6. Host Side Coding

This chapter describes the connection of a host model with the pool of CCPP-Physics schemes through the CCPP-Framework.

In several places, references are made to an Interoperable Physics Driver (IPD). The IPD was originally developed by EMC and later expanded by NOAA GFDL with the goal of connecting GFS physics to various models. A top motivation for its development was the dycore test that led to the selection of FV3 as the dycore for the UFS. Designed in a fundamentally different way than the CCPP, the IPD will be phased out in the near future in favor of the CCPP as a single way to interface with physics in the UFS. To enable a smooth transition, several of the CCPP components must interact with the IPD and, as such, parts of the CCPP code in the UFS currently carry the tag “IPD”.

6.1. Variable Requirements on the Host Model Side

All variables required to communicate between the host model and the physics, as well as to communicate between physics schemes, need to be allocated by the host model. An exception is variables errflg, errmsg, loop_cnt, blk_no, and thrd_no, which are allocated by the CCPP-Framework, as explained in Section 6.4.1. A list of all variables required for the current pool of physics can be found in ccpp/framework/doc/DevelopersGuide/CCPP_VARIABLES_XYZ.pdf (XYZ: SCM, FV3).

At present, only two types of variable definitions are supported by the CCPP-Framework:

  • Standard Fortran variables (character, integer, logical, real) defined in a module or in the main program. For character variables, a fixed length is required. All others can have a kind attribute of a kind type defined by the host model.

  • Derived data types (DDTs) defined in a module or the main program. While the use of DDTs as arguments to physics schemes in general is discouraged (see Section 2.2), it is perfectly acceptable for the host model to define the variables requested by physics schemes as components of DDTs and pass these components to CCPP by using the correct local_name (e.g., myddt%thecomponentIwant; see Section 6.2.)

6.2. Metadata for Variable in the Host Model

To establish the link between host model variables and physics scheme variables, the host model must provide metadata information similar to those presented in Section 2.2. The host model can have multiple metadata files (.meta) with multiple metadata sections in each file ([ccpp-arg-table]) or just one. The host model Fortran files contain three-line snippets to indicate the location for insertion of the metadata information contained in the corresponding section in the .meta file.

!!> \section arg_table_example_vardefs
!! \htmlinclude example_vardefs.html
!!

For each variable required by the pool of CCPP-Physics schemes, one and only one entry must exist on the host model side. The connection between a variable in the host model and in the physics scheme is made through its standard_name.

The following requirements must be met when defining metadata for variables in the host model (see also Listing 6.1 and Listing 6.2 for examples of host model metadata).

  • The standard_name must match that of the target variable in the physics scheme.

  • The type, kind, shape and size of the variable (as defined in the host model Fortran code) must match that of the target variable.

  • The attributes units, rank, type and kind in the host model metadata must match those in the physics scheme metadata.

  • The attributes optional and intent must be set to F and none, respectively.

  • The local_name of the variable must be set to the name the host model cap uses to refer to the variable.

  • The metadata section that exposes a DDT to the CCPP (as opposed to the section that describes the components of a DDT) must be in the same module where the memory for the DDT is allocated. If the DDT is a module variable, then it must be exposed via the module’s metadata section, which must have the same name as the module.

  • Metadata sections describing module variables must be placed inside the module.

  • Metadata sections describing components of DDTs must be placed immediately before the type definition and have the same name as the DDT.

    module example_vardefs

      implicit none

!!> \section arg_table_example_vardefs
!! \htmlinclude example_vardefs.html
!!

      integer, parameter           :: r15 = selected_real_kind(15)
      integer                      :: ex_int
      real(kind=8), dimension(:,:) :: ex_real1
      character(len=64)            :: errmsg
      logical                      :: errflg

!!> \section arg_table_example_ddt
!! \htmlinclude example_ddt.html
!!

      type ex_ddt
        logical              :: l
        real, dimension(:,:) :: r
      end type ex_ddt

      type(ex_ddt) :: ext

    end module example_vardefs

Listing 6.1: Example host model file with reference to metadata. In this example, both the definition and the declaration (memory allocation) of a DDT ext (of type ex_ddt ) are in the same module.

