[SPARK-10353][MLLIB] BLAS gemm not scaling when beta = 0.0 for some subset of matrix multiplications

[brkyvz]

@mengxr @jkbradley @rxin

It would be great if this fix made it into RC3!

[mengxr]

@brkyvz Could you whether whether those branches are already implemented in f2jBLAS and reference BLAS? If they are already there, I think we should let them handle those special cases.

[brkyvz]

dgemm does handle the easy cases (alpha == 0.0 && beta == 1.0). dscal doesn't handle beta == 1.0. I still think we should keep them, because not all paths use BLAS

[brkyvz]

The most important change here is https://github.com/apache/spark/pull/8525/files#diff-6d42e8ee906809226b36f06c9d2a8e8fR414

[mengxr]

LGTM pending Jenkins.

[srowen]

+1 I thought we fixed this but it was something similar but not identical. I did not see similar occurrences.

[SparkQA]

Test build #41796 has finished for PR 8525 at commit 4378ce8.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #41802 has finished for PR 8525 at commit b535b53.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[mengxr]

Merged into master, branch-1.5, and branch-1.4. @brkyvz Could you send another PR to fix branch-1.3? Thanks! It seems that on 1.3, we also have issues with if alpha == 0 return ....