【Hackathon 5th No.104】move fusion_repeated_fc_relu/fusion_squared_mat_sub to phi #58300

zeroRains · 2023-10-21T14:32:57Z

PR types

Others

PR changes

Others

Description

move fusion_repeated_fc_relu/fusion_squared_mat_sub to phi
#57262

paddle-bot · 2023-10-21T14:33:02Z

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.

zeroRains · 2023-10-21T14:40:30Z

scaled_dp_attention_functor.h的L330-L333

请问，上面的链接对应的代码部分会因为softmax_sum_max和update_out_blk是定义在#if defined(__AVX512F__)中的函数，导致编译失败，这种应该怎么处理

yuanlehome · 2023-10-26T05:19:53Z

这个是当时在fluid下添加这个算子的PR #54139 ，看起来由于这个条件满足才会去编这个算子

在phi下可能也需要类似的控制一下，具体怎么做我先了解一下，晚点回复你～

yuanlehome · 2023-10-26T07:18:06Z

scaled_dp_attention_functor.h的L330-L333

请问，上面的链接对应的代码部分会因为softmax_sum_max和update_out_blk是定义在#if defined(__AVX512F__)中的函数，导致编译失败，这种应该怎么处理

这样写试试

if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "cpu/self_dp_attention_kernel.cc")
endif()

zeroRains · 2023-10-26T11:35:00Z

scaled_dp_attention_functor.h的L330-L333
请问，上面的链接对应的代码部分会因为softmax_sum_max和update_out_blk是定义在#if defined(__AVX512F__)中的函数，导致编译失败，这种应该怎么处理

这样写试试
if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "cpu/self_dp_attention_kernel.cc")
endif()

出问题的部分

额，已添加(已执行make clean+cmake)，但还是说was not declared in this scope

yuanlehome · 2023-10-26T13:11:25Z

scaled_dp_attention_functor.h的L330-L333
请问，上面的链接对应的代码部分会因为softmax_sum_max和update_out_blk是定义在#if defined(__AVX512F__)中的函数，导致编译失败，这种应该怎么处理

这样写试试
if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "cpu/self_dp_attention_kernel.cc")
endif()
出问题的部分

额，已添加(已执行make clean+cmake)，但还是说was not declared in this scope

这样写～

if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")
endif()

zeroRains · 2023-10-26T13:46:43Z

scaled_dp_attention_functor.h的L330-L333
请问，上面的链接对应的代码部分会因为softmax_sum_max和update_out_blk是定义在#if defined(__AVX512F__)中的函数，导致编译失败，这种应该怎么处理

这样写试试
if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "cpu/self_dp_attention_kernel.cc")
endif()
出问题的部分
额，已添加(已执行make clean+cmake)，但还是说was not declared in this scope
这样写～
if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")
endif()

也不行，(:з」∠)

yuanlehome · 2023-10-27T03:00:54Z

在remove item前后打印一下 kernel_cc，看一下是否真的被移出去了

zeroRains · 2023-10-27T03:16:03Z

remove item前后打印一下 kernel_cc，看一下是否真的被移出去了

他编译的时候走的是if分支，所以没有移出去

yuanlehome · 2023-10-27T03:27:33Z

参考下这个，设置一下编译属性

zeroRains · 2023-10-27T05:46:52Z

参考下这个，设置一下编译属性

。。。应该没写错吧，好像也不太行。。。

yuanlehome · 2023-10-27T05:49:46Z

写在if分支里。。

if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")
endif()

不过我估计也不行

zeroRains · 2023-10-27T06:38:29Z

写在if分支里。。

if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")
endif()

不过我估计也不行

确实不行

paddle-ci-bot · 2023-10-29T03:11:24Z

Sorry to inform you that bfe2cae's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

zeroRains · 2023-10-30T12:31:28Z

写在if分支里。。

if(WITH_AVX
   AND AVX512F_FOUND
   AND AVX512F_FLAG
   AND WITH_MKL)
  # nothing
else()
  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")
endif()

不过我估计也不行

确实不行

好像写在phi目录下的CMakeList中就可以编译通过了

yuanlehome · 2023-10-30T13:34:22Z

paddle/phi/CMakeLists.txt

+    kernels/fusion/cpu/self_dp_attention_kernel.cc
+    PROPERTIES COMPILE_FLAGS "-Wno-maybe-uninitialized  -mfma ${AVX512F_FLAG}")
+else()
+  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")


这一行需要写到paddle/phi/kernels目录下，不然"fusion/cpu/self_dp_attention_kernel.cc"路径不对

是否可以将路径改成“"kernels/fusion/cpu/self_dp_attention_kernel.cc"”

