The result of the weights calculation in the build definitions making use of Intrinsics are different from the other builds.

[sfilipi]

I am adding a test in this PR and it is failing in the following definitions:

• core30 Build_Debug_Intrinsics
• core30 Build_Release_Intrinsics
• CentOS Build_Debug_Intrinsics
• Ubuntu Build_Release_Intrinsics

The weight calculation for MulticlassLR in those build definitions are different from the ones in the other definitions.

weights in windows x64
[0, 0.5247429, -2.43042159, -0.490746975],
[0.210617781, 0, 0, -0.370147228],
[1.62008985E-06, -0.002228372, 2.49340868, 1.77681744]

weights in windows core30_debug_Intrinsics:
[0, 0.621862, -2.38398933, -0.427630424],
[0.108599916, 0, 0, -0.372884184],
[-0.106881194, -0.0351369865, 2.42675567, 1.80688608]

[sfilipi]

@Anipik how can i run those builds locally? is there a flag to run with .net core 3.0 and intrinsics?

[Anipik]

Yes, git clean the repo and build the repo with flag build -debug-Intrinsics.
If you want to run tests using test explorer, changed the build configuration in the VS. (Dont Forget to restart VS)

[sfilipi]

After looking more into this, the difference yielded between the two types of implementations, for this dataset and pipeline is comparable to the diverge we see in the weights calculated by varying the number of threads in the trainer.

Outcome of the measurements for the weights calculations:

Multithred Debug Configuration. Number of thread = 8 (This is the baseline)

weights	0	1	2	3
0	0	0.6630079	-2.386228	-0.468426347
1	0.217111193	0	0	-0.329844654
2	3.59202E-06	-0.084489964	2.41805434	1.79730487

Single thread Debug Configuration

weights	0	1	2	3
0	0	0.7623239	-2.34227	-0.51741457
1	0.07323902	0	0	-0.301424
2	-0.09961926	-0.143525153	2.37595	1.785325

Stats about the differences of weights with the Multithred Debug Configuration

Max	0.143872173
Average	0.054212062

Multithread Debug-Intrinsics Configuration.. Number of threads = 8.

weights	0	1	2	3
0	0	0.7623239	-2.34227	-0.51741457
1	0.07323902	0	0	-0.301424
2	-0.09961926	-0.143525153	2.37595	1.785325

Stats about the differences of weights with the Multithred Debug Configuration

Max	0.138265
Average	0.038507767

Single thread Debug-Intrinsics Configuration..

weights	0	1	2	3
0	0	0.472711831	-2.44474339	-0.461801231
1	0.238091767	0	0	-0.335685045
2	0	0	2.53461838	1.78326273

Stats about the differences of weights with the Multithred Debug Configuration

Max	0.190296069
Average	0.052316625

Stats about the differences of weights between Single thread Debug Configuration and Single thread Debug-Intrinsics Configuration..

Max	0.289612069
Average	0.106527889

[rosebyte]

Exact trained weights are not guaranteed across instruction sets, build configurations or thread counts because floating-point operations may be evaluated in a different order.