2011-02-23 10:52:22 +00:00
|
|
|
/*
|
2008-04-16 22:40:48 +00:00
|
|
|
* ***** BEGIN GPL LICENSE BLOCK *****
|
2002-10-12 11:37:38 +00:00
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License
|
|
|
|
* as published by the Free Software Foundation; either version 2
|
2008-04-16 22:40:48 +00:00
|
|
|
* of the License, or (at your option) any later version.
|
2002-10-12 11:37:38 +00:00
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
* GNU General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU General Public License
|
|
|
|
* along with this program; if not, write to the Free Software Foundation,
|
2010-02-12 13:34:04 +00:00
|
|
|
* Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
|
2002-10-12 11:37:38 +00:00
|
|
|
*
|
|
|
|
* The Original Code is Copyright (C) 2001-2002 by NaN Holding BV.
|
|
|
|
* All rights reserved.
|
|
|
|
*
|
|
|
|
* The Original Code is: all of this file.
|
|
|
|
*
|
|
|
|
* Contributor(s): none yet.
|
|
|
|
*
|
2008-04-16 22:40:48 +00:00
|
|
|
* ***** END GPL LICENSE BLOCK *****
|
2002-10-12 11:37:38 +00:00
|
|
|
*/
|
|
|
|
|
2011-02-27 20:37:56 +00:00
|
|
|
/** \file blender/blenlib/intern/BLI_ghash.c
|
|
|
|
* \ingroup bli
|
2013-06-22 22:30:56 +00:00
|
|
|
*
|
2014-12-17 09:51:18 +01:00
|
|
|
* A general (pointer -> pointer) chaining hash table
|
|
|
|
* for 'Abstract Data Types' (known as an ADT Hash Table).
|
2013-08-18 01:00:52 +00:00
|
|
|
*
|
2013-06-22 22:30:56 +00:00
|
|
|
* \note edgehash.c is based on this, make sure they stay in sync.
|
2011-02-27 20:37:56 +00:00
|
|
|
*/
|
|
|
|
|
2011-10-20 13:50:24 +00:00
|
|
|
#include <string.h>
|
|
|
|
#include <stdlib.h>
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
#include <stdarg.h>
|
2013-07-21 17:05:41 +00:00
|
|
|
#include <limits.h>
|
2011-10-20 13:50:24 +00:00
|
|
|
|
2010-08-17 11:10:48 +00:00
|
|
|
#include "MEM_guardedalloc.h"
|
2002-10-12 11:37:38 +00:00
|
|
|
|
2013-05-28 19:35:26 +00:00
|
|
|
#include "BLI_sys_types.h" /* for intptr_t support */
|
2011-01-07 18:36:47 +00:00
|
|
|
#include "BLI_utildefines.h"
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
#include "BLI_hash_mm2a.h"
|
2011-10-24 08:18:20 +00:00
|
|
|
#include "BLI_mempool.h"
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
|
|
|
#define GHASH_INTERNAL_API
|
2002-10-12 11:37:38 +00:00
|
|
|
#include "BLI_ghash.h"
|
2013-09-01 00:46:04 +00:00
|
|
|
#include "BLI_strict_flags.h"
|
2011-10-20 13:50:24 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
#define GHASH_USE_MODULO_BUCKETS
|
2013-05-08 12:55:51 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/* Also used by smallhash! */
|
2013-05-08 12:54:07 +00:00
|
|
|
const unsigned int hashsizes[] = {
|
2010-01-23 11:25:20 +00:00
|
|
|
5, 11, 17, 37, 67, 131, 257, 521, 1031, 2053, 4099, 8209,
|
2002-10-12 11:37:38 +00:00
|
|
|
16411, 32771, 65537, 131101, 262147, 524309, 1048583, 2097169,
|
|
|
|
4194319, 8388617, 16777259, 33554467, 67108879, 134217757,
|
|
|
|
268435459
|
|
|
|
};
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
# define GHASH_MAX_SIZE 27
|
2013-08-26 09:37:15 +00:00
|
|
|
#else
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
# define GHASH_BUCKET_BIT_MIN 2
|
|
|
|
# define GHASH_BUCKET_BIT_MAX 28 /* About 268M of buckets... */
|
2013-08-26 09:37:15 +00:00
|
|
|
#endif
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/**
|
|
|
|
* \note Max load #GHASH_LIMIT_GROW used to be 3. (pre 2.74).
|
|
|
|
* Python uses 0.6666, tommyhaslib even goes down to 0.5.
|
|
|
|
* Reducing our from 3 to 0.75 gives huge speedup (about twice quicker pure GHash insertions/lookup,
|
|
|
|
* about 25% - 30% quicker 'dynamic-topology' stroke drawing e.g.).
|
|
|
|
* Min load #GHASH_LIMIT_SHRINK is a quarter of max load, to avoid resizing to quickly.
|
|
|
|
*/
|
|
|
|
#define GHASH_LIMIT_GROW(_nbkt) (((_nbkt) * 3) / 4)
|
|
|
|
#define GHASH_LIMIT_SHRINK(_nbkt) (((_nbkt) * 3) / 16)
|
|
|
|
|
2002-10-12 11:37:38 +00:00
|
|
|
/***/
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/* WARNING! Keep in sync with ugly _gh_Entry in header!!! */
|
2013-08-18 00:36:04 +00:00
|
|
|
typedef struct Entry {
|
|
|
|
struct Entry *next;
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
void *key;
|
2013-08-18 00:36:04 +00:00
|
|
|
} Entry;
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
typedef struct GHashEntry {
|
|
|
|
Entry e;
|
|
|
|
|
|
|
|
void *val;
|
|
|
|
} GHashEntry;
|
|
|
|
|
|
|
|
typedef Entry GSetEntry;
|
|
|
|
|
|
|
|
#define GHASH_ENTRY_SIZE(_is_gset) \
|
|
|
|
((_is_gset) ? sizeof(GSetEntry) : sizeof(GHashEntry))
|
|
|
|
|
2013-08-18 14:15:51 +00:00
|
|
|
struct GHash {
|
2013-08-18 00:36:04 +00:00
|
|
|
GHashHashFP hashfp;
|
|
|
|
GHashCmpFP cmpfp;
|
|
|
|
|
|
|
|
Entry **buckets;
|
|
|
|
struct BLI_mempool *entrypool;
|
|
|
|
unsigned int nbuckets;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
unsigned int limit_grow, limit_shrink;
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
unsigned int cursize, size_min;
|
|
|
|
#else
|
|
|
|
unsigned int bucket_mask, bucket_bit, bucket_bit_min;
|
|
|
|
#endif
|
|
|
|
|
2013-08-18 00:36:04 +00:00
|
|
|
unsigned int nentries;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
unsigned int flag;
|
2013-08-18 14:15:51 +00:00
|
|
|
};
|
2013-08-18 00:36:04 +00:00
|
|
|
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_INLINE void ghash_entry_copy(
|
|
|
|
GHash *gh_dst, Entry *dst, GHash *gh_src, Entry *src,
|
|
|
|
GHashKeyCopyFP keycopyfp, GHashValCopyFP valcopyfp)
|
|
|
|
{
|
|
|
|
dst->key = (keycopyfp) ? keycopyfp(src->key) : src->key;
|
|
|
|
|
|
|
|
if ((gh_dst->flag & GHASH_FLAG_IS_GSET) == 0) {
|
|
|
|
if ((gh_src->flag & GHASH_FLAG_IS_GSET) == 0) {
|
|
|
|
((GHashEntry *)dst)->val = (valcopyfp) ? valcopyfp(((GHashEntry *)src)->val) : ((GHashEntry *)src)->val;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
((GHashEntry *)dst)->val = NULL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-18 00:36:04 +00:00
|
|
|
/* -------------------------------------------------------------------- */
|
|
|
|
/* GHash API */
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \name Internal Utility API
|
|
|
|
* \{ */
|
2009-02-12 18:05:34 +00:00
|
|
|
|
2013-08-26 13:41:13 +00:00
|
|
|
/**
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* Get the full hash for a key.
|
2013-08-26 13:41:13 +00:00
|
|
|
*/
|
|
|
|
BLI_INLINE unsigned int ghash_keyhash(GHash *gh, const void *key)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return gh->hashfp(key);
|
2013-08-26 13:41:13 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* Get the full hash for an entry.
|
2013-08-25 12:17:46 +00:00
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_INLINE unsigned int ghash_entryhash(GHash *gh, const Entry *e)
|
2013-08-24 13:04:03 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return gh->hashfp(e->key);
|
2013-08-24 13:04:03 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 16:16:38 +00:00
|
|
|
/**
|
2015-11-29 17:11:40 +11:00
|
|
|
* Get the bucket-index for an already-computed full hash.
|
2013-08-25 16:16:38 +00:00
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_INLINE unsigned int ghash_bucket_index(GHash *gh, const unsigned int hash)
|
|
|
|
{
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
return hash % gh->nbuckets;
|
|
|
|
#else
|
2015-04-06 09:03:32 +10:00
|
|
|
return hash & gh->bucket_mask;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2016-02-20 15:17:40 +01:00
|
|
|
/**
|
|
|
|
* Find the index of next used bucket, starting from \a curr_bucket (\a gh is assumed non-empty).
|
|
|
|
*/
|
|
|
|
BLI_INLINE unsigned int ghash_find_next_bucket_index(GHash *gh, unsigned int curr_bucket)
|
|
|
|
{
|
|
|
|
if (curr_bucket >= gh->nbuckets) {
|
|
|
|
curr_bucket = 0;
|
|
|
|
}
|
|
|
|
if (gh->buckets[curr_bucket]) {
|
|
|
|
return curr_bucket;
|
|
|
|
}
|
|
|
|
for (; curr_bucket < gh->nbuckets; curr_bucket++) {
|
|
|
|
if (gh->buckets[curr_bucket]) {
|
|
|
|
return curr_bucket;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (curr_bucket = 0; curr_bucket < gh->nbuckets; curr_bucket++) {
|
|
|
|
if (gh->buckets[curr_bucket]) {
|
|
|
|
return curr_bucket;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
BLI_assert(0);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/**
|
|
|
|
* Expand buckets to the next size up or down.