[ccpp-arg-table]
  name = arg_table_example_vardefs
  type = module
[ex_int]
  standard_name = example_int
  long_name = ex. int
  units = none
  dimensions = ()
  type = integer
  kind =
[ex_real]
  standard_name = example_real
  long_name = ex. real
  units = m
  dimensions = (horizontal_dimension,vertical_dimension)
  type = real
  kind = kind=8
[ex_ddt]
  standard_name = example_ddt
  long_name = ex. ddt
  units = DDT
  dimensions = (horizontal_dimension,vertical_dimension)
  type = ex_ddt
  kind =
[ext]
  standard_name = example_ddt_instance
  long_name = ex. ddt inst
  units = DDT
  dimensions = (horizontal_dimension,vertical_dimension)
  type = ex_ddt
  kind =
[errmsg]
  standard_name = ccpp_error_message
  long_name = error message for error handling in CCPP
  units = none
  dimensions = ()
  type = character
  kind = len=64
[errflg]
  standard_name = ccpp_error_flag
  long_name = error flag for error handling in CCPP
  units = flag
  dimensions = ()
  type = integer

########################################################################
[ccpp-arg-table]
  name = arg_table_example_ddt
  type = ddt
[ext%1]
  standard_name = example_flag
  long_name = ex. flag
  units = flag
  dimensions =
  type = logical
  kind =
[ext%r]
  standard_name = example_real3
  long_name = ex. real
  units = kg
  dimensions = (horizontal_dimension,vertical_dimension)
  type = real
  kind = r15
[ext%r(;,1)]
  standard_name = example_slice
  long_name = ex. slice
  units = kg
  dimensions = (horizontal_dimension,vertical_dimension)
  type = real
  kind = r15

Listing 6.2: Example host model metadata file ( .meta ).

6.3. CCPP Variables in the SCM and UFS Atmosphere Host Models

While the use of standard Fortran variables is preferred, in the current implementation of the CCPP in the UFS Atmosphere and in the SCM almost all data is contained in DDTs for organizational purposes. In the case of the SCM, DDTs are defined in gmtb_scm_type_defs.f90 and GFS_typedefs.F90, and in the case of the UFS Atmosphere, they are defined in both GFS_typedefs.F90 and CCPP_typedefs.F90. The current implementation of the CCPP in both host models uses the following set of DDTs:

  • GFS_init_type variables to allow proper initialization of GFS physics

  • GFS_statein_type prognostic state data provided by dycore to physics

  • GFS_stateout_type prognostic state after physical parameterizations

  • GFS_sfcprop_type surface properties read in and/or updated by climatology, obs, physics

  • GFS_coupling_type fields from/to coupling with other components, e.g., land/ice/ocean

  • GFS_control_type control parameters input from a namelist and/or derived from others

  • GFS_grid_type grid data needed for interpolations and length-scale calculations

  • GFS_tbd_type data not yet assigned to a defined container

  • GFS_cldprop_type cloud properties and tendencies needed by radiation from physics

  • GFS_radtend_type radiation tendencies needed by physics

  • GFS_diag_type fields targeted for diagnostic output to disk

  • GFS_interstitial_type fields used to communicate variables among schemes in the slow physics group required to replace interstitial code in GFS_{physics, radiation}_driver.F90 in CCPP

  • GFS_data_type combined type of all of the above except GFS_control_type and GFS_interstitial_type

  • CCPP_interstitial_type fields used to communicate variables among schemes in the fast physics group

The DDT descriptions provide an idea of what physics variables go into which data type. GFS_diag_type can contain variables that accumulate over a certain amount of time and are then zeroed out. Variables that require persistence from one timestep to another should not be included in the GFS_diag_type nor the GFS_interstitial_type DDTs. Similarly, variables that need to be shared between groups cannot be included in the GFS_interstitial_type DDT. Although this memory management is somewhat arbitrary, new variables provided by the host model or derived in an interstitial scheme should be put in a DDT with other similar variables.

Each DDT contains a create method that allocates the data defined using the metadata. For example, the GFS_stateout_type contains:

type GFS_stateout_type

   !-- Out (physics only)
   real (kind=kind_phys), pointer :: gu0 (:,:)   => null()  !< updated zonal wind
   real (kind=kind_phys), pointer :: gv0 (:,:)   => null()  !< updated meridional wind
   real (kind=kind_phys), pointer :: gt0 (:,:)   => null()  !< updated temperature
   real (kind=kind_phys), pointer :: gq0 (:,:,:) => null()  !< updated tracers

   contains
     procedure :: create  => stateout_create  !<   allocate array data
 end type GFS_stateout_type

In this example, gu0, gv0, gt0, and gq0 are defined in the host-side metadata section, and when the subroutine stateout_create is called, these arrays are allocated and initialized to zero. With the CCPP, it is possible to not only refer to components of DDTs, but also to slices of arrays with provided metadata as long as these are contiguous in memory. An example of an array slice from the GFS_stateout_type looks like:

[ccpp-arg-table]
  name = GFS_stateout_type
  type = ddt
[gq0(:,:,index_for_snow_water)]
  standard_name = snow_water_mixing_ratio_updated_by_physics
  long_name = moist (dry+vapor, no condensates) mixing ratio of snow water updated by physics
  units = kg kg-1
  dimensions = (horizontal_dimension,vertical_dimension)
  type = real
  kind = kind_phys

Array slices can be used by physics schemes that only require certain values from an array.

