[1.x] Backport of intgemm #17559

[kpuatamazon]

Description

This is a backport of intgemm from master to the v1.x branch. intgemm was merged into master in #17559. It is an 8-bit matrix multiply library for x86 CPUs.

Changes from the master cherry-pick:

• Fixing CMakeLists.txt so that target_link_libraries attaches to mxnet_static
• In the tests, @pytest.mark.parametrize is broken with @with_seed() so I used manual for loops
• Add Makefile compilation with apology in advance to @leezu

Checklist

Essentials

• PR's title starts with a category (e.g. [BUGFIX], [MODEL], [TUTORIAL], [FEATURE], [DOC], etc)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented

[mxnet-bot]

Hey @kpuatamazon , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [website, windows-cpu, edge, clang, centos-gpu, windows-gpu, unix-cpu, unix-gpu, centos-cpu, sanity, miscellaneous]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[samskalicky]

@pengzhao-intel @ZhennanQin @ciyongch Can you guys hook us up with someone who might be able to help with the issues around intrinsics?

[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/intgemm.cc:1:
[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/intgemm.h:46:
[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/sse2_gemm.h:3:
[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/kernels.h:4:
[2020-09-07T21:41:15.193Z] build/3rdparty/intgemm/intgemm/intrinsics.h:543:10: error: argument to '__builtin_ia32_psllwi512_mask' must be a constant integer
[2020-09-07T21:41:15.193Z]   return _mm512_slli_epi16(a, b);
[2020-09-07T21:41:15.193Z]          ^~~~~~~~~~~~~~~~~~~~~~~
[2020-09-07T21:41:15.193Z] /usr/lib/llvm-3.9/bin/../lib/clang/3.9.1/include/avx512bwintrin.h:1767:12: note: expanded from macro '_mm512_slli_epi16'
[2020-09-07T21:41:15.193Z]   (__m512i)__builtin_ia32_psllwi512_mask((__v32hi)(__m512i)(A), (int)(B), \
[2020-09-07T21:41:15.193Z]            ^                                                    ~~~~~~~~

https://jenkins.mxnet-ci.amazon-ml.com/blue/rest/organizations/jenkins/pipelines/mxnet-validation/pipelines/clang/branches/PR-19099/runs/3/nodes/42/steps/128/log/?start=0
There seems to be some issues compiling this PR on the 1.x branch

[pengzhao-intel]

@pengzhao-intel @ZhennanQin @ciyongch Can you guys hook us up with someone who might be able to help with the issues around intrinsics?

[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/intgemm.cc:1:
[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/intgemm.h:46:
[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/sse2_gemm.h:3:
[2020-09-07T21:41:15.193Z] In file included from build/3rdparty/intgemm/intgemm/kernels.h:4:
[2020-09-07T21:41:15.193Z] build/3rdparty/intgemm/intgemm/intrinsics.h:543:10: error: argument to '__builtin_ia32_psllwi512_mask' must be a constant integer
[2020-09-07T21:41:15.193Z]   return _mm512_slli_epi16(a, b);
[2020-09-07T21:41:15.193Z]          ^~~~~~~~~~~~~~~~~~~~~~~
[2020-09-07T21:41:15.193Z] /usr/lib/llvm-3.9/bin/../lib/clang/3.9.1/include/avx512bwintrin.h:1767:12: note: expanded from macro '_mm512_slli_epi16'
[2020-09-07T21:41:15.193Z]   (__m512i)__builtin_ia32_psllwi512_mask((__v32hi)(__m512i)(A), (int)(B), \
[2020-09-07T21:41:15.193Z]            ^                                                    ~~~~~~~~

https://jenkins.mxnet-ci.amazon-ml.com/blue/rest/organizations/jenkins/pipelines/mxnet-validation/pipelines/clang/branches/PR-19099/runs/3/nodes/42/steps/128/log/?start=0
There seems to be some issues compiling this PR on the 1.x branch

@kpuatamazon write this code and he will be the right people for this issue :)

[pengzhao-intel]

CC @anko-intel for any quick thoughts ~~~

[samskalicky]

@kpuatamazon write this code and he will be the right people for this issue :)

:D agreed, we're getting close to feature freeze on the 18th so any help would be awesome!

[kpuatamazon]

Hi sorry I work Mondays on this.

The issues are all around older compilers and getting Makefile to work.

gcc below 5 doesn't support avx2 intrinstics well, so I want to disable intgemm compilation below that. Any comments on how best to test for this, keeping in mind that gcc can have many different CXX names? I was thinking about a compilation test. But at least in 2.x, MXNet fetches intgemm so I figured the Makefile would do the same with a target that does a git checkout. Which means any intgemm compilation tests are not available at the time Makefile is deciding whether to download stuff. Currently I have a rather hacky test in the Makefile. Is there somewhere in the main repo to put a compilation test?

Regarding debug and _mm512_slli_epi16 that should be fixed now. It requires a constant shift argument. Compiler versions differ in their willingness to see a constant through a function call, especially in debug. I've changed it to a template argument to force the constant to be passed through function calls.

Windows should be fixed. It just didn't have pytest so import pytest was failing. I've removed the dependency on pytest.