|
|
|
|
*/
|
|
|
|
static void ghash_buckets_resize(GHash *gh, const unsigned int nbuckets)
|
2013-08-25 16:16:38 +00:00
|
|
|
{
|
2013-08-26 13:41:13 +00:00
|
|
|
Entry **buckets_old = gh->buckets;
|
|
|
|
Entry **buckets_new;
|
|
|
|
const unsigned int nbuckets_old = gh->nbuckets;
|
|
|
|
unsigned int i;
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_assert((gh->nbuckets != nbuckets) || !gh->buckets);
|
|
|
|
// printf("%s: %d -> %d\n", __func__, nbuckets_old, nbuckets);
|
2013-08-26 13:41:13 +00:00
|
|
|
|
|
|
|
gh->nbuckets = nbuckets;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
#else
|
|
|
|
gh->bucket_mask = nbuckets - 1;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
buckets_new = (Entry **)MEM_callocN(sizeof(*gh->buckets) * gh->nbuckets, __func__);
|
|
|
|
|
|
|
|
if (buckets_old) {
|
|
|
|
if (nbuckets > nbuckets_old) {
|
|
|
|
for (i = 0; i < nbuckets_old; i++) {
|
2015-11-29 17:11:40 +11:00
|
|
|
for (Entry *e = buckets_old[i], *e_next; e; e = e_next) {
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned hash = ghash_entryhash(gh, e);
|
|
|
|
const unsigned bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
e_next = e->next;
|
|
|
|
e->next = buckets_new[bucket_index];
|
|
|
|
buckets_new[bucket_index] = e;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
for (i = 0; i < nbuckets_old; i++) {
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
2015-11-29 17:11:40 +11:00
|
|
|
for (Entry *e = buckets_old[i], *e_next; e; e = e_next) {
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned hash = ghash_entryhash(gh, e);
|
|
|
|
const unsigned bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
e_next = e->next;
|
|
|
|
e->next = buckets_new[bucket_index];
|
|
|
|
buckets_new[bucket_index] = e;
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
/* No need to recompute hashes in this case, since our mask is just smaller, all items in old bucket i
|
|
|
|
* will go in same new bucket (i & new_mask)! */
|
|
|
|
const unsigned bucket_index = ghash_bucket_index(gh, i);
|
|
|
|
BLI_assert(!buckets_old[i] || (bucket_index == ghash_bucket_index(gh, ghash_entryhash(gh, buckets_old[i]))));
|
2015-11-29 17:11:40 +11:00
|
|
|
Entry *e;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
for (e = buckets_old[i]; e && e->next; e = e->next);
|
|
|
|
if (e) {
|
|
|
|
e->next = buckets_new[bucket_index];
|
|
|
|
buckets_new[bucket_index] = buckets_old[i];
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
}
|
2013-08-26 13:41:13 +00:00
|
|
|
}
|
2013-08-25 16:16:38 +00:00
|
|
|
}
|
2013-08-26 13:41:13 +00:00
|
|
|
|
|
|
|
gh->buckets = buckets_new;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
if (buckets_old) {
|
|
|
|
MEM_freeN(buckets_old);
|
|
|
|
}
|
2013-08-25 16:16:38 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* Check if the number of items in the GHash is large enough to require more buckets,
|
|
|
|
* or small enough to require less buckets, and resize \a gh accordingly.
|
2013-08-25 12:17:46 +00:00
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
static void ghash_buckets_expand(
|
|
|
|
GHash *gh, const unsigned int nentries, const bool user_defined)
|
2013-08-18 03:41:39 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
unsigned int new_nbuckets;
|
|
|
|
|
|
|
|
if (LIKELY(gh->buckets && (nentries < gh->limit_grow))) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
new_nbuckets = gh->nbuckets;
|
|
|
|
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
while ((nentries > gh->limit_grow) &&
|
|
|
|
(gh->cursize < GHASH_MAX_SIZE - 1))
|
|
|
|
{
|
|
|
|
new_nbuckets = hashsizes[++gh->cursize];
|
|
|
|
gh->limit_grow = GHASH_LIMIT_GROW(new_nbuckets);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
while ((nentries > gh->limit_grow) &&
|
|
|
|
(gh->bucket_bit < GHASH_BUCKET_BIT_MAX))
|
|
|
|
{
|
|
|
|
new_nbuckets = 1u << ++gh->bucket_bit;
|
|
|
|
gh->limit_grow = GHASH_LIMIT_GROW(new_nbuckets);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
if (user_defined) {
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
gh->size_min = gh->cursize;
|
|
|
|
#else
|
|
|
|
gh->bucket_bit_min = gh->bucket_bit;
|
|
|
|
#endif
|
2013-08-26 13:41:13 +00:00
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
|
|
|
if ((new_nbuckets == gh->nbuckets) && gh->buckets) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
gh->limit_grow = GHASH_LIMIT_GROW(new_nbuckets);
|
|
|
|
gh->limit_shrink = GHASH_LIMIT_SHRINK(new_nbuckets);
|
|
|
|
ghash_buckets_resize(gh, new_nbuckets);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ghash_buckets_contract(
|
|
|
|
GHash *gh, const unsigned int nentries, const bool user_defined, const bool force_shrink)
|
|
|
|
{
|
|
|
|
unsigned int new_nbuckets;
|
|
|
|
|
|
|
|
if (!(force_shrink || (gh->flag & GHASH_FLAG_ALLOW_SHRINK))) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (LIKELY(gh->buckets && (nentries > gh->limit_shrink))) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
new_nbuckets = gh->nbuckets;
|
|
|
|
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
while ((nentries < gh->limit_shrink) &&
|
2015-03-23 22:51:12 +11:00
|
|
|
(gh->cursize > gh->size_min))
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
{
|
|
|
|
new_nbuckets = hashsizes[--gh->cursize];
|
|
|
|
gh->limit_shrink = GHASH_LIMIT_SHRINK(new_nbuckets);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
while ((nentries < gh->limit_shrink) &&
|
2015-03-23 22:51:12 +11:00
|
|
|
(gh->bucket_bit > gh->bucket_bit_min))
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
{
|
|
|
|
new_nbuckets = 1u << --gh->bucket_bit;
|
|
|
|
gh->limit_shrink = GHASH_LIMIT_SHRINK(new_nbuckets);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
if (user_defined) {
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
gh->size_min = gh->cursize;
|
|
|
|
#else
|
|
|
|
gh->bucket_bit_min = gh->bucket_bit;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
if ((new_nbuckets == gh->nbuckets) && gh->buckets) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
gh->limit_grow = GHASH_LIMIT_GROW(new_nbuckets);
|
|
|
|
gh->limit_shrink = GHASH_LIMIT_SHRINK(new_nbuckets);
|
|
|
|
ghash_buckets_resize(gh, new_nbuckets);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Clear and reset \a gh buckets, reserve again buckets for given number of entries.
|
|
|
|
*/
|
|
|
|
BLI_INLINE void ghash_buckets_reset(GHash *gh, const unsigned int nentries)
|
|
|
|
{
|
|
|
|
MEM_SAFE_FREE(gh->buckets);
|
|
|
|
|
|
|
|
#ifdef GHASH_USE_MODULO_BUCKETS
|
|
|
|
gh->cursize = 0;
|
|
|
|
gh->size_min = 0;
|
|
|
|
gh->nbuckets = hashsizes[gh->cursize];
|
|
|
|
#else
|
|
|
|
gh->bucket_bit = GHASH_BUCKET_BIT_MIN;
|
|
|
|
gh->bucket_bit_min = GHASH_BUCKET_BIT_MIN;
|
|
|
|
gh->nbuckets = 1u << gh->bucket_bit;
|
|
|
|
gh->bucket_mask = gh->nbuckets - 1;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
gh->limit_grow = GHASH_LIMIT_GROW(gh->nbuckets);
|
|
|
|
gh->limit_shrink = GHASH_LIMIT_SHRINK(gh->nbuckets);
|
|
|
|
|
|
|
|
gh->nentries = 0;
|
|
|
|
|
|
|
|
ghash_buckets_expand(gh, nentries, (nentries != 0));
|
2013-08-18 03:41:39 +00:00
|
|
|
}
|
2012-03-11 23:47:41 +00:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Internal lookup function.
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* Takes hash and bucket_index arguments to avoid calling #ghash_keyhash and #ghash_bucket_index multiple times.
|
2013-08-25 12:17:46 +00:00
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_INLINE Entry *ghash_lookup_entry_ex(
|
|
|
|
GHash *gh, const void *key, const unsigned int bucket_index)
|
2013-08-18 03:41:39 +00:00
|
|
|
{
|
|
|
|
Entry *e;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/* If we do not store GHash, not worth computing it for each entry here!
|
|
|
|
* Typically, comparison function will be quicker, and since it's needed in the end anyway... */
|
|
|
|
for (e = gh->buckets[bucket_index]; e; e = e->next) {
|
|
|
|
if (UNLIKELY(gh->cmpfp(key, e->key) == false)) {
|
|
|
|
return e;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
2012-03-11 23:47:41 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/**
|
|
|
|
* Internal lookup function, returns previous entry of target one too.
|
2015-11-29 17:11:40 +11:00
|
|
|
* Takes bucket_index argument to avoid calling #ghash_keyhash and #ghash_bucket_index multiple times.
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* Useful when modifying buckets somehow (like removing an entry...).
|
|
|
|
*/
|
|
|
|
BLI_INLINE Entry *ghash_lookup_entry_prev_ex(
|
2015-11-29 17:49:56 +11:00
|
|
|
GHash *gh, const void *key,
|
|
|
|
Entry **r_e_prev, const unsigned int bucket_index)
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
{
|
|
|
|
/* If we do not store GHash, not worth computing it for each entry here!
|
|
|
|
* Typically, comparison function will be quicker, and since it's needed in the end anyway... */
|
2015-11-29 17:49:56 +11:00
|
|
|
for (Entry *e_prev = NULL, *e = gh->buckets[bucket_index]; e; e_prev = e, e = e->next) {
|
2014-09-25 14:29:23 +02:00
|
|
|
if (UNLIKELY(gh->cmpfp(key, e->key) == false)) {
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
*r_e_prev = e_prev;
|
2013-08-18 03:41:39 +00:00
|
|
|
return e;
|
|
|
|
}
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
|
|
|
*r_e_prev = NULL;
|
2013-08-18 03:41:39 +00:00
|
|
|
return NULL;
|
2002-10-12 11:37:38 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Internal lookup function. Only wraps #ghash_lookup_entry_ex
|
|
|
|
*/
|
2013-08-18 03:41:39 +00:00
|
|
|
BLI_INLINE Entry *ghash_lookup_entry(GHash *gh, const void *key)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2013-08-18 03:41:39 +00:00
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
return ghash_lookup_entry_ex(gh, key, bucket_index);
|
2004-01-07 06:13:43 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 20:00:19 +00:00
|
|
|
static GHash *ghash_new(GHashHashFP hashfp, GHashCmpFP cmpfp, const char *info,
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int nentries_reserve, const unsigned int flag)
|
2013-08-25 20:00:19 +00:00
|
|
|
{
|
|
|
|
GHash *gh = MEM_mallocN(sizeof(*gh), info);
|
|
|
|
|
|
|
|
gh->hashfp = hashfp;
|
|
|
|
gh->cmpfp = cmpfp;
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
gh->buckets = NULL;
|
|
|
|
gh->flag = flag;
|
2013-08-25 20:00:19 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
ghash_buckets_reset(gh, nentries_reserve);
|
|
|
|
gh->entrypool = BLI_mempool_create(GHASH_ENTRY_SIZE(flag & GHASH_FLAG_IS_GSET), 64, 64, BLI_MEMPOOL_NOP);
|
2013-08-25 20:00:19 +00:00
|
|
|
|
|
|
|
return gh;
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Internal insert function.
