【IR】Support Sparsecootensor and Sparsecsrtensor #63015

risemeup1 · 2024-03-25T15:33:18Z

PR Category

Others

PR Types

New features

Description

此PR做了以下两件事情：

扩充了op_gen的能力，基于 sparse_ops.yaml 及 sparse_backward.yaml 自动生成 PIR 的 Sparse 算子定义
基于 PT 打通重点算子单测

paddle-bot · 2024-03-25T15:33:26Z

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 · 2024-03-25T15:33:28Z

❌ 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.

chen2016013 · 2024-03-29T08:41:44Z

paddle/fluid/pir/dialect/op_generator/api_gen.py

-                op_compat_item = op_compat_parser.get_compat(
-                    op['forward']['name']
-                )
+            for op in op_yaml_items:


Suggested change

for op in op_yaml_items:

for op in ops:

chen2016013 · 2024-03-29T08:41:58Z

paddle/fluid/pir/dialect/op_generator/api_gen.py

        for yaml_file in op_yaml_files:
+            op_yaml_items = []


Suggested change

op_yaml_items = []

paddle/fluid/pir/dialect/op_generator/api_gen.py

chen2016013 · 2024-03-29T08:57:29Z

paddle/fluid/pir/dialect/op_generator/op_gen.py

@@ -1367,20 +1382,56 @@ def AutoCodeGen(
                # =================================== #
                #      gen interface list str         #
                # =================================== #
+                op_class_name_suffix = 'Sp' if op_info.is_sparse_op else ''


这部分逻辑出现了2遍，考虑封装成一个函数

chen2016013 · 2024-03-29T09:01:12Z

paddle/fluid/pir/dialect/op_generator/op_gen.py

-                            op_info_items[op_info.op_phi_name[0]],
-                            all_op_info_items[op_info.backward_name],
+                            op_info_items[
+                                op_info.op_phi_name[0] + sparse_op_name_suffix


这里建议封装成encode/decode形式，并添加注释，便于理解

paddle-ci-bot · 2024-04-06T03:15:38Z

Sorry to inform you that b2c5d88's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

…e/Paddle into support_sparse_op_in_pir

… support_sparse_op_in_pir

chen2016013

LGTM

Aurelius84 · 2024-04-16T06:55:06Z

paddle/fluid/ir_adaptor/translator/op_translator.cc

@@ -141,6 +142,14 @@ inline std::string OpNameCompatibleMapping(std::string op_name) {
  auto& op_normalizer = OpNameNormalizer::instance();
  return op_normalizer[op_name];
 }
+inline bool isSparseString(const std::string& str) {
+  std::regex pattern("^sparse_[a-zA-Z0-9_]+$");
+  if (std::regex_match(str, pattern)) {


Suggested change

if (std::regex_match(str, pattern)) {

return std::regex_match(str, pattern);

这样会更简洁？

Aurelius84 · 2024-04-16T06:56:39Z

paddle/fluid/pir/dialect/op_generator/api_gen.py

@@ -662,6 +665,8 @@ def _gen_amp_logic(self, op_info, op_name, is_mutable_attr):
            return 'VLOG(5) << " No AMP for grad apis. ";'
        if op_name.endswith('_') or op_name == 'cast':
            return f'VLOG(5) << "No AMP for {op_name} because it is a inplace or cast api.";'
+        if op_info.is_sparse_op:
+            op_name += "sp_" if op_name[-1] == "_" else "_sp"


类似这样的专有后缀，建议在文件头部通过 kSparseOpPrefix 或者 kSparseOpSuffix来定义；

Aurelius84 · 2024-04-16T06:58:45Z

paddle/fluid/pir/dialect/operator/ir/ir_sparse.cc

+      non_zero_elements_(non_zero_elements),
+      coalesced_(coalesced) {}
+
+IrSparseCooTensor::IrSparseCooTensor(const IrSparseCooTensor& other) {


IrSparseCooTensor(const IrSparseCooTensor& other) = default;

是不是就可以？可以复用下C++的一些默认构造的写法（前提是里面不能有指针类型成员）

Aurelius84 · 2024-04-16T06:58:56Z

paddle/fluid/pir/dialect/operator/ir/ir_sparse.cc

+  coalesced_ = other.coalesced();
+}
+
+IrSparseCooTensor& IrSparseCooTensor::operator=(


同上，可以了解下C++的编译器默认生成的系列构造

zhangbo9674 · 2024-04-16T09:32:34Z

paddle/fluid/pir/dialect/operator/ir/ir_sparse.h

@@ -0,0 +1,209 @@
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.


