[phi] Decoupled phi from fluid tracking issue

[GreatV]

问题描述 Please describe your issue

项目说明

用于跟踪 Call-for-Contributions: 飞桨算子体系重构 - PHI算子库独立编译 并记录该计划下各个子任务的计划和进度。

主要工作

当前 PHI 解耦独立编译基本可以分为以下两个阶段的工作。

头文件依赖解耦（整体进展 103/103）

按 merge 的时间顺序，排名不分先后：@GreatV (8), @jzhang533 (1), @chenwhql (6), @huangjiyi (53), @AndPuQing (7), @Silv3S (1), @jczaja (1)， @Asthestarsfalll (2) @Patrick-Star125 (5) @DrRyanHuang (1),@engineer1109(3), @RedContritio (3) @YuanRisheng (8) @gouzil（2）@jinyouzhi(1) @jiahy0825(1)

编号	细分头文件	存量数	难度	相关PR	作者
	======= collective ========
1	paddle/fluid/distributed/collective/ProcessGroup.h	4	困难	#51651	@YuanRisheng
2	paddle/fluid/distributed/collective/ProcessGroupNCCL.h	1	困难	#51651	@YuanRisheng
	======= eager ========
3	paddle/fluid/eager/api/utils/global_utils.h ✅ (2022/11/08)	3	简单	#47739	@GreatV
	======= framework ========
4	paddle/fluid/framework/convert_utils.h ✅ (2022/11/17)	17	中等	#48001 #48168	@huangjiyi
5	paddle/fluid/framework/custom_operator.h	1	中等		@jiahy0825
6	paddle/fluid/framework/data_layout.h ✅ (2022/11/09)	2	简单	#47770	@huangjiyi
7	paddle/fluid/framework/data_type.h ✅ (2022/11/09)	9	简单	#47776	@GreatV
8	paddle/fluid/framework/eigen.h ✅ (2022/11/08)	9	简单	#47675	@jzhang533
9	paddle/fluid/framework/expect.h ✅ (2022/11/05)	1	简单	#46342	@chenwhql
10	paddle/fluid/framework/generator.h ✅ (2022/11/10)	9	简单	#47822	@GreatV
11	paddle/fluid/framework/gpu_utils.h ✅ (2022/11/16)	1	中等	#48020	@huangjiyi
12	paddle/fluid/framework/infershape_utils.h ✅ (2022/11/09)	2	中等	#47714	@chenwhql
13	paddle/fluid/framework/lod_tensor.h ✅ (2022/11/11)	2	简单	#47840	@GreatV
14	paddle/fluid/framework/mixed_vector.h ✅	5	中等	#50282	@huangjiyi
15	paddle/fluid/framework/op_registry.h ✅ (2022/11/11)	5	简单	#47868	@GreatV
16	paddle/fluid/framework/op_version_registry.h ✅ (2022/11/09)	1	简单	#47714	@chenwhql
17	paddle/fluid/framework/operator.h	2	中等	#51472 #50624	@jinyouzhi @RedContritio
18	paddle/fluid/framework/tensor.h ✅ (2023/2/20)	4	简单	#50475	@RedContritio
19	paddle/fluid/framework/tensor_util.h ✅	43	困难	#49885 #50139 #50264 #50538 #50608	@engineer1109 @huangjiyi
20	paddle/fluid/framework/threadpool.h ✅ (2022/11/21)	1	中等	#48075	@AndPuQing
21	paddle/fluid/framework/variable.h ✅	2	中等	#50407	@huangjiyi
22	paddle/fluid/framework/version.h ✅	1	困难	#50608	@huangjiyi
	======= imperative ========
23	paddle/fluid/imperative/gradient_accumulator.h ✅	1	中等	#50385	@huangjiyi
	======= memory ========
24	paddle/fluid/memory/allocation/allocator.h	3	困难	#50419	@YuanRisheng
25	paddle/fluid/memory/buffer.h	2	中等	#51288	@YuanRisheng
26	paddle/fluid/memory/malloc.h	18	困难	#50419	@YuanRisheng
27	paddle/fluid/memory/memcpy.h	47	困难	#50870	@YuanRisheng
28	paddle/fluid/memory/memory.h	18	困难	#50870	@YuanRisheng
	======= operators ========
29	paddle/fluid/operators/amp/fp16_type_traits.h ✅ (2022/11/18)	5	中等	#48051	@huangjiyi
30	paddle/fluid/operators/conv_op.h ✅ (2022/11/15)	1	中等	#47914	@GreatV
31	paddle/fluid/operators/detection/bbox_util.h ✅ (2022/12/8)	1	中等	#48761	@Patrick-Star125