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* Takes hash and bucket_index arguments to avoid calling #ghash_keyhash and #ghash_bucket_index multiple times.
|
2013-08-25 12:17:46 +00:00
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_INLINE void ghash_insert_ex(
|
|
|
|
GHash *gh, void *key, void *val, const unsigned int bucket_index)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHashEntry *e = BLI_mempool_alloc(gh->entrypool);
|
|
|
|
|
2013-08-18 00:36:04 +00:00
|
|
|
BLI_assert((gh->flag & GHASH_FLAG_ALLOW_DUPES) || (BLI_ghash_haskey(gh, key) == 0));
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
2013-08-18 00:36:04 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
e->e.next = gh->buckets[bucket_index];
|
|
|
|
e->e.key = key;
|
2013-08-26 13:41:13 +00:00
|
|
|
e->val = val;
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
gh->buckets[bucket_index] = (Entry *)e;
|
2011-08-12 02:23:06 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
ghash_buckets_expand(gh, ++gh->nentries, false);
|
2013-08-26 13:41:13 +00:00
|
|
|
}
|
2011-08-12 02:23:06 +00:00
|
|
|
|
2015-04-06 19:55:08 +10:00
|
|
|
/**
|
|
|
|
* Insert function that takes a pre-allocated entry.
|
|
|
|
*/
|
|
|
|
BLI_INLINE void ghash_insert_ex_keyonly_entry(
|
|
|
|
GHash *gh, void *key, const unsigned int bucket_index,
|
|
|
|
Entry *e)
|
|
|
|
{
|
|
|
|
BLI_assert((gh->flag & GHASH_FLAG_ALLOW_DUPES) || (BLI_ghash_haskey(gh, key) == 0));
|
|
|
|
|
|
|
|
e->next = gh->buckets[bucket_index];
|
|
|
|
e->key = key;
|
|
|
|
gh->buckets[bucket_index] = e;
|
|
|
|
|
|
|
|
ghash_buckets_expand(gh, ++gh->nentries, false);
|
|
|
|
}
|
|
|
|
|
2013-08-26 13:41:13 +00:00
|
|
|
/**
|
|
|
|
* Insert function that doesn't set the value (use for GSet)
|
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_INLINE void ghash_insert_ex_keyonly(
|
|
|
|
GHash *gh, void *key, const unsigned int bucket_index)
|
2013-08-26 13:41:13 +00:00
|
|
|
{
|
2015-04-06 09:03:32 +10:00
|
|
|
Entry *e = BLI_mempool_alloc(gh->entrypool);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
2013-08-26 13:41:13 +00:00
|
|
|
BLI_assert((gh->flag & GHASH_FLAG_ALLOW_DUPES) || (BLI_ghash_haskey(gh, key) == 0));
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
BLI_assert((gh->flag & GHASH_FLAG_IS_GSET) != 0);
|
|
|
|
|
|
|
|
e->next = gh->buckets[bucket_index];
|
2013-08-26 13:41:13 +00:00
|
|
|
e->key = key;
|
2015-04-06 09:03:32 +10:00
|
|
|
gh->buckets[bucket_index] = e;
|
2011-08-12 02:23:06 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
ghash_buckets_expand(gh, ++gh->nentries, false);
|
2013-08-26 13:41:13 +00:00
|
|
|
}
|
2013-08-25 20:00:19 +00:00
|
|
|
|
2013-08-26 13:41:13 +00:00
|
|
|
BLI_INLINE void ghash_insert(GHash *gh, void *key, void *val)
|
|
|
|
{
|
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
|
|
|
|
ghash_insert_ex(gh, key, val, bucket_index);
|
|
|
|
}
|
|
|
|
|
|
|
|
BLI_INLINE bool ghash_insert_safe(
|
2015-11-29 16:52:05 +11:00
|
|
|
GHash *gh, void *key, void *val, const bool override,
|
|
|
|
GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp)
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
{
|
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
GHashEntry *e = (GHashEntry *)ghash_lookup_entry_ex(gh, key, bucket_index);
|
|
|
|
|
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
|
|
|
|
|
|
|
if (e) {
|
|
|
|
if (override) {
|
2015-11-29 16:52:05 +11:00
|
|
|
if (keyfreefp) {
|
|
|
|
keyfreefp(e->e.key);
|
|
|
|
}
|
|
|
|
if (valfreefp) {
|
|
|
|
valfreefp(e->val);
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
e->e.key = key;
|
|
|
|
e->val = val;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
ghash_insert_ex(gh, key, val, bucket_index);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-11-29 16:52:05 +11:00
|
|
|
BLI_INLINE bool ghash_insert_safe_keyonly(
|
|
|
|
GHash *gh, void *key, const bool override,
|
|
|
|
GHashKeyFreeFP keyfreefp)
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
{
|
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
2015-04-06 09:03:32 +10:00
|
|
|
Entry *e = ghash_lookup_entry_ex(gh, key, bucket_index);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
|
|
|
BLI_assert((gh->flag & GHASH_FLAG_IS_GSET) != 0);
|
|
|
|
|
|
|
|
if (e) {
|
|
|
|
if (override) {
|
2015-11-29 16:52:05 +11:00
|
|
|
if (keyfreefp) {
|
|
|
|
keyfreefp(e->key);
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
e->key = key;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
ghash_insert_ex_keyonly(gh, key, bucket_index);
|
|
|
|
return true;
|
|
|
|
}
|
2011-08-12 02:23:06 +00:00
|
|
|
}
|
2013-08-25 21:02:31 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Remove the entry and return it, caller must free from gh->entrypool.
|
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
static Entry *ghash_remove_ex(
|
2015-11-29 16:52:05 +11:00
|
|
|
GHash *gh, const void *key,
|
|
|
|
GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp,
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int bucket_index)
|
2013-08-25 21:02:31 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
Entry *e_prev;
|
|
|
|
Entry *e = ghash_lookup_entry_prev_ex(gh, key, &e_prev, bucket_index);
|
2013-08-25 21:02:31 +00:00
|
|
|
|
2015-03-23 22:51:12 +11:00
|
|
|
BLI_assert(!valfreefp || !(gh->flag & GHASH_FLAG_IS_GSET));
|
2013-08-25 21:02:31 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
if (e) {
|
2015-11-29 16:52:05 +11:00
|
|
|
if (keyfreefp) {
|
|
|
|
keyfreefp(e->key);
|
|
|
|
}
|
|
|
|
if (valfreefp) {
|
|
|
|
valfreefp(((GHashEntry *)e)->val);
|
|
|
|
}
|
2013-08-25 21:02:31 +00:00
|
|
|
|
2015-11-29 16:52:05 +11:00
|
|
|
if (e_prev) {
|
|
|
|
e_prev->next = e->next;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
gh->buckets[bucket_index] = e->next;
|
|
|
|
}
|
2013-08-25 21:02:31 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
ghash_buckets_contract(gh, --gh->nentries, false, false);
|
2013-08-25 21:02:31 +00:00
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return e;
|
2013-08-25 21:02:31 +00:00
|
|
|
}
|
|
|
|
|
2016-02-20 15:17:40 +01:00
|
|
|
/**
|
|
|
|
* Remove a random entry and return it (or NULL if empty), caller must free from gh->entrypool.
|
|
|
|
*/
|
|
|
|
static Entry *ghash_pop(GHash *gh, GHashIterState *state)
|
|
|
|
{
|
|
|
|
unsigned int curr_bucket = state->curr_bucket;
|
|
|
|
if (gh->nentries == 0) {
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Note: using first_bucket_index here allows us to avoid potential huge number of loops over buckets,
|
|
|
|
* in case we are poping from a large ghash with few items in it... */
|
|
|
|
curr_bucket = ghash_find_next_bucket_index(gh, curr_bucket);
|
|
|
|
|
|
|
|
Entry *e = gh->buckets[curr_bucket];
|
|
|
|
BLI_assert(e);
|
|
|
|
|
|
|
|
ghash_remove_ex(gh, e->key, NULL, NULL, curr_bucket);
|
|
|
|
|
|
|
|
state->curr_bucket = curr_bucket;
|
|
|
|
return e;
|
|
|
|
}
|
|
|
|
|
2013-08-25 21:02:31 +00:00
|
|
|
/**
|
|
|
|
* Run free callbacks for freeing entries.
|
|
|
|
*/
|
2015-11-29 16:52:05 +11:00
|
|
|
static void ghash_free_cb(
|
|
|
|
GHash *gh,
|
|
|
|
GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp)
|
2013-08-25 21:02:31 +00:00
|
|
|
{
|
|
|
|
unsigned int i;
|
|
|
|
|
2015-03-21 16:40:17 +11:00
|
|
|
BLI_assert(keyfreefp || valfreefp);
|
|
|
|
BLI_assert(!valfreefp || !(gh->flag & GHASH_FLAG_IS_GSET));
|
2013-08-25 21:02:31 +00:00
|
|
|
|
|
|
|
for (i = 0; i < gh->nbuckets; i++) {
|
|
|
|
Entry *e;
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
for (e = gh->buckets[i]; e; e = e->next) {
|
2015-11-29 16:52:05 +11:00
|
|
|
if (keyfreefp) {
|
|
|
|
keyfreefp(e->key);
|
|
|
|
}
|
|
|
|
if (valfreefp) {
|
|
|
|
valfreefp(((GHashEntry *)e)->val);
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Copy the GHash.
