| Age | Commit message (Collapse) | Author |
|---|
|
Use i686 as TARGET_ARCH for 32bit core2 (and corei7 and atom) builds.
In most cases, i586 and i686 are equivalent values for TARGET_ARCH, however
one important exception is glibc. When configured for i686, glibc enables
optimised string functions (SSE, SSE2, etc), which are not used when
building for i586.
The benefits of i686 optimised string functions vary depending on the
application and the CPU, however in some cases the improvements are
significant. In one test, a 50% increase in FPS was seen when running the
'smashcat' benchmark [1] in a qtwebkit browser on an Intel Atom based SoC.
The gain seems to comes from a 3x improvement in memcpy performance when
copying graphics buffer lines (5120 bytes, or 1280 x 4 bytes/pixel), from
the CPU to GPU. Note that very large memcpy's (e.g. 32MB) on the same
machine show no particular performance increase between i586 and i686.
[1] http://www.smashcat.org/av/canvas_test/
Warning: The change in TARGET_ARCH means that _i586 architecture specific
over-rides will no longer take effect. Both oe-core and meta-oe have been
updated to replace _i586 over-rides with _x86, however other layers may
still need review and updating.
Signed-off-by: Andre McCurdy <armccurdy@gmail.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Core2 has both a 32b and a 64b variant. Currently, core2 implies 32b,
while core2_64 is the 64b version. This implicit 32b mode will become
confusing in later architectures, such as corei7, where it would be
natural for people to assume "corei7" meant 64 bit.
Rather than carrying forward an implicit 32b mode and rather than
changing the naming scheme part way through the architecture hiearchy,
make the 32b and 64b variant explicit in the tune name by changing core2
to core2-32. This patch also standardises on using '-' in the names.
Signed-off-by: Darren Hart <dvhart@linux.intel.com>
Cc: Richard Purdie <richard.purdie@linuxfoundation.org>
Cc: Paul Eggleton <paul.eggleton@intel.com>
Cc: Tom Zanussi <tom.zanussi@intel.com>
Cc: Nitin Kamble <nitin.a.kamble@intel.com>
Cc: Mark Hatle <mark.hatle@windriver.com>
Cc: Bruce Ashfield <bruce.ashfield@windriver.com>
Cc: Martin Jansa <martin.jansa@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Inherit the PACKAGE_EXTRA_ARCHS from i586 and only explicitly add core2
here.
Signed-off-by: Darren Hart <dvhart@linux.intel.com>
Cc: Richard Purdie <richard.purdie@linuxfoundation.org>
Cc: Paul Eggleton <paul.eggleton@intel.com>
Cc: Tom Zanussi <tom.zanussi@intel.com>
Cc: Nitin Kamble <nitin.a.kamble@intel.com>
Cc: Mark Hatle <mark.hatle@windriver.com>
Cc: Bruce Ashfield <bruce.ashfield@windriver.com>
Cc: Martin Jansa <martin.jansa@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
-march specifies which ISA to use. -mtune specifies which cpu-type to
optimize instruction ordering for, but not which ISA to use. There are
times when it may make sense to specify mtune=generic and use a more
specific march, such as core2, but the opposite makes little sense at
all: use cpu-type specific ISA, but order the instructions
generically. While the -mtune is implied by -march, gcc does not verify
it is using -mtune=core2 with:
gcc -Q -march=core2 --help=target
Explicitly specify -mtune=core2 to be sure.
Add a comment header describing the CPUs targeted by this tune file.
Signed-off-by: Darren Hart <dvhart@linux.intel.com>
Cc: Richard Purdie <richard.purdie@linuxfoundation.org>
Cc: Paul Eggleton <paul.eggleton@intel.com>
Cc: Tom Zanussi <tom.zanussi@intel.com>
Cc: Nitin Kamble <nitin.a.kamble@intel.com>
Cc: Mark Hatle <mark.hatle@windriver.com>
Cc: Bruce Ashfield <bruce.ashfield@windriver.com>
Cc: Martin Jansa <martin.jansa@gmail.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
* number of TUNE_CCARGS conditionals is important if we add
extra space with each one in "else" branch
I'm building for 2 MACHINEs one is cortexa9, second is cortexa8
few months ago we added TUNE_CCARGS[vardepvalue] in bitbake.conf
http://git.openembedded.org/openembedded-core/commit/?id=03f1e34ea3ce80931e9c3cd2ab22824f28a7233b
which fixed some cases (like mentioned tune-xscale and tune-arm926ejs)
where both had unused TUNE_CCARGS when common DEFAULTTUNE was used.
