topic Comparing compilers in Software Tuning, Performance Optimization & Platform Monitoring

Comparing compilers

magicfoot — Wed, 18 Jan 2012 13:05:31 GMT

In terms of the "not so obvious" performance benefit gained when using an Intel compiler on an Intel platform, are there any suitable numerical benchmarks that can be used for a comparison of different compilers? My comparison(using numerical algorithms)of Intel's Composer against other popular compilers shows that Intel Composer out performs the others by nearly 50% when using nothing more than /O2. No SSE, etc used.

I have not yet started profiling the code on both platforms as to why this should be, but am suspicious that the improvement when using Intel Composer is so large as this is too good to be true.I have some detail off the web that shows similar results.

http://www.open-mag.com/features/Vol_15/IntelC/intelc.htm

How does the Intel compiler build code that is so much quicker than others ?

Comparing compilers

A_T_Intel — Wed, 18 Jan 2012 18:59:24 GMT

The Intel Compiler team optimizes well known numerical algorithms such as LINPACK and FFTs. It may be that the algorithm you are testing has been highly optimized for performance on Intel Architecture.

At optimization level 2 (O2)you can encounter "dead-code elimiation".I would verify that the numerical algorithm of interest in your test was not elimiated by the compiler.

Comparing compilers

TimP — Wed, 18 Jan 2012 19:06:09 GMT

The most frequent gain with Intel compilers at default options is due to auto-vectorization, but that is associated with default SSE2 code generation. Most other compilers which support auto-vectorization require an option set to turn it on. You quote a web site from the days before gnu C supported auto-vectorization.

Comparing compilers

rom828b — Wed, 18 Jan 2012 19:21:49 GMT

The newer compilers also know about AVX if you are on a SNB system.

Comparing compilers

SergeyKostrov — Thu, 19 Jan 2012 15:14:30 GMT

Quoting magicfoot

Did you tryLINPACK?

Comparing compilers

TimP — Thu, 19 Jan 2012 17:22:05 GMT

Intel relies on MKL to optimize linpack, so the compilers may be no better than others. Do you compile your BLAS libraries yourself? Or, do you mean the quality of the compiler's interface to MKL?

Comparing compilers

SergeyKostrov — Fri, 20 Jan 2012 04:34:04 GMT

Quoting magicfoot

...I have some detail off the web that shows similar results.

http://www.open-mag.com/features/Vol_15/IntelC/intelc.htm

How does the Intel compiler build code that is so much quicker than others?

http://www.open-mag.com/features/Vol_15/IntelC/intelc.htm

I read the article and, unfortunately, itis full of inconsistencies. It is not clear whenthe articlewas
written.Taking into account that testers have usedPentium III computers, Visual C++ v6.0 and
Intel C++ compiler v5.0 the articleis describing10 years old events!

>>...
>>That becomes even more interesting when Microsoft Visual Studio.NET is brought into the equation.
>>While still a beta product, tests of the new complier has so far proven it to be about 25% faster than the
>>current MS Visual C++.
>>...

I'm not surprised becauseVisual C++ v6.0 was introduced in 1998and it couldn'tcompile a declaration
like:

typedef union _declspec( align( 16 ) )__m128
{
...
} __m128;

Asupport for SSE instructions was introduced in a first version ofVisual Studio .NET in 2001.

Comparing compilers

SergeyKostrov — Sat, 21 Jan 2012 02:34:30 GMT

Quoting TimP (Intel)

Intel relies on MKL to optimize linpack, so the compilers may be no better than others. Do you compile your BLAS libraries yourself?

[SergeyK] No.

Or, do you mean the quality of the compiler's interface to MKL?

[SergeyK] No.

Best regards,
Sergey

Comparing compilers

TimP — Mon, 23 Jan 2012 11:28:16 GMT

Quoting Sergey Kostrov

typedef union _declspec( align( 16 ) )__m128
{
...
} __m128;

Asupport for SSE instructions was introduced in a first version ofVisual Studio .NET in 2001.

