4.9 Optimization Calls
The following procedures are included to aid in optimization of the
QDP implementation
Marking discarded data
| Syntax | void QDP_discard_T(Type *a);
| | Meaning | Indicates data in a is no longer needed.
| | Type | I, R, C, V, H, D, M, P
| | Example | QDP_discard_M(utemp);
|
|
The field is not destroyed and memory is not released. For that
purpose, see QDP_destroy. This call allows the implementation
to ignore the data dependency that may have been created from a shift
which allows the shift to be efficiently reused later.
It is a runtime error to attempt to use discarded data as an rvalue
(source operand or incremented destination) in any subsequent operation.
However, once the field is used as an lvalue (fully replaced destination),
data integrity is automatically reinstated.
This should be called on the result of a shift once the result is no longer
needed.
It is safe to call it for any field as long as the data is no longer needed.
It is highy recommended for the sake of efficiency that one always call
this as soon as the result of a shift is no longer needed.
It can also be used to make sure that a field is not being used elsewhere
without first being set since using a discarded field without first setting it
produces an error.
Block size operations
| Syntax | int QDP_get_block_size(void);
| | Meaning | Returns the block size.
|
|
| Syntax | void QDP_set_block_size(int bs);
| | Meaning | Sets the block size.
|
|
These functions allow one to set the block size used for the operations
on arrays of QDP fields. Instead of running over all sites of each field
individually, it can run over a few sites for each field in turn and then
move to the next few sites on each field and so on until all sites are done.
The number of sites done at a time is determined by the block size.
Memory alignment
| Syntax | int QDP_get_mem_align(void);
| | Meaning | Returns the default memory alignment.
|
|
| Syntax | void QDP_set_mem_align(int align);
| | Meaning | Sets the default memory alignment.
|
|
Some architectures might perform better for fields with certain alignments.
This allows one to set the default alignment used when fields are created.
It only affects newly created fields.
Currently the default is set to 16 bytes if not set by the user which
should be sufficient for most current platforms.
It is expected that the QDP library will automatically use whatever value
is necessary to utilize the optimized QLA routines for architectures
which QLA has been optimized for.
Therefore the user normally would not need to set this and it is recommended
that they don't unless they specifically find it useful for a platform
not already optimized in QLA.
The user can also use the macro QDP_ALIGN_DEFAULT to specify the
default alignment provided by QMP for that machine.
Memory attributes
| Syntax | int QDP_get_mem_flags(void);
| | Meaning | Returns the default memory flags
|
|
| Syntax | void QDP_set_mem_flags(int flags);
| | Meaning | Sets the default memory flags
| | Example | QDP_set_mem_flags(QDP_MEM_FAST|QDP_MEM_COMMS);
|
|
This allows the user to set the characteristics of the memory used for
newly created fields if the architecture supports them.
The QDP memory flags have the same meaning as the corresponding QMP flags.
The integer value is the combination (bitwise or) of the flags given below.
The actual effect of each combination is specific to the implementation and
the architecture.
| Memory Flag | Meaning
|
| QDP_MEM_NONCACHE | Use noncached memory.
|
| QDP_MEM_COMMS | Use memory that is optimized for communucations.
|
| QDP_MEM_FAST | Use fastest memory available (such as EDRAM on QCDOC).
|
| QDP_MEM_DEFAULT | Use default value for architecture (from QMP).
|