with cortexa[89] it's different, because cortexa9 has one extra TUNE_CCARGS
TUNE_CCARGS += "${@bb.utils.contains("TUNE_FEATURES", "cortexa9", "-mtune=cortex-a9", "", d)}"
which adds extra *space* even when not used because of '+=' and as result:
$ bitbake-diffsigs tmp-eglibc/sstate-diff/1366797730/*/armv7*/adapterbase/*do_configure*
basehash changed from f986789fb8fb3579ed6a3492cc8a8d10 to c851b5f838d945ee13072e9ad6725dca
Variable TUNE_CCARGS value changed from
' -march=armv7-a -mthumb-interwork -mfloat-abi=softfp -mfpu=neon '
to
' -march=armv7-a -mthumb-interwork -mfloat-abi=softfp -mfpu=neon '
Hash for dependent task gcc-runtime_4.7.bb.do_populate_sysroot changed from bdeabf7a86958b9110b566344b7916de to 2be5618e6bc8c57ec9db5659bf217915
Hash for dependent task eglibc_2.17.bb.do_populate_sysroot changed from b4f40fc62dde684acd0a574532a55360 to 97fcb426603d4a1c1099c0504d2ebf7d
Hash for dependent task glib-2.0_2.34.3.bb.do_populate_sysroot changed from fd2f90b83098c34e88d649d70f6ea4f5 to ebd740bb94ea3eb0a914efda6fc82c4a
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Signed-off-by: Saul Wold <sgw@linux.intel.com>
|
|
We perform a basic cleanup of the IA32 architecture and related
tunings in order to match the rules and descriptions within the
new tuning README file.
A number of small issues were corrected in the "c3" tuning to
bring it inline with the README.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
rpmbuild can not handle the PACKAGE_ARCH of these kinds:
x86_64-x32, core2-64, core2-64-x32
With these kinds of PACKAGE_ARCH the --target parameter of rpmbuild
becomes like: core2-64-x32-poky-linux-gnux32 ; And rpmbuild extracts
%_target (arch) wrongly as core2 generating these kinds of rpms with
incorrect filenames: zip-3.0-r0.core2.rpm
So this commit fixes the issue by making PACKAGE_ARCH like this:
x86_64_x32, core2_64, core2_64_x32
Now --target parameter of rpmbuild becomes like:
core2_64_x32-poky-linux-gnux32 ; And rpmbuild extracts %_target (arch)
correctly as core2_64_x32 generating these kinds of rpms with correct
filenames: zip-3.0-r0.core2_64_x32.rpm
Signed-off-by: Nitin A Kamble <nitin.a.kamble@intel.com>
|
|
This ensures that on a multilib system the two executable formats
don't conflict.
Signed-off-by: Nitin A Kamble <nitin.a.kamble@intel.com>
|
|
Signed-off-by: Nitin A Kamble <nitin.a.kamble@intel.com>
|
|
Using i686 doesn't work well with locale generation and doesn't gain anything
so revert to the i586 default.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
The current approach causes duplicate values to appear in the TUNE_ARCH
field and this patch addresses that.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
files and tune features
These changes revolve around the idea of tune features. These are represented by
'flag' strings that are included in the TUNE_FEATURES variable.
Any string included in TUNE_FEATURES should also add a TUNEVALID[<name>] entry so
we can know which flags are available in TUNE_FEATURES and have documentation about
what the flags do. We will add sanity code to error if flags are listed in
TUNE_FEATURES but are not documented in TUNEVALID.
A given tune configuration will want to define one or more predetermined sets of
_FEATURE flag lists. These are defined in the form TUNE_FEATURES_tune-<name>.
For defined tune configuation, <name> should be added to the AVAILTUNE list so that
we can determine what tune configurations are available. Flags cannot be used in this
case as with TUNEVALID since its useful to be able to build up tune lists from other
TUNE_FEATURES_tune-yyy options.
A given tune configuration may also define PACKAGE_EXTRA_ARCHS_tune-<name> and
BASE_LIB_tune-<name> to control the multilib location. All options can be overridden
by the distro or local user configuration.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
Andre McCurdy
Darren Hart
Martin Jansa
Mark Hatle
Nitin A Kamble
Richard Purdie