6.4. CCPP API

The CCPP Application Programming Interface (API) is comprised of a set of clearly defined methods used to communicate variables between the host model and the physics and to run the physics. The bulk of the CCPP API is located in the CCPP-Framework, and is described in file ccpp_api.F90. Some aspects of the API differ between the dynamic and static build. In particular, subroutines ccpp_physics_init, ccpp_physics_finalize, and ccpp_physics_run (described below) are made public from ccpp_api.F90 for the dynamic build, and are contained in ccpp_static_api.F90 for the static build. Moreover, these subroutines take an additional argument (suite_name) for the static build. File ccpp_static_api.F90 is auto-generated when the script ccpp_prebuild.py is run for the static build.

6.4.1. Data Structure to Transfer Variables between Dynamics and Physics

The roles of cdata structure in dealing with data exchange are not the same between the dynamic and the static builds of the CCPP. For the dynamic build, the cdata structure handles the data exchange between the host model and the physics schemes. cdata is a DDT containing a list of pointers to variables and their metadata and is persistent in memory.

For both the dynamic and static builds, the cdata structure is used for holding five variables that must always be available to the physics schemes. These variables are listed in a metadata table in ccpp/framework/src/ccpp_types.meta (Listing 6.3).

  • Error flag for handling in CCPP (errmsg).

  • Error message associated with the error flag (errflg).

  • Loop counter for subcycling loops (loop_cnt).

  • Number of block for explicit data blocking in CCPP (blk_no).

  • Number of thread for threading in CCPP (thrd_no).

[ccpp-arg-table]
  name = ccpp_t
  type = scheme
[errflg]
  standard_name = ccpp_error_flag
  long_name = error flag for error handling in CCPP
  units = flag
  dimensions = ()
  type = integer
[errmsg]
  standard_name = ccpp_error_message
  long_name = error message for error handling in CCPP
  units = none
  dimensions = ()
  type = character
  kind = len=512
[loop_cnt]
  standard_name = ccpp_loop_counter
  long_name = loop counter for subcycling loops in CCPP
  units = index
  dimensions = ()
  type = integer
[blk_no]
  standard_name = ccpp_block_number
  long_name = number of block for explicit data blocking in CCPP
  units = index
  dimensions = ()
  type = integer
[thrd_no]
  standard_name = ccpp_thread_number
  long_name = number of thread for threading in CCPP
  units = index
  dimensions = ()
  type = integer

Listing 6.3: Mandatory variables provided by the CCPP-Framework from ccpp/framework/src/ccpp_types.meta . These variables must not be defined by the host model.

Two of the variables are mandatory and must be passed to every physics scheme: errmsg and errflg. The variables loop_cnt, blk_no, and thrd_no can be passed to the schemes if required, but are not mandatory. For the static build of the CCPP, the cdata structure is only used to hold these five variables, since the host model variables are directly passed to the physics without the need for an intermediate data structure.

Note that cdata is not restricted to being a scalar but can be a multidimensional array, depending on the needs of the host model. For example, a model that uses a one-dimensional array of blocks for better cache-reuse may require cdata to be a one-dimensional array of the same size. Another example of a multi-dimensional array of cdata is in the SCM, which uses a one-dimensional cdata array for N independent columns.

Due to a restriction in the Fortran language, there are no standard pointers that are generic pointers, such as the C language allows. The CCPP system therefore has an underlying set of pointers in the C language that are used to point to the original data within the host application cap. The user does not see this C data structure, but deals only with the public face of the Fortran cdata DDT. The type ccpp_t is defined in ccpp/framework/src/ccpp_types.meta and declared in ccpp/framework/src/ccpp_types.F90.

6.4.2. Adding and Retrieving Information from cdata (dynamic build option)

Subroutines ccpp_field_add and ccpp_field_get are part of the CCPP-Framework and are used (in the dynamic build only) to load and retrieve information to and from cdata. The calls to ccpp_field_add are auto-generated by the script ccpp_prebuild.py and inserted onto the host model code via include files (i.e. FV3/CCPP_layer/ccpp_fields_slow_physics.inc) before it is compiled.