所以REMOVE_ITEM操作和set_source_files_properties分开写在不同的目录下

EMMM，我的意思是在paddle/phi的CMakeLists中改成这样，是否可行

yuanlehome · 2023-10-30T13:34:54Z

paddle/phi/CMakeLists.txt

+   AND AVX512F_FOUND
+   AND AVX512F_FLAG
+   AND WITH_MKL)
+  set_source_files_properties(


kernels/funcs/scaled_dp_attention_functor.h文件就不需要set_source_files_properties了吧，.cc才是编译单元

好的我去除一下

yuanlehome · 2023-10-30T13:36:14Z

paddle/fluid/operators/fused/CMakeLists.txt

-  set_target_properties(
-    self_dp_attention_op
-    PROPERTIES COMPILE_FLAGS "-Wno-maybe-uninitialized -mfma ${AVX512F_FLAG}")
+


这个if分支已经空了，为了不删掉？

忘记了，我改一下

yuanlehome · 2023-10-30T13:53:45Z

paddle/phi/CMakeLists.txt

  set_source_files_properties(
    kernels/fusion/cpu/self_dp_attention_kernel.cc
    PROPERTIES COMPILE_FLAGS "-Wno-maybe-uninitialized  -mfma ${AVX512F_FLAG}")
 else()
-  list(REMOVE_ITEM kernel_cc "fusion/cpu/self_dp_attention_kernel.cc")
+  list(REMOVE_ITEM kernel_cc "kernels/fusion/cpu/self_dp_attention_kernel.cc")


把这个else分支删掉，把if判断取反，把REMOVE_ITEM写在kernel/CMakeLists.txt里。kernel_cc不在当前这个文件里。。。

哦哦，明白了

zeroRains · 2023-11-06T06:01:36Z

你能把self_dp_attention的修改拆出来单独提交一个PR吗？另外两个op的修改还放在这个PR～

好吧

zeroRains · 2023-11-06T08:09:16Z

你能把self_dp_attention的修改拆出来单独提交一个PR吗？另外两个op的修改还放在这个PR～

这是self_dp_attention的新pr链接#58715
晚点我再把self_dp_attention的修改从这个pr里移除

yuanlehome

LGTM

zyfncg · 2023-11-07T06:42:57Z

paddle/phi/api/yaml/fused_ops.yaml

+  infer_meta :
+    func : FusionRepeatedFCReluInferMeta
+  kernel :
+    func : fusion_repeated_fc_relu


需要对齐原来的逻辑，指定kernel的data_type 为 x

zyfncg · 2023-11-07T06:43:08Z

paddle/phi/api/yaml/fused_ops.yaml

+  infer_meta :
+    func : FusionSquaredMatSubInferMeta
+  kernel :
+    func : fusion_squared_mat_sub


zyfncg · 2023-11-07T06:44:37Z

paddle/phi/infermeta/fusion.cc

+  out->set_dims({i_dims[0], w_dims[sz - 1][1]});
+  out->share_lod(x);


dtype 也设置下

zyfncg · 2023-11-07T06:44:59Z

paddle/phi/infermeta/fusion.cc

+  squared_x->set_dims(x_dims);
+  squared_y->set_dims(y_dims);
+  squared_xy->set_dims({x_dims[0], y_dims[1]});
+  out->set_dims({x_dims[0], y_dims[1]});


设置 dtype

CLAassistant · 2023-11-07T07:09:23Z

All committers have signed the CLA.

rollback add the type Update fusion.cc

yuanlehome

LGTM

phlrain

LGTM for check_dygraph

XiaoguangHu01

LGTM

sunzhongkai588

LGTM, no docs changes

yuanlehome · 2023-11-09T07:14:48Z

这个冲突了，麻烦处理下再合入吧，另一个已经合入了

…_sub to phi (PaddlePaddle#58300)

move self_dp_attention_op to phi, but have some bug in __AVX512F__

bfe2cae

paddle-bot bot added the contributor External developers label Oct 21, 2023

zeroRains added 2 commits October 22, 2023 03:56

move fusion_repeated_fc_relu to phi without make

88428d0

move fusion_squared_mat_sub_op to phi without make

ab3fb95

luotao1 mentioned this pull request Oct 24, 2023

【PaddlePaddle Hackathon 5th】开源贡献个人挑战赛 #57262

Open

zeroRains added 3 commits October 30, 2023 12:42

move self_dp_attention_op to phi

5ac78b2

fix the test param

5602523

fix conflict

69df2f7

yuanlehome reviewed Oct 30, 2023

View reviewed changes

zeroRains force-pushed the 104 branch from b8dcfcf to 69df2f7 Compare October 30, 2023 13:35

