llvm dev - May 2018 - Is there any relationship between IR instruction and execution time

Hi Yin,

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Yin Liu
via llvm-dev
> Hello,
> 
> As is known to all, there is a relationship between program's
instructions and its execution time. In other words, we can estimate the
execution time based on the number of program > instructions.
> 
> I'm curious about what the relationship between IR instruction and
execution time. I know the number of program instructions and the execution time
is highly related to the
> platform and architecture, while the IR instruction is independent and
intermediate. But, intuitively, there may be some relationship between IR
instruction and execution time.
> 
> Would it be possible to give me some advice about it?
What instructions finally get emitted by the compiler is highly dependent on the
specified target.  As you pointed out, IR is relatively abstract, and can at
best only generate a "rough" estimate to timing.  Maybe that loss of
fidelity is acceptable in your case.  Be aware that there are also target
specific optimizations that operate after the IR is lowered to a target friendly
representation.  Any early approximation of IR performance will be less accurate
after target specific optimization passes are ran.    For more accurate results,
you will need to wait until the IR is lowered to the target architecture and
emitted as assembly or object code.  But it seems that might be too late for
what you are looking for.  In any case, if you do want to analyze the assembly
code, then look no further than llvm's Machine Code Analyzer(MCA).  This
tool takes an assembly code as input and generates throughput and latency
information.  For more details see:
https://llvm.org/docs/CommandGuide/llvm-mca.html

-Matt

Hi Matt,

Thanks you so much for the reply!

I've tried the llvm-mca, it is helpful.
I was wondering whether the llvm-mca support the assembly code for the ARM?

I cross-compile the test file for ARM like that: clang test.c -O2 -target
arm-linux-gnueabihf -static -S -o test.s

If I want to check the performance using llvm-mca, is there any option of
"-mcpu" for ARM ?


Thanks,
Yin


2018-05-07 21:52 GMT-04:00 <Matthew.Davis at sony.com>:
> Hi Yin,
>
> From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Yin
Liu via
> llvm-dev
> > Hello,
> >
> > As is known to all, there is a relationship between program's
> instructions and its execution time. In other words, we can estimate the
> execution time based on the number of program > instructions.
> >
> > I'm curious about what the relationship between IR instruction and
> execution time. I know the number of program instructions and the execution
> time is highly related to the
> > platform and architecture, while the IR instruction is independent and
> intermediate. But, intuitively, there may be some relationship between IR
> instruction and execution time.
> >
> > Would it be possible to give me some advice about it?
>
> What instructions finally get emitted by the compiler is highly dependent
> on the specified target.  As you pointed out, IR is relatively abstract,
> and can at best only generate a "rough" estimate to timing. 
Maybe that
> loss of fidelity is acceptable in your case.  Be aware that there are also
> target specific optimizations that operate after the IR is lowered to a
> target friendly representation.  Any early approximation of IR performance
> will be less accurate after target specific optimization passes are ran.
> For more accurate results, you will need to wait until the IR is lowered to
> the target architecture and emitted as assembly or object code.  But it
> seems that might be too late for what you are looking for.  In any case, if
> you do want to analyze the assembly code, then look no further than
llvm's
> Machine Code Analyzer(MCA).  This tool takes an assembly code as input and
> generates throughput and latency information.  For more details see:
> https://llvm.org/docs/CommandGuide/llvm-mca.html
>
> -Matt
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180508/cbc57e23/attachment.html>

Hi Yin,

MCA does support the –mcpu and –mtriple options.  We have one arm test in
llvm/test/tools/llvm-mca/ARM for a cortex-9, which is an Out of Order chip.
Hope that helps!

-Matt

From: Yin Liu <yinliu.tiger at gmail.com>
Sent: Tuesday, May 8, 2018 2:49 PM
To: Davis, Matthew <Matthew.Davis at sony.com>
Cc: llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Is there any relationship between IR instruction and
execution time

Hi Matt,

Thanks you so much for the reply!

I've tried the llvm-mca, it is helpful.
I was wondering whether the llvm-mca support the assembly code for the ARM?

I cross-compile the test file for ARM like that: clang test.c -O2 -target
arm-linux-gnueabihf -static -S -o test.s

If I want to check the performance using llvm-mca, is there any option of
"-mcpu" for ARM ?


Thanks,
Yin


2018-05-07 21:52 GMT-04:00 <Matthew.Davis at sony.com<mailto:Matthew.Davis
at sony.com>>:
Hi Yin,

From: llvm-dev <llvm-dev-bounces at lists.llvm.org<mailto:llvm-dev-bounces
at lists.llvm.org>> On Behalf Of Yin Liu via
llvm-dev
> Hello,
>
> As is known to all, there is a relationship between program's
instructions and its execution time. In other words, we can estimate the
execution time based on the number of program > instructions.
>
> I'm curious about what the relationship between IR instruction and
execution time. I know the number of program instructions and the execution time
is highly related to the
> platform and architecture, while the IR instruction is independent and
intermediate. But, intuitively, there may be some relationship between IR
instruction and execution time.
>
> Would it be possible to give me some advice about it?
What instructions finally get emitted by the compiler is highly dependent on the
specified target.  As you pointed out, IR is relatively abstract, and can at
best only generate a "rough" estimate to timing.  Maybe that loss of
fidelity is acceptable in your case.  Be aware that there are also target
specific optimizations that operate after the IR is lowered to a target friendly
representation.  Any early approximation of IR performance will be less accurate
after target specific optimization passes are ran.    For more accurate results,
you will need to wait until the IR is lowered to the target architecture and
emitted as assembly or object code.  But it seems that might be too late for
what you are looking for.  In any case, if you do want to analyze the assembly
code, then look no further than llvm's Machine Code Analyzer(MCA).  This
tool takes an assembly code as input and generates throughput and latency
information.  For more details see:
https://llvm.org/docs/CommandGuide/llvm-mca.html

-Matt

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180509/149bf0a9/attachment.html>