A typical call to ccpp_field_add is below, where the first argument is the instance of cdata to which the information should be added, the second argument is the standard_name of the variable, the third argument is the corresponding host model variable, the fourth argument is an error flag, the fifth argument is the units of the variable, and the last (optional) argument is the position within cdata in which the variable is expected to be stored.

call ccpp_field_add(cdata, 'y_wind_updated_by_physics', GFS_Data(cdata%blk_no)%Stateout%gv0, ierr=ierr, units='m s-1', index=886)

For DDTs, the interface to CCPP_field_add is slightly different:

call ccpp_field_add(cdata, 'GFS_cldprop_type_instance', '', c_loc(GFS_Data(cdata%blk_no)%Cldprop), ierr=ierr, index=1)

where the first argument and second arguments bear the same meaning as in the first example, the third argument is the units (can be left empty or set to “DDT”), the fourth argument is the C pointer to the variable in memory, the fifth argument is an error flag, and the last (optional) argument is the position within cdata as in the first example.

Each new variable added to cdata is always placed at the next free position, and a check is performed to confirm that this position corresponds to the expected one, which in this example is 886. A mismatch will occur if a developer manually adds a call to ccpp_field_add, in which case a costly binary search is applied every time a variable is retrieved from memory. Adding calls manually is not recommended as all calls to ccpp_fields_add should be auto-generated.

The individual physics caps used in the dynamic build, which are auto-generated using the script ccpp_prebuild.py, contain calls to ccpp_field_get to pull data from the cdata DDT as a Fortran pointer to a variable that will be passed to the individual physics scheme.

6.4.3. Initializing and Finalizing the CCPP

At the beginning of each run, the cdata structure needs to be set up. Similarly, at the end of each run, it needs to be terminated. This is done with subroutines ccpp_init and ccpp_finalize. These subroutines should not be confused with ccpp_physics_init and ccpp_physics_finalize, which were described in Chapter 5.

Note that optional arguments are denoted with square brackets.

6.4.3.1. Suite Initialization Subroutine

The suite initialization subroutine, ccpp_init, takes three mandatory and two optional arguments. The mandatory arguments are the name of the suite (of type character), the name of the cdata variable that must be allocated at this point, and an integer used for the error status. Note that the suite initialization routine ccpp_init parses the SDF corresponding to the given suite name and initializes the state of the suite and its schemes. This process must be repeated for every element of a multi-dimensional cdata. For performance reasons, it is possible to avoid repeated reads of the SDF and to have a single state of the suite shared between the elements of cdata. To do so, specify an optional argument variable called cdata_target = X in the call to ccpp_init, where X refers to the instance of cdata that has already been initialized.

For a given suite name XYZ, the name of the suite definition file is inferred as suite_XYZ.xml, and the file is expected to be present in the current run directory. It is possible to specify the optional argument is_filename=.true. to ccpp_init, which will treat the suite name as an actual file name (with or without the path to it).

Typical calls to ccpp_init are below, where ccpp_suite is the name of the suite, and ccpp_sdf_filepath the actual SDF filename, with or without a path to it.

call ccpp_init(trim(ccpp_suite), cdata, ierr)
call ccpp_init(trim(ccpp_suite), cdata2, ierr, [cdata_target=cdata])
call ccpp_init(trim(ccpp_sdf_filepath), cdata, ierr, [is_filename=.true.])

6.4.3.2. Suite Finalization Subroutine

The suite finalization subroutine, ccpp_finalize, takes two arguments, the name of the cdata variable that must be de-allocated at this point, and an integer used for the error status. A typical call to ccpp_finalize is below:

call ccpp_finalize(cdata, ierr)

If a specific data instance was used in a call to ccpp_init, as in the above example in Section 6.4.3.1, then this data instance must be finalized last:

call ccpp_finalize(cdata2, ierr)
call ccpp_finalize(cdata, ierr)

6.4.4. Running the physics

The physics is invoked by calling subroutine ccpp_physics_run. This subroutine is part of the CCPP API and is included with the CCPP-Framework (for the dynamic build) or auto-generated (for the static build). This subroutine is capable of executing the physics with varying granularity, that is, a single scheme (dynamic build only), a single group, or an entire suite can be run with a single subroutine call. Typical calls to ccpp_physics_run are below, where scheme_name and group_name are optional and mutually exclusive (dynamic build), and where suite_name is mandatory and group_name is optional (static build).

