Support code generation for op fill_any #54378

huangjiyi · 2023-06-06T06:34:49Z

PR types

Others

PR changes

Others

Description

Support code generation for op fill_any
算子定义生成体系建设--静态图算子自动生成-第二期 #53267

paddle-bot · 2023-06-06T06:34:55Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

huangjiyi · 2023-06-06T06:42:59Z

test/legacy_test/test_fill_any_op.py

-def fill_any_wrapper(x, value_float=0, value_int=0):
-    return paddle._legacy_C_ops.fill_any(
-        x, "value_float", value_float, "value_int", value_int
-    )
+def fill_any_wrapper(x, value=0):
+    return paddle._legacy_C_ops.fill_any(x, "value", value)


改这里是因为原来的 attr value 在 OpMaker 里有定义两个：value_float 和 value_int，但实际用到的只有 value_float（因为 fill_sig.cc 的映射只有 value_float，即传给 FillKernel 的 value 只有 value_float），所以既然 value_int 没用到就删了，然后为了与 kernel 对齐，value_float 改成了 value

目前看fill_any仅仅用于旧动态图算子，但是不确定旧版本保存的模型是否含有该算子，所以这里删除op的attribute存在潜在的兼容性问题，这个PR需要其他人同意后才可以merge。

huangjiyi · 2023-06-06T06:48:23Z

本次 PR 自动生成的 op 是 fill_any 而不是 fill，因为 fill_any 的算子名会被映射成 fill，所以在 yaml 里配置的都是 fill，另外 fluid/operators 下还有一个 fill_op.cc，感觉这里定义的才是真正的 fill

huangjiyi · 2023-06-06T12:57:27Z

@heavyrain-lzy，帮忙 Review 一下 ~

heavyrain-lzy

LGTM

heavyrain-lzy · 2023-06-06T13:18:02Z

test/legacy_test/test_fill_any_op.py

-def fill_any_wrapper(x, value_float=0, value_int=0):
-    return paddle._legacy_C_ops.fill_any(
-        x, "value_float", value_float, "value_int", value_int
-    )
+def fill_any_wrapper(x, value=0):
+    return paddle._legacy_C_ops.fill_any(x, "value", value)


目前看fill_any仅仅用于旧动态图算子，但是不确定旧版本保存的模型是否含有该算子，所以这里删除op的attribute存在潜在的兼容性问题，这个PR需要其他人同意后才可以merge。

zyfncg · 2023-06-07T02:27:00Z

test/legacy_test/test_fill_any_op.py

-        x, "value_float", value_float, "value_int", value_int
-    )
+def fill_any_wrapper(x, value=0):
+    return paddle._legacy_C_ops.fill_any(x, "value", value)


这里的fill_any_wrapper是不可以直接去掉了？

我删了之后跑单测会报下面的错误：

871: Traceback (most recent call last): 871: File "/parallel/Paddle/build/test/legacy_test/test_fill_any_op.py", line 40, in test_check_output 871: self.check_output() 871: File "/parallel/Paddle/build/test/legacy_test/eager_op_test.py", line 2176, in check_output 871: check_cinn=check_cinn, 871: File "/parallel/Paddle/build/test/legacy_test/eager_op_test.py", line 2042, in check_output_with_place 871: dygraph_checker.check() 871: File "/parallel/Paddle/build/test/legacy_test/eager_op_test.py", line 1806, in check 871: self.calculate_output() 871: File "/parallel/Paddle/build/test/legacy_test/eager_op_test.py", line 1878, in calculate_output 871: dygraph_outs = self.op_test._calc_python_api_output(place) 871: File "/parallel/Paddle/build/test/legacy_test/eager_op_test.py", line 1068, in _calc_python_api_output 871: % self.op_type 871: AssertionError: Detect there is KernelSignature for `fill_any` op, please set the `self.python_api` if you set check_dygraph = True

zyfncg · 2023-06-07T02:27:49Z

paddle/phi/api/yaml/op_compat.yaml

+  scalar :
+    value :
+      data_type : float
+      support_tensor : true


fill_any的value之前没有配置support_tensor，感觉这里也可以不加support_tensor

Paddle/paddle/fluid/operators/generator/generate_op.py

Lines 118 to 119 in 397758f

if attr_item['is_support_tensor'] is False:

attr_item['tensor_name'] = scalar_config['tensor_name']

