TAU for HPF
The initial target for the TAU tools was the object-parallel programming
language pC++. We retargeted some of the TAU tools (fancy,
cagey, and racy; see below) for High Performance
Fortran (HPF). The other TAU tools (cosy, classy, and
speedy) are pC++ / C++ specific. We plan to apply breezy
to HPF as well.
The current TAU for HPF implementation is based on the HPF compiler available
from The Portland Group, Inc. (PGI).
Their pghpf compiler already supports profiling, so in order to
use the TAU for HPF tools, the user only has to link his/her HPF program
with the TAU implementation of the profiling runtime system functions
(which generates TAU compatible profiling data output as well as the
file, function, and callgraph information for fancy
and cagey, if necessary). However, the static information can
also be generated by a small HPF file and function scanner (which is
implemented as an extra phase after the normal HPF compilation steps).
This is the preferred way, as it allows program browsing before/without
running the program and it keeps the runtime overhead small.
The TAU tools did not have to be changed.
The example used in the screendumps below is PDE1, one program of
the GENESIS Distributed Memory Benchmark suite written in Subset HPF.
The benchmark solves the Poisson-Equation on a 3-dimensional
grid by parallel red-black succesive over relaxation with Chebyshev
acceleration. The benchmark ran on a 8 processor SGI PowerChallenge.
FANCY
FANCY (File ANd Class displaY)
lets you browse through the files used in the source text of the
application, and lets you display the source text of subroutine and
functions. The header of the selected routine is shown in orange (not
visible in the screendump), the body is shown in yellow.
Within the body of a routine calls to other routines
are shown in light blue, calls to F90 and HPF intrinsics in purple.
CAGEY
CAGEY (CAll Graph Extended displaY)
lets you browse through the static callgraph of the application.
The currently selected functions is marked with a red box.
Names of F90 and HPF intrinsics are shown within
<...>
and can be hidden on request.
RACY
RACY (Routine and data ACcess profile displaY)
is a parallel profile data viewer. After compiling an application
for profiling and running it, racy lets you browse through the
function profile data generated. In addition to the summary function
profile display in the racy main window, node and function
summary windows are supported.
We would like to thank PGI, as without their support this work would not
be possible.
Next: TAU Publications
Up: TAU Home Page
Prev: TAU Overview
mohr@cs.uoregon.edu
April 13, 1995