Dynamic build:

call ccpp_physics_run(cdata, [group_name], [scheme_name], ierr=ierr)

Static build:

call ccpp_physics_run(cdata, suite_name, [group_name], ierr=ierr)

6.4.5. Initializing and Finalizing the Physics

Many (but not all) physical parameterizations need to be initialized, which includes functions such as reading lookup tables, reading input datasets, computing derived quantities, broadcasting information to all MPI ranks, etc. Initialization procedures are typically done for the entire domain, that is, they are not subdivided by blocks. Similarly, many (but not all) parameterizations need to be finalized, which includes functions such as deallocating variables, resetting flags from initialized to non-initiaIized, etc. Initialization and finalization functions are each performed once per run, before the first call to the physics and after the last call to the physics, respectively.

The initialization and finalization can be invoked for a single parameterization (only in dynamic build), for a single group, or for the entire suite. In all cases, subroutines ccpp_physics_init and ccpp_physics_finalize are used and the arguments passed to those subroutines determine the type of initialization.

These subroutines should not be confused with ccpp_init and ccpp_finalize, which were explained previously.

6.4.5.1. Subroutine ccpp_physics_init

This subroutine is part of the CCPP API and is included with the CCPP-Framework (for the dynamic build) or auto-generated (for the static build). It cannot contain thread-dependent information but can have block-dependent information. Typical calls to ccpp_physics_init are below.

Dynamic build:

call ccpp_physics_init(cdata, [group_name], [scheme_name], ierr=ierr)

Static build:

call ccpp_physics_init(cdata, suite_name, [group_name], ierr=ierr)

6.4.5.2. Subroutine ccpp_physics_finalize

This subroutine is part of the CCPP API and is included with the CCPP-Framework (for the dynamic build) or auto-generated (for the static build). Typical calls to ccpp_physics_finalize are below.

Dynamic build:

call ccpp_physics_finalize(cdata, [group_name], [scheme_name], ierr=ierr)

Static build:

call ccpp_physics_finalize(cdata, suite_name, [group_name], ierr=ierr)

6.5. Host Caps

The purpose of the host model cap is to abstract away the communication between the host model and the CCPP-Physics schemes. While CCPP calls can be placed directly inside the host model code (as is done for the relatively simple SCM), it is recommended to separate the cap in its own module for clarity and simplicity (as is done for the UFS Atmosphere). While the details of implementation will be specific to each host model, the host model cap is responsible for the following general functions:

  • Allocating memory for variables needed by physics

    • All variables needed to communicate between the host model and the physics, and all variables needed to communicate among physics schemes, need to be allocated by the host model. The latter, for example for interstitial variables used exclusively for communication between the physics schemes, are typically allocated in the cap.

  • Allocating the cdata structure(s)

    • For the dynamic build, the cdata structure handles the data exchange between the host model and the physics schemes, while for the static build, cdata is utilized in a reduced capacity.

  • Calling the suite initialization subroutine

    • The suite must be initialized using ccpp_init.

  • Populating the cdata structure(s)

    • For the dynamic build, each variable required by the physics schemes must be added to the cdata structure (or to each element of a multi-dimensional cdata) on the host model side using subroutine ccpp_field_add. This is an automated task accomplished by inserting a preprocessor directive

      #include ccpp_modules.inc
      

      at the top of the cap (before implicit none) to load the required modules and a second preprocessor directive

      #include ccpp_fields.inc
      

      after the cdata variable and the variables required by the physics schemes are allocated and after the call to ccpp_init for this cdata variable. For the static build, this step can be skipped because the autogenerated caps for the physics (groups and suite caps) are automatically given memory access to the host model variables and they can be used directly, without the need for a data structure containing pointers to the actual variables (which is what cdata is).

      Note

      The CCPP-Framework supports splitting physics schemes into different sets that are used in different parts of the host model. An example is the separation between slow and fast physics processes for the GFDL microphysics implemented in the UFS Atmosphere: while the slow physics are called as part of the usual model physics, the fast physics are integrated in the dynamical core. The separation of physics into different sets is determined in the CCPP prebuild configuration for each host model (see Chapter 5.1, and Figure 8.1), which allows to create multiple include files (e.g. ccpp_fields_slow_physics.inc and ccpp_fields_fast_physics.inc that can be used by different cdata structures in different parts of the model). This is a highly advanced feature and developers seeking to take further advantage of it should consult with GMTB first.

