[PIR] 解决废弃算子mul op 在新 IR 下的translator问题

[DrRyanHuang]

PR types

Others

PR changes

Others

Description

Others

• [NewIR] 在op_translator.cc支持mul算子的映射处理 #56521

[kangguangli]

@kangguangli 非常不好意思，打扰您了

Op mul arg out should be optional if it can be empty

想问一下您，我是需要重写一下 OpTranscriber::GenerateOperationOutput 方法吗?

把这一行注释掉嘛?

      IR_ENFORCE(info.optional,
                 "Op %s arg %s should be optional if it can be empty",
                 op_desc.Type(),
                 legacy_output_name);

2023-09-05 21:26:30 ======================================================================
2023-09-05 21:26:30 ERROR: test_check_output (test_mul_op.TestMulOp2)
2023-09-05 21:26:30 ----------------------------------------------------------------------
2023-09-05 21:26:30 Traceback (most recent call last):
2023-09-05 21:26:30   File "/workspace/Paddle/build/test/legacy_test/test_mul_op.py", line 90, in test_check_output
2023-09-05 21:26:30     self.check_output(check_dygraph=False)
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/eager_op_test.py", line 2605, in check_output
2023-09-05 21:26:30     res = self.check_output_with_place(
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/eager_op_test.py", line 2459, in check_output_with_place
2023-09-05 21:26:30     static_checker.check()
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/eager_op_test.py", line 2117, in check
2023-09-05 21:26:30     self.calculate_output()
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/eager_op_test.py", line 2125, in calculate_output
2023-09-05 21:26:30     outs, fetch_list = self.op_test._calc_output(
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/eager_op_test.py", line 1513, in _calc_output
2023-09-05 21:26:30     self._check_ir_output(place, program, feed_map, fetch_list, outs)
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/eager_op_test.py", line 1395, in _check_ir_output
2023-09-05 21:26:30     ir_outs = executor.run(
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/paddle/fluid/executor.py", line 1635, in run
2023-09-05 21:26:30     res = self._run_impl(
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/paddle/fluid/executor.py", line 1812, in _run_impl
2023-09-05 21:26:30     program, new_exe = self._executor_cache.get_program_and_executor(
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/paddle/fluid/executor.py", line 890, in get_program_and_executor
2023-09-05 21:26:30     return self._get_cached_program_and_executor(
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/paddle/fluid/executor.py", line 1011, in _get_program_and_executor
2023-09-05 21:26:30     new_exe = _StandaloneExecutor(place, plan, scope)
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/paddle/fluid/executor.py", line 787, in __init__
2023-09-05 21:26:30     self._new_exe = self._create_new_executor()
2023-09-05 21:26:30   File "/workspace/Paddle/build/python/paddle/fluid/executor.py", line 811, in _create_new_executor
2023-09-05 21:26:30     new_exe = core.StandaloneExecutor(self._place, self._plan, self._scope)
2023-09-05 21:26:30 RuntimeError: Error occured at: ../paddle/fluid/ir_adaptor/translator/op_translator.cc:434 :
2023-09-05 21:26:30 Op mul arg out should be optional if it can be empty

@DrRyanHuang
这种情况理论上需要添加参数映射，错误原因是PT模块安装OpYamlInfo从OpDesc中查找需要的信息，OpYamlInfo标记的输出是out，但是实际上OpDesc中存储的是Out，我们需要标记这种映射关系。

可以尝试修改op_compat.yaml，如下：

- op : matmul_with_flatten (mul)
  backward : matmul_with_flatten_grad (mul_grad) 
  inputs :
    {x : X, y : Y}
  outputs :
    out : Out
  extra :
    attrs : [bool use_mkldnn = false, float scale_x = 1.0f, 'float[] scale_y = {1.0f}',
             float scale_out = 1.0f, bool force_fp32_output = false]

这里写的可以参考下。我们对mul的处理比较特殊，如果这样后续有问题的话，可以再@下我。

[paddle-ci-bot]

Sorry to inform you that ebce876's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[paddle-ci-bot]

Sorry to inform you that 51b9e8c's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[paddle-ci-bot]

Sorry to inform you that 62b9089's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[Aurelius84]

comment 里的建议下个PR 统一fix下~~

[kangguangli]

LGTM overall. But there're two problems that we should consider.

1. insert two much reshape, every pair of mul/mul_grad will lead to 6-10 additional reshape op insertions, is there some way to decrease this number?
2. too much redundant code. Try to compress by some kind of abstraction like function.

[zyfncg]

LGTM for op_compat