As you seem to emphasize use of ancient versions of Visual Studio (requiring obsolete versions of Windows), I'll mention that there was an SSE header service pack for VS6 prior to introduction of VS .net. Even now, a decade later, MSVC "support" for parallel SSE instructions doesn't include auto-vectorization, as all other currently maintained compilers do.
The old article you quote didn't likely include optimization by intrinsics among the demonstrated MSVC performance improvements subsequent to VS6. As I recall, VS2005 was the first MSVC to offer scalar SSE2 code generation. Without support for vectorization, there wasn't sufficient incentive to support mixed SSE and x87 code for P-III and Athlon32. VS2003 didn't support X64, although it was still on support 5 years ago. Performance improvements included splitting all 64-bit data moves into 32-bit pairs to compensate for lack of alignment support, along with several other scalar code optimizations.
MSVC with /fp:fast continues reasonably competitive with respect to scalar code optimization. Note that /fp:fast in MSVC (more aggressive than default) is roughly equivalent to /fp:source in ICL (less aggressive than default).

Comparing compilers

SergeyKostrov — Mon, 23 Jan 2012 13:39:06 GMT

Quoting TimP (Intel)

Quoting Sergey Kostrov

typedef union _declspec( align( 16 ) )__m128
{
...
} __m128;

Asupport for SSE instructions was introduced in a first version ofVisual Studio .NET in 2001.

As you seem to emphasize use of ancient versions of Visual Studio (requiring obsolete versions of Windows),

[SergeyK] VS6could be installed and it workson ALL 32-bit versions of Windows, that is,
95, 98, Millennium, NT3.5\4.0, 2000, XP,etc, and I wouldn't use VS6for the
development unless a customer needs a support for it.I told in another thread that some
companies are still using VS6.

I'll mention that there was an SSE header service pack for VS6 prior to introduction of VS .net.

[SergeyK]That's good to know and I'll check it.
...

Comparing compilers

SergeyKostrov — Tue, 24 Jan 2012 00:25:05 GMT

Quoting TimP (Intel)

...
I'll mention that there was an SSE header service pack for VS6 prior to introduction of VS .net.
...

Tim,

Thank you formentioning it! As far as I remember it was called a Visual C++ 6.0Processor Pack.

Best regards,
Sergey

Comparing compilers

SergeyKostrov — Wed, 25 Jan 2012 02:42:23 GMT

Quoting Sergey Kostrov

Quoting TimP (Intel)
...
I'll mention that there was an SSE header service pack for VS6 prior to introduction of VS .net.
...

Tim,

Thank you formentioning it! As far as I remember it was called a Visual C++ 6.0Processor Pack.

Best regards,
Sergey

Here is a link to that "magic" update:

http://msdn.microsoft.com/en-us/vstudio/aa718349.aspx

Comparing compilers

gladiolus — Fri, 03 Feb 2012 10:34:36 GMT

The Intel Fortran compiler generates machine code to support these requirements in accordance with your selected /fpe option setting.. In other words, the generated code must support the trapping mode..

Comparing compilers

SergeyKostrov — Fri, 03 Feb 2012 14:49:54 GMT

Quoting Sergey Kostrov

...
I'm not surprised becauseVisual C++ v6.0 was introduced in 1998and it couldn'tcompile a declaration
like:

typedef union _declspec( align( 16 ) )__m128
{
...
} __m128;

Asupport for SSE instructions was introduced in a first version ofVisual Studio .NET in 2001.

This is a follow up on the topic.

As soon as a Visual C++ 6.0Processor Pack is installed the compiler could compile a declaration
mentioned above and there isa support for some subset of SSE intrinsic functions.

Comparing compilers

TimP — Fri, 03 Feb 2012 15:35:28 GMT

Quoting Sergey Kostrov

Quoting Sergey Kostrov
...
I'm not surprised becauseVisual C++ v6.0 was introduced in 1998and it couldn'tcompile a declaration
like:

typedef union _declspec( align( 16 ) )__m128
{
...
} __m128;

Asupport for SSE instructions was introduced in a first version ofVisual Studio .NET in 2001.

This is a follow up on the topic.

As soon as a Visual C++ 6.0Processor Pack is installed the compiler could compile a declaration
mentioned above and there isa support for some subset of SSE intrinsic functions.

It supports the SSE instructions which were implemented on P-III and Athlon-32. It would not support SSE2 instructions, so you would not run into the lack of full SSE2 support on original Turion.
Intel compilers no longer support SSE for those early SSE CPUs.