  • Providing interfaces to call the CCPP

    • The cap must provide functions or subroutines that can be called at the appropriate places in the host model time integration loop and that internally call ccpp_init, ccpp_physics_init, ccpp_physics_run, ccpp_physics_finalize and ccpp_finalize, and handle any errors returned See Listing 6.4.

module example_ccpp_host_cap

 use ccpp_api,           only: ccpp_t, ccpp_init, ccpp_finalize
 use ccpp_static_api,    only: ccpp_physics_init, ccpp_physics_run,     &
                               ccpp_physics_finalize

  implicit none
  ! CCPP data structure
  type(ccpp_t), save, target :: cdata
  public :: physics_init, physics_run, physics_finalize
contains

 subroutine physics_init(ccpp_suite_name)
   character(len=*), intent(in) :: ccpp_suite_name
   integer :: ierr
   ierr = 0

   ! Initialize the CCPP framework, parse SDF
   call ccpp_init(trim(ccpp_suite_name), cdata, ierr=ierr)
   if (ierr/=0) then
     write(*,'(a)') "An error occurred in ccpp_init"
     stop
   end if

   ! Initialize CCPP physics (run all _init routines)
   call ccpp_physics_init(cdata, suite_name=trim(ccpp_suite_name),      &
                          ierr=ierr)
   ! error handling as above

 end subroutine physics_init

 subroutine physics_run(ccpp_suite_name, group)
   ! Optional argument group can be used to run a group of schemes      &
   ! defined in the SDF. Otherwise, run entire suite.
   character(len=*),           intent(in) :: ccpp_suite_name
   character(len=*), optional, intent(in) :: group

   integer :: ierr
   ierr = 0

   if (present(group)) then
      call ccpp_physics_run(cdata, suite_name=trim(ccpp_suite_name),    &
                            group_name=group, ierr=ierr)
   else
      call ccpp_physics_run(cdata, suite_name=trim(ccpp_suite_name),    &
                            ierr=ierr)
   end if
   ! error handling as above

 end subroutine physics_run

 subroutine physics_finalize(ccpp_suite_name)
   character(len=*), intent(in) :: ccpp_suite_name
   integer :: ierr
   ierr = 0

   ! Finalize CCPP physics (run all _finalize routines)
   call ccpp_physics_finalize(cdata, suite_name=trim(ccpp_suite_name),  &
                              ierr=ierr)
   ! error handling as above
   call ccpp_finalize(cdata, ierr=ierr)
   ! error handling as above

 end subroutine physics_finalize

end module example_ccpp_host_cap

Listing 6.4: Fortran template for a CCPP host model cap from ccpp/framework/doc/DevelopersGuide/host_cap_template.F90.

The following sections describe two implementations of host model caps to serve as examples. For each of the functions listed above, a description for how it is implemented in each host model is included.

6.5.1. SCM Host Cap

The only build type supported for the SCM v4 is the static build. The cap functions are mainly implemented in:

gmtb-scm/scm/src/gmtb_scm.F90

With smaller parts in:

gmtb-scm/scm/src/gmtb_scm_type_defs.f90

gmtb-scm/scm/src/gmtb_scm_setup.f90

gmtb-scm/scm/src/gmtb_scm_time_integration.f90

The host model cap is responsible for:

  • Allocating memory for variables needed by physics

    All variables and constants required by the physics have metadata provided on the host-side, arg_table_physics_type and arg_table_gmtb_scm_physical_constants, which are implemented in gmtb_scm_type_defs.f90 and gmtb_scm_physical_constants.f90. To mimic the UFS Atmosphere and to hopefully reduce code maintenance, currently, the SCM uses GFS DDTs as sub-types within the physics DDT.

    In gmtb_scm_type_defs.f90, the physics DDT has a create type-bound procedure (see subroutine physics_create and type physics_type), which allocates GFS sub-DDTs and other physics variables and initializes them with zeros. physics%create is called from gmtb_scm.F90 after the initial SCM state has been set up.

  • Allocating the cdata structure

    The SCM uses a one-dimensional cdata array for N independent columns, i.e. in gmtb_scm.F90:

    allocate(cdata(scm_state%n_cols))

  • Calling the suite initialization subroutine

    Within scm_state%n_cols loop in gmtb_scm.F90 after initial SCM state setup and before first timestep, the suite initialization subroutine ccpp_init is called for each column with own instance of cdata, and takes three arguments, the name of the runtime SDF, the name of the cdata variable that must be allocated at this point, and ierr.

  • Populating the cdata structure

    Within the same scm_state%n_cols loop, but after the ccpp_init call, the cdata structure is filled in with real initialized values:

  • physics%Init_parm (GFS DDT for setting up suite) are filled in from scm_state%

  • call GFS_suite_setup(): similar to GFS_initialize() in the UFS Atmosphere, is called and includes:

  • %init/%create calls for GFS DDTs

  • initialization for other variables in physics DDT

  • init calls for legacy non-ccpp schemes

  • call physics%associate(): to associate pointers in physics DDT with targets in scm_state, which contains variables that are modified by the SCM “dycore” (i.e. forcing).

