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

Author: yuanlehome

README.md

@@ -15,7 +15,7 @@ English | [简体中文](./README_cn.md) | [日本語](./README_ja.md)
 Welcome to the PaddlePaddle GitHub.
 
 PaddlePaddle, as the first independent R&D deep learning platform in China, has been officially open-sourced to professional communities since 2016. It is an industrial platform with advanced technologies and rich features that cover core deep learning frameworks, basic model libraries, end-to-end development kits, tools & components as well as service platforms.
-PaddlePaddle is originated from industrial practices with dedication and commitments to industrialization. It has been widely adopted by a wide range of sectors including manufacturing, agriculture, enterprise service, and so on while serving more than 8 million developers, 220,000 companies and generating 800,000 models. With such advantages, PaddlePaddle has helped an increasing number of partners commercialize AI.
+PaddlePaddle is originated from industrial practices with dedication and commitments to industrialization. It has been widely adopted by a wide range of sectors including manufacturing, agriculture, enterprise service, and so on while serving more than 10.7 million developers, 235,000 companies and generating 860,000 models. With such advantages, PaddlePaddle has helped an increasing number of partners commercialize AI.
 
 ## Installation
 

README_cn.md

@@ -14,7 +14,7 @@
 
 欢迎来到 PaddlePaddle GitHub
 
-飞桨(PaddlePaddle)以百度多年的深度学习技术研究和业务应用为基础，是中国首个自主研发、功能完备、 开源开放的产业级深度学习平台，集深度学习核心训练和推理框架、基础模型库、端到端开发套件和丰富的工具组件于一体。目前，飞桨累计开发者800万，服务企业22万家，基于飞桨开源深度学习平台产生了80万个模型。飞桨助力开发者快速实现AI想法，快速上线AI业务。帮助越来越多的行业完成AI赋能，实现产业智能化升级。
+飞桨(PaddlePaddle)以百度多年的深度学习技术研究和业务应用为基础，是中国首个自主研发、功能完备、 开源开放的产业级深度学习平台，集深度学习核心训练和推理框架、基础模型库、端到端开发套件和丰富的工具组件于一体。目前，飞桨累计开发者1070万，服务企业23.5万家，基于飞桨开源深度学习平台产生了86万个模型。飞桨助力开发者快速实现AI想法，快速上线AI业务。帮助越来越多的行业完成AI赋能，实现产业智能化升级。
 
 ## 安装
 

README_ja.md

@@ -15,7 +15,7 @@
 PaddlePaddle GitHub へようこそ。
 
 PaddlePaddle は中国初の独立系 R&D ディープラーニングプラットフォームとして、2016年からプロのコミュニティに正式にオープンソース化されました。コアとなる深層学習フレームワーク、基本モデルライブラリ、エンドツーエンドの開発キット、ツール＆コンポーネント、さらにサービスプラットフォームを網羅する、高度な技術と豊富な機能を備えた産業プラットフォームです。
-PaddlePaddle は、工業化に対するコミットメントを持つ工業的実践から生まれたものです。製造業、農業、企業サービスなど幅広い分野で採用され、800万人以上の開発者、22万以上の企業、80万以上のモデルを生み出しています。それにより PaddlePaddle は、ますます多くのパートナーの AI 商用化を支援しています。
+PaddlePaddle は、工業化に対するコミットメントを持つ工業的実践から生まれたものです。製造業、農業、企業サービスなど幅広い分野で採用され、1070万人以上の開発者、23.5万以上の企業、86万以上のモデルを生み出しています。それにより PaddlePaddle は、ますます多くのパートナーの AI 商用化を支援しています。
 
 ## インストール
 

paddle/cinn/common/broadcast_tree.cc

@@ -185,8 +185,8 @@ using Pattern2Placement = std::unordered_map<symbol::DimExpr, symbol::DimExpr>;
 Pattern2Placement ConstructCstrLhsEqRhsReplacement(
     const symbol::Broadcastable<symbol::DimExpr>& broadcastable_condition) {
   auto [lhs, rhs] = *broadcastable_condition;
-  if (lhs.isa<std::string>()) return Pattern2Placement{{lhs, rhs}};
   if (rhs.isa<std::string>()) return Pattern2Placement{{rhs, lhs}};
+  if (lhs.isa<std::string>()) return Pattern2Placement{{lhs, rhs}};
   return Pattern2Placement{{lhs, rhs}};
 }
 
