| Re: [eigen] Extracting the executed computations |
[ Thread Index |
Date Index
| More lists.tuxfamily.org/eigen Archives
]
- To: eigen <eigen@xxxxxxxxxxxxxxxxxxx>
- Subject: Re: [eigen] Extracting the executed computations
- From: Gael Guennebaud <gael.guennebaud@xxxxxxxxx>
- Date: Thu, 7 Feb 2013 09:34:25 +0100
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20120113; h=x-received:mime-version:in-reply-to:references:from:date:message-id :subject:to:content-type; bh=s+ZCuVK3bH87SZ9FbD/W8ElnLXfXTuqAOrBeiLpaIAI=; b=v0T3lJekcgFw5SxeB4G9tdoY8bEdt68RZBsNomEman5kP9h6TlUySzgsnLkbLW66SU 6ziumx+oObHEdP/FVUibSRe6yE6ySyDiLanHAZD9u+tVYeTs24YQSgh+6eNgXyuDJMjs g0Mg+Cz5fIhaFbEMMRM2HwoxV29pOyzzbbuF5Af4Oow46cgZ6Hg4hU7SXw5jl1cLLufL wnctrZOdN/+Af3NJdfYDDexP9wTzZtTkEFu2W+4SYkvb38qgrUWTpFWc35/Kwfii47Bs QnAVkl00XRVtTi1+vY1vWwDls904UaBros7+BslA1qyjVHaoq774rxhxSa5KhaxQD+yk skfg==
Hi,
what I usually is to generate the assembly (g++ -S -DNDEBUG -O2
file.cpp -o file.s) and examine it. To help finding the starting point
I add the following comment just before the expression I'm interested
in:
EIGEN_ASM_COMMENT("HERE_BEGINS_MY_EXPRESSION")
A = 2*A+B;
Then you can (quite) easily follow the generated SIMD code, the loops,
the function calls..
A more meaningful output could be obtained by putting such
EIGEN_ASM_COMMENT() in Eigen's code itself. For instance, we could put
ASM comments like:
EIGEN_ASM_COMMENT("EIGEN_BEGIN: vectorized inner unrolling")
EIGEN_ASM_COMMENT("EIGEN_BEGIN: scalar linear evaluation")
EIGEN_ASM_COMMENT("EIGEN: temporary creation")
EIGEN_ASM_COMMENT("EIGEN: call GEMM")
EIGEN_ASM_COMMENT("EIGEN_BEGIN: vectorized unrolled reduction")
....
cheers,
Gael
On Wed, Feb 6, 2013 at 9:20 PM, Cowing-Zitron, Christopher
<ccowingzitron@xxxxxxxx> wrote:
> Hello,
>
> I'm interested in the internals of Eigen's expression optimizations, in
> order to both optimize my usage of Eigen and to improve my overall
> programming techniques. Given some matrix expression such as D=C+(A*B)^-1, I
> know Eigen doesn't actually execute anything until the equals sign. Rather,
> Eigen's expression templates generate an optimized set of calculations,
> combining steps to avoid temporaries and aliasing, for instance. Eigen also
> rearranges the computations to make them more amenable to compiler
> optimizations like loop unrolling and autovectorization. But while I
> understand the general ideas, and the significance of each individual step,
> the complexity of both Eigen's and the compiler's optimizations are such
> that I have a hard time seeing the whole picture. Is it possible to have an
> expression somehow print out or save the final computation sequence chosen
> by the compiler? If that's not possible with Eigen on its own, is there some
> third-party tool, i.e. a debugger or disassembler, that I can use in
> conjunction with Eigen which will record in human-readable format the
> executed computations? Thanks for your help,
>
> -- Chris