Date:   Sat, 4 Dec 2021 12:55:45 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Jiaxun Yang <jiaxun.yang@flygoat.com>
Cc:     x86@kernel.org, tglx@linutronix.de, mingo@redhat.com, bp@alien8.de,
        dave.hansen@linux.intel.com, hpa@zytor.com,
        chang.seok.bae@intel.com, linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH 00/10] x86: Allocate AVX512 xstate ondemand
Message-ID: <20211204115545.GW16608@worktop.programming.kicks-ass.net>
References: <20211203003636.11417-1-jiaxun.yang@flygoat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20211203003636.11417-1-jiaxun.yang@flygoat.com>
Precedence: bulk

On Fri, Dec 03, 2021 at 12:36:26AM +0000, Jiaxun Yang wrote:
> This series makes allocation of AVX512 xstate buffer ondemand.
> It can save some memory (~2k for a thread not using AVX512).
> 
> Also we are going to have heterogeneous processors that
> only some cores support AVX512, it can be helpful when
> dealing with such processors.
> 
> Patch 1~6 are some preparations.
> Patch 7 moves reallocation of xstate buffer to save_fpregs_to_fpstate.
> Patch 8~10 are some cleanup and enablement work.
> 
> Tested on a Skylake-X system with AVX512 intensive y-cruncher and numpy,
> the performance impact seems neglectable.
> 
> Any sugguestions are welcomed.

If we're going to do asymmetric avx512 it needs to come with a prctl()
interface like AMX.