Dispatch to mv rather than mm in the case that tensor1.ndim == 1 and tensor2.ndim == 2

[lezcano]

Stack from ghstack:

• Avoid copies in matmul #76828
• Micro-optimisations for matmul 2.0: Electric boogaloo #75197
• Go through the dispatcher in matmul_out for the 1D-1D case #75196
• -> Dispatch to mv rather than mm in the case that tensor1.ndim == 1 and tensor2.ndim == 2 #75195
• Fix mv/addmv on CUDA when dealing with vectors of size=1 and stride=0 #75279
• Refactor the API of the matmul implementation #75194

This should hopefully be faster, it makes the calling code simpler, and
it solves a bug when using matmul with the out= parameter (before it
would throw an incorrect error).

[facebook-github-bot]

🔗 Helpful links

• 🧪  See artifacts and rendered test results at hud.pytorch.org/pr/75195
• 📄  Preview Python docs built from this PR
• 📄  Preview C++ docs built from this PR
• ❓Need help or want to give feedback on the CI? Visit our office hours

💊 CI failures summary and remediations

As of commit debf0bf (more details on the Dr. CI page):

Expand to see more

---

• 1/1 failures introduced in this PR

---

🕵️ 1 new failure recognized by patterns

The following CI failures do not appear to be due to upstream breakages

pull / pytorch-xla-linux-bionic-py3.7-clang8 / test (xla, 1, 1, linux.2xlarge) (1/1)

Step: "Test" (full log | diagnosis details | 🔁 rerun)

2022-05-04T20:36:24.5572937Z �[0;31m[ FAILED ] �[mAtenXlaTensorTest.TestMatmul_1x2

2022-05-04T20:36:24.5513452Z �[0;32m[ RUN      ] �[mXlaUtilCacheTest.BasicTest
2022-05-04T20:36:24.5514102Z �[0;32m[       OK ] �[mXlaUtilCacheTest.BasicTest (0 ms)
2022-05-04T20:36:24.5517062Z �[0;32m[----------] �[m1 test from XlaUtilCacheTest (0 ms total)
2022-05-04T20:36:24.5517392Z 
2022-05-04T20:36:24.5570100Z �[0;32m[----------] �[mGlobal test environment tear-down
2022-05-04T20:36:24.5570786Z �[0;32m[==========] �[m623 tests from 8 test suites ran. (512555 ms total)
2022-05-04T20:36:24.5571192Z �[0;32m[  PASSED  ] �[m621 tests.
2022-05-04T20:36:24.5571554Z �[0;32m[  SKIPPED ] �[m1 test, listed below:
2022-05-04T20:36:24.5572040Z �[0;32m[  SKIPPED ] �[mAtenXlaTensorTest.TestGroupNormBackward
2022-05-04T20:36:24.5572501Z �[0;31m[  FAILED  ] �[m1 test, listed below:
2022-05-04T20:36:24.5572937Z �[0;31m[  FAILED  ] �[mAtenXlaTensorTest.TestMatmul_1x2
2022-05-04T20:36:24.5573171Z 
2022-05-04T20:36:24.5573268Z  1 FAILED TEST
2022-05-04T20:36:24.7276308Z + cleanup
2022-05-04T20:36:24.7276526Z + retcode=1
2022-05-04T20:36:24.7276685Z + set +x
2022-05-04T20:36:24.7569431Z ##[error]Process completed with exit code 1.
2022-05-04T20:36:24.7765995Z ##[group]Run pytorch/pytorch/.github/actions/get-workflow-job-id@master
2022-05-04T20:36:24.7766245Z with:
2022-05-04T20:36:24.7766611Z   github-token: ***
2022-05-04T20:36:24.7766792Z env:

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[lezcano]

@ezyang This PR shows (or will show, once the CI of the previous PRs is green) that this is the line that's causing trouble.

We get this error: https://github.com/pytorch/pytorch/runs/5820894454?check_suite_focus=true in the distributed build. I have tried to track down what could be causing it, and following the execution of the trace that it throws, I don't see how that state could be reached. It's also particularly odd that this just happens in some random test in a distributed setting...

[ezyang]

Can you repro it locally?

I looked over the patch and it looks... fine-ish? The contiguous changes look a little skeevy

[lezcano]

I could not reproduce it locally. This PR should just be one line. I pushed some further changes to start debugging it via CI, but this is certainly not the best approach.

Once we had access to sshing into the jobs, but J believe that's not working atm...

[ngimel]

@pytorchbot merge this