Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20211025200852.3002369-1-kaleshsingh@google.com>
 <20211025200852.3002369-7-kaleshsingh@google.com> <20211026151451.7f3e09a4@gandalf.local.home>
 <CAC_TJveHgsPZw7p7BWOgQw6h8GNU_Pv_WUjNmw3AUq+wnSzk6Q@mail.gmail.com>
 <20211026201846.08990d1d@rorschach.local.home> <CAC_TJve-mKSojaXtukdFeQKvPz-8TQtS=pgGD0Z18Wt6yJi7dg@mail.gmail.com>
 <20211026211511.403d76ca@rorschach.local.home> <CAC_TJvdwqQAKrVs3w6NcQNBT+bAgdyqR+8Zt_An7R9AQSSthGA@mail.gmail.com>
 <20211026222123.5e206fcf@rorschach.local.home> <20211026231557.1eedad9b@rorschach.local.home>
In-Reply-To: <20211026231557.1eedad9b@rorschach.local.home>
From:   Kalesh Singh <kaleshsingh@google.com>
Date:   Tue, 26 Oct 2021 21:04:29 -0700
Message-ID: <CAC_TJveyKhZ3jyTZ33jbxSKTxxJ1WG+NUUUbf11pJ8Gqk7UeaQ@mail.gmail.com>
Subject: Re: [PATCH v4 6/8] tracing/histogram: Optimize division by a power of 2
To:     Steven Rostedt <rostedt@goodmis.org>
Cc:     surenb@google.com, hridya@google.com, namhyung@kernel.org,
        kernel-team@android.com, Jonathan Corbet <corbet@lwn.net>,
        Ingo Molnar <mingo@redhat.com>, Shuah Khan <shuah@kernel.org>,
        Masami Hiramatsu <mhiramat@kernel.org>,
        Tom Zanussi <zanussi@kernel.org>, linux-doc@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-kselftest@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Oct 26, 2021 at 8:16 PM Steven Rostedt <rostedt@goodmis.org> wrote:
>
> On Tue, 26 Oct 2021 22:21:23 -0400
> Steven Rostedt <rostedt@goodmis.org> wrote:
>
> > I'm sure there's an algorithm somewhere that can give as the real max.
>
> You got me playing with this more ;-)
>
> OK, I added the rounding in the wrong place. I found that we can make
> the max_div to be the same as the shift! The bigger the shift, the
> bigger the max!

Nice! :)
>
>         mult = (1 << shift) / div;
>         max_div = (1 << shift)
>
> But the rounding needs to be with the mult / shift:
>
>         return (val * mult + ((1 << shift) - 1)) >> shift;
>
>
> When val goes pass 1 << shift, then the error will be off by more than
> one.
Did you mean, val should be such that when we do the (val * mult) we
only get rounding errors less than (1 << shift)?

I think we also need to flip the delta now since we round down initially:

    delta =  (1 << shift) - (mult * div)

Thanks,
Kalesh
>
> -- Steve