[Operator] fix index_select bug [MooreThreads]

[Salv1a]

PR Category

Operator

Type of Change

Bug Fix

Description

We found an issue with the definition of cols_offsets in the index_select operator. When N is larger than BLOCK_N, the indices cannot cover all the input columns.

Issue

• Test 1: With the input dimension of REDUCTION_SHAPES=(4096,n), we forced BLOCK_N=2048. When dim=0, test_reduction_ops.py::test_accuracy_index_select failed.
• Test 2: We increased the original input dimension to REDUCTION_SHAPES=(4096*m,n). When dim=0, test_reduction_ops.py::test_accuracy_index_select failed.

  ======================================== short test summary info =========================================
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype0-0-shape0] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype0-0-shape1] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype0-0-shape2] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype0-0-shape3] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype0-0-shape4] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype1-0-shape0] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype1-0-shape1] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype1-0-shape2] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype1-0-shape3] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype1-0-shape4] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype2-0-shape0] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype2-0-shape1] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype2-0-shape2] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype2-0-shape3] - AssertionError
  FAILED tests/test_reduction_ops.py::test_accuracy_index_select[dtype2-0-shape4] - AssertionError
  =============================== 15 failed, 15 passed, 1 warning in 11.51s ================================

Progress

• [x] Change is properly reviewed (1 reviewer required, 2 recommended).
• [x] Change is responded to an issue.
• [x] Change is fully covered by a UT.

Advice

We suggest to increase the input dimensions in the test cases, ensuring that the input dimensions are larger than the dimensions of the Triton sub-blocks, in order to improve test accuracy.

[Salv1a]

have to add from ..utils import libentry to test reduction ops case?

[StrongSpoon]

have to add from ..utils import libentry to test reduction ops case?

not necessary, but recommend rebasing on the latest master branch and running ci again.

[Salv1a]

have to add from ..utils import libentry to test reduction ops case?

not necessary, but recommend rebasing on the latest master branch and running ci again.

already rebased, please approve ci

[Salv1a]

what's wrong...?

[tongxin]

what's wrong...?

We'll take a look. Thanks.

[StrongSpoon]

what's wrong...?

add unit test parameter to cover the case you mentioned, and coverage check will pass.

[Salv1a]

what's wrong...?

add unit test parameter to cover the case you mentioned, and coverage check will pass.

Already modified the op test parameter, but in an inelegant way: we did not modify REDUCTION_SHAPE to avoid potential errors. Maybe you should consider increasing REDUCTION_SHAPE to ensure the accuracy of the tests.