这里如果不加 support_tensor，generate_op.py 会报错说需要配置 tensor_name，如果把 tensor_name 也配置成 value 的话，生成的 OpMaker 会多一个 Input，然后编译的时候会报 input 数量有问题

class FillAnyOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor), input 0 of fill_any op."); AddOutput("Out", "(Tensor), output 0 of fill_any op."); AddInput("value", "attribute 0 for fill_any op from 0D Tensor.") .AsDispensable(); AddAttr<float>("value", "(float), attribute 0 for fill_any op.") .SetDefault(0);

Paddle/paddle/fluid/operators/generator/generate_op.py

Lines 118 to 119 in 397758f

if attr_item['is_support_tensor'] is False:

attr_item['tensor_name'] = scalar_config['tensor_name']

这里如果不加 support_tensor，generate_op.py 会报错说需要配置 tensor_name，如果把 tensor_name 也配置成 value 的话，生成的 OpMaker 会多一个 Input，然后编译的时候会报 input 数量有问题

class FillAnyOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor), input 0 of fill_any op."); AddOutput("Out", "(Tensor), output 0 of fill_any op."); AddInput("value", "attribute 0 for fill_any op from 0D Tensor.") .AsDispensable(); AddAttr<float>("value", "(float), attribute 0 for fill_any op.") .SetDefault(0);

尝试修改一下这里的逻辑：

if attr_item['is_support_tensor'] is False and 'tensor_name' in scalar_config: attr_item['tensor_name'] = scalar_config['tensor_name']

尝试修改一下这里的逻辑：

if attr_item['is_support_tensor'] is False and 'tensor_name' in scalar_config: attr_item['tensor_name'] = scalar_config['tensor_name']

这样虽然不会报错，但是生成的 OpMaker 中依然会多一个 Input：

class FillAnyOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor), input 0 of fill_any op."); AddOutput("Out", "(Tensor), output 0 of fill_any op."); AddInput("ValueTensor", "attribute 0 for fill_any op from 0D Tensor.") .AsDispensable(); AddAttr<float>("value", "(float), attribute 0 for fill_any op.") .SetDefault(0);

好像是因为下面的代码：

Paddle/paddle/fluid/operators/generator/templates/operator_utils.c.j2

Lines 75 to 79 in 397758f

{% if typename is scalar and

("is_support_tensor" not in attr or attr["is_support_tensor"] is false) %}

AddInput("{{attr | to_scalar_tensor_name}}", "attribute {{i}} for {{op_name}} op from 0D Tensor.")

.AsDispensable();

AddAttr<{{attr["data_type"]}}>("{{name}}", "({{attr["data_type"]}}), attribute {{i}} for {{op_name}} op.")

这里只要 is_support_tensor 设置 false（包括不设置）就会生成一个 Input

确实存在问题，那暂时保持support_tensor : true

huangjiyi · 2023-06-07T07:12:48Z

@zyfncg，Review 意见都回复了，帮忙再 Review 一下 ~

heavyrain-lzy · 2023-06-13T09:19:41Z

麻烦解决一下conflict

huangjiyi · 2023-06-14T02:39:04Z

@heavyrain-lzy，帮忙再 Review 一下 ~

heavyrain-lzy

LGTM

huangjiyi · 2023-06-14T02:53:01Z

@luotao1，帮忙豁免一下 CI ~

update

6c92eca

paddle-bot bot added contributor External developers status: proposed labels Jun 6, 2023

huangjiyi commented Jun 6, 2023

View reviewed changes

huangjiyi changed the title ~~Support code generation for op fill~~ Support code generation for op fill_any Jun 6, 2023

huangjiyi changed the title ~~Support code generation for op fill_any~~ Support code generation for op ~~fill~~ fill_any Jun 6, 2023

