[20.10 FEAT] Algebra lib optimizations - OpenBlas (0.3.10)

[bugproxy]

The OpenBlas library is the backend for many machine learning algorithms using the Numpy Python framework.

The library has been enhanced to support z13 and partially z14.

Upgrade to newest version 0.3.10

[Frank Heimes (fheimes)]

v0.3.10 is released: https://www.openblas.net/

[bugproxy (bugproxy)]

------- Comment From <email address hidden> 2020-07-27 05:16 EDT-------
OpenBLAS release 0.3.10 has two bugs in the build system and the test suite that may cause build failures on s390x and other architectures. For both issues we have fixes merged upstream. These patches apply cleanly on top of 0.3.10 and we recommend including them (they can be dropped when adopting releases >0.3.10, since they will be included in those upcoming releases).

Bugs:
- A subset of the test suite fails on systems with 4 or fewer CPUs because of an incorrect calculation of the number of threads to use for testing (common code, affects all architecture)
- Builds with gcc10 and dynamic architecture support fail with link failures because of incorrect comparison of the compiler version (only affects s390x, no issue for gcc 9 and earlier)

Issue tracked on github:
https://github.com/xianyi/OpenBLAS/issues/2668

Fixes:
Pull request https://github.com/xianyi/OpenBLAS/pull/2669
with commit a2d13ea61183 ("Fix gcc version detection for zarch")

and pull request https://github.com/xianyi/OpenBLAS/pull/2672
with commits 478898b37a91 ("cpp_thread_test/dgemv: cap concurrency to number of hw threads on small systems")
and de838c38ef9d ("cpp_thread_test/dgemv: fail early if concurrency is zero")

[Steve Langasek (vorlon)]

openblas 0.3.10 is now in groovy.

[bugproxy (bugproxy)]

------- Comment From <email address hidden> 2020-08-18 04:06 EDT-------
IBM Bugzilla status-> closed, Fix Released with groovy

[bugproxy (bugproxy)]

------- Comment From <email address hidden> 2020-08-31 08:23 EDT-------
Please enable "DYNAMIC_ARCH=1" for s390x to enable choosing the best performing implementation at run-time (i.e., s390x should be added to ENABLE_DYNAMIC_ARCHS). Current builds of openblas version 0.3.10+ds-3 for groovy instead rely on generic floating-point instructions (TARGET=ZARCH_GENERIC) and thus do not use the significant performance potential of the SIMD instructions that were introduced with z13 and extended with z14 (cf. https://launchpadlibrarian.net/491360113/buildlog_ubuntu-groovy-s390x.openblas_0.3.10+ds-3_BUILDING.txt.gz)

While the mechanism for selecting the best implementation at run-time had some deficits on s390x in the past, we have revised it in OpenBLAS 0.3.10 and consider it the best choice for portable binaries.

Please note that there is a small bug in the s390-specific part of the Makefile triggered with gcc-10 and DYNAMIC_ARCH=1, which has been fixed upstream in June. The fix cleanly applies on top of 0.3.10 -- https://github.com/xianyi/OpenBLAS/pull/2669

[bugproxy (bugproxy)]

------- Comment From <email address hidden> 2020-08-31 09:41 EDT-------
IBM Bugzilla status-> closed, Fix Released with groovy