build not portable despite using SAGE_FAT_BINARY=yes, NTL/openblas/numpy-related

[mkoeppe]

We fix several issues involving SAGE_FAT_BINARY:

• numpy intrinsics need a special build flag in this mode

• ntl needs to disable use of a problematic instruction set

• After ci-cygwin*.yml: delegate to tox, add more stages, use more specific SAGE_LOCAL #31064, the GH Actions tests for cygwin did not pass SAGE_FAT_BINARY to the actual build, randomly causing SIGILLs when a partial build was passed to the next stage through the sage-local artifact

Depends on #31493

CC: @dimpase @orlitzky @kiwifb @embray @darijgr @tscrim @kliem @seblabbe

Component: porting

Keywords: sd111

Author: Jonathan Kliem, Matthias Koeppe

Branch/Commit: d6dced2

Reviewer: Matthias Koeppe, Jonathan Kliem

Issue created by migration from https://trac.sagemath.org/ticket/29537

[dimpase]

comment:3

could it be some kind of hardware mismatch?

[mkoeppe]

comment:4

Possibly, I'll investigate

[mkoeppe]

comment:5

I will try with the configuration that Erik's https://github.com/sagemath/sage-windows/blob/master/Makefile
uses

[embray]

comment:7

I looked at one of the build logs on GitHub and it does look like SAGE_FAT_BINARY=yes is being passed down, so other than that I'm not aware of any issue that would cause this unless there's a new issue with NTL I haven't seen before.

[embray]

comment:8

Possibly related: #29109

[mkoeppe]

Changed keywords from none to sd111

[mkoeppe]

comment:11

It's quite possible that this also affects our Docker builds.

[mkoeppe]

comment:12

A report regarding the Docker builds: https://groups.google.com/g/sage-devel/c/Dbd4nFi8R7I (but no NTL involvement visible)

[kliem]

comment:13

From NTL/doc/tour-win.html:

152 Configuration flags.
153 </b>
154 <p>
155 
156 Also in directory "<tt>include/NTL</tt>" is a file called "<tt>config.h</tt>".
157 You can edit this file to override some of NTL's default options
158 for basic configuration and performance.
159 Again, the defaults should be good for
160 Windows with MSVC++.

In config.h there is a couple of flags that needs to be set by hand apparently.

Those look like they might make problems:

195 #if 0
196 #define NTL_CLEAN_INT
197 
198 /*
199  *   This will disallow the use of some non-standard integer arithmetic
200  *   that may improve performance somewhat.
201  *
202  */
203 
204 #endif
205 
206 #if 1
207 #define NTL_CLEAN_PTR
208 
209 /*
210  *   This will disallow the use of some non-standard pointer arithmetic
211  *   that may improve performance somewhat.
212  *
213  */
214 
215 #endif
216 
217 #if 1
218 #define NTL_SAFE_VECTORS
219 
220 /*
221  * This will compile NTL in "safe vector" mode, only assuming
222  * the relocatability property for trivial types and types
223  * explicitly declared relocatable.  See vector.txt for more details.
224  */
225 
226 #endif
227 
228 #if 0
229 #define NTL_ENABLE_AVX_FFT
230 
231 /*
232  * This will compile NTL in a way that enables an AVX implemention
233  * of the small-prime FFT.
234  */
235 
236 #endif
237 
238 
239 #if 0
240 #define NTL_AVOID_AVX512
241  
242 /*
243  * This will compile NTL in a way that avoids 512-bit operations,
244  * even if AVX512 is available.
245  */
246 
247 #endif
248  
249 #if 0
250 #define NTL_RANGE_CHECK
251 
252 /*
253  *   This will generate vector subscript range-check code.
254  *   Useful for debugging, but it slows things down of course.
255  *
256  */
257 
258 #endif

[kliem]

comment:14

So I guess we have to patch NTL as to avoid those flags by configuration.

[kliem]

comment:15

The relevant configuration is done in src/DoConfig, which is unfortunately perl (yet another language whichs syntax I do not understand).

But one might even define those things as environment variables and this might be passed through??

In src/DoConfig there is the following list:

 55 %ConfigFlag = (
 56 
 57 'NTL_LEGACY_NO_NAMESPACE' => 'off',
 58 'NTL_LEGACY_INPUT_ERROR'  => 'off',
 59 'NTL_DISABLE_LONGDOUBLE'  => 'off',
 60 'NTL_DISABLE_LONGLONG'    => 'off',
 61 'NTL_DISABLE_LL_ASM'      => 'off',
 62 'NTL_MAXIMIZE_SP_NBITS'   => 'off',
 63 'NTL_LEGACY_SP_MULMOD'    => 'off',
 64 'NTL_THREADS'             => 'on',
 65 'NTL_TLS_HACK'            => 'on',
 66 'NTL_EXCEPTIONS'          => 'off',
 67 'NTL_STD_CXX11'           => 'on',
 68 'NTL_STD_CXX14'           => 'off',
 69 'NTL_DISABLE_MOVE_ASSIGN' => 'on',
 70 'NTL_DISABLE_MOVE'        => 'off',
 71 'NTL_THREAD_BOOST'        => 'on',
 72 'NTL_GMP_LIP'             => 'on',
 73 'NTL_GF2X_LIB'            => 'off',
 74 'NTL_X86_FIX'             => 'off',
 75 'NTL_NO_X86_FIX'          => 'off',
 76 'NTL_NO_INIT_TRANS'       => 'on',
 77 'NTL_CLEAN_INT'           => 'off',
 78 'NTL_CLEAN_PTR'           => 'on',
 79 'NTL_SAFE_VECTORS'        => 'on',
 80 'NTL_RANGE_CHECK'         => 'off',
 81 'NTL_ENABLE_AVX_FFT'      => 'off',
 82 'NTL_AVOID_AVX512'        => 'off',
 83 
 84 
 85 'NTL_SPMM_ULL'            => 'off',
 86 'NTL_AVOID_BRANCHING'     => 'off',
 87 'NTL_FFT_BIGTAB'          => 'off',
 88 'NTL_FFT_LAZYMUL'         => 'off',
 89 'NTL_TBL_REM'             => 'off',
 90 'NTL_CRT_ALTCODE'         => 'off',
 91 'NTL_CRT_ALTCODE_SMALL'   => 'off',
 92 'NTL_GF2X_NOINLINE'       => 'off',
 93 'NTL_GF2X_ALTCODE'        => 'off',
 94 'NTL_GF2X_ALTCODE1'       => 'off',
 95 
 96 
 97 );

What makes me thing that it might be AVX512 related, is that we haven't encountered this before. Github workflows is the only place where I encountered processors capable of AVX512.

[mkoeppe]

comment:16

I agree that NTL_AVOID_AVX512 is a top suspect here.
Perhaps it can be passed to configure, like we already pass NATIVE=off when SAGE_FAT_BINARY=yes?

[mkoeppe]

comment:71

OK, testing with dash at https://github.com/mkoeppe/sage/actions/runs/662065807

[mkoeppe]

comment:72

This worked! https://github.com/mkoeppe/sage/runs/2134463553

I'll push the fixes to this ticket.

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

d0df00c	.github/workflows/extract-sage-local.sh: Use sage-rebase.sh --all
da978f8	.github/workflows/{extract-sage-local.sh, ci-cygwin-*.yml): Use /bin/dash for scripts invoking rebaseall

[sagetrac-git]

Changed commit from 2037d68 to da978f8

[mkoeppe]

Changed reviewer from https://github.com/mkoeppe/sage/actions/runs/647547421 to https://github.com/mkoeppe/sage/actions/runs/662602246

[mkoeppe]

comment:74

Now testing at https://github.com/mkoeppe/sage/actions/runs/662602246

[mkoeppe]

comment:75

With the rebasing fixes applied, I am getting again to sage-iii (https://github.com/mkoeppe/sage/runs/2140904387?check_suite_focus=true)

  [sagelib-9.3.beta9]   Generating auto-generated sources
  [sagelib-9.3.beta9]   Building interpreters for fast_callable
  [sagelib-9.3.beta9]   -> First build of interpreters
  [sagelib-9.3.beta9]   running build_cython
  [sagelib-9.3.beta9]   Enabling Cython debugging support
  [sagelib-9.3.beta9]   /cygdrive/d/a/sage/sage/build/pkgs/sagelib/spkg-install: line 58: 61038 Illegal instruction     (core dumped) python3 -u setup.py --no-user-cfg build install

which is again coming from numpy, I think.
Help....

[mkoeppe]

Changed reviewer from https://github.com/mkoeppe/sage/actions/runs/662602246 to Matthias Koeppe, ..., https://github.com/mkoeppe/sage/actions/runs/662602246

[mkoeppe]

comment:76

But the current branch is already an improvement, so let's get this in...

[kliem]

comment:77

LGTM.

[kliem]

Changed reviewer from Matthias Koeppe, ..., https://github.com/mkoeppe/sage/actions/runs/662602246 to Matthias Koeppe, Jonathan Kliem

[mkoeppe]

comment:78

Thanks!

[mkoeppe]

comment:79

Follow-up in #31521

[vbraun]

comment:80

Merge conflict

    STDOUT: Auto-merging tox.ini
    STDOUT: Auto-merging .github/workflows/ci-cygwin-standard.yml
    STDOUT: CONFLICT (content): Merge conflict in .github/workflows/ci-cygwin-standard.yml
    STDOUT: Auto-merging .github/workflows/ci-cygwin-minimal.yml
    STDOUT: CONFLICT (content): Merge conflict in .github/workflows/ci-cygwin-minimal.yml
    STDOUT: Automatic merge failed; fix conflicts and then commit the result.

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

7081189	tox.ini, build/bin/write-dockerfile.sh: Use configure --enable-download-from-upstream-url --enable-experimental-packages instead of setting SAGE_SPKG directly
d6dced2	Merge #31493

[sagetrac-git]

Changed commit from da978f8 to d6dced2

[mkoeppe]

Dependencies: #31493

[vbraun]

Changed branch from public/29537 to d6dced2