huangjiyi changed the title ~~Support code generation for op ~~fill~~ fill_any~~ Support code generation for op fill_any Jun 6, 2023

update

ddbc1bd

heavyrain-lzy previously approved these changes Jun 6, 2023

View reviewed changes

zyfncg reviewed Jun 7, 2023

View reviewed changes

luotao1 assigned luotao1 and heavyrain-lzy Jun 7, 2023

luotao1 added the HappyOpenSource 快乐开源活动issue与PR label Jun 7, 2023

paddle-bot bot removed the status: proposed label Jun 7, 2023

huangjiyi requested a review from zyfncg June 9, 2023 05:48

zyfncg previously approved these changes Jun 13, 2023

View reviewed changes

luotao1 mentioned this pull request Jun 13, 2023

算子定义生成体系建设--静态图算子自动生成-第二期 #53267

Closed

Merge branch 'develop' into autogen_fill

918998f

huangjiyi dismissed stale reviews from zyfncg and heavyrain-lzy via 918998f June 13, 2023 12:00

heavyrain-lzy approved these changes Jun 14, 2023

View reviewed changes

luotao1 merged commit 4277f61 into PaddlePaddle:develop Jun 14, 2023

huangjiyi deleted the autogen_fill branch January 9, 2024 12:04

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support code generation for op fill_any #54378

Support code generation for op fill_any #54378

huangjiyi commented Jun 6, 2023 •

edited

Loading

paddle-bot bot commented Jun 6, 2023

huangjiyi Jun 6, 2023 •

edited

Loading

heavyrain-lzy Jun 6, 2023

huangjiyi Jun 6, 2023

huangjiyi commented Jun 6, 2023 •

edited

Loading

huangjiyi commented Jun 6, 2023

heavyrain-lzy left a comment

heavyrain-lzy Jun 6, 2023

zyfncg Jun 7, 2023

huangjiyi Jun 7, 2023 •

edited

Loading

zyfncg Jun 7, 2023

huangjiyi Jun 7, 2023 •

edited

Loading

heavyrain-lzy Jun 13, 2023

huangjiyi Jun 13, 2023

heavyrain-lzy Jun 13, 2023

huangjiyi commented Jun 7, 2023

heavyrain-lzy commented Jun 13, 2023

huangjiyi commented Jun 14, 2023

heavyrain-lzy left a comment

huangjiyi commented Jun 14, 2023

	if attr_item['is_support_tensor'] is False:
	attr_item['tensor_name'] = scalar_config['tensor_name']

	{% if typename is scalar and
	("is_support_tensor" not in attr or attr["is_support_tensor"] is false) %}
	AddInput("{{attr \| to_scalar_tensor_name}}", "attribute {{i}} for {{op_name}} op from 0D Tensor.")
	.AsDispensable();
	AddAttr<{{attr["data_type"]}}>("{{name}}", "({{attr["data_type"]}}), attribute {{i}} for {{op_name}} op.")

Support code generation for op fill_any #54378

Support code generation for op fill_any #54378

Conversation

huangjiyi commented Jun 6, 2023 • edited Loading

PR types

PR changes

Description

paddle-bot bot commented Jun 6, 2023

huangjiyi Jun 6, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

huangjiyi commented Jun 6, 2023 • edited Loading

huangjiyi commented Jun 6, 2023

heavyrain-lzy left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

huangjiyi Jun 7, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

huangjiyi Jun 7, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

huangjiyi commented Jun 7, 2023

heavyrain-lzy commented Jun 13, 2023

huangjiyi commented Jun 14, 2023

heavyrain-lzy left a comment

Choose a reason for hiding this comment

huangjiyi commented Jun 14, 2023

huangjiyi commented Jun 6, 2023 •

edited

Loading

huangjiyi Jun 6, 2023 •

edited

Loading

huangjiyi commented Jun 6, 2023 •

edited

Loading

huangjiyi Jun 7, 2023 •

edited

Loading

huangjiyi Jun 7, 2023 •

edited

Loading