[NPU] Support npu kernel for flatten_contiguous_range op, test=develop #34642

Liu-xiandong · 2021-08-05T09:28:29Z

PR types

New features

PR changes

OPs

Describe

描述：
1.完成flatten_contiguous_range在NPU上的移植，支持float32，float64，uint8, int8, int, int64数据类型

2.通过单测

3.单测在NPU上执行

4.关于int8数据类型，之前代码没有将int8和NPU中的INT8对应，因而NPU识别不了INT8类型。在npu_op_runner.cc中增加了映射关系。

###根据review意见进行了修改
1.增加了python API单测，用于测试相关接口函数。

2.关于flatten_函数。test_flatten_contiguout_range_op.py中在静态图和动态图中分别测试了flatten_函数。静态图中的flatten_函数转化成了flatten函数，NPU中的输入输出采用不同的地址空间，程序可以运行。而动态图中，NPU中的输入输出占用了同一块地址空间。NPU不支持这种inplace操作。特此说明。

paddle-bot-old · 2021-08-05T09:28:33Z

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

…e/Paddle into add_flatten_contiguous_range_npu_op

xingfeng01 · 2021-08-06T04:00:19Z

paddle/fluid/operators/flatten_op_npu.cc

+
+REGISTER_OP_NPU_KERNEL(
+    flatten_contiguous_range,
+    ops::FlattenContiguousRangeNPUKernel<paddle::platform::NPUDeviceContext,


需要 int16 吗？

现有的paddle没有支持int16，所以NPU上也没有支持这个数据类型。

xingfeng01 · 2021-08-06T04:00:48Z

paddle/fluid/framework/ir/CMakeLists.txt

 pass_library(graph_to_program_pass base)
 pass_library(graph_viz_pass base)
-pass_library(lock_free_optimize_pass base)
+pass_library(lock_free_optimize_pass base DEPS string_helper)


这个改动是做什么呢？

这个和当前的PR #34656 会冲突，不需要修改，等前面那个PR合入就好了

一些编译上的错误需要修改，这个应该是不需要提交的，我改一下

xingfeng01 · 2021-08-06T04:02:03Z

paddle/fluid/operators/npu_op_runner.cc

    DTYPE_2_ACL_DTYPE = {
        {framework::proto::VarType::BOOL, ACL_BOOL},
        {framework::proto::VarType::UINT8, ACL_UINT8},
+        {framework::proto::VarType::INT8, ACL_INT8},


这里只增加了uint8,其它的需要支持吗？

其他的已经有映射了，所以是不需要的

xingfeng01 · 2021-08-06T04:02:46Z

python/paddle/fluid/tests/unittests/npu/test_flatten_contiguous_range_op_npu.py

+        }
+
+
+class TestFlattenOp_uint8(TestFlattenOp):


测试好像没有覆盖所有的类型？

测试中覆盖了float64,float32,int,uint8,int8,int64。跟要求的6个数据类型是对应的。

xingfeng01 · 2021-08-06T04:03:34Z

paddle/fluid/operators/expand_op_npu.cc

            "must be less than or equal to %d, but the value received is %d.",
            MAX_RANK_SUPPORTED, rank));
-    switch (rank) { REP_EXPAND_TEMPLATE(MAX_RANK_SUPPORTED) }
+    switch (rank) {


这个改动是所有op都要做的吗？

这个和当前的PR #34656 会冲突，不需要修改，等前面那个PR合入就好了

qili93 · 2021-08-06T04:15:10Z

paddle/fluid/framework/ir/CMakeLists.txt

 pass_library(graph_to_program_pass base)
 pass_library(graph_viz_pass base)
-pass_library(lock_free_optimize_pass base)
+pass_library(lock_free_optimize_pass base DEPS string_helper)


这个和当前的PR #34656 会冲突，不需要修改，等前面那个PR合入就好了

qili93 · 2021-08-06T04:15:22Z

paddle/fluid/operators/expand_op_npu.cc

            "must be less than or equal to %d, but the value received is %d.",
            MAX_RANK_SUPPORTED, rank));
-    switch (rank) { REP_EXPAND_TEMPLATE(MAX_RANK_SUPPORTED) }
+    switch (rank) {


这个和当前的PR #34656 会冲突，不需要修改，等前面那个PR合入就好了

qili93 · 2021-08-06T04:20:26Z

python/paddle/fluid/tests/unittests/npu/test_flatten_contiguous_range_op_npu.py

@@ -0,0 +1,228 @@
+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.


