webinar presentation by Kevin Oleary

[TimP]

https://software.intel.com/en-us/video/vectorize-or-performance-dies-tune-for-the-latest-avx-simd?utm_source=TEC+Spring+Webinars+-Thank+you+Emails%232&utm_medium=Email&utm_campaign=cmd_11761-2&utm_content=Link3

[TimP]

I couldn't find in the presentation a reference to the apparent requirement for setting -debug:inline-debug-info to enable vectorization advisor (in the usual case where there is in-lining).  That option isn't available in VS project properties.  There appear to be different consequences of in-lining among compiler versions, and between VS GUI and command line, not under control of the documented or VS properties options.

The presentation alluded to a provision in the beta version to predict AVX512 performance on an AVX2 platform, but it didn't appear to be implemented on Windows.

The presentation solicits email comments, but the email address rejects.

I haven't seen adequate explanation of the threading advisor for current releases. 

[TimP]

When running Advisor on SSE4 double data, it reports Float32 due to (correct) use of SSE 128-bit moves.  This is somewhat shocking to me.

I see no rhyme or reason to how Advisor chooses whether to report VL of 4 or 8 for AVX (and to quote "efficiency" relative to that choice).

[Kirill_R_Intel]

Hi Tim,

The "-debug:inline-debug-info" is Intel Compiler option and can be passed in "C++/command line" section in project properties directly.

"Provisioning AVX512" works on Windows. For that you should build with "-ax" compiler option and enable analyzing non-executed code paths in Advisor project properties. See for details: https://software.intel.com/en-us/blogs/2016/02/02/explore-intel-avx-512-code-paths-while-not-having-compatible-hardware .

What email address does reject? Maybe there is a typo.

Threading advisor doesn't have major features in 2017 release, functionality mostly the same as in previous version.

What is wrong with VL reporting on AVX? Can you attach Advisor project for our reference? Efficiency is relative to VL, if speedup estimates to 2x and VL is 4, efficiency is 50%. If speedup is 2x and VL=8, efficiency is 25%.

Regards,

Kirill