public inbox for [email protected]  

help / color / mirror / Atom feed

From: Joel Jacobson <[email protected]>
To: Dean Rasheed <[email protected]>
Cc: Dagfinn Ilmari Mannsåker <[email protected]>
Cc: pgsql-hackers <[email protected]>
Subject: Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
Date: Tue, 09 Jul 2024 22:28:36 +0200
Message-ID: <[email protected]> (raw)
In-Reply-To: <CAEZATCXoemvhECHiL8Ug1MQxxtU0WqZfYqE853fDr_PvUpFerA@mail.gmail.com>
References: <[email protected]>
	<[email protected]>
	<[email protected]>
	<[email protected]>
	<[email protected]>
	<CAEZATCV2qPTGo2Fd8xDs06Q7iU5aorgSa9+Fw9zkuQv1y15rcw@mail.gmail.com>
	<[email protected]>
	<CAEZATCWEou=NjpvS-r1WxymbAYihsdKQoQBj1HfBeocoNcv5Ww@mail.gmail.com>
	<[email protected]>
	<[email protected]>
	<CAEZATCW+N9qaE=5_1xcB0isMrirojN=Bd0yY4Pw67iji4swYiQ@mail.gmail.com>
	<[email protected]>
	<[email protected]>
	<[email protected]>
	<[email protected]>
	<CAEZATCUA+GBYvzbqSr7LPDDfmr24jQE1R0iyYfjFk82xeuBCRw@mail.gmail.com>
	<[email protected]>
	<CAEZATCUnn2JK4gSrymvmkhy_qsGybKv0-n0CTnLcfwfVNfTE7g@mail.gmail.com>
	<[email protected]>
	<[email protected]>
	<CAEZATCXVB15_RfzROFiO_7Fh3voOuOe7FotbvON-6a+ManuGJg@mail.gmail.com>
	<[email protected]>
	<CAEZATCWY_h7jTzsQnZY3bChNF85W4KLYV-rRgy=cc4QAUXEUdg@mail.gmail.com>
	<[email protected]>
	<[email protected]>
	<CAEZATCUmrq4pgJdQNQendiPkmyM=qdnZv1g-vtoMyrOyCi6xdA@mail.gmail.com>
	<[email protected]>
	<CAEZATCW3CDcVAy1EoYzBGyVPfdpryEbOTgfK2L9A5oEUhJuJPQ@mail.gmail.com>
	<CAEZATCXoemvhECHiL8Ug1MQxxtU0WqZfYqE853fDr_PvUpFerA@mail.gmail.com>

On Tue, Jul 9, 2024, at 14:01, Dean Rasheed wrote:
> One thing I noticed while testing the earlier patches on this thread
> was that they were significantly faster if they used unsigned integers
> rather than signed integers. I think the reason is that operations
> like "x / 10000" and "x % 10000" use fewer CPU instructions (on every
> platform, according to godbolt.org) if x is unsigned.
>
> In addition, this reduces the number of times the digit array needs to
> be renormalised, which seems to be the biggest factor.
>
> Another small optimisation that seems to be just about worthwhile is
> to pull the first digit of var1 out of the main loop, so that its
> contributions can be set directly in dig[], rather than being added to
> it. This allows palloc() to be used to allocate dig[], rather than
> palloc0(), and only requires the upper part of dig[] to be initialised
> to zeros, rather than all of it.
>
> Together, these seem to give a decent speed-up:
..
> Attachments:
> * optimise-mul_var.patch

I've reviewed the patch now.
Code is straightforward, and comments easy to understand.
LGTM.

Regards,
Joel

view thread (23+ messages)

reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients using the --to and --cc options:
  reply via email

  To: [email protected]
  Cc: [email protected], [email protected], [email protected]
  Subject: Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
  In-Reply-To: <[email protected]>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox