Enabling large tensor support for binary broadcast operators

[ChaiBapchya]

Description

In continuation of #16714
Add tests

• arctan2
• hypot

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Code is well-documented:
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• tests/nightly/test_large_array.py
• src/operator/tensor/elemwise_binary_broadcast_op.h

[access2rohit]

LGTM ! Also paste the output for tests run.
Can we also do large vector tests too for arctan2 ? I assume hypt is hypotenuse so maybe that requires 2D matrix

[ChaiBapchya]

Build flag

python -c "from mxnet.runtime import feature_list; print(feature_list())"
[✖ CUDA, ✖ CUDNN, ✖ NCCL, ✖ CUDA_RTC, ✖ TENSORRT, ✔ CPU_SSE, ✔ CPU_SSE2, ✔ CPU_SSE3, ✔ CPU_SSE4_1, ✔ CPU_SSE4_2, ✖ CPU_SSE4A, ✔ CPU_AVX, ✖ CPU_AVX2, ✔ OPENMP, ✖ SSE, ✔ F16C, ✖ JEMALLOC, ✔ BLAS_OPEN, ✖ BLAS_ATLAS, ✖ BLAS_MKL, ✖ BLAS_APPLE, ✔ LAPACK, ✔ MKLDNN, ✔ OPENCV, ✖ CAFFE, ✖ PROFILER, ✔ DIST_KVSTORE, ✖ CXX14, ✔ INT64_TENSOR_SIZE, ✔ SIGNAL_HANDLER, ✖ DEBUG, ✖ TVM_OP]

Output of test run

ubuntu@ip-172-31-27-86:~/workspace/incubator-mxnet$ nosetests tests/nightly/test_large_array.py:test_binary_broadcast
.
----------------------------------------------------------------------
Ran 1 test in 39.432s

OK

[ChaiBapchya]

LGTM ! Also paste the output for tests run.
Can we also do large vector tests too for arctan2 ? I assume hypt is hypotenuse so maybe that requires 2D matrix

hypot can work for vector too (it takes 2 values as input which basically act as 2 sides of the triangle) - shape of the 2 values has to be the same / broadcastable

[ChaiBapchya]

Vector test

nosetests tests/nightly/test_large_vector.py:test_binary_broadcast
.
----------------------------------------------------------------------
Ran 1 test in 20.125s

OK

[access2rohit]

Vector test

nosetests tests/nightly/test_large_vector.py:test_binary_broadcast
.
----------------------------------------------------------------------
Ran 1 test in 20.125s

OK

Did also you run the whole suite for tensor and vector ?

[ChaiBapchya]

python -m nose tests/nightly/test_large_array.py:test_tvm_add
.
----------------------------------------------------------------------
Ran 1 test in 10.660s

OK

However, tvm_add test fails when individual dimension size > 2**32.
for eg in case of test_large_vector.py SHAPE = (4300000000) test fails.

Verified with @yzhliu that change needs to be done on the tvm side. It's a non-trivial change. and thus the test_tvm_add will be added in subsequent separate PR.

[ChaiBapchya]

Vector test

nosetests tests/nightly/test_large_vector.py:test_binary_broadcast
.
----------------------------------------------------------------------
Ran 1 test in 20.125s

OK

Did also you run the whole suite for tensor and vector ?

Yes. As noted previously, on 480GB CPU test_large_array needs to be run twice (first half tests followed by second half tests approx) due to memory free-ing issue.

[apeforest]

LGTM