Code: Select all
-O1 -march=pentium4 -pipe -fno-omit-frame-pointer -finline-functions -fno-delayed-branch -fcse-follow-jumps -mnopush-args -freorder-blocks -falign-jumps -minline-all-stringop -fno-merge-constants -falign-labels -frename-registers -fno-defer-pop -fsched-spec -fno-signaling-nans
"-fno-signaling-nans" might screw up math libraries that require proper reporting of floating point exceptions. I would omit it if you do scientific work ...metal leper wrote:Are any of these dangerous? I got -mieee-fp and -malign-double as yes too, but I read that I shouldn't use theseCode: Select all
-O1 -march=pentium4 -pipe -fno-omit-frame-pointer -finline-functions -fno-delayed-branch -fcse-follow-jumps -mnopush-args -freorder-blocks -falign-jumps -minline-all-stringop -fno-merge-constants -falign-labels -frename-registers -fno-defer-pop -fsched-spec -fno-signaling-nans
Scott -- I'm curious how mainline performs on all these tests with the
patch below.
Code: Select all
diff -rup orig/egcc-CVS20040502/gcc/c-cppbuiltin.c egcc-CVS20040502/gcc/c-cppbuiltin.c
--- orig/egcc-CVS20040502/gcc/c-cppbuiltin.c Thu Apr 1 20:03:20 2004
+++ egcc-CVS20040502/gcc/c-cppbuiltin.c Sun May 2 10:49:49 2004
@@ -384,6 +384,10 @@ c_cpp_builtins (cpp_reader *pfile)
if (optimize)
cpp_define (pfile, "__OPTIMIZE__");
+ /* GCC's builtins tend to do better than GLIBC's inlines. */
+ cpp_define (pfile, "__NO_STRING_INLINES");
+ cpp_define (pfile, "__NO_MATH_INLINES");
+
if (fast_math_flags_set_p ())
cpp_define (pfile, "__FAST_MATH__");
if (flag_really_no_inline)Code: Select all
GCC Mainline, Pentium 4
default math with -mfpmath=sse
test no patch patched no patch patched
---- -------- -------- -------- --------
alma 64.3 42.4 68.2 44.7
evo 52.9 52.8 79.5 79.8
fft 34.6 32.0 34.5 33.4
huff 16.0 16.0 16.7 16.3
lin 26.2 22.9 25.9 23.7
mat1 7.5 7.5 8.0 7.9
mole 32.8 31.3 30.6 22.4
tree 25.0 24.5 25.7 25.6
-------- -------- -------- --------
total 259.3 229.5 289.0 253.9Code: Select all
GCC 3.4.0-r2, Pentium III
Score | So? | Switch (annotation)
------------------------------------------------------------------------------
46.7 | Yes | -fstrict-aliasing (-O2)
35.0 | Yes | -mno-align-stringops
28.9 | Yes | -freorder-functions (-O2 GCC 3.3)
28.7 | Maybe | -funsafe-math-optimizations (fast math)
28.6 | Yes | -minline-all-stringops
28.2 | Yes | -fsched-interblock (-O2 GCC 3.3)
28.1 | Maybe | -fpeephole2 (-O2)
28.0 | Maybe | -frename-registers (-O3)
27.4 | Yes | -fno-crossjumping (! -O1)
27.2 | Yes | -falign-functions (-O2 GCC 3.4)
26.9 | Maybe | -fgcse (-O2)
26.6 | Yes | -fdelete-null-pointer-checks (-O2)
26.5 | Maybe | -finline-functions (-O3)
26.1 | Yes | -fmove-all-movables
26.1 | Maybe | -fno-math-errno (fast math)
26.1 | Yes | -fexpensive-optimizations (-O2)
26.1 | Maybe | -finline-limit
25.9 | Yes | -mno-push-args
25.6 | Maybe | -fforce-mem (-O2)
25.4 | Yes | -fno-delayed-branch (! -O1)
25.4 | Maybe | -fpeel-loops
25.3 | Maybe | -maccumulate-outgoing-args
25.3 | Maybe | -fno-trapping-math (fast math)
25.2 | Maybe | -frerun-cse-after-loop (-O2)
25.1 | Yes | -falign-loops (-O2 GCC 3.3)
24.9 | Maybe | -fno-signaling-nans (fast math)
24.8 | Yes | -fsched-spec (-O2 GCC 3.3)
24.3 | Yes | -malign-double
23.9 | Maybe | -fcaller-saves (-O2)
23.0 | Yes | -falign-labels (-O2 GCC 3.3)
23.0 | Maybe | -freorder-blocks (-O2)
22.8 | Maybe | -frerun-loop-opt (-O2)
22.8 | Yes | -fno-thread-jumps (! -O1)