yuanlehome reviewed Oct 30, 2023

View reviewed changes

zeroRains added 2 commits October 30, 2023 13:48

move some useless code

5ad60ed

remove some useless code

b807aa5

yuanlehome reviewed Oct 30, 2023

View reviewed changes

zeroRains changed the title ~~【Hackathon 5th No.104】move self_dp_attention/fusion_repeated_fc_relu/fusion_squared_mat_sub to phi~~ 【Hackathon 5th No.104】move fusion_repeated_fc_relu/fusion_squared_mat_sub to phi Nov 6, 2023

zeroRains added 2 commits November 6, 2023 08:14

fix the conflict

464af3c

reoback self_dp_attention

eb0d2c9

zeroRains force-pushed the 104 branch from 649b518 to eb0d2c9 Compare November 6, 2023 08:58

change name

d1870a1

yuanlehome previously approved these changes Nov 7, 2023

View reviewed changes

zyfncg reviewed Nov 7, 2023

View reviewed changes

zeroRains dismissed yuanlehome’s stale review via f916a07 November 7, 2023 07:09

squash the wrong commit

f9653a6

rollback add the type Update fusion.cc

zeroRains force-pushed the 104 branch from b7a883b to f9653a6 Compare November 8, 2023 02:43

zyfncg previously approved these changes Nov 9, 2023

View reviewed changes

yuanlehome previously approved these changes Nov 9, 2023

View reviewed changes

phlrain previously approved these changes Nov 9, 2023

View reviewed changes

XiaoguangHu01 previously approved these changes Nov 9, 2023

View reviewed changes

sunzhongkai588 previously approved these changes Nov 9, 2023

View reviewed changes

Merge branch 'develop' into 104

bf395f1

zeroRains dismissed stale reviews from sunzhongkai588, XiaoguangHu01, phlrain, yuanlehome, and zyfncg via bf395f1 November 9, 2023 07:18

luotao1 approved these changes Nov 10, 2023

View reviewed changes

luotao1 merged commit 5fed7ad into PaddlePaddle:develop Nov 10, 2023

danleifeng pushed a commit to danleifeng/Paddle that referenced this pull request Nov 14, 2023

【Hackathon 5th No.104】move fusion_repeated_fc_relu/fusion_squared_mat…

b1c63c7

…_sub to phi (PaddlePaddle#58300)

zeroRains deleted the 104 branch November 24, 2023 14:33

SecretXV pushed a commit to SecretXV/Paddle that referenced this pull request Nov 28, 2023

【Hackathon 5th No.104】move fusion_repeated_fc_relu/fusion_squared_mat…

ab5c642

…_sub to phi (PaddlePaddle#58300)

		out->set_dims({i_dims[0], w_dims[sz - 1][1]});
		out->share_lod(x);

【Hackathon 5th No.104】move fusion_repeated_fc_relu/fusion_squared_mat_sub to phi #58300

【Hackathon 5th No.104】move fusion_repeated_fc_relu/fusion_squared_mat_sub to phi #58300

Conversation

zeroRains commented Oct 21, 2023 • edited Loading

PR types

PR changes

Description

paddle-bot bot commented Oct 21, 2023

zeroRains commented Oct 21, 2023

yuanlehome commented Oct 26, 2023

yuanlehome commented Oct 26, 2023 • edited Loading

zeroRains commented Oct 26, 2023 • edited Loading

yuanlehome commented Oct 26, 2023 • edited Loading

zeroRains commented Oct 26, 2023

yuanlehome commented Oct 27, 2023

zeroRains commented Oct 27, 2023

yuanlehome commented Oct 27, 2023

zeroRains commented Oct 27, 2023

yuanlehome commented Oct 27, 2023 • edited Loading

zeroRains commented Oct 27, 2023

paddle-ci-bot bot commented Oct 29, 2023

zeroRains commented Oct 30, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zeroRains Oct 30, 2023 • edited Loading

Choose a reason for hiding this comment

yuanlehome Oct 30, 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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zeroRains commented Nov 6, 2023

zeroRains commented Nov 6, 2023 • edited Loading

yuanlehome left a comment

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

Choose a reason for hiding this comment

CLAassistant commented Nov 7, 2023 • edited Loading

yuanlehome left a comment

Choose a reason for hiding this comment

phlrain left a comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

sunzhongkai588 left a comment

Choose a reason for hiding this comment

yuanlehome commented Nov 9, 2023

zeroRains commented Oct 21, 2023 •

edited

Loading

yuanlehome commented Oct 26, 2023 •

edited

Loading

zeroRains commented Oct 26, 2023 •

edited

Loading

yuanlehome commented Oct 26, 2023 •

edited

Loading

yuanlehome commented Oct 27, 2023 •

edited

Loading

zeroRains commented Oct 30, 2023 •

edited

Loading

zeroRains Oct 30, 2023 •

edited

Loading

yuanlehome Oct 30, 2023 •

edited

Loading

zeroRains commented Nov 6, 2023 •

edited

Loading

CLAassistant commented Nov 7, 2023 •

edited

Loading