New ir support legacy kernel instruction #55880

phlrain · 2023-08-01T09:28:49Z

PR types

New features

PR changes

Others

Description

在new ir executor，BetaRun（已是默认接口）流程中支持fluid op

Others

Pcard-67164

… new_ir_remove_fetch_list

… new_ir_support_legacy_kernel_instruction

paddle-bot · 2023-08-01T09:28:54Z

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.

paddle-bot · 2023-08-01T09:28:56Z

❌ The PR is not created using PR's template. You can refer to this Demo.
Please use PR's template, it helps save our maintainers' time so that more developers get helped.

… new_ir_remove_fetch_list

… new_ir_support_legacy_kernel_instruction

…e/Paddle into new_ir_support_legacy_kernel_instruction

… new_ir_support_legacy_kernel_instruction

zhangbo9674 · 2023-08-07T06:37:09Z

paddle/fluid/framework/new_executor/instruction/legacy_kernel_instruction.h

+namespace paddle {
+namespace framework {
+class Scope;
+class Value;


Value 属于 ir namespace

zhangbo9674 · 2023-08-07T06:37:48Z

paddle/fluid/framework/new_executor/instruction/instruction_base.h

@@ -21,6 +21,8 @@

 #include "paddle/fluid/framework/new_executor/new_executor_defs.h"
 #include "paddle/fluid/platform/event.h"
+#include "paddle/ir/core/operation.h"


前置声明 Operation 是不是就可以

zhangbo9674 · 2023-08-07T07:34:19Z

paddle/fluid/framework/new_executor/instruction/legacy_kernel_instruction.h

+
+  phi::InferMetaContext infer_meta_context_;
+
+  paddle::framework::ExecutionContext* kernel_context_{nullptr};


只不是直接叫 execution_context 更加直观，以此和 phi 的 kernel_context 进行区别

zhangbo9674 · 2023-08-07T07:34:28Z

paddle/fluid/framework/new_executor/instruction/legacy_kernel_instruction.h

+  std::shared_ptr<framework::RuntimeContext> runtime_context_;
+  std::shared_ptr<paddle::framework::OperatorBase> operator_base_;
+
+  phi::Kernel* phi_kernel_{nullptr};  // not owned


fluid 算子也需要 phi_kernel ？么

Aurelius84

LGTM overall，comment可以单独提PR fix

Aurelius84 · 2023-08-08T07:13:16Z

paddle/fluid/framework/new_executor/instruction/instruction_base.h

@@ -21,6 +21,7 @@

 #include "paddle/fluid/framework/new_executor/new_executor_defs.h"
 #include "paddle/fluid/platform/event.h"
+#include "paddle/ir/core/value.h"


这里其实不需要包含value.h头文件，Value已经有前置声明

Aurelius84 · 2023-08-08T07:14:38Z

paddle/fluid/framework/new_executor/instruction/instruction_util.cc

+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/framework/new_executor/instruction/instruction_util.h"


从规范上，instruction_util.h 要放在第一个来include

另外文件名是否直接 utils.h 就可以，不需要再加instruction前缀了，已经很长了~~

Aurelius84 · 2023-08-08T07:16:07Z

paddle/fluid/framework/new_executor/instruction/instruction_util.cc

+    const std::unordered_map<const paddle::framework::Variable*, std::string>&
+        variable_2_var_name) {
+  std::vector<int> ids;
+  std::string var_name = value_2_var_name.at(value);


Suggested change

std::string var_name = value_2_var_name.at(value);

auto& var_name = value_2_var_name.at(value);

Aurelius84 · 2023-08-08T07:21:08Z

paddle/fluid/framework/new_executor/instruction/instruction_util.cc

+  // computing. They execute serially in device thread and block CUDA kernel
+  // launching in other GPU OPs. To improve performance, set them as kGpuSync
+  // and so that they would be dispatched to host thread.
+  auto op_attributes = op->attributes();


Suggested change

auto op_attributes = op->attributes();

auto& op_attributes = op->attributes();

消除一次隐式copy 构造

Aurelius84 · 2023-08-08T07:21:33Z

paddle/fluid/framework/new_executor/instruction/instruction_util.cc

+    const platform::Place& place,
+    const std::string& execution_stream,
+    const int stream_priority) {
+  auto op_attributes = op->attributes();


Suggested change

auto op_attributes = op->attributes();

auto& op_attributes = op->attributes();

auto 与 auto& 是有区别的。

Aurelius84 · 2023-08-08T07:26:25Z

paddle/fluid/framework/new_executor/instruction/legacy_kernel_instruction.cc

+    const std::unordered_map<const paddle::framework::Variable*, std::string>&
+        variable_2_var_name)
+    : InstructionBase(id, place) {
+  auto op_attributes = op->attributes();


Suggested change

auto op_attributes = op->attributes();

auto& op_attributes = op->attributes();

