NCSDK2 without AVX

[idata]

I've just downloaded the new NCSDK 2 with the instructions from https://software.intel.com/en-us/neural-compute-stick/get-started and the install failed because the CPU on which I need to run it doesn't have AVX support (Intel Atom). That being said, are there ways to still use the sdk without that flag and if so, how should I proceed ? Is it possible to recompile some stuff for my specific machine ?

[idata]

Hi @Oppiie

What are the specs of the system you're trying to install on?

Best Regards,

Sahira

[idata]

Hi @Sahira_at_Intel,

cat cpuinfo

processor : 0

vendor_id : GenuineIntelcpu family : 6model : 76model name : Intel(R) Atom(TM) x5-Z8350 CPU @ 1.44GHzstepping : 4microcode : 0x40ccpu MHz : 480.000cache size : 1024 KBphysical id : 0siblings : 4core id : 0cpu cores : 4apicid : 0initial apicid : 0fpu : yesfpu_exception : yescpuid level : 11wp : yesflags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology tsc_reliable nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm sse4_1 sse4_2 movbe popcnt tsc_deadline_timer aes rdrand lahf_lm 3dnowprefetch epb pti tpr_shadow vnmi flexpriority ept vpid tsc_adjust smep erms dtherm ida aratbugs : cpu_meltdown spectre_v1 spectre_v2bogomips : 2880.00clflush size : 64cache_alignment : 64address sizes : 36 bits physical, 48 bits virtualpower management:

If you need other specific info, please let me know which info and I'll provide it. Thanks

[idata]

Hi @Oppiie

Thanks for providing that info. What exactly is the error message you were getting?

Best Regards,

Sahira

[idata]

Hi,

The source is at https://github.com/movidius/ncsdk/blob/ncsdk2/install.sh

Lines 39 to 45.

Thanks

[idata]

Hi @Oppiie

Tensorflow 1.6 and later started using AVX instructions in their prebuild binaries. So CPUs with no AVX instructions like the Atom processor would have to use Tensorflow 1.5 or below or build from source. I hope this was helpful!

Best Regards,

Sahira

[idata]

Hi @Sahira_at_Intel

Ok so if I build TF 1.12 for my machine (without AVX flag), I can simply remove that check from Install.sh and it should all work as intended ?

This is a lot of work since I have to cross-compile TF from Windows for Ubuntu 16.04, so I'd like to have some sort of confirmation that it should work past that. ^^'

Thank you,

Oppiie

[idata]

Hi @Oppiie

I would recommend using the latest Tensorflow version that doesn't use AVX instructions. TF 1.12 is a little old and there have been some errors with running versions that old. Otherwise I think it should work.

Sincerely,

Sahira

[idata]

Hi @Sahira_at_Intel,

I meant the r1.12 branch of Tensorflow, which is much more recent than r1.5 (the latest being r1.13). Is that ok ?

Thanks

[idata]

Hello @Sahira_at_Intel ,

I managed to compile Tensorflow r1.12 from source and removed the AVX check from the Install.sh for the Movidius stick. It seems to be working fine and I ran inference using the NCS.

Thanks,

Mat

[idata]

Hi @Oppiie

Glad to see it's working!

Best Regards,

Sahira