Vax C90-EL SERIES Service Manual download pdf (Page 85)

#pragma Directives [3]

For debugging purposes, the names of slave functions are in the following form:

__tsk_name_nnn

The name is the name of the original function that contained the parallel region,

and nnn is a 3-digit number indicating how many parallel regions preceded

this one in the current function. If the name is longer than 200 characters, only

the first 200 characters are used when the slave function’s name is created.

Numbers are assigned beginning with 000 and are incremented by 1 for each

additional parallel region in that function. This counter is reset to 000 for each

new function. For example, __tsk_fctn_002 is the name of the slave function

generated for the third parallel region in function fctn.

3.8.1.1.3 Unitasked Code

The unitasked code contains only the original parallel region. No tasking code

is present in the unitasked code. The unitasked code is executed by the master

task if it determines, at run time, that tasking should not be performed. (For

example, if the expression for the if argument evaluates false or if tasking

is already being done at a higher level in the code.) If the unitasked code is

executed, no slave processors are used. The term unitasked code should not be

confused with nontasked code. Nontasked code refers to any program that does

not have parallel regions, or to the part of the tasked program that is executed

outside of any parallel region.

3.8.1.2 Tasking Initialization

A function that potentially executes tasking code requires some initialization.

This initialization sets up the arguments for each slave function. These arguments

are usually the variables that the slave task shares with the master task and

with other slave tasks.

The execution cost for this code is usually small. For functions that contain

a large number of distinct parallel regions, however, the cost may become

significant.

3.8.1.3 Vectorization with Tasking (Stripmining)

The chunksize, numchunks, guided, and vector arguments to the

taskloop directive can be specified to allow stripmining. Stripmining refers to

the partitioning of long vector for loops into shorter vector chunks and the

processing of these chunks through tasked iterations. While stripmining can

cause wall-clock time to decrease, it usually increases CPU time because of the

reduced vector performance resulting from the smaller vector lengths involved.

S–2179–36 65

1 2 ... 80 81 82 83 84 85 86 87 88 89 90 ... 215 216

No comments

Vax C90-EL SERIES Service Manual Page 85