@@ -295,4 +295,54 @@ BroadcastTree ConstructBroadcastTree(const BroadcastLeaf& leaves) {
   return ConstructBroadcastBranch(broadcastable_condition.value(), leaves);
 }
 
+namespace {
+
+std::string ToTxtStringImpl(const BroadcastBranch<BroadcastTree>& branch) {
+  std::stringstream ss;
+  const auto& [cstr, lhs_eq_rhs, lhs_eq_one, rhs_eq_one] = branch.tuple();
+  const auto& [lhs, rhs] = *cstr;
+  const auto& Put = [&](const std::string& key, const auto& value) {
+    ss << "\"" << key << "\": ";
+    ss << ToTxtString(value);
+    ss << ",\n ";
+  };
+  ss << "{";
+  ss << "\"$lhs\": " << lhs << ",\n ";
+  ss << "\"$rhs\": " << rhs << ",\n ";
+  Put("$lhs == $rhs", lhs_eq_rhs);
+  Put("$lhs == 1", lhs_eq_one);
+  Put("$rhs == 1", rhs_eq_one);
+  ss << "}";
+  return ss.str();
+}
+
+std::string ToTxtStringImpl(const BroadcastLeaf& leaf) {
+  std::stringstream ss;
+  ss << "[";
+  for (const auto& dim_exprs : *leaf) {
+    ss << "[";
+    int j = 0;
+    for (const auto& dim_expr : dim_exprs) {
+      if (j++) {
+        ss << ",";
+      }
+      ss << dim_expr;
+    }
+    ss << "]";
+  }
+  ss << "]";
+  return ss.str();
+}
+
+}  // namespace
+
+std::string ToTxtString(const BroadcastTree& tree) {
+  return std::visit([&](const auto& impl) { return ToTxtStringImpl(impl); },
+                    tree.variant());
+}
+
+std::ostream& operator<<(std::ostream& os, const BroadcastTree& tree) {
+  os << ToTxtString(tree);
+}
+
 }  // namespace cinn::common

paddle/cinn/common/broadcast_tree.h

@@ -31,4 +31,8 @@ using BroadcastTree = adt::Tree<BroadcastBranch, BroadcastLeaf>;
 
 BroadcastTree ConstructBroadcastTree(const BroadcastLeaf& leaves);
 
+std::string ToTxtString(const BroadcastTree&);
+
+std::ostream& operator<<(std::ostream& os, const BroadcastTree& tree);
+
 }  // namespace cinn::common

paddle/cinn/hlir/dialect/operator/transforms/CMakeLists.txt

@@ -11,9 +11,7 @@ if(NOT CINN_ONLY)
       cinn_runtime_dialect
       pir_compiler)
 
-  cc_library(
-    cinn_transforms
-    SRCS ${cinn_transforms_srcs}
-    DEPS ${cinn_transforms_deps})
+  cinn_cc_library(cinn_transforms SRCS ${cinn_transforms_srcs} DEPS
+                  ${cinn_transforms_deps})
 
 endif()

paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.h

@@ -138,9 +138,8 @@ bool ProcessOp(paddle::dialect::ExpandOp op, pir::PatternRewriter* rewriter) {
     pir::ShapeConstraintIRAnalysis& shape_analysis =
         pir::ShapeAnalysisManager::Instance().Get(
             op.x().defining_op()->GetParentProgram());
-    CHECK(shape_analysis.value_id_to_shapeordata_.find(GetValueId(&value)) !=
-          shape_analysis.value_id_to_shapeordata_.end());
-    return shape_analysis.value_id_to_shapeordata_.at(GetValueId(&value));
+
+    return shape_analysis.GetShapeOrDataForValue(value);
   };
   std::optional<pir::Value> opt_generated_shape =
       GetOutOfRewritedGenerateShapeOp(

paddle/cinn/hlir/dialect/operator/transforms/fuse_shape_ops_into_generate_shape_op_pass.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.h"
+#include "paddle/cinn/hlir/dialect/operator/transforms/insert_broadcast_pass.h"
 
