[OpenBLAS_jll] Include patches to fix CASUM in 1.10

[imciner2]

This updates 1.10 to contain two new OpenBLAS patches that come from upstream (they are included in newer OpenBLAS versions already). Specifically the patches do:

• Use AVX512 kernels on modern processors
• Fix incorrect CASUM computation in fallback kernel

This fixes the reported linear algebra issue: JuliaLang/LinearAlgebra.jl#1406.

I believe I updated all the checksums/files to make this work, specifically I did

• Updated version number in stdlib/OpenBLAS_jll/Project.toml
• Updated version number and sha in deps/openblas.version
• Refresh checksums by running make -f contrib/refresh_checksums.mk -j openblas

cc @giordano, @ViralBShah

[ViralBShah]

The build is failing on linux.

[imciner2]

It is showing this error:

In file included from [buildroot]/src/processor.cpp:963:
[buildroot]/src/processor_x86.cpp: In function 'std::vector<jl_target_spec_t> jl_get_llvm_clone_targets()':
[buildroot]/src/processor_x86.cpp:1164:31: error: 'JL_TARGET_CLONE_BFLOAT16' was not declared in this scope; did you mean 'JL_TARGET_CLONE_FLOAT16'?
 1164 |                 t.en.flags |= JL_TARGET_CLONE_BFLOAT16;
      |                               ^~~~~~~~~~~~~~~~~~~~~~~~
      |                               JL_TARGET_CLONE_FLOAT16
make[1]: *** [Makefile:234: processor.o] Error 1

which shouldn't be a consequence of this PR. It looks like that is a consequence of 7a28ad5 being backported without the BFloat16 define - cc @gbaraldi.

[ViralBShah]

@imciner2 Is this issue only on 1.10? This is severe enough to require a new 1.10 release, and the original issue suggested that this did not happen on 1.11 (and hence unlikely on 1.12 also).

[imciner2]

@imciner2 Is this issue only on 1.10? This is severe enough to require a new 1.10 release, and the original issue suggested that this did not happen on 1.11 (and hence unlikely on 1.12 also).

You mean the linear algebra issue here? This is only in OpenBLAS pre-0.3.25(ish), and 1.11+ are all on newer versions than that, so they aren't affected. The original reporter even compared against 1.11 in the initial report to show it was broken and somehow fixed.

[imciner2]

The errors are now the same test failures as on the branch normally, so they aren't caused by this change. It should be good to merge.

[DilumAluthge]

The errors are now the same test failures as on the branch normally, so they aren't caused by this change.

I think that this statement is correct.

On backports-release-1.10 (#58889), I see the following failures:

1. Check ---> llvmpasses ---> Failed Tests (1): Julia :: remove-addrspaces.ll
2. Test ---> [seen on multiple platforms] ---> testset intrinsics ---> Got exception outside of a @test Failed to parse LLVM assembly: <string>:1:8: error: expected type define ptr addrspace(1) @entry(ptr addrspace(1) %0) #0 {
3. Test ---> [seen on multiple platforms] ---> testset Unicode ---> [...] Evaluated: 16 == 0

On this PR, I see the above three failures. But I also see the following on this PR only:

1. Check ---> asan ---> failing with timeout.

However, I suspect that the asan timeout is spurious. So I've retried the job. Let's see if it passes on retry.

[DilumAluthge]

After a retry, asan is green.

Per #59346 (comment), it seems that the failures on this PR are the same failures that already exist on backports-release-1.10.

@ViralBShah Is this good to merge?

[ViralBShah]

We should probably kick off a new release to include this.

[DilumAluthge]

We should probably kick off a new release to include this.

It's in progress. #58889

[DilumAluthge]

We should probably kick off a new release to include this.

It's in progress. #58889

I'll run the backports script again either today or this weekend.

[KristofferC]

@ViralBShah, you added the 1.12 backport label but 1.12 is on OpenBLAS 0.3.29 so this should not be backported to there, I guess?

Based on

This is only in OpenBLAS pre-0.3.25(ish), and 1.11+ are all on newer versions than that, so they aren't affected.

I'll remove the backport label.

[ViralBShah]

Yeah, I must have done that in error. I believe it was only needed on 1.10 and not later.