|
|
|
|
*/
|
|
|
|
static GHash *ghash_copy(GHash *gh, GHashKeyCopyFP keycopyfp, GHashValCopyFP valcopyfp)
|
|
|
|
{
|
|
|
|
GHash *gh_new;
|
|
|
|
unsigned int i;
|
|
|
|
/* This allows us to be sure to get the same number of buckets in gh_new as in ghash. */
|
|
|
|
const unsigned int reserve_nentries_new = MAX2(GHASH_LIMIT_GROW(gh->nbuckets) - 1, gh->nentries);
|
|
|
|
|
|
|
|
BLI_assert(!valcopyfp || !(gh->flag & GHASH_FLAG_IS_GSET));
|
|
|
|
|
|
|
|
gh_new = ghash_new(gh->hashfp, gh->cmpfp, __func__, 0, gh->flag);
|
|
|
|
ghash_buckets_expand(gh_new, reserve_nentries_new, false);
|
|
|
|
|
|
|
|
BLI_assert(gh_new->nbuckets == gh->nbuckets);
|
|
|
|
|
|
|
|
for (i = 0; i < gh->nbuckets; i++) {
|
|
|
|
Entry *e;
|
|
|
|
|
|
|
|
for (e = gh->buckets[i]; e; e = e->next) {
|
|
|
|
Entry *e_new = BLI_mempool_alloc(gh_new->entrypool);
|
|
|
|
ghash_entry_copy(gh_new, e_new, gh, e, keycopyfp, valcopyfp);
|
2013-08-25 21:02:31 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/* Warning!
|
|
|
|
* This means entries in buckets in new copy will be in reversed order!
|
|
|
|
* This shall not be an issue though, since order should never be assumed in ghash. */
|
|
|
|
|
|
|
|
/* Note: We can use 'i' here, since we are sure that 'gh' and 'gh_new' have the same number of buckets! */
|
|
|
|
e_new->next = gh_new->buckets[i];
|
|
|
|
gh_new->buckets[i] = e_new;
|
2013-08-25 21:02:31 +00:00
|
|
|
}
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
gh_new->nentries = gh->nentries;
|
|
|
|
|
|
|
|
return gh_new;
|
2013-08-25 21:02:31 +00:00
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \} */
|
2011-08-12 02:23:06 +00:00
|
|
|
|
2013-08-18 03:41:39 +00:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \name Public API
|
|
|
|
* \{ */
|
2013-08-18 03:41:39 +00:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Creates a new, empty GHash.
|
|
|
|
*
|
|
|
|
* \param hashfp Hash callback.
|
|
|
|
* \param cmpfp Comparison callback.
|
|
|
|
* \param info Identifier string for the GHash.
|
|
|
|
* \param nentries_reserve Optionally reserve the number of members that the hash will hold.
|
|
|
|
* Use this to avoid resizing buckets if the size is known or can be closely approximated.
|
|
|
|
* \return An empty GHash.
|
|
|
|
*/
|
2013-08-24 13:04:03 +00:00
|
|
|
GHash *BLI_ghash_new_ex(GHashHashFP hashfp, GHashCmpFP cmpfp, const char *info,
|
|
|
|
const unsigned int nentries_reserve)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return ghash_new(hashfp, cmpfp, info, nentries_reserve, 0);
|
2013-08-18 03:41:39 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Wraps #BLI_ghash_new_ex with zero entries reserved.
|
|
|
|
*/
|
2013-08-24 13:04:03 +00:00
|
|
|
GHash *BLI_ghash_new(GHashHashFP hashfp, GHashCmpFP cmpfp, const char *info)
|
|
|
|
{
|
|
|
|
return BLI_ghash_new_ex(hashfp, cmpfp, info, 0);
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/**
|
|
|
|
* Copy given GHash. Keys and values are also copied if relevant callback is provided, else pointers remain the same.
|
|
|
|
*/
|
|
|
|
GHash *BLI_ghash_copy(GHash *gh, GHashKeyCopyFP keycopyfp, GHashValCopyFP valcopyfp)
|
|
|
|
{
|
|
|
|
return ghash_copy(gh, keycopyfp, valcopyfp);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2015-09-14 02:21:15 +10:00
|
|
|
* Reserve given amount of entries (resize \a gh accordingly if needed).
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
*/
|
|
|
|
void BLI_ghash_reserve(GHash *gh, const unsigned int nentries_reserve)
|
|
|
|
{
|
|
|
|
ghash_buckets_expand(gh, nentries_reserve, true);
|
|
|
|
ghash_buckets_contract(gh, nentries_reserve, true, false);
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* \return size of the GHash.
|
|
|
|
*/
|
2015-04-13 13:45:14 +10:00
|
|
|
unsigned int BLI_ghash_size(GHash *gh)
|
2013-08-18 03:41:39 +00:00
|
|
|
{
|
2015-04-13 13:45:14 +10:00
|
|
|
return gh->nentries;
|
2013-08-18 03:41:39 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Insert a key/value pair into the \a gh.
|
|
|
|
*
|
|
|
|
* \note Duplicates are not checked,
|
|
|
|
* the caller is expected to ensure elements are unique unless
|
|
|
|
* GHASH_FLAG_ALLOW_DUPES flag is set.
|
|
|
|
*/
|
2013-08-18 03:41:39 +00:00
|
|
|
void BLI_ghash_insert(GHash *gh, void *key, void *val)
|
|
|
|
{
|
2013-08-26 13:41:13 +00:00
|
|
|
ghash_insert(gh, key, val);
|
2013-08-18 03:41:39 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2013-08-25 12:17:46 +00:00
|
|
|
* Inserts a new value to a key that may already be in ghash.
|
|
|
|
*
|
2013-08-18 03:41:39 +00:00
|
|
|
* Avoids #BLI_ghash_remove, #BLI_ghash_insert calls (double lookups)
|
2013-08-25 20:00:19 +00:00
|
|
|
*
|
|
|
|
* \returns true if a new key has been added.
|
2013-08-18 03:41:39 +00:00
|
|
|
*/
|
2013-08-25 20:00:19 +00:00
|
|
|
bool BLI_ghash_reinsert(GHash *gh, void *key, void *val, GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp)
|
2013-08-18 03:41:39 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return ghash_insert_safe(gh, key, val, true, keyfreefp, valfreefp);
|
2011-08-12 02:23:06 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Lookup the value of \a key in \a gh.
|
|
|
|
*
|
|
|
|
* \param key The key to lookup.
|
|
|
|
* \returns the value for \a key or NULL.
|
|
|
|
*
|
|
|
|
* \note When NULL is a valid value, use #BLI_ghash_lookup_p to differentiate a missing key
|
|
|
|
* from a key with a NULL value. (Avoids calling #BLI_ghash_haskey before #BLI_ghash_lookup)
|
|
|
|
*/
|
2013-08-18 01:00:52 +00:00
|
|
|
void *BLI_ghash_lookup(GHash *gh, const void *key)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHashEntry *e = (GHashEntry *)ghash_lookup_entry(gh, key);
|
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
2013-08-18 01:00:52 +00:00
|
|
|
return e ? e->val : NULL;
|
|
|
|
}
|
|
|
|
|
2014-04-22 00:55:30 +10:00
|
|
|
/**
|
|
|
|
* A version of #BLI_ghash_lookup which accepts a fallback argument.
|
|
|
|
*/
|
|
|
|
void *BLI_ghash_lookup_default(GHash *gh, const void *key, void *val_default)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHashEntry *e = (GHashEntry *)ghash_lookup_entry(gh, key);
|
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
2014-04-22 00:55:30 +10:00
|
|
|
return e ? e->val : val_default;
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Lookup a pointer to the value of \a key in \a gh.
|
|
|
|
*
|
|
|
|
* \param key The key to lookup.
|
|
|
|
* \returns the pointer to value for \a key or NULL.
|
|
|
|
*
|
2014-08-11 13:36:35 +10:00
|
|
|
* \note This has 2 main benefits over #BLI_ghash_lookup.
|
2013-08-25 12:17:46 +00:00
|
|
|
* - A NULL return always means that \a key isn't in \a gh.
|
|
|
|
* - The value can be modified in-place without further function calls (faster).
|
|
|
|
*/
|
2013-08-18 01:00:52 +00:00
|
|
|
void **BLI_ghash_lookup_p(GHash *gh, const void *key)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHashEntry *e = (GHashEntry *)ghash_lookup_entry(gh, key);
|
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
2013-08-18 01:00:52 +00:00
|
|
|
return e ? &e->val : NULL;
|
|
|
|
}
|
|
|
|
|
2015-04-06 19:55:08 +10:00
|
|
|
/**
|
|
|
|
* Ensure \a key is exists in \a gh.
|
|
|
|
*
|
|
|
|
* This handles the common situation where the caller needs ensure a key is added to \a gh,
|
|
|
|
* constructing a new value in the case the key isn't found.
|
|
|
|
* Otherwise use the existing value.
|
|
|
|
*
|
|
|
|
* Such situations typically incur multiple lookups, however this function
|
|
|
|
* avoids them by ensuring the key is added,
|
|
|
|
* returning a pointer to the value so it can be used or initialized by the caller.
|
|
|
|
*
|
|
|
|
* \returns true when the value didn't need to be added.
|
|
|
|
* (when false, the caller _must_ initialize the value).
|
|
|
|
*/
|
|
|
|
bool BLI_ghash_ensure_p(GHash *gh, void *key, void ***r_val)
|
|
|
|
{
|
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
GHashEntry *e = (GHashEntry *)ghash_lookup_entry_ex(gh, key, bucket_index);
|
|
|
|
const bool haskey = (e != NULL);
|
|
|
|
|
|
|
|
if (!haskey) {
|
|
|
|
e = BLI_mempool_alloc(gh->entrypool);
|
|
|
|
ghash_insert_ex_keyonly_entry(gh, key, bucket_index, (Entry *)e);
|
|
|
|
}
|
|
|
|
|
|
|
|
*r_val = &e->val;
|
|
|
|
return haskey;
|
|
|
|
}
|
|
|
|
|
2015-05-11 09:27:05 +10:00
|
|
|
/**
|
|
|
|
* A version of #BLI_ghash_ensure_p copies the key on insertion.
