Eliminate unnecessary ScatterND

[gramalingam]

Identify ScatterND(data, indices, updates) that can be replaced by Identity(updates). This is generated by the translation of x[:, ...] = y in PyTorch. The specific pattern is that the updated indices take the form [[0], ..., [S-1]] for the first dimension, where S is the size of the first dimension of the updated-data tensor. In effect, the scatter-update ends up being an assignment of a new value to the entire tensor.

[codecov]

❌ 10 Tests Failed:

Tests completed	Failed	Passed	Skipped
13236	10	13226	2528

View the top 3 failed test(s) by shortest run time

onnxscript.backend.onnx_export_test.TestOnnxBackEnd::test_export2python_produces_correct_onnx_script_model_0316_test_ceil_example

Stack Traces | 0.003s run time

onnxscript\backend\onnx_export_test.py:137: in extract_functions
    mod = importlib.import_module(import_name)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib\importlib\__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E   ModuleNotFoundError: No module named 'tests.onnx_backend_test_code.test_ceil_example'

The above exception was the direct cause of the following exception:
.nox\test_torch_nightly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\backend\onnx_export_test.py:271: in test_export2python_produces_correct_onnx_script_model
    functions = extract_functions(backend_test.name, code, self.test_folder)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\backend\onnx_export_test.py:139: in extract_functions
    raise AssertionError(
E   AssertionError: Unable to import 'tests.onnx_backend_test_code.test_ceil_example' (e=No module named 'tests.onnx_backend_test_code.test_ceil_example') (file: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_ceil_example.py', absolute path: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_ceil_example.py', current folder: D:\a\onnxscript\onnxscript
E   ---- CONTENT --
E   import numpy
E   from onnx import TensorProto
E   from onnx.helper import make_tensor
E   from onnxscript import script, external_tensor
E   from onnxscript.values import Opset
E   from onnxscript.onnx_types import FLOAT
E   from onnxscript.onnx_opset import opset13
E   
E   @script()
E   def bck_test_ceil_example(x: FLOAT[2]) -> (FLOAT[2]):
E       y = opset13.Ceil(x)
E       return y

onnxscript.backend.onnx_export_test.TestOnnxBackEnd::test_export2python_produces_correct_onnx_script_model_0743_test_max_int64

Stack Traces | 0.003s run time

onnxscript\backend\onnx_export_test.py:137: in extract_functions
    mod = importlib.import_module(import_name)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib\importlib\__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E   ModuleNotFoundError: No module named 'tests.onnx_backend_test_code.test_max_int64'

The above exception was the direct cause of the following exception:
.nox\test_torch_nightly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\backend\onnx_export_test.py:271: in test_export2python_produces_correct_onnx_script_model
    functions = extract_functions(backend_test.name, code, self.test_folder)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\backend\onnx_export_test.py:139: in extract_functions
    raise AssertionError(
E   AssertionError: Unable to import 'tests.onnx_backend_test_code.test_max_int64' (e=No module named 'tests.onnx_backend_test_code.test_max_int64') (file: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_max_int64.py', absolute path: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_max_int64.py', current folder: D:\a\onnxscript\onnxscript
E   ---- CONTENT --
E   import numpy
E   from onnx import TensorProto
E   from onnx.helper import make_tensor
E   from onnxscript import script, external_tensor
E   from onnxscript.values import Opset
E   from onnxscript.onnx_types import INT64
E   from onnxscript.onnx_opset import opset13
E   
E   @script()
E   def bck_test_max_int64(data_0: INT64[3], data_1: INT64[3]) -> (INT64[3]):
E       result = opset13.Max(data_0, data_1)
E       return result

onnxscript.backend.onnx_export_test.TestOnnxBackEnd::test_export2python_produces_correct_onnx_script_model_1392_test_tile_precomputed

Stack Traces | 0.003s run time

onnxscript\backend\onnx_export_test.py:137: in extract_functions
    mod = importlib.import_module(import_name)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib\importlib\__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E   ModuleNotFoundError: No module named 'tests.onnx_backend_test_code.test_tile_precomputed'

The above exception was the direct cause of the following exception:
.nox\test_torch_nightly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\backend\onnx_export_test.py:271: in test_export2python_produces_correct_onnx_script_model
    functions = extract_functions(backend_test.name, code, self.test_folder)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
onnxscript\backend\onnx_export_test.py:139: in extract_functions
    raise AssertionError(
E   AssertionError: Unable to import 'tests.onnx_backend_test_code.test_tile_precomputed' (e=No module named 'tests.onnx_backend_test_code.test_tile_precomputed') (file: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_tile_precomputed.py', absolute path: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_tile_precomputed.py', current folder: D:\a\onnxscript\onnxscript
E   ---- CONTENT --
E   import numpy
E   from onnx import TensorProto
E   from onnx.helper import make_tensor
E   from onnxscript import script, external_tensor
E   from onnxscript.values import Opset
E   from onnxscript.onnx_types import FLOAT, INT64
E   from onnxscript.onnx_opset import opset13
E   
E   @script()
E   def bck_test_tile_precomputed(x: FLOAT[2,2], y: INT64[2]) -> (FLOAT[4,4]):
E       z = opset13.Tile(x, y)
E       return z

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}

[gramalingam]

Can the torchlib implementation of slice_scatter be optimized to do the same, I wonder? I thought this part was potentially inside pytorch decomposition of x[:, :, :, s] = ..., but perhaps it can be done in slice_scatter itself?

[justinchuby]

Can the torchlib implementation of slice_scatter be optimized to do the same, I wonder? I thought this part was potentially inside pytorch decomposition of x[:, :, :, s] = ..., but perhaps it can be done in slice_scatter itself?

Defininately!

[justinchuby]

The torchlib implementation also seems off: SymInt? start=None, SymInt? end=None means they can be None even though the comment says otherwise. We should improve torchlib for this op. Related #2372

[gramalingam]

The torchlib implementation also seems off: SymInt? start=None, SymInt? end=None means they can be None even though the comment says otherwise. We should improve torchlib for this op. Related #2372

Right. I was puzzled by the comments. Not entirely sure what's happening there.

[titaiwangms]

I don't see these ops in the repro: pytorch/pytorch#157289

[titaiwangms]

Maybe you can delete

onnxscript/onnxscript/rewriter/collapse_slices.py

Line 134 in 7b89760

remove_redundant_scatternd = pattern.RewriteRule(

[gramalingam]

I think we can consolidate the rules separately. (I am thinking of trying out Copilot to do it.)

[gramalingam]

May need to make other dimensions symbolic. Otherwise, all of these ops will be constant-folded, and the indices becomes a constant.