32	paddle/fluid/operators/detection/poly_util.h ✅ (2022/12/5)	1	中等	#48499	@Patrick-Star125
33	paddle/fluid/operators/dropout_impl.cu.h ✅ (2022/12/20)	2	中等	#49139	@huangjiyi
34	paddle/fluid/operators/eigen/eigen_function.h ✅ (2022/11/11)	5	简单	#47842	@huangjiyi
35	paddle/fluid/operators/elementwise/elementwise_op_function.h ✅ (2022/11/15)	1	简单	#47870	@huangjiyi
36	paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h ✅ (2022/11/15)	3	简单	#47870	@huangjiyi
37	paddle/fluid/operators/gather_scatter_kernel.h ✅ (2022/12/19)	8	中等	#49132	@huangjiyi
38	paddle/fluid/operators/jit/kernels.h	6	困难	#50911	@gouzil
39	paddle/fluid/operators/layer_norm_kernel.cu.h ✅	2	中等	#50506	@huangjiyi
40	paddle/fluid/operators/layout_utils.h ✅	2	中等	#50355	@huangjiyi
41	paddle/fluid/operators/math.h ✅ (2022/11/17)	6	中等	#48062	@huangjiyi
42	paddle/fluid/operators/math/cross_entropy.h ✅ (2022/11/21)	3	中等	#48160	@huangjiyi
43	paddle/fluid/operators/math/im2col.h ✅ (2022/11/23)	9	中等	#48174	@huangjiyi
44	paddle/fluid/operators/math/inclusive_scan.h ✅ (2022/12/12)	2	中等	#48980	@Patrick-Star125
45	paddle/fluid/operators/math/matrix_bit_code.h ✅ (2022/12/19)	2	中等	#49131	@huangjiyi
46	paddle/fluid/operators/math/maxouting.h ✅ (2022/12/19)	2	中等	#49131	@huangjiyi
47	paddle/fluid/operators/math/sequence_padding.h	2	中等	#50639	@huangjiyi
48	paddle/fluid/operators/math/sequence_scale.h	2	中等	#49668	@DrRyanHuang
49	paddle/fluid/operators/math/softmax.h ✅ (2022/12/15)	6	中等	#48970	@huangjiyi
50	paddle/fluid/operators/math/softmax_impl.h ✅ (2022/12/15)	2	中等	#48970	@huangjiyi
51	paddle/fluid/operators/math/vol2col.h ✅(2022/11/22)	5	中等	#48175	@huangjiyi
52	paddle/fluid/operators/mkldnn/axpy_handler.h ✅(2022/11/15)	1	中等	#47779	@Silv3S
53	paddle/fluid/operators/norm_utils.cu.h ✅(2022/12/12)	2	中等	#48930	@huangjiyi
54	paddle/fluid/operators/ops_extra_info.h ✅ (2022/12/9)	1	中等	#48926	@huangjiyi
55	paddle/fluid/operators/strided_memcpy.h ✅	4	困难	#50346	@huangjiyi
56	paddle/fluid/operators/top_k_function_cuda.h ✅ (2022/11/28)	4	中等	#48251	@Asthestarsfalll
57	paddle/fluid/operators/transpose_op.cu.h ✅ (2022/11/30)	1	中等	#48286	@Patrick-Star125
58	paddle/fluid/operators/utils.h	5	中等	#50624	@RedContritio
59	paddle/fluid/platform/bfloat16.h ✅ (2022/11/11)	4	简单	#47831	@AndPuQing
60	paddle/fluid/platform/collective_helper.h	3	中等	#51651	@YuanRisheng
61	paddle/fluid/platform/complex.h ✅ (2022/11/15)	4	简单	#47926	@huangjiyi
62	paddle/fluid/platform/cpu_info.h✅ (2022/11/28)	5	中等	#48403	@AndPuQing
63	paddle/fluid/platform/cuda_graph_with_memory_pool.h ✅ (2022/12/20)	2	中等	#49139	@huangjiyi
64	paddle/fluid/platform/cudnn_workspace_helper.h	6	中等	已完成	未知
	======= device ========
65	paddle/fluid/platform/device_context.h	24	困难	#50865 #51541 #51559	@huangjiyi
66	paddle/fluid/platform/device/custom/enforce_custom.h ✅ (2022/11/29)	1	中等	#48422	@Asthestarsfalll
67	paddle/fluid/platform/device/device_wrapper.h	4	中等	#50749	@gouzil
68	paddle/fluid/platform/device/gpu/cuda/cuda_graph.h ✅ (2022/12/08)	1	中等	#48686	@huangjiyi