|
|
|
|
*/
|
|
|
|
bool BLI_ghash_ensure_p_ex(
|
|
|
|
GHash *gh, const void *key, void ***r_val,
|
|
|
|
GHashKeyCopyFP keycopyfp)
|
|
|
|
{
|
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
GHashEntry *e = (GHashEntry *)ghash_lookup_entry_ex(gh, key, bucket_index);
|
|
|
|
const bool haskey = (e != NULL);
|
|
|
|
|
|
|
|
if (!haskey) {
|
|
|
|
/* keycopyfp(key) is the only difference to BLI_ghash_ensure_p */
|
|
|
|
e = BLI_mempool_alloc(gh->entrypool);
|
|
|
|
ghash_insert_ex_keyonly_entry(gh, keycopyfp(key), bucket_index, (Entry *)e);
|
|
|
|
}
|
|
|
|
|
|
|
|
*r_val = &e->val;
|
|
|
|
return haskey;
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Remove \a key from \a gh, or return false if the key wasn't found.
|
|
|
|
*
|
|
|
|
* \param key The key to remove.
|
|
|
|
* \param keyfreefp Optional callback to free the key.
|
|
|
|
* \param valfreefp Optional callback to free the value.
|
|
|
|
* \return true if \a key was removed from \a gh.
|
|
|
|
*/
|
2015-05-11 12:39:08 +10:00
|
|
|
bool BLI_ghash_remove(GHash *gh, const void *key, GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp)
|
2011-08-12 02:23:06 +00:00
|
|
|
{
|
2013-08-18 03:41:39 +00:00
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
Entry *e = ghash_remove_ex(gh, key, keyfreefp, valfreefp, bucket_index);
|
2013-08-25 21:02:31 +00:00
|
|
|
if (e) {
|
|
|
|
BLI_mempool_free(gh->entrypool, e);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
return false;
|
2011-08-12 02:23:06 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-05-28 19:33:14 +00:00
|
|
|
/* same as above but return the value,
|
|
|
|
* no free value argument since it will be returned */
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Remove \a key from \a gh, returning the value or NULL if the key wasn't found.
|
|
|
|
*
|
|
|
|
* \param key The key to remove.
|
|
|
|
* \param keyfreefp Optional callback to free the key.
|
|
|
|
* \return the value of \a key int \a gh or NULL.
|
|
|
|
*/
|
2015-05-11 12:39:08 +10:00
|
|
|
void *BLI_ghash_popkey(GHash *gh, const void *key, GHashKeyFreeFP keyfreefp)
|
2012-05-28 19:33:14 +00:00
|
|
|
{
|
2013-08-18 03:41:39 +00:00
|
|
|
const unsigned int hash = ghash_keyhash(gh, key);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int bucket_index = ghash_bucket_index(gh, hash);
|
|
|
|
GHashEntry *e = (GHashEntry *)ghash_remove_ex(gh, key, keyfreefp, NULL, bucket_index);
|
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
2013-08-25 21:02:31 +00:00
|
|
|
if (e) {
|
|
|
|
void *val = e->val;
|
|
|
|
BLI_mempool_free(gh->entrypool, e);
|
|
|
|
return val;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
return NULL;
|
2012-05-28 19:33:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* \return true if the \a key is in \a gh.
|
|
|
|
*/
|
2013-03-04 19:27:51 +00:00
|
|
|
bool BLI_ghash_haskey(GHash *gh, const void *key)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2013-08-18 01:00:52 +00:00
|
|
|
return (ghash_lookup_entry(gh, key) != NULL);
|
|
|
|
}
|
2011-08-12 02:23:06 +00:00
|
|
|
|
2016-02-20 15:17:40 +01:00
|
|
|
/**
|
|
|
|
* Remove a random entry from \a ghp, returning true if a key/value pair could be removed, false otherwise.
|
|
|
|
*
|
|
|
|
* \param r_key: The removed key.
|
|
|
|
* \param r_val: The removed value.
|
|
|
|
* \param state: Used for efficient removal.
|
|
|
|
* \return true if there was somethjing to pop, false if ghash was already empty.
|
|
|
|
*/
|
|
|
|
bool BLI_ghash_pop(
|
|
|
|
GHash *gh, GHashIterState *state,
|
|
|
|
void **r_key, void **r_val)
|
|
|
|
{
|
|
|
|
GHashEntry *e = (GHashEntry *)ghash_pop(gh, state);
|
|
|
|
|
|
|
|
BLI_assert(!(gh->flag & GHASH_FLAG_IS_GSET));
|
|
|
|
|
|
|
|
if (e) {
|
|
|
|
*r_key = e->e.key;
|
|
|
|
*r_val = e->val;
|
|
|
|
|
|
|
|
BLI_mempool_free(gh->entrypool, e);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
*r_key = *r_val = NULL;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Reset \a gh clearing all entries.
|
|
|
|
*
|
|
|
|
* \param keyfreefp Optional callback to free the key.
|
|
|
|
* \param valfreefp Optional callback to free the value.
|
2013-08-25 16:16:38 +00:00
|
|
|
* \param nentries_reserve Optionally reserve the number of members that the hash will hold.
|
2013-08-25 12:17:46 +00:00
|
|
|
*/
|
2013-08-25 16:16:38 +00:00
|
|
|
void BLI_ghash_clear_ex(GHash *gh, GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp,
|
|
|
|
const unsigned int nentries_reserve)
|
2013-08-18 01:00:52 +00:00
|
|
|
{
|
2013-08-25 21:02:31 +00:00
|
|
|
if (keyfreefp || valfreefp)
|
|
|
|
ghash_free_cb(gh, keyfreefp, valfreefp);
|
2013-08-18 01:00:52 +00:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
ghash_buckets_reset(gh, nentries_reserve);
|
2013-08-25 16:16:38 +00:00
|
|
|
BLI_mempool_clear_ex(gh->entrypool, nentries_reserve ? (int)nentries_reserve : -1);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Wraps #BLI_ghash_clear_ex with zero entries reserved.
|
|
|
|
*/
|
|
|
|
void BLI_ghash_clear(GHash *gh, GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp)
|
|
|
|
{
|
|
|
|
BLI_ghash_clear_ex(gh, keyfreefp, valfreefp, 0);
|
2011-08-12 02:23:06 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Frees the GHash and its members.
|
|
|
|
*
|
|
|
|
* \param gh The GHash to free.
|
|
|
|
* \param keyfreefp Optional callback to free the key.
|
|
|
|
* \param valfreefp Optional callback to free the value.
|
|
|
|
*/
|
2011-12-17 00:52:36 +00:00
|
|
|
void BLI_ghash_free(GHash *gh, GHashKeyFreeFP keyfreefp, GHashValFreeFP valfreefp)
|
|
|
|
{
|
2013-08-25 21:02:31 +00:00
|
|
|
BLI_assert((int)gh->nentries == BLI_mempool_count(gh->entrypool));
|
|
|
|
if (keyfreefp || valfreefp)
|
|
|
|
ghash_free_cb(gh, keyfreefp, valfreefp);
|
2012-03-11 23:47:41 +00:00
|
|
|
|
2010-01-23 11:25:20 +00:00
|
|
|
MEM_freeN(gh->buckets);
|
2009-02-12 18:05:34 +00:00
|
|
|
BLI_mempool_destroy(gh->entrypool);
|
2002-10-12 11:37:38 +00:00
|
|
|
MEM_freeN(gh);
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Sets a GHash flag.
|
|
|
|
*/
|
2013-08-24 20:16:14 +00:00
|
|
|
void BLI_ghash_flag_set(GHash *gh, unsigned int flag)
|
2013-08-18 00:36:04 +00:00
|
|
|
{
|
|
|
|
gh->flag |= flag;
|
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/**
|
|
|
|
* Clear a GHash flag.
|
|
|
|
*/
|
2013-08-24 20:16:14 +00:00
|
|
|
void BLI_ghash_flag_clear(GHash *gh, unsigned int flag)
|
2013-08-18 00:36:04 +00:00
|
|
|
{
|
2013-08-24 20:16:14 +00:00
|
|
|
gh->flag &= ~flag;
|
2013-08-18 00:36:04 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \} */
|
|
|
|
|
2013-08-18 00:36:04 +00:00
|
|
|
|
|
|
|
/* -------------------------------------------------------------------- */
|
|
|
|
/* GHash Iterator API */
|
2002-10-12 11:37:38 +00:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \name Iterator API
|
|
|
|
* \{ */
|
|
|
|
|
2013-08-25 16:28:48 +00:00
|
|
|
/**
|
|
|
|
* Create a new GHashIterator. The hash table must not be mutated
|
|
|
|
* while the iterator is in use, and the iterator will step exactly
|
|
|
|
* BLI_ghash_size(gh) times before becoming done.
|
|
|
|
*
|
|
|
|
* \param gh The GHash to iterate over.
|
|
|
|
* \return Pointer to a new DynStr.
|
|
|
|
*/
|
2011-12-17 00:52:36 +00:00
|
|
|
GHashIterator *BLI_ghashIterator_new(GHash *gh)
|
|
|
|
{
|
2012-03-11 23:47:41 +00:00
|
|
|
GHashIterator *ghi = MEM_mallocN(sizeof(*ghi), "ghash iterator");
|
2013-07-21 17:05:41 +00:00
|
|
|
BLI_ghashIterator_init(ghi, gh);
|
2004-01-07 06:13:43 +00:00
|
|
|
return ghi;
|
|
|
|
}
|
2013-08-25 16:28:48 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Init an already allocated GHashIterator. The hash table must not
|
|
|
|
* be mutated while the iterator is in use, and the iterator will
|
|
|
|
* step exactly BLI_ghash_size(gh) times before becoming done.
|
|
|
|
*
|
|
|
|
* \param ghi The GHashIterator to initialize.
|
|
|
|
* \param gh The GHash to iterate over.
|
|
|
|
*/
|
2015-11-26 12:35:27 +11:00
|
|
|
void BLI_ghashIterator_init(GHashIterator *ghi, GHash *gh)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2012-03-11 23:47:41 +00:00
|
|
|
ghi->gh = gh;
|
|
|
|
ghi->curEntry = NULL;
|
2013-07-21 17:05:41 +00:00
|
|
|
ghi->curBucket = UINT_MAX; /* wraps to zero */
|
2014-04-21 02:50:07 +10:00
|
|
|
if (gh->nentries) {
|
2014-08-07 11:19:55 +10:00
|
|
|
do {
|
2014-04-21 02:50:07 +10:00
|
|
|
ghi->curBucket++;
|
|
|
|
if (UNLIKELY(ghi->curBucket == ghi->gh->nbuckets))
|
|
|
|
break;
|
|
|
|
ghi->curEntry = ghi->gh->buckets[ghi->curBucket];
|
2014-08-07 11:19:55 +10:00
|
|
|
} while (!ghi->curEntry);
|
2014-04-21 02:50:07 +10:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Steps the iterator to the next index.
