Redesign of parallelism and checkpointing

[gdevenyi]

Looping and checkpointing is all bad here right now. Rearranging this properly will enable better local parallelism, and enable clean dependency chains of job submission.

TABS MEAN RUN IN PARALLEL

Right now:
Second-level

• subject
• compute delin
• compute composite of delin
• compute composite of affine
  • generate jacobian of nlin
  • generate jacobian of delin
  • generate jacobian of affine
• compute relative jacobian
• compute absolute jacobian

First level

• subject
• otsu mask
• subject scans
• get delin
• delin composite
• affine composite
  • jacobian nlin
  • jacobian delin
  • jacobian absolute
• compute relative jacobian
• compute absolute jacobian

---

Should be
Second Level

• subject
  • compute delin
  • COMPLETE
• subject
  • compute composite of delin
  • compute composite of affine
  • COMPLETE
• subject
  • compute nlin jacobian
  • compute delin jacobian
  • compute affine jacobian
  • COMPLETE
• subject
  • compute relative jacobian
  • compute absolute jacobian
  • COMPLETE
• IF COMMON SPACE
  • resample nlin to common space
  • resample relative to common space
  • resample absolute to common space

First-level

• subject

  • generate otsu masks
  • COMPLETE

• subject

  • scan
    • generate delin
    • COMPLETE
  • scan
    • generate delin composite
    • generate affine composite
    • COMPLETE
  • scan
    • generate nlin jacobian
    • generate delin jacobian
    • generate affine jacobian
    • COMPLETE
  • scan
    • compute relative jacobian
    • compute absolute jacobian
    • COMPLETE
  • scan
    • resample nlin jacobian to unbiased common space
    • resample relative jacobian to unbiased common space
    • resample absolute jacobian to unbiased common space
    • COMPLETE

• IF common space

  • scan
    • resample nlin jacobian to target common space
    • resample relative jacobian to target common space
    • resample absolute jacobian to target common space
    • COMPLETE

• blur all jacobians