diff options
author | schmonz <schmonz@pkgsrc.org> | 2018-07-31 06:34:11 +0000 |
---|---|---|
committer | schmonz <schmonz@pkgsrc.org> | 2018-07-31 06:34:11 +0000 |
commit | dc108b4290f2e535abc06fe2d9375beaa4564beb (patch) | |
tree | eb1df219f8d4dc9aa17f0e21d0a80b572d9c7308 /math/djbsort | |
parent | 7642779285f3df3be182e0f3a6c9d76b479ea2d0 (diff) | |
download | pkgsrc-dc108b4290f2e535abc06fe2d9375beaa4564beb.tar.gz |
Update to 20180729. From the changelog:
Algorithm
Rewrite of the core int32/avx2 implementation for (1) higher speed and
(2) reduced memory consumption. Stack allocation is now at most a few
kilobytes, even for gigantic arrays.
Internally, the sorting algorithm is now mostly bitonic to simplify
indexing, although odd-even speedups are still applied when
convenient. Lanes are complemented to take the down-up decision out of
the inner loops.
As in previous djbsort versions, data is sorted first in vector lanes
and then transposed for final merges, reducing the overall number of
vector permutations. Unlike previous versions, transposition is done
in-place. The transposition in this version is bit-reversal on the outer
6 bits (bottom 3 bits and the top 3 bits), but leaves intermediate bits
alone. Non-power-of-2 array sizes are handled by an extra, more
traditional, merge step.
Sizes 2, 3, 4, 5, 6, 7, 8, 16, 32 are now special-cased. Non-power-of-2
sizes below 256 are padded to the next power of 2.
Portable implementations: The out-of-place int32/portable1 and
int32/portable2 implementations are now gone; the in-place
int32/portable3 and int32/portable4 implementations remain.
C API
float32_sort is now supported. The arithmetic in the reduction from
float32 to int32 is int32 31-bit right shift, uint32 1-bit right shift,
xor; this is slightly more efficient than the reduction from float32 to
uint32 from 2001 Herf.
Compiling
Tests now have more variation (without much slowdown): the uint32 test
cases now deviate from int32 in more than the sign; float32 uses
floating-point numbers that aren't integers; int32 does more loops for
small cases, and some larger cases.
Internals
API for 2-input sorting is now MINMAX macro operating on two
inputs in place.
Better inline assembly from Jason Donenfeld for 2-input sorting: more
flexibility in compiler's register allocation.
The package version number is now automatically copied to version.c as
the implementation version number for implementations that don't provide
version.c.
Verification
minmax now supports more peephole optimizations for complemented bitonic
sorting and for padding: xor(s,xor(s,t)) ⇒ t; xor(-1,s) ⇒ invert(s);
Reverse(Reverse(s)) ⇒ s; signedmin(invert(s),invert(t)) ⇒
invert(signedmax(s,t)); signedmax(invert(s),invert(t)) ⇒
invert(signedmin(s,t)); invert(s)[high:low] ⇒ invert(s[high:low]);
s[bits-1:0] ⇒ s; s[high:low][high2:low2] ⇒ s[high2+low:low2+low];
Concat(...)[high:low] ⇒ ...[high-pos:low-pos] when possible;
Reverse(s)[high:low] ⇒ Reverse(s[...]) when possible; eliminate
signedmin/signedmax when one input is the minimum or maximum constant.
verifymany now includes the implementation version number on
verified lines.
Diffstat (limited to 'math/djbsort')
-rw-r--r-- | math/djbsort/Makefile | 5 | ||||
-rw-r--r-- | math/djbsort/distinfo | 10 | ||||
-rw-r--r-- | math/djbsort/pseudo-PLIST | 1 |
3 files changed, 8 insertions, 8 deletions
diff --git a/math/djbsort/Makefile b/math/djbsort/Makefile index 75c01bb345d..9f33cfc3799 100644 --- a/math/djbsort/Makefile +++ b/math/djbsort/Makefile @@ -1,7 +1,6 @@ -# $NetBSD: Makefile,v 1.3 2018/07/31 06:24:57 schmonz Exp $ +# $NetBSD: Makefile,v 1.4 2018/07/31 06:34:11 schmonz Exp $ -DISTNAME= djbsort-20180717 -PKGREVISION= 1 +DISTNAME= djbsort-20180729 CATEGORIES= math MASTER_SITES= ${HOMEPAGE} diff --git a/math/djbsort/distinfo b/math/djbsort/distinfo index e3898b35099..d3d2459fd49 100644 --- a/math/djbsort/distinfo +++ b/math/djbsort/distinfo @@ -1,9 +1,9 @@ -$NetBSD: distinfo,v 1.2 2018/07/31 05:38:56 schmonz Exp $ +$NetBSD: distinfo,v 1.3 2018/07/31 06:34:11 schmonz Exp $ -SHA1 (djbsort-20180717.tar.gz) = 8f6efd8f66aa23707d54c083b9fb28371a409c55 -RMD160 (djbsort-20180717.tar.gz) = 64d92cccf13d8a997ea25ee77491f54b6a6b3ce1 -SHA512 (djbsort-20180717.tar.gz) = c26a871c46006580343b2876203264c8de0db613ee90ba5dd050f46a71a346d9340ecd06220dc1c0a46a6c1cd910af4bd2cb3fcd39588a62cc2a4a996fc9113e -Size (djbsort-20180717.tar.gz) = 25055 bytes +SHA1 (djbsort-20180729.tar.gz) = 7b64b9c5b3c6428cb6bc16638876f57ffff3c58b +RMD160 (djbsort-20180729.tar.gz) = 65455661524b5d2cdd21c1c329a9065cfc9b3b0d +SHA512 (djbsort-20180729.tar.gz) = b46c36514ce0c2c1716421d4da216d2e7ddb80e8e9a1b97b274ab0b3fe7271ea946a65485e81d4dc76c080658b41ef1e2fa64a225594532f1be98dc29535b2a6 +Size (djbsort-20180729.tar.gz) = 29347 bytes SHA1 (patch-build) = 0f1ae219495cc63fc301aa1f9ae1c1655da3aaa3 SHA1 (patch-test) = b5ba29b15d7baa896bf9e47a1522e88e73df4c9e SHA1 (patch-upgrade) = 2bec25c36e1ed28c67e4967ac929780d28bb2e68 diff --git a/math/djbsort/pseudo-PLIST b/math/djbsort/pseudo-PLIST index 1ce5958d13e..1bedcbd3ead 100644 --- a/math/djbsort/pseudo-PLIST +++ b/math/djbsort/pseudo-PLIST @@ -1,4 +1,5 @@ include/djbsort_cpucycles.h +include/float32_sort.h include/int32_sort.h include/uint32_sort.h lib/libdjbsort.a |