Aurelius84 · 2023-08-08T07:27:43Z

paddle/fluid/framework/new_executor/instruction/legacy_kernel_instruction.cc

+                        .data();
+  auto kernel_result = phi::KernelFactory::Instance().SelectKernelOrThrowError(
+      kernel_name, kernel_key);
+  phi_kernel_ = new phi::Kernel(kernel_result.kernel);


析构函数里也要加上 delete phi_kernel_

Aurelius84 · 2023-08-08T07:28:50Z

paddle/fluid/framework/new_executor/instruction/phi_kernel_instruction.h

@@ -73,6 +63,8 @@ class PhiKernelInstruction : public InstructionBase {
  phi::KernelContext kernel_context_;

  phi::Kernel* phi_kernel_{nullptr};  // not owned
+
+  std::string phi_op_name_;


这里加了这个成员，似乎没有处理它的构造？

Aurelius84 · 2023-08-08T07:29:45Z

paddle/fluid/framework/new_executor/new_ir_interpreter.cc

-                                                 variable_2_var_name_));
+
+      if (op_name == "pd.fused_softmax_mask_upper_triangle" ||
+          op_name == "pd.fused_softmax_mask_upper_triangle_grad") {


后续这里得单独抽离放到一个函数里，作为单独的分支。下面有定义 LegacyOpList，其实可以复用？

phlrain added 15 commits July 28, 2023 13:34

new ir remove fetch list

06d6ede

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

f4d5f2b

… new_ir_remove_fetch_list

fix pattern rewrite bug

774196e

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

d2f5ac7

… new_ir_remove_fetch_list

try to remove constant fold

7e60294

revert code

4b38bad

add pattern rewrite test flag

354b1f9

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

156d46e

… new_ir_remove_fetch_list

fix multi fetch

edf3ce2

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

b070b68

… new_ir_remove_fetch_list

remove usless code

c7206c1

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

aa3af1a

… new_ir_remove_fetch_list

new ir support legacy kernel instraction

b80e3d4

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

da6b3d4

… new_ir_support_legacy_kernel_instruction

new ir support legacy kernel instruction

755eace

phlrain added 13 commits August 1, 2023 10:23

add scope prefix

adf66ea

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

95c53f3

… new_ir_remove_fetch_list

update

e8f64bc

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

f770a71

… new_ir_support_legacy_kernel_instruction

Merge commit 'refs/pull/55796/head' of https://github.com/PaddlePaddl…

2f8760d

…e/Paddle into new_ir_support_legacy_kernel_instruction

update

e97beed

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

ee3f22d

… new_ir_support_legacy_kernel_instruction

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

02ea55a

… new_ir_support_legacy_kernel_instruction

update

3eeec33

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

8580a9d

… new_ir_support_legacy_kernel_instruction

update

cdf32c7

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

b117bd5

… new_ir_support_legacy_kernel_instruction

fix

79a57bf

phlrain added 5 commits August 6, 2023 03:37

revert channel shuffl test

33947b0

polish code

14c4e9b

try to fix windows compile error

55f6306

polish code

a17ab14

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

54546df

… new_ir_support_legacy_kernel_instruction

zhangbo9674 reviewed Aug 7, 2023

View reviewed changes

phlrain added 3 commits August 7, 2023 07:48

update

1f713ed

update

ca68a2b

revert op test

8ec5558

zhangbo9674 approved these changes Aug 8, 2023

View reviewed changes

Aurelius84 approved these changes Aug 8, 2023

View reviewed changes

phlrain merged commit f9c2f4c into PaddlePaddle:develop Aug 8, 2023
27 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

New ir support legacy kernel instruction #55880

New ir support legacy kernel instruction #55880

phlrain commented Aug 1, 2023 •

edited

Loading

paddle-bot bot commented Aug 1, 2023

paddle-bot bot commented Aug 1, 2023

zhangbo9674 Aug 7, 2023

zhangbo9674 Aug 7, 2023

zhangbo9674 Aug 7, 2023

zhangbo9674 Aug 7, 2023

Aurelius84 left a comment

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023

Aurelius84 Aug 8, 2023


		phi::InferMetaContext infer_meta_context_;

		paddle::framework::ExecutionContext* kernel_context_{nullptr};

	std::string var_name = value_2_var_name.at(value);
	auto& var_name = value_2_var_name.at(value);

	auto op_attributes = op->attributes();
	auto& op_attributes = op->attributes();

New ir support legacy kernel instruction #55880

New ir support legacy kernel instruction #55880

Conversation

phlrain commented Aug 1, 2023 • edited Loading

PR types

PR changes

Description

Others

paddle-bot bot commented Aug 1, 2023

paddle-bot bot commented Aug 1, 2023

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

Aurelius84 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

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

phlrain commented Aug 1, 2023 •

edited

Loading