|
|
|
|
*
|
|
|
|
* \param ghi The iterator.
|
|
|
|
*/
|
|
|
|
void BLI_ghashIterator_step(GHashIterator *ghi)
|
|
|
|
{
|
|
|
|
if (ghi->curEntry) {
|
|
|
|
ghi->curEntry = ghi->curEntry->next;
|
|
|
|
while (!ghi->curEntry) {
|
|
|
|
ghi->curBucket++;
|
|
|
|
if (ghi->curBucket == ghi->gh->nbuckets)
|
|
|
|
break;
|
|
|
|
ghi->curEntry = ghi->gh->buckets[ghi->curBucket];
|
|
|
|
}
|
2008-05-27 13:22:17 +00:00
|
|
|
}
|
|
|
|
}
|
2013-08-25 16:28:48 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Free a GHashIterator.
|
|
|
|
*
|
|
|
|
* \param ghi The iterator to free.
|
|
|
|
*/
|
2011-12-17 00:52:36 +00:00
|
|
|
void BLI_ghashIterator_free(GHashIterator *ghi)
|
|
|
|
{
|
2010-01-23 11:25:20 +00:00
|
|
|
MEM_freeN(ghi);
|
2004-01-07 06:13:43 +00:00
|
|
|
}
|
|
|
|
|
2014-04-08 15:50:38 +10:00
|
|
|
/* inline functions now */
|
|
|
|
#if 0
|
2013-08-25 16:28:48 +00:00
|
|
|
/**
|
|
|
|
* Retrieve the key from an iterator.
|
|
|
|
*
|
|
|
|
* \param ghi The iterator.
|
|
|
|
* \return The key at the current index, or NULL if the
|
|
|
|
* iterator is done.
|
|
|
|
*/
|
2011-12-17 00:52:36 +00:00
|
|
|
void *BLI_ghashIterator_getKey(GHashIterator *ghi)
|
|
|
|
{
|
2014-04-08 15:50:38 +10:00
|
|
|
return ghi->curEntry->key;
|
2004-01-07 06:13:43 +00:00
|
|
|
}
|
2013-08-25 16:28:48 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Retrieve the value from an iterator.
|
|
|
|
*
|
|
|
|
* \param ghi The iterator.
|
|
|
|
* \return The value at the current index, or NULL if the
|
|
|
|
* iterator is done.
|
|
|
|
*/
|
2011-12-17 00:52:36 +00:00
|
|
|
void *BLI_ghashIterator_getValue(GHashIterator *ghi)
|
|
|
|
{
|
2014-04-08 15:50:38 +10:00
|
|
|
return ghi->curEntry->val;
|
2004-01-07 06:13:43 +00:00
|
|
|
}
|
|
|
|
|
2013-09-02 02:58:53 +00:00
|
|
|
/**
|
|
|
|
* Retrieve the value from an iterator.
|
|
|
|
*
|
|
|
|
* \param ghi The iterator.
|
|
|
|
* \return The value at the current index, or NULL if the
|
|
|
|
* iterator is done.
|
|
|
|
*/
|
|
|
|
void **BLI_ghashIterator_getValue_p(GHashIterator *ghi)
|
|
|
|
{
|
2014-04-08 15:50:38 +10:00
|
|
|
return &ghi->curEntry->val;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Determine if an iterator is done (has reached the end of
|
|
|
|
* the hash table).
|
|
|
|
*
|
|
|
|
* \param ghi The iterator.
|
|
|
|
* \return True if done, False otherwise.
|
|
|
|
*/
|
|
|
|
bool BLI_ghashIterator_done(GHashIterator *ghi)
|
|
|
|
{
|
|
|
|
return ghi->curEntry == NULL;
|
2013-09-02 02:58:53 +00:00
|
|
|
}
|
2014-04-08 15:50:38 +10:00
|
|
|
#endif
|
2013-09-02 02:58:53 +00:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \} */
|
|
|
|
|
|
|
|
|
|
|
|
/** \name Generic Key Hash & Comparison Functions
|
|
|
|
* \{ */
|
|
|
|
|
2004-01-07 06:13:43 +00:00
|
|
|
/***/
|
|
|
|
|
2013-08-24 20:30:08 +00:00
|
|
|
#if 0
|
|
|
|
/* works but slower */
|
2011-12-17 00:52:36 +00:00
|
|
|
unsigned int BLI_ghashutil_ptrhash(const void *key)
|
|
|
|
{
|
2009-01-17 00:51:42 +00:00
|
|
|
return (unsigned int)(intptr_t)key;
|
2002-10-12 11:37:38 +00:00
|
|
|
}
|
2013-08-24 20:30:08 +00:00
|
|
|
#else
|
|
|
|
/* based python3.3's pointer hashing function */
|
|
|
|
unsigned int BLI_ghashutil_ptrhash(const void *key)
|
|
|
|
{
|
|
|
|
size_t y = (size_t)key;
|
|
|
|
/* bottom 3 or 4 bits are likely to be 0; rotate y by 4 to avoid
|
|
|
|
* excessive hash collisions for dicts and sets */
|
|
|
|
y = (y >> 4) | (y << (8 * sizeof(void *) - 4));
|
|
|
|
return (unsigned int)y;
|
|
|
|
}
|
|
|
|
#endif
|
2014-09-25 06:15:52 +10:00
|
|
|
bool BLI_ghashutil_ptrcmp(const void *a, const void *b)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2014-09-25 06:15:52 +10:00
|
|
|
return (a != b);
|
2002-10-12 11:37:38 +00:00
|
|
|
}
|
|
|
|
|
2014-04-15 14:39:23 +10:00
|
|
|
unsigned int BLI_ghashutil_uinthash_v4(const unsigned int key[4])
|
|
|
|
{
|
|
|
|
unsigned int hash;
|
|
|
|
hash = key[0];
|
|
|
|
hash *= 37;
|
|
|
|
hash += key[1];
|
|
|
|
hash *= 37;
|
|
|
|
hash += key[2];
|
|
|
|
hash *= 37;
|
|
|
|
hash += key[3];
|
|
|
|
return hash;
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
unsigned int BLI_ghashutil_uinthash_v4_murmur(const unsigned int key[4])
|
|
|
|
{
|
2015-03-22 21:13:43 +01:00
|
|
|
return BLI_hash_mm2((const unsigned char *)key, sizeof(int) * 4 /* sizeof(key) */, 0);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
}
|
2014-04-15 14:39:23 +10:00
|
|
|
|
2014-09-25 06:15:52 +10:00
|
|
|
bool BLI_ghashutil_uinthash_v4_cmp(const void *a, const void *b)
|
2014-07-30 07:30:18 +10:00
|
|
|
{
|
2014-09-25 06:15:52 +10:00
|
|
|
return (memcmp(a, b, sizeof(unsigned int[4])) != 0);
|
2014-07-30 07:30:18 +10:00
|
|
|
}
|
|
|
|
|
2014-04-15 14:17:54 +10:00
|
|
|
unsigned int BLI_ghashutil_uinthash(unsigned int key)
|
|
|
|
{
|
|
|
|
key += ~(key << 16);
|
|
|
|
key ^= (key >> 5);
|
|
|
|
key += (key << 3);
|
|
|
|
key ^= (key >> 13);
|
|
|
|
key += ~(key << 9);
|
|
|
|
key ^= (key >> 17);
|
|
|
|
|
|
|
|
return key;
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned int BLI_ghashutil_inthash_p(const void *ptr)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2008-08-17 17:08:00 +00:00
|
|
|
uintptr_t key = (uintptr_t)ptr;
|
2005-12-21 22:21:43 +00:00
|
|
|
|
|
|
|
key += ~(key << 16);
|
|
|
|
key ^= (key >> 5);
|
|
|
|
key += (key << 3);
|
|
|
|
key ^= (key >> 13);
|
|
|
|
key += ~(key << 9);
|
|
|
|
key ^= (key >> 17);
|
|
|
|
|
2011-04-21 15:53:30 +00:00
|
|
|
return (unsigned int)(key & 0xffffffff);
|
2005-12-21 22:21:43 +00:00
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
unsigned int BLI_ghashutil_inthash_p_murmur(const void *ptr)
|
|
|
|
{
|
|
|
|
uintptr_t key = (uintptr_t)ptr;
|
|
|
|
|
|
|
|
return BLI_hash_mm2((const unsigned char *)&key, sizeof(key), 0);
|
|
|
|
}
|
|
|
|
|
2015-04-11 23:36:37 +10:00
|
|
|
unsigned int BLI_ghashutil_inthash_p_simple(const void *ptr)
|
|
|
|
{
|
|
|
|
return GET_UINT_FROM_POINTER(ptr);
|
|
|
|
}
|
|
|
|
|
2014-09-25 06:15:52 +10:00
|
|
|
bool BLI_ghashutil_intcmp(const void *a, const void *b)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2014-09-25 06:15:52 +10:00
|
|
|
return (a != b);
|
2005-12-21 22:21:43 +00:00
|
|
|
}
|
|
|
|
|
2012-12-11 13:57:58 +00:00
|
|
|
/**
|
|
|
|
* This function implements the widely used "djb" hash apparently posted
|
|
|
|
* by Daniel Bernstein to comp.lang.c some time ago. The 32 bit
|
|
|
|
* unsigned hash value starts at 5381 and for each byte 'c' in the
|
2014-05-14 15:00:47 +10:00
|
|
|
* string, is updated: ``hash = hash * 33 + c``. This
|
2012-12-11 13:57:58 +00:00
|
|
|
* function uses the signed value of each byte.
|
|
|
|
*
|
|
|
|
* note: this is the same hash method that glib 2.34.0 uses.
