Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   =?ISO-8859-1?Q?R=E9mi?= Denis-Courmont <remi@remlab.net>
To:     Atish Patra <atishp@atishpatra.org>
Cc:     linux-riscv@lists.infradead.org, linux-kernel@vger.kernel.org,
        Aurelien Jarno <aurelien@aurel32.net>,
        Mathieu Malaterre <malat@debian.org>,
        Jan Newger <jannewger@google.com>
Subject: Re: [PATCH v4 00/10] riscv: Allow userspace to directly access perf counters
Date:   Wed, 19 Jul 2023 17:45:57 +0300
Message-ID: <5951331.lOV4Wx5bFT@basile.remlab.net>
Organization: Remlab
In-Reply-To: <CAOnJCU+W9u+HnMfjU_y9W7331Zx07eQs9R1twV0=KsBAeSMSjw@mail.gmail.com>
References: <20230703124647.215952-1-alexghiti@rivosinc.com>
 <8756384.Zvl9fuB8X0@basile.remlab.net>
 <CAOnJCU+W9u+HnMfjU_y9W7331Zx07eQs9R1twV0=KsBAeSMSjw@mail.gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8BIT
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Le keskiviikkona 19. heinäkuuta 2023, 1.48.49 EEST Atish Patra a écrit :
> > Isn't RDTIM susceptible to interference from power management and CPU
> > frequency scaling? I suppose that RDCYCLE may behave differently depending
> > on PM in *some* designs, but that would still be way better than RDTIME
> > for the purpose.
> 
> Yes. But that's what it is probably using for other ISAs ?

At least on AArch64, it is using either Linux perf cycle counter, or if that 
is disabled at build time, the raw PMU cycle counter - which obviously leads 
to SIGILL on Linux, just like this MR would do with RDCYCLE.

Again, I do not *personally* have objections to disabling RDCYCLE for 
userspace (somebody else does, but that's neither my nor your problem). I do 
have objections to the wording of some of the commit messages though.

> My point was it should just do whatever it does for other ISA. RISC-V is no
> special in that regard.

Sure. My point is that RDTIME may be great for, so to say, system-level 
benchmarks. For FFmpeg that could something like how long it takes to 
transcode a video. But it doesn't seem to make much sense for 
microbenchmarking of single threaded tightly optimised loops, as opposed to 
RDCYCLE (or a wrapper for RDCYCLE).

-- 
Rémi Denis-Courmont
http://www.remlab.net/