Cypress CSC-1200T Guia do Utilizador descarregar pdf (Página 99)

Chapter 9. Programming tools 99

performance information and instruction counts.

PAT is able to analyze programs written in Fortran 90, C, C++ and HPF.

The executable only needs to relinked, no recompiling is necessary. The

linker option -l pat along with the PAT speciﬁc cld ﬁle pat.cld are

required.

As an example, suppose that a Fortran 90 program in the ﬁle prog.f90

is to be analyzed. The following commands can be used:

t3e% f90 -c prog.f90

t3e% f90 prog.o -o prog -l pat pat.cld

C, C++ and HPF programs are linked similarly. A log ﬁle of the type

pdf.1234 is produced during the execution.

Timing information is then displayed with the command

t3e% pat -T prog pdf.1234

Sample output for a four PE run looks like

Elapsed Time 4.229 sec 4 PEs

User Time (ave) 3.441 sec 81%

System Time (ave) 0.023 sec 1%

The PAT option -m produces performance counter statistics:

t3e% pat -m prog pdf.1234

Performance counters for FpOps

Values given are in MILLIONS.

PE cycles operations ops/sec dcache misses/sec

misses

0 425.34 152.45 134.39 5.48 4.83

1 1574.81 152.40 36.29 5.61 1.34

2 1574.87 152.40 36.28 5.62 1.34

3 1575.15 152.40 36.28 5.62 1.34

The column ops/sec contains the ﬂoating point performance given in

Mﬂop/s for each PE. A high cache miss rate can be caused by less than

optimal program design. Instead of ﬂoating point operations it is pos-

sible to measure integer performance by setting

t3e% setenv PAT_SEL INTOPS

Memory load and store operations can also be monitored.

Proﬁle information can be obtained with the option -p. Normally only

subroutine-level proﬁle is available. To produce a line-level proﬁle the

1 2 ... 94 95 96 97 98 99 100 101 102 103 104 ... 123 124

Sem comentários

Cypress CSC-1200T Guia do Utilizador Página 99