时间是2021

qili93 · 2021-08-06T04:21:07Z

python/paddle/fluid/tests/unittests/npu/test_flatten_contiguous_range_op_npu.py

+
+    def test_check_grad(self):
+        pass
+        #self.check_grad_with_place(self.place, ['X'], 'Out')


删掉注释

qili93 · 2021-08-06T04:22:10Z

python/paddle/fluid/tests/unittests/npu/test_flatten_contiguous_range_op_npu.py

+            "stop_axis": self.stop_axis
+        }
+
+


参考test_flatten_contiguout_range_op.py 增加对Python API的单测

… add_flatten_contiguous_range_npu_op

qili93 · 2021-08-06T10:44:24Z

python/paddle/fluid/tests/unittests/npu/test_flatten_contiguous_range_op_npu.py

+        self.assertRaises(ValueError, test_InputError)
+
+        def test_Negative():
+            paddle.disable_static()


改成 paddle.disable_static(paddle.NPUPlace(0)): 不然这个就是跑在CPU上的

qili93 · 2021-08-06T10:44:34Z

python/paddle/fluid/tests/unittests/npu/test_flatten_contiguous_range_op_npu.py

+        x = x.astype('float32')
+
+        def test_Negative():
+            paddle.disable_static()


qili93

LGTM

qili93

LGTM

qili93 and others added 3 commits August 6, 2021 10:52

fix npu compile error, test=develop

d986e4f

Merge commit 'refs/pull/34656/head' of https://github.com/PaddlePaddl…

ce71970

…e/Paddle into add_flatten_contiguous_range_npu_op

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

9ee888c

Liu-xiandong force-pushed the add_flatten_contiguous_range_npu_op branch from 728f3ce to 9ee888c Compare August 6, 2021 03:54

xingfeng01 reviewed Aug 6, 2021

View reviewed changes

qili93 reviewed Aug 6, 2021

View reviewed changes

Liu-xiandong added 5 commits August 6, 2021 13:00

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

0da867c

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

58b2b6b

… add_flatten_contiguous_range_npu_op

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

c70db0f

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

e92e5b0

… add_flatten_contiguous_range_npu_op

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

548adcb

qili93 reviewed Aug 6, 2021

View reviewed changes

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

ee141b5

qili93 previously approved these changes Aug 9, 2021

View reviewed changes

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

a9595e7

Liu-xiandong dismissed qili93’s stale review via a9595e7 August 9, 2021 12:45

Liu-xiandong and others added 4 commits August 10, 2021 10:01

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop

f3f1ce5

Merge branch 'develop' into add_flatten_contiguous_range_npu_op

5c97462

Update flatten_op_npu.cc

c9f1676

Update flatten_op_npu.cc

5287906

qili93 approved these changes Aug 10, 2021

View reviewed changes

qili93 merged commit 79be842 into PaddlePaddle:develop Aug 10, 2021

		@@ -0,0 +1,228 @@
		# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.

		}


		class TestFlattenOp_uint8(TestFlattenOp):

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop #34642

[NPU] Support npu kernel for flatten_contiguous_range op, test=develop #34642

Uh oh!

Conversation

Liu-xiandong commented Aug 5, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR types

PR changes

Describe

Uh oh!

paddle-bot-old bot commented Aug 5, 2021

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

qili93 left a comment

Choose a reason for hiding this comment

Uh oh!

qili93 left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Liu-xiandong commented Aug 5, 2021 •

edited

Loading