69	paddle/fluid/platform/device/gpu/cuda/cudnn_helper.h ✅ (2022/11/28)	2	中等	#48415	@huangjiyi
70	paddle/fluid/platform/device/gpu/cuda/cudnn_desc.h ✅ (2022/11/28)	1	中等	#48415	@huangjiyi
71	paddle/fluid/platform/device/gpu/gpu_device_function.h ✅ (2022/11/22)	16	中等	#48097 #48117	@huangjiyi
72	paddle/fluid/platform/device/gpu/gpu_dnn.h ✅ (2022/11/28)	14	中等	#48415	@huangjiyi
73	paddle/fluid/platform/device/gpu/gpu_info.h ✅ (2022/12/8)	10	中等	#48811	@Patrick-Star125
74	paddle/fluid/platform/device/gpu/gpu_launch_config.h ✅ (2022/11/10)	14	简单	#47808	@AndPuQing
75	paddle/fluid/platform/device/gpu/gpu_primitives.h✅ (2022/11/18)	56	困难	#48015 #48063	@GreatV
76	paddle/fluid/platform/device/gpu/nccl_helper.h	4	中等	#51651	@YuanRisheng
77	paddle/fluid/platform/device/gpu/rocm/miopen_helper.h ✅ (2022/11/28)	2	中等	#48415	@huangjiyi
78	paddle/fluid/platform/device/xpu/xpu_header.h ✅ (2022/11/28)	2	中等	#48420	@huangjiyi
79	paddle/fluid/platform/device/xpu/xpu_op_list.h ✅ (2022/11/28)	1	中等	#48420	@huangjiyi
80	paddle/fluid/platform/device_memory_aligment.h ✅ (2022/12/05)	2	中等	#48694	@huangjiyi
	======= dynload ========
81	paddle/fluid/platform/dynload/cublas.h ✅ (2022/11/09)	1	简单	#47778	@huangjiyi
82	paddle/fluid/platform/dynload/cudnn.h ✅ (2022/11/09)	1	简单	#47787	@huangjiyi
83	paddle/fluid/platform/dynload/mklml.h ✅ (2022/11/09)	3	简单	#47787	@huangjiyi
84	paddle/fluid/platform/dynload/rocblas.h ✅ (2022/11/09)	1	简单	#47787	@huangjiyi
85	paddle/fluid/platform/enforce.h ✅ (2022/11/24)	24	中等	#48049	@AndPuQing
86	paddle/fluid/platform/errors.h ✅ (2022/11/10)	1	简单	#47815	@huangjiyi
87	paddle/fluid/platform/fast_divmod.h ✅ (2022/11/10)	4	简单	#47815	@huangjiyi
88	paddle/fluid/platform/flags.h ✅ (2022/12/09)	4	困难	#48696	@AndPuQing
89	paddle/fluid/platform/float16.h ✅ (2022/11/10)	5	简单	#47817	@huangjiyi
90	paddle/fluid/platform/for_range.h ✅ (2022/11/10)	6	简单	#47817	@huangjiyi
91	paddle/fluid/platform/function_traits.h ✅ (2022/11/17)	1	中等	#48065	@GreatV
92	paddle/fluid/platform/init.h	1	困难	#51511	@YuanRisheng
93	paddle/fluid/platform/mkldnn_utils.h ✅ (2022/11/9)	1	中等	#46768	@jczaja
94	paddle/fluid/platform/place.h ✅ (2022/11/10)	9	简单	#47823	@huangjiyi
	======= profiler ========
95	paddle/fluid/platform/profiler.h	5	困难	#50191	@YuanRisheng
96	paddle/fluid/platform/profiler/event_tracing.h	11	困难	#50183	@YuanRisheng
97	paddle/fluid/platform/profiler/supplement_tracing.h	9	困难	#50721	@YuanRisheng
98	paddle/fluid/platform/transform.h ✅	8	中等	#50498	@huangjiyi
	======= cmake ========
99	paddle/fluid/operators/generated_op.cc ✅ (2022/11/09)	1	中等	#47714	@chenwhql
100	paddle/fluid/operators/generated_sparse_op.cc ✅ (2022/11/09)	1	中等	#47714	@chenwhql
101	paddle/fluid/operators/ops_extra_info.cc ✅ (2022/11/09)	1	中等	#47714	@chenwhql
	======= others ========
102	paddle/fluid/platform/macros.h ✅ (2022/11/21)	1	简单	#48069	@AndPuQing
103	paddle/fluid/platform/device/gpu/rocm/miopen_desc.h✅ (2022/11/28)	1	中等	#48415	@huangjiyi