|
|
|
|
*/
|
2014-04-15 14:17:54 +10:00
|
|
|
unsigned int BLI_ghashutil_strhash_n(const char *key, size_t n)
|
|
|
|
{
|
|
|
|
const signed char *p;
|
|
|
|
unsigned int h = 5381;
|
|
|
|
|
|
|
|
for (p = (const signed char *)key; n-- && *p != '\0'; p++) {
|
|
|
|
h = (h << 5) + h + (unsigned int)*p;
|
|
|
|
}
|
|
|
|
|
|
|
|
return h;
|
|
|
|
}
|
|
|
|
unsigned int BLI_ghashutil_strhash_p(const void *ptr)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
2012-12-11 13:57:58 +00:00
|
|
|
const signed char *p;
|
|
|
|
unsigned int h = 5381;
|
2012-03-11 23:47:41 +00:00
|
|
|
|
2012-12-11 13:57:58 +00:00
|
|
|
for (p = ptr; *p != '\0'; p++) {
|
2013-05-08 12:55:51 +00:00
|
|
|
h = (h << 5) + h + (unsigned int)*p;
|
2012-03-11 23:47:41 +00:00
|
|
|
}
|
|
|
|
|
2012-12-11 13:57:58 +00:00
|
|
|
return h;
|
2002-10-12 11:37:38 +00:00
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
unsigned int BLI_ghashutil_strhash_p_murmur(const void *ptr)
|
|
|
|
{
|
|
|
|
const unsigned char *key = ptr;
|
|
|
|
|
|
|
|
return BLI_hash_mm2(key, strlen((const char *)key) + 1, 0);
|
|
|
|
}
|
2014-09-25 06:15:52 +10:00
|
|
|
bool BLI_ghashutil_strcmp(const void *a, const void *b)
|
2011-12-17 00:52:36 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return (a == b) ? false : !STREQ(a, b);
|
2002-10-12 11:37:38 +00:00
|
|
|
}
|
2011-12-16 23:56:18 +00:00
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
GHashPair *BLI_ghashutil_pairalloc(const void *first, const void *second)
|
|
|
|
{
|
|
|
|
GHashPair *pair = MEM_mallocN(sizeof(GHashPair), "GHashPair");
|
|
|
|
pair->first = first;
|
|
|
|
pair->second = second;
|
|
|
|
return pair;
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned int BLI_ghashutil_pairhash(const void *ptr)
|
|
|
|
{
|
|
|
|
const GHashPair *pair = ptr;
|
|
|
|
unsigned int hash = BLI_ghashutil_ptrhash(pair->first);
|
|
|
|
return hash ^ BLI_ghashutil_ptrhash(pair->second);
|
|
|
|
}
|
|
|
|
|
2014-09-25 06:15:52 +10:00
|
|
|
bool BLI_ghashutil_paircmp(const void *a, const void *b)
|
2013-08-25 12:17:46 +00:00
|
|
|
{
|
|
|
|
const GHashPair *A = a;
|
|
|
|
const GHashPair *B = b;
|
|
|
|
|
2014-09-25 06:15:52 +10:00
|
|
|
return (BLI_ghashutil_ptrcmp(A->first, B->first) ||
|
|
|
|
BLI_ghashutil_ptrcmp(A->second, B->second));
|
2013-08-25 12:17:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void BLI_ghashutil_pairfree(void *ptr)
|
|
|
|
{
|
|
|
|
MEM_freeN(ptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
/** \} */
|
|
|
|
|
|
|
|
|
2013-08-25 20:00:19 +00:00
|
|
|
/** \name Convenience GHash Creation Functions
|
2013-08-25 12:17:46 +00:00
|
|
|
* \{ */
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHash *BLI_ghash_ptr_new_ex(const char *info, const unsigned int nentries_reserve)
|
2013-08-24 15:14:50 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_ghash_new_ex(BLI_ghashutil_ptrhash, BLI_ghashutil_ptrcmp, info, nentries_reserve);
|
2013-08-24 15:14:50 +00:00
|
|
|
}
|
2012-05-16 00:51:36 +00:00
|
|
|
GHash *BLI_ghash_ptr_new(const char *info)
|
|
|
|
{
|
2013-08-24 15:14:50 +00:00
|
|
|
return BLI_ghash_ptr_new_ex(info, 0);
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHash *BLI_ghash_str_new_ex(const char *info, const unsigned int nentries_reserve)
|
2013-08-24 15:14:50 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_ghash_new_ex(BLI_ghashutil_strhash_p, BLI_ghashutil_strcmp, info, nentries_reserve);
|
2012-05-16 00:51:36 +00:00
|
|
|
}
|
|
|
|
GHash *BLI_ghash_str_new(const char *info)
|
|
|
|
{
|
2013-08-24 15:14:50 +00:00
|
|
|
return BLI_ghash_str_new_ex(info, 0);
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHash *BLI_ghash_int_new_ex(const char *info, const unsigned int nentries_reserve)
|
2013-08-24 15:14:50 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_ghash_new_ex(BLI_ghashutil_inthash_p, BLI_ghashutil_intcmp, info, nentries_reserve);
|
2012-05-16 00:51:36 +00:00
|
|
|
}
|
|
|
|
GHash *BLI_ghash_int_new(const char *info)
|
|
|
|
{
|
2013-08-24 15:14:50 +00:00
|
|
|
return BLI_ghash_int_new_ex(info, 0);
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GHash *BLI_ghash_pair_new_ex(const char *info, const unsigned int nentries_reserve)
|
2013-08-24 15:14:50 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_ghash_new_ex(BLI_ghashutil_pairhash, BLI_ghashutil_paircmp, info, nentries_reserve);
|
2012-05-16 00:51:36 +00:00
|
|
|
}
|
|
|
|
GHash *BLI_ghash_pair_new(const char *info)
|
|
|
|
{
|
2013-08-24 15:14:50 +00:00
|
|
|
return BLI_ghash_pair_new_ex(info, 0);
|
2012-05-16 00:51:36 +00:00
|
|
|
}
|
|
|
|
|
2013-08-25 12:17:46 +00:00
|
|
|
/** \} */
|
2013-08-25 20:00:19 +00:00
|
|
|
|
|
|
|
|
|
|
|
/* -------------------------------------------------------------------- */
|
|
|
|
/* GSet API */
|
|
|
|
|
|
|
|
/* Use ghash API to give 'set' functionality */
|
|
|
|
|
|
|
|
/** \name GSet Functions
|
|
|
|
* \{ */
|
|
|
|
GSet *BLI_gset_new_ex(GSetHashFP hashfp, GSetCmpFP cmpfp, const char *info,
|
|
|
|
const unsigned int nentries_reserve)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return (GSet *)ghash_new(hashfp, cmpfp, info, nentries_reserve, GHASH_FLAG_IS_GSET);
|
2013-08-25 20:00:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
GSet *BLI_gset_new(GSetHashFP hashfp, GSetCmpFP cmpfp, const char *info)
|
|
|
|
{
|
|
|
|
return BLI_gset_new_ex(hashfp, cmpfp, info, 0);
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
/**
|
|
|
|
* Copy given GSet. Keys are also copied if callback is provided, else pointers remain the same.
|
|
|
|
*/
|
|
|
|
GSet *BLI_gset_copy(GSet *gs, GHashKeyCopyFP keycopyfp)
|
|
|
|
{
|
|
|
|
return (GSet *)ghash_copy((GHash *)gs, keycopyfp, NULL);
|
|
|
|
}
|
|
|
|
|
2015-04-13 13:45:14 +10:00
|
|
|
unsigned int BLI_gset_size(GSet *gs)
|
2013-08-25 20:00:19 +00:00
|
|
|
{
|
2015-04-13 13:45:14 +10:00
|
|
|
return ((GHash *)gs)->nentries;
|
2013-08-25 20:00:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Adds the key to the set (no checks for unique keys!).
|
|
|
|
* Matching #BLI_ghash_insert
|
|
|
|
*/
|
|
|
|
void BLI_gset_insert(GSet *gs, void *key)
|
|
|
|
{
|
|
|
|
const unsigned int hash = ghash_keyhash((GHash *)gs, key);
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
const unsigned int bucket_index = ghash_bucket_index((GHash *)gs, hash);
|
|
|
|
ghash_insert_ex_keyonly((GHash *)gs, key, bucket_index);
|
2013-08-25 20:00:19 +00:00
|
|
|
}
|
|
|
|
|
2014-05-29 14:00:11 +10:00
|
|
|
/**
|
|
|
|
* A version of BLI_gset_insert which checks first if the key is in the set.
|
|
|
|
* \returns true if a new key has been added.
|
|
|
|
*
|
|
|
|
* \note GHash has no equivalent to this because typically the value would be different.
|
|
|
|
*/
|
|
|
|
bool BLI_gset_add(GSet *gs, void *key)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return ghash_insert_safe_keyonly((GHash *)gs, key, false, NULL);
|
2014-05-29 14:00:11 +10:00
|
|
|
}
|
|
|
|
|
2013-08-25 20:00:19 +00:00
|
|
|
/**
|
|
|
|
* Adds the key to the set (duplicates are managed).
|
|
|
|
* Matching #BLI_ghash_reinsert
|
|
|
|
*
|
|
|
|
* \returns true if a new key has been added.
|
|
|
|
*/
|
|
|
|
bool BLI_gset_reinsert(GSet *gs, void *key, GSetKeyFreeFP keyfreefp)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return ghash_insert_safe_keyonly((GHash *)gs, key, true, keyfreefp);
|
2013-08-25 20:00:19 +00:00
|
|
|
}
|
|
|
|
|
2015-05-11 12:39:08 +10:00
|
|
|
bool BLI_gset_remove(GSet *gs, const void *key, GSetKeyFreeFP keyfreefp)
|
2013-08-25 20:00:19 +00:00
|
|
|
{
|
|
|
|
return BLI_ghash_remove((GHash *)gs, key, keyfreefp, NULL);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
bool BLI_gset_haskey(GSet *gs, const void *key)
|
|
|
|
{
|
|
|
|
return (ghash_lookup_entry((GHash *)gs, key) != NULL);
|
|
|
|
}
|
|
|
|
|
2016-02-20 15:17:40 +01:00
|
|
|
/**
|
|
|
|
* Remove a random entry from \a gsp, returning true if a key could be removed, false otherwise.
|
|
|
|
*
|
|
|
|
* \param r_key: The removed key.
|
|
|
|
* \param state: Used for efficient removal.
|
|
|
|
* \return true if there was somethjing to pop, false if gset was already empty.
