rockbox/apps/codecs/lib/udiv32_armv4.S, branch master

Improvements to specialized dividers for APE codec:

2010-01-28T02:28:52+00:00

* Use Newton-Raphson divider on ARMv5e and ARMv6, about 7% speedup on Gigabeat S.
* On ARMv4 targets using IRAM, remove insane filter buffer from IRAM, fill available IRAM with LUT of reciprocals for small divisors - speedup varies according to target and available IRAM, APE normal sample is approx. 109% RT on e200.
* Rename apps/codecs/lib/udiv32_armv4.S to apps/codecs/lib/udiv32_arm.S, which includes dividers for all ARM targets specialized for APE.


git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24354 a1c6a512-1295-4272-9138-f99709370657

Invert divisor earlier in udiv32_arm, allowing the div0 test to be done before entering the 32-bit divide portion of the code, and making the handling of div0 simpler.

2010-01-03T15:57:03+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24166 a1c6a512-1295-4272-9138-f99709370657

Use long jump to reach __div0 from udiv32_arm if building with IRAM and without EABI.

2010-01-03T04:48:19+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24152 a1c6a512-1295-4272-9138-f99709370657

More comments for udiv32_armv4.S, reduce zero divisor test to one cycle for the skipped branch by setting flags when inverting divisor, 32-bit numerators are handled by calling the 31-bit divider and fixing the results.

2010-01-03T04:30:13+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24151 a1c6a512-1295-4272-9138-f99709370657

Add missing EOF newline.

2010-01-02T15:25:34+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24143 a1c6a512-1295-4272-9138-f99709370657

Remove special cases from udiv32_armv4.S, except for zero divisor and large numerator. Improvement of 1.23MHz on e200 with ape normal.

2010-01-02T15:15:21+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24142 a1c6a512-1295-4272-9138-f99709370657

Add 31/31-bit unsigned division in apps/codecs/lib/udiv_arm.S, with 2 cycles / iteration, falling back to previous 32-bit, 3 cycle / iteration code when needed (well under 1% of divisions in sample file). APE normal sample is now 96.90% realtime, approx 1.3% improved vs svn. TODO: unify divisor normalization for both trial subtraction routines, possibly use divisor bits to select 31- vs 32-bit division.

2009-12-31T08:32:15+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24130 a1c6a512-1295-4272-9138-f99709370657

ARMv4 unsigned integer division: Using an overflow-safe comparison method in the main calculation allows to put back the 1.5 cyle (average) optimisation. Shaved off another instruction, as we don't need the remainder. * Use the very efficient ffs algorithm from ffs-arm.S for dividing by a power of 2.

2008-11-06T21:21:33+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@19032 a1c6a512-1295-4272-9138-f99709370657

This optimisation breaks for very large divisors (MSB set), so remove it.

2008-11-05T07:36:39+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@19012 a1c6a512-1295-4272-9138-f99709370657

Further optimised (vs. libgcc) unsigned 32 bit division for ARMv4 (based on the ARMv5(+) version from libgcc), in IRAM on PP for better performance on PP5002, and put into the codeclib for possible reuse. APE -c1000 is now usable on both PP502x and PP5002 (~138% realtime, they're on par now). Gigabeat F/X should also see an APE speedup.

2008-11-05T00:10:05+00:00

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@19009 a1c6a512-1295-4272-9138-f99709370657