PHI 独立编译

当头文件的依赖清理完毕，我们可以修改当前 Paddle 的编译方式，将 PHI 作为独立的 Library 对象进行编译，这需要对于编译原理和 CMake 语法比较熟悉的同学先设计编译改动方案，经过飞桨研发工程师和社区开发者一起评审后再推进执行。

[paddle-bot]

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

[GreatV]

@luotao1 @chenwhql 麻烦看一下是否有遗漏或者分解地太细了。

[luotao1]

头文件依赖解耦具体说明 （ 整理 by @GreatV ）

首先，paddle/phi 目录下仍然有许多文件在 include paddle/fluid 目录下的文件。

基于本文撰写时 Paddle repo 的最新 commit d5e7d20 在 paddle/phi 下搜索 #include "paddle/fluid 并去掉 tests 子目录，可以匹配到 355 个文件中包含 583 个结果，首要任务就是要将这 583 个 include 去掉。

使用如下命令查找include的结果，共583个结果：

grep -In -r --exclude-dir=tests "paddle/fluid/" paddle/phi

paddle/phi/core/kernel_factory.cc:20:#include "paddle/fluid/platform/device/xpu/xpu_op_list.h"
paddle/phi/core/string_tensor.cc:17:#include "paddle/fluid/memory/malloc.h"
paddle/phi/core/tensor_utils.cc:23:#include "paddle/fluid/memory/malloc.h"
paddle/phi/core/tensor_utils.cc:24:#include "paddle/fluid/memory/memcpy.h"
paddle/phi/core/tensor_utils.cc:25:#include "paddle/fluid/platform/device_context.h"
paddle/phi/core/cuda_stream.h:32:#include "paddle/fluid/platform/enforce.h"
paddle/phi/core/dense_tensor_impl.cc:15:#include "paddle/fluid/memory/malloc.h"
paddle/phi/core/dense_tensor_impl.cc:23:#include "paddle/fluid/platform/mkldnn_utils.h"
paddle/phi/core/serialization.cc:21:#include "paddle/fluid/framework/tensor_util.h"
paddle/phi/core/serialization.cc:22:#include "paddle/fluid/framework/version.h"
paddle/phi/core/dense_tensor.cc:40:#include "paddle/fluid/memory/malloc.h"
paddle/phi/core/device_context.cc:18:#include "paddle/fluid/platform/device/gpu/cuda/cuda_graph.h"
paddle/phi/core/selected_rows_impl.cc:20:#include "paddle/fluid/memory/memcpy.h"
paddle/phi/api/lib/backward_api.cc.tmp:16:#include "paddle/fluid/eager/api/utils/global_utils.h"
paddle/phi/api/lib/backward_api.cc.tmp:17:#include "paddle/fluid/platform/profiler/event_tracing.h"
paddle/phi/api/lib/backward_api.cc.tmp:18:#include "paddle/fluid/platform/profiler/supplement_tracing.h"
paddle/phi/api/lib/context_pool.cc:20:#include "paddle/fluid/platform/init.h"
paddle/phi/api/lib/api.cc:18:#include "paddle/fluid/platform/profiler/event_tracing.h"
paddle/phi/api/lib/api.cc:19:#include "paddle/fluid/platform/profiler/supplement_tracing.h"
paddle/phi/api/lib/op_meta_info.cc:22:#include "paddle/fluid/framework/custom_operator.h"
... 

