From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757402Ab0LBKNT (ORCPT ); Thu, 2 Dec 2010 05:13:19 -0500 Received: from mail30t.wh2.ocn.ne.jp ([125.206.180.136]:8191 "HELO mail30t.wh2.ocn.ne.jp" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with SMTP id S1751517Ab0LBKNR (ORCPT ); Thu, 2 Dec 2010 05:13:17 -0500 Subject: [PATCH 2/3] lib: Improve EWMA efficiency by using bitshifts To: linville@tuxdriver.com From: Bruno Randolf Cc: randy.dunlap@oracle.com, br1@thinktube.com, peterz@infradead.org, blp@cs.stanford.edu, linux-wireless@vger.kernel.org, linux-kernel@vger.kernel.org, Lars_Ericsson@telia.com, j@w1.fi, stefanr@s5r6.in-berlin.de, kosaki.motohiro@jp.fujitsu.com, akpm@linux-foundation.org, kevin.granade@gmail.com, johannes@sipsolutions.net Date: Thu, 02 Dec 2010 19:12:36 +0900 Message-ID: <20101202101236.22988.1691.stgit@localhost6.localdomain6> In-Reply-To: <20101202101231.22988.33396.stgit@localhost6.localdomain6> References: <20101202101231.22988.33396.stgit@localhost6.localdomain6> User-Agent: StGit/0.15 MIME-Version: 1.0 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 7bit X-SF-Loop: 1 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Using bitshifts instead of division and multiplication should improve performance. That requires weight and factor to be powers of two, but i think this is something we can live with. Thanks to Peter Zijlstra for the improved formula! Signed-off-by: Bruno Randolf --- include/linux/average.h | 4 +--- lib/average.c | 29 +++++++++++++++++++++-------- 2 files changed, 22 insertions(+), 11 deletions(-) diff --git a/include/linux/average.h b/include/linux/average.h index 7706e40..c6028fd 100644 --- a/include/linux/average.h +++ b/include/linux/average.h @@ -1,8 +1,6 @@ #ifndef _LINUX_AVERAGE_H #define _LINUX_AVERAGE_H -#include - /* Exponentially weighted moving average (EWMA) */ /* For more documentation see lib/average.c */ @@ -26,7 +24,7 @@ extern struct ewma *ewma_add(struct ewma *avg, unsigned long val); */ static inline unsigned long ewma_read(const struct ewma *avg) { - return DIV_ROUND_CLOSEST(avg->internal, avg->factor); + return avg->internal >> avg->factor; } #endif /* _LINUX_AVERAGE_H */ diff --git a/lib/average.c b/lib/average.c index f1d1b46..f438b40 100644 --- a/lib/average.c +++ b/lib/average.c @@ -24,18 +24,31 @@ * ewma_init() - Initialize EWMA parameters * @avg: Average structure * @factor: Factor to use for the scaled up internal value. The maximum value - * of averages can be ULONG_MAX/(factor*weight). + * of averages can be ULONG_MAX/(factor*weight). For performance reasons + * factor has to be a power of 2. * @weight: Exponential weight, or decay rate. This defines how fast the - * influence of older values decreases. Has to be bigger than 1. + * influence of older values decreases. For performance reasons weight has + * to be a power of 2. * * Initialize the EWMA parameters for a given struct ewma @avg. */ void ewma_init(struct ewma *avg, unsigned long factor, unsigned long weight) { - WARN_ON(weight <= 1 || factor == 0); + int n; + + /* get bitshift for weight */ + for (n = 0; !(weight & 1); n++) + weight = weight >> 1; + WARN_ON(weight > 1 || n < 1); + avg->weight = n; + + /* get bitshift for factor */ + for (n = 0; !(factor & 1); n++) + factor = factor >> 1; + WARN_ON(factor > 1); + avg->factor = n; + avg->internal = 0; - avg->weight = weight; - avg->factor = factor; } EXPORT_SYMBOL(ewma_init); @@ -49,9 +62,9 @@ EXPORT_SYMBOL(ewma_init); struct ewma *ewma_add(struct ewma *avg, unsigned long val) { avg->internal = avg->internal ? - (((avg->internal * (avg->weight - 1)) + - (val * avg->factor)) / avg->weight) : - (val * avg->factor); + (((avg->internal << avg->weight) - avg->internal) + + (val << avg->factor)) >> avg->weight : + (val << avg->factor); return avg; } EXPORT_SYMBOL(ewma_add);