  • Actual cdata fill in through ccpp_field_add calls:

#include “ccpp_fields.inc”

This include file is auto-generated from ccpp/scripts/ccpp_prebuild.py, which parses tables in gmtb_scm_type_defs.f90.

  • Providing interfaces to call the CCPP

  • Calling ccpp_physics_init()

Within the same scm_state%n_cols loop but after cdata is filled, the physics initialization routines (*_init()) associated with the physics suite, group, and/or schemes are called at each column.

  • Calling ccpp_physics_run()

At the first timestep, if the forward scheme is selected (i.e. scm_state%time_scheme == 1), call do_time_step() to apply forcing and ccpp_physics_run() calls at each column; if the leapfrog scheme is selected (i.e. scm_state%time_scheme == 2), call ccpp_physics_run() directly at each column.

At a later time integration, call do_time_step() to apply forcing and ccpp_physics_run() calls at each column. Since there is no need to execute anything between physics groups in the SCM, the ccpp_physics_run call is only given cdata and an error flag as arguments.

  • Calling ccpp_physics_finalize() and ccpp_finalize()

ccpp_physics_finalize() and ccpp_finalize() are called after the time loop at each column.

6.5.2. UFS Atmosphere Host Cap

For the UFS Atmosphere, there are slightly different versions of the host cap implementation depending on the desired build type (dynamic or static). Only the static build is publicly supported at this time, as the dynamic build will be deprecated soon. As discussed in Chapter 8, these modes are controlled via appropriate strings included in the MAKEOPTS build-time argument. Within the source code, the three modes are executed within appropriate pre-processor directive blocks:

For any build that uses CCPP (dynamic orstatic):

#ifdef CCPP
#endif

For static (often nested within #ifdef CCPP):

#ifdef STATIC
#endif

The following text describes how the host cap functions listed above are implemented for the dynamic build only. Where the other modes of operation differ in their implementation, it will be called out.

  • Allocating memory for variables needed by physics

  • Within the atmos_model_init subroutine of atmos_model.F90, the following statement is executed

allocate(IPD_Data)

IPD_Data is of IPD_data_type, which is defined in IPD_typedefs.F90 as a synonym for GFS_data_type defined in GFS_typedefs.F90. This data type contains GFS-related DDTs (GFS_statein_type, GFS_stateout_type, GFS_sfcprop_type, etc.) as sub-types, which are defined in GFS_typedefs.F90.

  • Allocating the cdata structures

  • For the current implementation of the UFS Atmosphere, which uses a subset of fast physics processes tightly coupled to the dynamical core, three instances of cdata exist within the host model: cdata_tile to hold data for the fast physics, cdata_domain to hold data needed for all UFS Atmosphere blocks for the slow physics, and cdata_block, an array of cdata DDTs with dimensions of (number of blocks, number of threads) to contain data for individual block/thread combinations for the slow physics. All are defined as module-level variables in the CCPP_data module of CCPP_data.F90. The cdata_block array is allocated (since the number of blocks and threads is unknown at compile-time) as part of the ‘init’ step of the CCPP_step subroutine in CCPP_driver.F90. Note: Although the cdata containers are not used to hold the pointers to the physics variables for the static mode, they are still used to hold other CCPP-related information for that mode.

  • Calling the suite initialization subroutine

  • Corresponding to the three instances of cdata described above, the ccpp_init subroutine is called within three different contexts, all originating from the atmos_model_init subroutine of atmos_model.F90:

  • For cdata_tile (used for the fast physics), the ccpp_init call is made from the atmosphere_init subroutine of atmosphere.F90. Note: when fast physics is used, this is the first call to ccpp_init, so it reads in the SDF and initializes the suite in addition to setting up the fields for cdata_tile.

  • For cdata_domain and cdata_block used in the rest of the physics, the ‘init’ step of the CCPP_step subroutine in CCPP_driver.F90 is called. Within that subroutine, ccpp_init is called once to set up cdata_domain and within a loop for every block/thread combination to set up the components of the cdata_block array. Note: as mentioned in the CCPP API Section 6.4, when fast physics is used, the SDF has already been read and the suite is already setup, so this step is skipped and the suite information is simply copied from what was already initialized (cdata_tile) using the cdata_target optional argument.

