| Name: | typebind - optimizes procedure for certain data type
|
| Synopsis: | proc dict typebind -> proc
|
| Description: |
The operator is called on a procedure and optimizes the function
calls in this procedure by removing type checking of the overloaded
functions according to the information in the dictionary. The dictionary
specifies the names of the functions for which the type of the arguments
is knon and the expected combination of argument types. typebind uses this
information to replace the function calls in the procedure by the appropriate
variant. This reduces the number of interpreter cycles required to evaluate
the procedure. However, the bigger effect is the reduction of run time.
The operator typebind can be used to optimize SLI code for speed in
situations where it is guaranteed that a particular variant of a
function is needed. Here, replacing the overloaded function by the
specific variant removes the overhead of type checking. A typical
situation in which the data types of the arguments are known is a
function call inside a function with a specified signature (see
example below). Another situation is machine generated SLI code where
the machine, for example a compiler, is requested to emit optimized
code restricted to particular data type.
/incr [/doubletype]
{
1.0 add
}
<< /add [/doubletype /doubletype] >> typebind
def
The example below achieves a reduction of interpreter cycles by 20%
but a reduction of rune time by 60%.
|
| Parameters: |
proc, is the procedure to be optimized
dict, contains as keys the names of the operators that should
be constrained to a certain combination of argument types.
The value associated to the key is an array specifying the
data type of each argument.
|
| Examples: |
The following is the closed form expression of a postsynaptic potential at
at time t
(
weight * E/tau_syn * 1./C_m
* ( (exp(-t/tau_m)-exp(-t/tau_syn)) / (1./tau_syn - 1./tau_m)^2.
- t*exp(-t/tau_syn) / (1./tau_syn - 1./tau_m)
)
)
CompileMath /v Set
The expression operates on doubles. Therefore we optimize with
/v load
<<
/add [/doubletype /doubletype]
/sub [/doubletype /doubletype]
/mul [/doubletype /doubletype]
/div [/doubletype /doubletype]
/pow [/doubletype /doubletype]
/exp [/doubletype ]
/neg [/doubletype ]
>> typebind /v_d Set
We can further optimize by replacing all symbolic constants by the
corresponding values
/v_d load
<<
/weight 3.2
/tau_syn 0.5
/tau_m 10.0
/C_m 250.0
/t 1.345
/E E
>> Inline /v_c Set
Let us now explore the required number of interpreter cycles and the run time
of the three versions
clic v cloc
clic v_d cloc
clic v_c cloc
tic 500000 {v pop} repeat toc
tic 500000 {v_d pop} repeat toc
tic 500000 {v_c pop} repeat toc
The results are:
v v_d v_c
----------------------------------
cycles 138 112 97
reduction factor 1.2 1.4
20% 30%
----------------------------------
time 20.8 8.3 6.5
reduction factor 2.5 3.2
60% 69%
----------------------------------
|
| Remarks: |
typebind may be implemented by Inline, see the documentation of variant.
|
| Author: | Diesmann
|
| FirstVersion: | 090210
| SeeAlso: | variant Inline CompileMath def clic tic |
|
| Source: | /home/abuild/rpmbuild/BUILD/nest-2.4.1/lib/sli/typeinit.sli
|