Vax C90-EL SERIES Service Manual download pdf (Page 86)

Cray Standard C/C++ Reference Manual

Besides executing the vector loop, each processor must execute some loop

overhead code for each chunk of the work.

If you specify the chunksize, numchunks, guided,orvector arguments, the

compiler rewrites the loop as two nested loops. The inner loop becomes a for

loop that processes the iterations in a vector chunk. This inner loop will be

vectorized, subject to the normal rules for vectorization of loops. The outer loop

becomes tasking overhead code that processes each of the chunks.

Usually, when you specify the ivdep directive prior to the taskloop directive

the ivdep is applied to the original, unitasked version of the loop. When the

ivdep directive appears prior to a taskloop directive containing a chunksize,

numchunks, guided,orvector argument, the inner for loops (as previously

described) in both the master and slave code and the original, unitasked version

of the loop are treated as though the ivdep directive had been specified prior

to these loops.

3.8.1.4 Vectorization Messages and Tasking

When vectorization messages are enabled by specification of the -h report=v

option on the cc command line, information is provided regarding all loops

in the master, slave, and unitasked versions of the tasked loop. In the master

code and slave function, the tasked loop no longer appears as a loop; thus, no

vectorization messages are generated. However, the original loop is unaltered in

the unitasked code, and a vectorization message will appear for the loop. Any

loops inside a tasked loop produce a vectorization message for each occurrence

in the master, slave, and unitasked versions of the tasked loop. The loops

generated from stripmining (see Section 3.8.1.3, page 65) also cause vectorization

messages to be generated.

The line numbers shown in the vectorization messages for loops in the tasked

code are the line numbers corresponding to the equivalent loop in the original

code, with the m suffix for messages from the master code and an s suffix for

messages from the slave code. Line numbers for messages from the unitasked

code have no suffix.

3.8.1.5 Reserved Semaphores and Shared Registers

Because the following semaphores and shared registers are used by tasking on

Cray PVP systems, you must not use them in your program:

• Semaphore registers 0-15

• Shared B registers 0-3

66 S–2179–36

1 2 ... 81 82 83 84 85 86 87 88 89 90 91 ... 215 216

No comments

Vax C90-EL SERIES Service Manual Page 86