  • Populating the cdata structures

  • When the dynamic mode is used, the cdata structures are filled with pointers to variables that are used by physics and whose memory is allocated by the host model. This is done using ccpp_field_add statements contained in the autogenerated include files. For the fast physics, this include file is named ccpp_fields_fast_physics.inc and is placed after the call to ccpp_init for cdata_tile in the atmosphere_init subroutine of atmosphere.F90. For populating cdata_domain and cdata_block, IPD data types are initialized in the atmos_model_init subroutine of atmos_model.F90. The Init_parm DDT is filled directly in this routine and IPD_initialize (pointing to GFS_initialize and for populating diagnostics and restart DDTs) is called in order to fill the GFS DDTs that are used in the physics. Once the IPD data types are filled, they are passed to the ‘init’ step of the CCPP_step subroutine in CCPP_driver.F90 where ccpp_field_add statements are included in ccpp_fields_slow_physics.inc after the calls to ccpp_init for the cdata_domain and cdata_block containers.

  • Note: for the static mode, filling of the cdata containers with pointers to physics variables is not necessary. This is because the autogenerated caps for the physics groups (that contain calls to the member schemes) can fill in the argument variables without having to retrieve pointers to the actual data. This is possible because the metadata about host model variables (that are known at ccpp_prebuild time) contain all the information needed about the location (DDTs and local names) to pass into the autogenerated caps for their direct use.

  • Providing interfaces to call the CCPP

  • Calling ccpp_physics_init

  • In order to call the initialization routines for the physics, ccpp_physics_init is called in the atmosphere_init subroutine of atmosphere.F90 after the included ccpp_field_add calls for the fast physics. For the slow physics, the ‘physics_init’ step of the CCPP_step subroutine in CCPP_driver.F90 is invoked immediately after the call to the ‘init’ step in the atmos_model_init subroutine of atmos_model.F90. Within the ‘physics_init’ step, calls to ccpp_physics_init for all blocks are executed.

  • Note: for the static mode, ccpp_physics_init is autogenerated and contained within ccpp_static_api.F90. As mentioned in the CCPP API Section 6.4 , it can be called to initialize groups as defined in the SDFs or the suite as a whole, depending on whether a group name is passed in as an optional argument.

  • Calling ccpp_physics_run

  • For actually running the physics within the FV3 time loop, ccpp_physics_run is called from a couple of different places in the FV3 source code. For the fast physics, ccpp_physics_run is called for the fast physics group from the Lagrangian_to_Eulerian subroutine of fv_mapz.F90 within the dynamical core. For the rest of the physics, the subroutine update_atmos_radiation_physics in atmos_model.F90 is called as part of the FV3 time loop. Within that subroutine, the various physics steps (defined as groups within a SDF) are called one after the other. The ‘time_vary’ step of the CCPP_step subroutine within CCPP_driver.F90 is called. Since this step is called over the entire domain, the call to ccpp_physics_run is done once using cdata_domain and the time_vary group. The ‘radiation’, ‘physics’, and ‘stochastics’ steps of the CCPP_step subroutine are called next. For each of these steps within CCPP_step, there is a loop over the number of blocks for calling ccpp_physics_run with the appropriate group and component of the cdata_block array for the current block and thread.

  • Note: The execution of calls to ccpp_physics_run is different for the three build types. For the static mode, ccpp_physics_run is called from ccpp_static_api.F90 and contains autogenerated caps for groups and the suite as a whole as defined in the SDFs.

  • calling ccpp_physics_finalize and ccpp_finalize

  • At the conclusion of the FV3 time loop, calls to finalize the physics are executed. For the fast physics, ccpp_physics_finalize is called from the atmosphere_end subroutine of atmosphere.F90. For the rest of the physics, the ‘finalize’ step of the CCPP_step subroutine in CCPP_driver.F90 is called from the atmos_model_end subroutine in atmos_model.F90. Within the ‘finalize’ step of CCPP_step, calls for ccpp_physics_finalize and ccpp_finalize are executed for every thread and block for cdata_block. Afterward, ccpp_finalize is called for cdata_domain and lastly, cdata_tile. (That is, the calls to ccpp_finalize are in reverse order than the calls to ccpp_initialize.) In addition, cdata_block is also deallocated in the ‘finalize’ step of CCPP_step.

  • Note: for the static mode, ccpp_physics_finalize is autogenerated and contained within ccpp_static_api.F90. As mentioned in the CCPP API Section 6.4, it can be called to finalize groups as defined in the current SDFs or the suite as a whole, depending on whether a group name is passed in as an optional argument.