Suggest change file name from ir_sparse to ir_sparse_tensor

zhangbo9674 · 2024-04-16T09:51:05Z

paddle/fluid/pir/dialect/op_generator/api_gen.py

@@ -383,6 +381,8 @@ def _gen_ret_type(self, op_info):
    def _gen_one_declare(
        self, op_info, op_name, is_mutable_attr, is_vector_mutable_attr
    ):
+        if op_info.is_sparse_op:
+            op_name += "sp_" if op_name[-1] == "_" else "_sp"


建议直接将带有_sp 后缀的名字作为OpInfo的一个属性

… support_sparse_op_in_pir

chen2016013 · 2024-04-26T06:15:57Z

paddle/fluid/framework/new_executor/garbage_collector/event_garbage_collector.cc

@@ -104,7 +104,18 @@ void InterpreterCoreEventGarbageCollector::Add(
            ->MoveMemoryHolder(),
        event,
        ctx);
-    var->GetMutable<phi::SelectedRows>()->mutable_rows()->clear();
+  } else if (var->IsType<phi::SparseCooTensor>()) {


同目录下no_event_garbage_collector.cc也需要补充SparseCooTensor相关逻辑

chen2016013 · 2024-04-26T06:22:50Z

paddle/fluid/ir_adaptor/translator/op_translator.cc

+    return op_info;
+  }
+  if (paddle::dialect::HaveOpToMultiKernelsMap(
+          OpNameCompatibleMapping(op_desc.Type())) &&


这部分处理sparse多kernel的逻辑和下面处理其他op多kernel的逻辑有很多重合，最好把重复部分抽出来形成一个函数或者合并一下，否则难以维护

zyfncg

LGTM for op_compat

vivienfanghuagood · 2024-04-26T12:22:37Z

cmake/external/eigen.cmake

@@ -56,14 +56,17 @@ endif()

 set(EIGEN_INCLUDE_DIR ${SOURCE_DIR})
 include_directories(${EIGEN_INCLUDE_DIR})
-
+if(NOT WIN32)
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=maybe-uninitialized")


为何加了这个？

* test * support_sparse_op_in_pir * support_sparse_op_in_pir * support suparse op in pir * test * fix * fix * support sparse op in pir * fix * fix * fix * fix * support translator of sparse tests * support translator of sparse tests * support translator of sparse tests * support translator of sparse tests * fix * fix * fix * fix

risemeup1 force-pushed the support_sparse_op_in_pir branch from 40a8c0f to 168b3eb Compare March 28, 2024 01:36

risemeup1 changed the title ~~test~~ 【IR】Generate sparse op from sparse_ops.yaml Mar 28, 2024

test

54b5b51

risemeup1 force-pushed the support_sparse_op_in_pir branch from 168b3eb to 54b5b51 Compare March 28, 2024 07:06

risemeup1 added 3 commits March 29, 2024 08:18

support_sparse_op_in_pir

ce6fab2

support_sparse_op_in_pir

b2c5d88

support suparse op in pir

873ca79

chen2016013 reviewed Mar 29, 2024

View reviewed changes

risemeup1 added 10 commits April 7, 2024 13:39

test

7a59131

fix

41e8a47

fix

52757c1

fix

a078561

Merge commit 'refs/pull/63015/head' of https://github.com/PaddlePaddl…

3c0cf6f

…e/Paddle into support_sparse_op_in_pir

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

f155e7f

… support_sparse_op_in_pir

support sparse op in pir

27f202f

fix

efaaad1

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

e4f2bb7

… support_sparse_op_in_pir

fix

7c70f62

chen2016013 previously approved these changes Apr 16, 2024

View reviewed changes

Aurelius84 reviewed Apr 16, 2024

View reviewed changes

zhangbo9674 reviewed Apr 16, 2024

View reviewed changes

fix

739e53c

risemeup1 dismissed chen2016013’s stale review via 739e53c April 23, 2024 06:40

risemeup1 added 3 commits April 24, 2024 07:34

fix

2620263

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

e893c3e

… support_sparse_op_in_pir

fix

7e356b8

risemeup1 added 4 commits April 25, 2024 08:04

support translator of sparse tests

ebb4e6d

support translator of sparse tests

796dc38

support translator of sparse tests

0e60f57

support translator of sparse tests

72d75ec

risemeup1 changed the title ~~【IR】Generate sparse op from sparse_ops.yaml~~ 【IR】Support SparseCooTensor and SparseCsrTensor Apr 25, 2024

risemeup1 changed the title ~~【IR】Support SparseCooTensor and SparseCsrTensor~~ 【IR】Support Sparsecootensor and Sparsecsrtensor Apr 25, 2024

risemeup1 added 4 commits April 25, 2024 09:02

support translator of sparse tests

464ed57

fix

bb58a9d

fix

05c3ef3

fix

f71526e

chen2016013 reviewed Apr 26, 2024

View reviewed changes

fix

0ae9778

zhangbo9674 approved these changes Apr 26, 2024

View reviewed changes

zyfncg approved these changes Apr 26, 2024

View reviewed changes

vivienfanghuagood reviewed Apr 26, 2024

View reviewed changes

risemeup1 merged commit fb3d141 into PaddlePaddle:develop Apr 26, 2024

	if (std::regex_match(str, pattern)) {
	return std::regex_match(str, pattern);

		@@ -0,0 +1,209 @@
		// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.

【IR】Support Sparsecootensor and Sparsecsrtensor #63015

【IR】Support Sparsecootensor and Sparsecsrtensor #63015

Uh oh!

Conversation

risemeup1 commented Mar 25, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Uh oh!

paddle-bot bot commented Mar 25, 2024

Uh oh!

paddle-bot bot commented Mar 25, 2024

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!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

paddle-ci-bot bot commented Apr 6, 2024

Uh oh!

chen2016013 left a comment

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!

zyfncg left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

risemeup1 commented Mar 25, 2024 •

edited

Loading