22.5 | Maybe | -fcse-follow-jumps (-O2)
22.2 | Maybe | -fno-merge-constants (! -O1)
21.6 | No | -fweb (-O3 GCC 3.4)
21.4 | Yes | -fno-defer-pop (! -O1)
21.0 | Maybe | -freduce-all-givs
20.6 | No | -funit-at-a-time (-O2 GCC 3.4)
20.4 | Maybe | -fcse-skip-blocks (-O2)
20.1 | Maybe | -foptimize-sibling-calls (-O2)
19.8 | Maybe | -funswitch-loops
19.8 | Maybe | -fno-omit-frame-pointer (! -O1)
19.6 | Maybe | -falign-jumps (-O2 GCC 3.3)
19.6 | Maybe | -mieee-fp
19.1 | Maybe | -fregmove (-O2)
18.0 | No | -fprefetch-loop-arrays
17.8 | Maybe | -fno-if-conversion2 (! -O1)
17.2 | No | -fstrength-reduce (-O2)
16.5 | No | -fno-cprop-registers (! -O1)
16.5 | No | -fno-if-conversion (! -O1)
15.9 | Maybe | -ffinite-math-only (fast math)
15.5 | No | -ftracer
14.6 | No | -fnew-ra
14.2 | No | -fschedule-insns2 (-O2)
11.7 | No | -fschedule-insns (-O2)
10.6 | No | -ffloat-store
10.5 | No | -funroll-all-loops
9.3 | No | -fno-inline
7.6 | No | -fomit-frame-pointer
7.4 | No | -funroll-loops
7.3 | No | -fno-loop-optimize (! -O1)
6.2 | No | -fbranch-target-load-optimize
0.0 | No | -fno-guess-branch-probability (! -O1)
0.0 | No | -fbranch-target-load-optimize2
0.0 | No | -mfpmath=387
0.0 | No | -mfpmath=sse
0.0 | No | -mfpmath=sse,387
0.0 | No | -momit-leaf-frame-pointer
Code: Select all
CFLAGS="-march=pentium3 -O2 -mno-align-stringops -minline-all-stringops -fno-crossjumping -fmove-all-movables -mno-push-args -fno-delayed-branch -malign-double -fno-thread-jumps -fno-defer-pop -Wall -pipe"-malign-double is dangerous -- that can caues binary incompatibility for linking between objects.wilburpan wrote:which suggests:Correct?Code: Select all
CFLAGS="-march=pentium3 -O2 -mno-align-stringops -minline-all-stringops -fno-crossjumping -fmove-all-movables -mno-push-args -fno-delayed-branch -malign-double -fno-thread-jumps -fno-defer-pop -Wall -pipe"
It used to do that here... but not any more...ikaro wrote:..as a side note, anyone got this running with gcc 3.4 ?
here it just segfaults....
One doubt, if you activate -O2 and then turn off some of the optimizations it carries (like -fcrossjumping), gcc uses the last flag you specified (-fno-crossjumping)wilburpan wrote: which suggests:Correct?Code: Select all
CFLAGS="-march=pentium3 -O2 -mno-align-stringops -minline-all-stringops -fno-crossjumping -fmove-all-movables -mno-push-args -fno-delayed-branch -malign-double -fno-thread-jumps -fno-defer-pop -Wall -pipe"
Code: Select all
-freduce-all-givs
"Register starvation" is a non-issue on late-generation CPUs; it was when floating point and SIMD units were first added to CPU dies with the same old registers.HeartBreakKid wrote:So has anyone who has run acovea thus far gotten a Yes for "-fomit-frame-pointer"? Most of the results I have seen so far appear to suggest that this is not as good as (we all) may have believed in the past. Personally, (P4 1300 Mhz, 284MB) I got better results with "-fno-omit-frame-pointer" and most of the results I have seen appear to agree with this. Anyone care to explain?
My scripts to run all the benches and process the results are near the top of the thread.heriophant wrote:Ok i just finished most of the thread and now realize there are more benchs that need to be read.
If the command i used in my first test only ran the huffbench, what is the command to run all the tests and then get the final output in a yes no maybe format?