使用如下命令查找匹配到的文件，共355个结果：

grep -Il -r --exclude-dir=tests "paddle/fluid/" paddle/phi

paddle/phi/core/kernel_factory.cc
paddle/phi/core/string_tensor.cc
paddle/phi/core/tensor_utils.cc
paddle/phi/core/cuda_stream.h
paddle/phi/core/dense_tensor_impl.cc
paddle/phi/core/serialization.cc
paddle/phi/core/dense_tensor.cc
paddle/phi/core/device_context.cc
paddle/phi/core/selected_rows_impl.cc
paddle/phi/api/lib/backward_api.cc.tmp
paddle/phi/api/lib/context_pool.cc
paddle/phi/api/lib/api.cc
paddle/phi/api/lib/op_meta_info.cc
paddle/phi/api/lib/CMakeLists.txt
paddle/phi/api/lib/dygraph_api.cc.tmp
paddle/phi/api/lib/dygraph_api.cc
paddle/phi/api/lib/utils/tensor_utils.h
paddle/phi/api/lib/utils/allocator.h
...

对包含的头文件去重，发现共有100个左右不同的头文件：

grep -Irn --exclude-dir=tests "paddle/fluid/*" paddle/phi | grep "paddle/fluid" | grep -v party | awk '{print $NF}' | sort | uniq -c | sort -nr

56 "paddle/fluid/platform/device/gpu/gpu_primitives.h"
47 "paddle/fluid/memory/memcpy.h"
43 "paddle/fluid/framework/tensor_util.h"
24 "paddle/fluid/platform/enforce.h"
23 "paddle/fluid/platform/device_context.h"
18 "paddle/fluid/memory/memory.h"
18 "paddle/fluid/memory/malloc.h"
17 "paddle/fluid/framework/convert_utils.h"
16 "paddle/fluid/platform/device/gpu/gpu_device_function.h"
14 "paddle/fluid/platform/device/gpu/gpu_launch_config.h"
14 "paddle/fluid/platform/device/gpu/gpu_dnn.h"
11 "paddle/fluid/platform/profiler/event_tracing.h"
10 "paddle/fluid/platform/device/gpu/gpu_info.h"
9 "paddle/fluid/platform/profiler/supplement_tracing.h"

...

具体的头文件可总结如下：

难度标记说明：结合任务说明中对于三类不同情况的介绍，根据所依赖模块的解耦复杂程度，将下列每个头文件所对应的解耦工作进行难度标记，分为三个等级：“简单”， “中等”，“困难（复杂）”，这里概要介绍一下，详细可以参考任务说明：

• 简单：具体情况包括，1）头文件无使用可以直接移除，2）头文件在phi下有可替换的头文件
• 中等：具体情况包括，1）需要先将部分函数迁移至phi目录，再更换头文件；2. 需要将部分phi中代码移回fluid
• 困难：具体情况包括：1）现阶段模块依赖复杂，解耦需要在phi目录下先设计实现相应子功能或接口用于替换；2）依赖的模块过于复杂，整体迁移成本很高

注：人工标记，可能存在一定误判，大家可以根据情况更正

[luotao1]

头文件依赖解耦已全部完成，感谢参与的小伙伴们！

@GreatV (8), @jzhang533 (1), @chenwhql (6), @huangjiyi (53), @AndPuQing (7), @Silv3S (1), @jczaja (1)， @Asthestarsfalll (2) @Patrick-Star125 (5) @DrRyanHuang (1),@engineer1109(3), @RedContritio (3) @YuanRisheng (8) @gouzil（2）@jinyouzhi(1) @jiahy0825(1)

结项小结等4月底完成 PHI 独立编译后发出。