Move masked_select broadcasting from codegen layer to native layer.

[gchanan]

Stack from ghstack:

• Specify _th_ ops in CUDAUnaryOps macros so they are easier to find. #37582 Specify th ops in CUDAUnaryOps macros so they are easier to find.
• Kill the ability to codegen tensor-based broadcasting. #37547 Kill the ability to codegen tensor-based broadcasting.
• Move addr broadcasting from codegen layer to native layer. #37546 Move addr broadcasting from codegen layer to native layer.
• Move broadcasting code for fmod, fmod_ from codegen layer. #37545 Move broadcasting code for fmod, fmod_ from codegen layer.
• Move baddbmm broadcasting from codegen layer to native layer. #37544 Move baddbmm broadcasting from codegen layer to native layer.
• Move masked_select broadcasting from codegen layer to native layer. #37543 Move masked_select broadcasting from codegen layer to native layer.

Differential Revision: D21315038

[dr-ci]

💊 Build failures summary and remediations

As of commit fb23d3f (more details on the Dr. CI page):

---

• 2/2 failures possibly* introduced in this PR
  • 1/2 non-CircleCI failure(s)

---

🕵️ 1 new failure recognized by patterns

The following build failures do not appear to be due to upstream breakages:

caffe2_onnx_ort1_py3_6_clang7_ubuntu16_04_test (1/1)

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

Apr 29 23:40:44 E AssertionError: False is not true

Apr 29 23:40:44         worker_coordinator = parallel_workers.init_workers(dummy_worker) 
Apr 29 23:40:44         worker_coordinator.start() 
Apr 29 23:40:44      
Apr 29 23:40:44         for _ in range(10): 
Apr 29 23:40:44             value = dequeue_value(queue) 
Apr 29 23:40:44             self.assertTrue( 
Apr 29 23:40:44                 value in [b'0', b'1'], 'Got unexpected value ' + str(value) 
Apr 29 23:40:44             ) 
Apr 29 23:40:44      
Apr 29 23:40:44 >       self.assertTrue(worker_coordinator.stop()) 
Apr 29 23:40:44 E       AssertionError: False is not true 
Apr 29 23:40:44  
Apr 29 23:40:44 ../.local/lib/python3.6/site-packages/caffe2/python/parallel_workers_test.py:73: AssertionError 
Apr 29 23:40:44 ----------------------------- Captured stdout call ----------------------------- 
Apr 29 23:40:44 Wait for workers to die: train 
Apr 29 23:40:44 Worker <Thread(parallel_workers worker id 0, started daemon 139679760369408)> failed to close while waiting 
Apr 29 23:40:44 Worker <Thread(parallel_workers worker id 1, started daemon 139679768762112)> failed to close while waiting 
Apr 29 23:40:44 All workers terminated: False 
Apr 29 23:40:44 - generated xml file: /var/lib/jenkins/workspace/caffe2_tests/python/result.xml - 
Apr 29 23:40:44 ---------- onnx coverage: ---------- 
Apr 29 23:40:44 Operators (passed/loaded/total): 0/0/177 

---

ci.pytorch.org: 1 failed

• Failed: pr/py3.6-clang7-rocmdeb-ubuntu16.04

---

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

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions on the GitHub issue tracker.

See how this bot performed.

This comment has been revised 2 times.

[bhosmer]

Haven't followed it thru the codegen but I'm assuming removal of the broadcast attribute in Declarations.cwrap eliminates the expand_outplace logic below... if so, does the codegen just generate a one-liner that returns s__th_masked_select(self, mask) now?

Tensor _th_masked_select(const Tensor & self, const Tensor & mask) {

    // DeviceGuard omitted
    Tensor b_self, b_mask;
    std::tie(b_self, b_mask) = expand_outplace(self, mask, "_th_masked_select");
    return s__th_masked_select(b_self, b_mask);
}

[gchanan]

that's basically correct except s__th_masked_select doesn't get generated anymore since it's not needed (the "s_" is for same size) -- so s__th_masked_select just becomes _th_masked_select and the old _th_masked_select moves to native.

[bhosmer]

that's basically correct except s__th_masked_select doesn't get generated anymore since it's not needed (the "s_" is for same size) -- so s__th_masked_select just becomes _th_masked_select and the old _th_masked_select moves to native.

Even better :)

BTW, what's the ONNX failure?

[facebook-github-bot]

@gchanan merged this pull request in f09eb39.

[gchanan]

BTW, what's the ONNX failure?

looks flaky.