home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
SGI Varsity Update 1998 August
/
SGI Varsity Update 1998 August.iso
/
docs6.4
/
relnotes
/
WorkShopMPF
/
ch5.z
/
ch5
Wrap
Text File
|
1998-07-29
|
6KB
|
198 lines
- 1 -
5. _K_n_o_w_n__p_r_o_b_l_e_m_s__a_n_d__W_o_r_k_a_r_o_u_n_d_s
5.1 _P_r_o_b_l_e_m_s__i_n__t_h_e__P_a_r_a_l_l_e_l__A_n_a_l_y_z_e_r__V_i_e_w
+o Updating a file with a change to one of the PFA
analysis parameters will cause a directive to be placed
immediately following the given loop, in order to
restore the previous value of the parameter for
subsequent loops. If the next loop has a directive for
the same parameter, it will generate a PFA warning to
the effect that the first directive is overriden by a
subsequent one. The warning should be ignored.
Another workaround is to remove the earlier directive
by hand after update, but before rebuilding.
+o A highlight for a variable name will miss an occurrence
of it whenever it has embedded spaces, or its name
appears immediately following an alphanumeric
continuation character in column 6 of a continuation
line, with no intervening spaces. The workaround is to
not embed blanks within a symbol, and to leave a blank
after the continuation character. Alternatively, the
WorkShop 2.4 Static Analyzer may be used to look for
variable uses.
+o Loops that are split into parallelizable and
unparallelizable portions are shown as
unparallelizable; consequently an inner loop nested in
the parallelizable portion will be shown as serial-
nested-in-parallel, although it appears to be nested in
an unparallelizable loop. This situation can be
resolved by looking at the transformed code, and the
Transformed Loops View.
5.2 _R_e_q_u_e_s_t_e_d__E_n_h_a_n_c_e_m_e_n_t_s__t_o__t_h_e__P_a_r_a_l_l_e_l__A_n_a_l_y_z_e_r__V_i_e_w
+o Save-as-Text currently saves the internal tables from
the program to a file, rather than a more user-friendly
formatted display of the data on screen. We welcome
your suggestions as to formatting the data for end-
users.
+o There should be a way to double-click in the
performance analyzer and have the function selected in
cvpav. The workaround is to select only from the
function list of cvpav.
+o _c_v_p_a_v needs a method to show loop nesting for source
files that are included in other files.
- 2 -
+o Performance annotations should allow the user to choose
between inclusive and exclusive times for the displays.
5.3 _C_o_m_p_i_l_e_r__P_r_o_b_l_e_m_s__M_a_n_i_f_e_s_t_e_d__i_n__W_o_r_k_S_h_o_p__P_r_o__M_P_F
+o When nested loops are strip-mined, with any of the
transformed loops parallelized, other loops may be
shown as unparallelizable, when, in fact, they were
best left serial. This will occur if there was a data
dependence that could be broken, such as a reduction.
The compiler should not issue the data dependence
message.
+o If a program uses a variable that has the same name as
a Fortran intrinsic, PFA will change its name to avoid
the conflict. Messages describing this substitution
are not emitted, but should be, and the highlighting
associated with a data dependence message uses the new
name, not the old name that is visible in the source.
+o No support is given for understanding interprocedural
analysis (IPA) or inlining.
+o Sometimes a loop is shown as serial but there is no
obstacle to parallelization shown. Known cases involve
complex control flow in an inner loop, either a
premature exit, or a label assignment, although they
should not affect the parallelization of the outer
loop.
5.4 _W_o_r_k_S_h_o_p__P_r_o_b_l_e_m_s__M_a_n_i_f_e_s_t_e_d__i_n__P_r_o__M_P_F
The following are problems in WorkShop that you may
encounter in Pro MPF:
+o The WorkShop Performance Analyzer produces misleading
data on MP executables (see the _W_o_r_k_S_h_o_p _R_e_l_e_a_s_e _N_o_t_e_s
for more information). The symptoms are that loop-
level performance counts for 64-bit experiments may
contain some double counting. This will show up as a
time greater than the total experiment time, and a
percentage written as a <n/a>.
+o The SourceView loop brackets should use a line-style
variation on monochrome displays.
+o SourceView does not always scroll properly to make
loops visible.
+o Double clicking on the icon describing a loop in the
loop list does not select the loop. It should, and the
- 3 -
workaround is to double click on the text, not the
icon.