|
|
|
|
*/
|
|
|
|
bool BLI_gset_pop(
|
|
|
|
GSet *gs, GSetIterState *state,
|
|
|
|
void **r_key)
|
|
|
|
{
|
|
|
|
GSetEntry *e = (GSetEntry *)ghash_pop((GHash *)gs, (GHashIterState *)state);
|
|
|
|
|
|
|
|
if (e) {
|
|
|
|
*r_key = e->key;
|
|
|
|
|
|
|
|
BLI_mempool_free(((GHash *)gs)->entrypool, e);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
*r_key = NULL;
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-25 20:00:19 +00:00
|
|
|
void BLI_gset_clear_ex(GSet *gs, GSetKeyFreeFP keyfreefp,
|
|
|
|
const unsigned int nentries_reserve)
|
|
|
|
{
|
|
|
|
BLI_ghash_clear_ex((GHash *)gs, keyfreefp, NULL,
|
|
|
|
nentries_reserve);
|
|
|
|
}
|
|
|
|
|
|
|
|
void BLI_gset_clear(GSet *gs, GSetKeyFreeFP keyfreefp)
|
|
|
|
{
|
|
|
|
BLI_ghash_clear((GHash *)gs, keyfreefp, NULL);
|
|
|
|
}
|
|
|
|
|
|
|
|
void BLI_gset_free(GSet *gs, GSetKeyFreeFP keyfreefp)
|
|
|
|
{
|
|
|
|
BLI_ghash_free((GHash *)gs, keyfreefp, NULL);
|
|
|
|
}
|
2014-08-12 13:26:29 +10:00
|
|
|
|
|
|
|
void BLI_gset_flag_set(GSet *gs, unsigned int flag)
|
|
|
|
{
|
|
|
|
((GHash *)gs)->flag |= flag;
|
|
|
|
}
|
|
|
|
|
|
|
|
void BLI_gset_flag_clear(GSet *gs, unsigned int flag)
|
|
|
|
{
|
|
|
|
((GHash *)gs)->flag &= ~flag;
|
|
|
|
}
|
|
|
|
|
2013-08-25 20:00:19 +00:00
|
|
|
/** \} */
|
|
|
|
|
|
|
|
|
|
|
|
/** \name Convenience GSet Creation Functions
|
|
|
|
* \{ */
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GSet *BLI_gset_ptr_new_ex(const char *info, const unsigned int nentries_reserve)
|
2013-08-25 20:00:19 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_gset_new_ex(BLI_ghashutil_ptrhash, BLI_ghashutil_ptrcmp, info, nentries_reserve);
|
2013-08-25 20:00:19 +00:00
|
|
|
}
|
|
|
|
GSet *BLI_gset_ptr_new(const char *info)
|
|
|
|
{
|
|
|
|
return BLI_gset_ptr_new_ex(info, 0);
|
|
|
|
}
|
|
|
|
|
2015-07-02 20:35:05 +02:00
|
|
|
GSet *BLI_gset_str_new_ex(const char *info, const unsigned int nentries_reserve)
|
|
|
|
{
|
|
|
|
return BLI_gset_new_ex(BLI_ghashutil_strhash_p, BLI_ghashutil_strcmp, info, nentries_reserve);
|
|
|
|
}
|
|
|
|
GSet *BLI_gset_str_new(const char *info)
|
|
|
|
{
|
|
|
|
return BLI_gset_str_new_ex(info, 0);
|
|
|
|
}
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
GSet *BLI_gset_pair_new_ex(const char *info, const unsigned int nentries_reserve)
|
2013-08-25 20:00:19 +00:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_gset_new_ex(BLI_ghashutil_pairhash, BLI_ghashutil_paircmp, info, nentries_reserve);
|
2013-08-25 20:00:19 +00:00
|
|
|
}
|
|
|
|
GSet *BLI_gset_pair_new(const char *info)
|
|
|
|
{
|
|
|
|
return BLI_gset_pair_new_ex(info, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/** \} */
|
2014-07-14 23:59:47 +10:00
|
|
|
|
|
|
|
|
|
|
|
/** \name Debugging & Introspection
|
|
|
|
* \{ */
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
|
|
|
#include "BLI_math.h"
|
2014-07-14 23:59:47 +10:00
|
|
|
|
|
|
|
/**
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
* \return number of buckets in the GHash.
|
|
|
|
*/
|
|
|
|
int BLI_ghash_buckets_size(GHash *gh)
|
|
|
|
{
|
|
|
|
return (int)gh->nbuckets;
|
|
|
|
}
|
|
|
|
int BLI_gset_buckets_size(GSet *gs)
|
|
|
|
{
|
|
|
|
return BLI_ghash_buckets_size((GHash *)gs);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Measure how well the hash function performs (1.0 is approx as good as random distribution),
|
|
|
|
* and return a few other stats like load, variance of the distribution of the entries in the buckets, etc.
|
2014-08-12 13:26:29 +10:00
|
|
|
*
|
|
|
|
* Smaller is better!
|
2014-07-14 23:59:47 +10:00
|
|
|
*/
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
double BLI_ghash_calc_quality_ex(
|
|
|
|
GHash *gh, double *r_load, double *r_variance,
|
|
|
|
double *r_prop_empty_buckets, double *r_prop_overloaded_buckets, int *r_biggest_bucket)
|
2014-07-14 23:59:47 +10:00
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
double mean;
|
2014-07-14 23:59:47 +10:00
|
|
|
unsigned int i;
|
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
if (gh->nentries == 0) {
|
|
|
|
if (r_load) {
|
|
|
|
*r_load = 0.0;
|
|
|
|
}
|
|
|
|
if (r_variance) {
|
|
|
|
*r_variance = 0.0;
|
|
|
|
}
|
|
|
|
if (r_prop_empty_buckets) {
|
|
|
|
*r_prop_empty_buckets = 1.0;
|
|
|
|
}
|
|
|
|
if (r_prop_overloaded_buckets) {
|
|
|
|
*r_prop_overloaded_buckets = 0.0;
|
|
|
|
}
|
|
|
|
if (r_biggest_bucket) {
|
|
|
|
*r_biggest_bucket = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0.0;
|
|
|
|
}
|
2014-07-14 23:59:47 +10:00
|
|
|
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
mean = (double)gh->nentries / (double)gh->nbuckets;
|
|
|
|
if (r_load) {
|
|
|
|
*r_load = mean;
|
|
|
|
}
|
|
|
|
if (r_biggest_bucket) {
|
|
|
|
*r_biggest_bucket = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (r_variance) {
|
|
|
|
/* We already know our mean (i.e. load factor), easy to compute variance.
|
|
|
|
* See http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Two-pass_algorithm
|
|
|
|
*/
|
|
|
|
double sum = 0.0;
|
|
|
|
for (i = 0; i < gh->nbuckets; i++) {
|
|
|
|
int count = 0;
|
|
|
|
Entry *e;
|
|
|
|
for (e = gh->buckets[i]; e; e = e->next) {
|
|
|
|
count++;
|
|
|
|
}
|
|
|
|
sum += ((double)count - mean) * ((double)count - mean);
|
2014-07-14 23:59:47 +10:00
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
*r_variance = sum / (double)(gh->nbuckets - 1);
|
2014-07-14 23:59:47 +10:00
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
|
|
|
|
{
|
2015-03-21 16:40:17 +11:00
|
|
|
uint64_t sum = 0;
|
|
|
|
uint64_t overloaded_buckets_threshold = (uint64_t)max_ii(GHASH_LIMIT_GROW(1), 1);
|
|
|
|
uint64_t sum_overloaded = 0;
|
|
|
|
uint64_t sum_empty = 0;
|
|
|
|
|
|
|
|
for (i = 0; i < gh->nbuckets; i++) {
|
|
|
|
uint64_t count = 0;
|
|
|
|
Entry *e;
|
|
|
|
for (e = gh->buckets[i]; e; e = e->next) {
|
|
|
|
count++;
|
|
|
|
}
|
|
|
|
if (r_biggest_bucket) {
|
|
|
|
*r_biggest_bucket = max_ii(*r_biggest_bucket, (int)count);
|
|
|
|
}
|
|
|
|
if (r_prop_overloaded_buckets && (count > overloaded_buckets_threshold)) {
|
|
|
|
sum_overloaded++;
|
|
|
|
}
|
|
|
|
if (r_prop_empty_buckets && !count) {
|
|
|
|
sum_empty++;
|
|
|
|
}
|
|
|
|
sum += count * (count + 1);
|
|
|
|
}
|
|
|
|
if (r_prop_overloaded_buckets) {
|
|
|
|
*r_prop_overloaded_buckets = (double)sum_overloaded / (double)gh->nbuckets;
|
|
|
|
}
|
|
|
|
if (r_prop_empty_buckets) {
|
|
|
|
*r_prop_empty_buckets = (double)sum_empty / (double)gh->nbuckets;
|
|
|
|
}
|
|
|
|
return ((double)sum * (double)gh->nbuckets /
|
|
|
|
((double)gh->nentries * (gh->nentries + 2 * gh->nbuckets - 1)));
|
|
|
|
}
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
}
|
|
|
|
double BLI_gset_calc_quality_ex(
|
|
|
|
GSet *gs, double *r_load, double *r_variance,
|
|
|
|
double *r_prop_empty_buckets, double *r_prop_overloaded_buckets, int *r_biggest_bucket)
|
|
|
|
{
|
|
|
|
return BLI_ghash_calc_quality_ex((GHash *)gs, r_load, r_variance,
|
|
|
|
r_prop_empty_buckets, r_prop_overloaded_buckets, r_biggest_bucket);
|
|
|
|
}
|
|
|
|
|
|
|
|
double BLI_ghash_calc_quality(GHash *gh)
|
|
|
|
{
|
|
|
|
return BLI_ghash_calc_quality_ex(gh, NULL, NULL, NULL, NULL, NULL);
|
2014-07-14 23:59:47 +10:00
|
|
|
}
|
|
|
|
double BLI_gset_calc_quality(GSet *gs)
|
|
|
|
{
|
GHash - code reorganization, performance enhancements, add a few missing utils to API.
This patch is the root of the GHash rework, all other diff will be based on it:
Reduce average load from 3.0 to 0.75
----------------------------------
This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.
Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).
Add optional shrinking
----------------------------------
I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.
Add code to use masking instead of modulo
----------------------------------
Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.
In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…
Add various missing API helpers
----------------------------------
I.e. a way to deep-copy a ghash/gset, and a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation).
Various code refactoring
----------------------------------
* Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
* Various re-shuffle and factorization in low-level internal code.
* Some work on hashing helpers, introducing some murmur2a-based hashing too.
Thanks a bunch to Campbell for the extensive review work. :)
Reviewers: sergey, campbellbarton
Subscribers: psy-fi, lukastoenne
Projects: #bf_blender
Maniphest Tasks: T43766
Differential Revision: https://developer.blender.org/D1178
2015-03-19 17:36:59 +01:00
|
|
|
return BLI_ghash_calc_quality_ex((GHash *)gs, NULL, NULL, NULL, NULL, NULL);
|
2014-07-14 23:59:47 +10:00
|
|
|
}
|
|
|
|
|
|
|
|
/** \} */
|