5.2 TLM and ADM generation in general

In this section we describe in a general fashion the parts of the code that are relevant for automatic differentiation using the software tool TAF.

Figure 5.2:  

The basic flow is depicted in 5.2. If CPP option ALLOW_AUTODIFF_TAMC is defined, the driver routine the_model_main, instead of calling the_main_loop, invokes the adjoint of this routine, adthe_main_loop, which is the toplevel routine in terms of automatic differentiation. The routine adthe_main_loop has been generated by TAF. It contains both the forward integration of the full model, the cost function calculation, any additional storing that is required for efficient checkpointing, and the reverse integration of the adjoint model.

[DESCRIBE IN A SEPARATE SECTION THE WORKING OF THE TLM]

In Fig. 5.2 the structure of adthe_main_loop has been strongly simplified to focus on the essentials; in particular, no checkpointing procedures are shown here. Prior to the call of adthe_main_loop, the routine ctrl_unpack is invoked to unpack the control vector or initialise the control variables. Following the call of adthe_main_loop, the routine ctrl_pack is invoked to pack the control vector (cf. Section 5.2.5). If gradient checks are to be performed, the option ALLOW_GRADIENT_CHECK is defined. In this case the driver routine grdchk_main is called after the gradient has been computed via the adjoint (cf. Section ).

• 5.2.1 General setup
• 5.2.2 Building the AD code
• 5.2.3 The AD build process in detail
  • 5.2.3.1 The list AD_FILES and .list files
  • 5.2.3.2 The list AD_FLOW_FILES and .flow files
  • 5.2.3.3 Store directives for 3-level checkpointing
  • 5.2.3.4 Changing the default AD tool flags: ad_options files
  • 5.2.3.5 Hand-written adjoint code
  
  
• 5.2.4 The cost function (dependent variable)
  • 5.2.4.1 Enabling the package
  • 5.2.4.2 Initialization
  • 5.2.4.3 Accumulation
  • 5.2.4.4 Finalize all contributions
  
  
• 5.2.5 The control variables (independent variables)
  • 5.2.5.1 genmake and CPP options
  • 5.2.5.2 Initialization
  • 5.2.5.3 Perturbation of the independent variables
  • 5.2.5.4 Output of adjoint variables and gradient
  • 5.2.5.5 Control variable handling for optimization applications

Copyright © 2002 Massachusetts Institute of Technology