
* convert EXT sigmas to a probability (watch that extNsigma is correctly defined)
* psphot forced fails for negative flux
  - do a run, plot a histogram of fluxes as various steps
  FIXED: this was caused by minimum limit of 0.0 in linear fitting

* capability to choose extended source fits based on |b|

* update psf-convolved fitting to work with sersic (iteration)
  DONE: but, needs to be sped up