Thanks a ton in advance
Heriophant
Code: Select all
*** alma ***
./acovea.bench: line 10: 7006 Segmentation fault runacovea -config gcc34_pentium4.acovea -bench ${bench}bench.c >${bench}.run 2>${bench}.err
real 0m13.828s
user 0m13.116s
sys 0m0.716s
*** evo ***
./acovea.bench: line 10: 7435 Segmentation fault runacovea -config gcc34_pentium4.acovea -bench ${bench}bench.c >${bench}.run 2>${bench}.err
real 0m34.124s
user 0m33.361s
sys 0m0.759s
*** fft ***
./acovea.bench: line 10: 7865 Segmentation fault runacovea -config gcc34_pentium4.acovea -bench ${bench}bench.c >${bench}.run 2>${bench}.err
real 0m9.894s
user 0m9.121s
sys 0m0.773s
*** huff ***
./acovea.bench: line 10: 8302 Segmentation fault runacovea -config gcc34_pentium4.acovea -bench ${bench}bench.c >${bench}.run 2>${bench}.err
real 0m25.244s
user 0m23.917s
sys 0m1.259s
*** lin ***
./acovea.bench: line 10: 8734 Segmentation fault runacovea -config gcc34_pentium4.acovea -bench ${bench}bench.c >${bench}.run 2>${bench}.err
real 0m12.139s
user 0m11.356s
sys 0m0.790s
*** mat1 ***
./acovea.bench: line 10: 9158 Segmentation fault runacovea -config gcc34_pentium4.acovea -bench ${bench}bench.c >${bench}.run 2>${bench}.err
real 0m18.248s
user 0m17.422s
sys 0m0.832s
*** tree ***
real 266m22.069s
user 263m50.781s
sys 2m8.694sDid anyone manage to finish benchmarking with a Pentium-M machine? I'd be very interested in knowning about the results.poisson wrote:I found another interesting Acovea application: what are the best optimizations for pentium-m? You know, such processor is an hybrid between pentium3 and pentium4, with 1M L2 cache. I started with "alma", but I don't like to stress my laptop
Does it hurt to include -fomit-frame-pointer though if I don't need to do any debugging? Everyone seems to say that -fomit-frame-pointer is one of the params that none would want to remove.Hypnos wrote:"Register starvation" is a non-issue on late-generation CPUs; it was when floating point and SIMD units were first added to CPU dies with the same old registers.HeartBreakKid wrote:So has anyone who has run acovea thus far gotten a Yes for "-fomit-frame-pointer"? Most of the results I have seen so far appear to suggest that this is not as good as (we all) may have believed in the past. Personally, (P4 1300 Mhz, 284MB) I got better results with "-fno-omit-frame-pointer" and most of the results I have seen appear to agree with this. Anyone care to explain?
-fomit-frame-pointer might help if you have heavily hand-optimized code that runs close to the hardware, like the kernel, or code that also uses a lot of special instructions like those in 3DNow.
-fomit-frame-pointer seems to introduce some overhead (according to the results here) and can cause some builds to break (Gentoo devs try to filter it out where it does).teedog wrote:Does it hurt to include -fomit-frame-pointer though if I don't need to do any debugging? Everyone seems to say that -fomit-frame-pointer is one of the params that none would want to remove.
Right, so now add the flags graded "Yes" and annotated "(! O1)" -- this will turn off detrimental optimizations implied by "-O1". Also, you might want to add "-mcpu=pentium4" even though it's redundant to "-march=pentium4", as the latter is filtered out by a number of ebuilds.heriophant wrote:
this is my start anyone care to finish it based on the results.
CFLAGS="-O3 -march=pentium4 -pipe
Thanks a ton in the hand holding until i learn what this all is and about
The latter (i.e., ONLY yes + ! O1). The CFLAGS above, you've introduced redundant flags (e.g., -fpeephole2 is implied by -O2, which is implied by -O3) and some dangerous flags (e.g., -malign-double).heriophant wrote:Ok just wanted to be clear as to your answer if im understanding you correctly this is how my flags should be set
with all the "yes" from the results.
CFLAGS="-O3 -march=pentium4 -mcpu=pentium4 -pipe -fpeephole2 -foptimize-sibling-calls -fno-thread-jumps -falign-jumps -fcse-skip-blocks -fno-omit-frame-pointer -fdelete-null-pointer-checks -fregmove -malign-double -fno-merge-constants -fno-delayed-branch -frerun-cse-after-loop -ffinite-math-only -frename-registers -freorder-functions"
or should it ONLY be the yes that flag !-O1
Anyways, the job was running as nice -n -18 ... let's hope).