 #include "paddle/cinn/hlir/dialect/operator/ir/cinn_op.h"
 #include "paddle/cinn/hlir/framework/pir/utils.h"
@@ -49,6 +49,14 @@ bool ProcessOp(pir::Operation* op, pir::PatternRewriter* rewriter) {
   }
   pir::Value x = op->operand_source(0);
   pir::Value y = op->operand_source(1);
+  pir::ShapeConstraintIRAnalysis& shape_analysis =
+      pir::ShapeAnalysisManager::Instance().Get(op->GetParentProgram());
+  const auto& x_shape = shape_analysis.GetShapeOrDataForValue(x);
+  const auto& y_shape = shape_analysis.GetShapeOrDataForValue(y);
+  if (x_shape.shape() == y_shape.shape() && x_shape.data() == y_shape.data()) {
+    return false;
+  }
+
   pir::Value output_dim_tensor = GetOutputDimTensor(rewriter, x, y);
   {
     pir::Value broadcasted_x =
@@ -66,7 +74,7 @@ bool ProcessOp(pir::Operation* op, pir::PatternRewriter* rewriter) {
 }  // namespace
 
 template <typename OPTYPE>
-class FullyInsertBroadcastPattern : public pir::OpRewritePattern<OPTYPE> {
+class InsertBroadcastPattern : public pir::OpRewritePattern<OPTYPE> {
  public:
   using pir::OpRewritePattern<OPTYPE>::OpRewritePattern;
 
@@ -76,42 +84,46 @@ class FullyInsertBroadcastPattern : public pir::OpRewritePattern<OPTYPE> {
   }
 };
 
-FullyInsertBroadcastPass::FullyInsertBroadcastPass()
-    : pir::PatternRewritePass("fully_insert_broadcast_pass", 1) {}
-
-pir::RewritePatternSet FullyInsertBroadcastPass::InitializePatterns(
-    pir::IrContext* context) {
-  pir::RewritePatternSet ps(context);
-  // elementwise ops
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::AddOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::SubtractOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::MultiplyOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::DivideOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::ElementwisePowOp>>(
-      context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::RemainderOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::FloorDivideOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::MaximumOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::MinimumOp>>(context);
-
-  // compare ops
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::LessThanOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::LessEqualOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::EqualOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::NotEqualOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::GreaterThanOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::GreaterEqualOp>>(context);
-
-  // bitwise ops
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::BitwiseOrOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::BitwiseXorOp>>(context);
-  ps.Add<FullyInsertBroadcastPattern<paddle::dialect::BitwiseNotOp>>(context);
-
-  return ps;
-}
+class InsertBroadcastPass : public pir::PatternRewritePass {
+ public:
+  InsertBroadcastPass() : pir::PatternRewritePass("insert_broadcast_pass", 1) {}
+
+  pir::RewritePatternSet InitializePatterns(pir::IrContext* context) override {
+    pir::RewritePatternSet ps(context);
+    // elementwise ops
+    ps.Add<InsertBroadcastPattern<paddle::dialect::AddOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::SubtractOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::MultiplyOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::DivideOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::ElementwisePowOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::RemainderOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::FloorDivideOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::MaximumOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::MinimumOp>>(context);
+
+    // compare ops
+    ps.Add<InsertBroadcastPattern<paddle::dialect::LessThanOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::LessEqualOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::EqualOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::NotEqualOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::GreaterThanOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::GreaterEqualOp>>(context);
+
+    // bitwise ops
+    ps.Add<InsertBroadcastPattern<paddle::dialect::BitwiseOrOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::BitwiseXorOp>>(context);
+    ps.Add<InsertBroadcastPattern<paddle::dialect::BitwiseNotOp>>(context);
+
+    return ps;
+  }
+
+  bool CanApplyOn(pir::Operation* op) const override {
+    return op->isa<pir::ModuleOp>() && op->num_regions() > 0;
+  }
+};
 
-bool FullyInsertBroadcastPass::CanApplyOn(pir::Operation* op) const {
-  return op->isa<pir::ModuleOp>() && op->num_regions() > 0;
+std::unique_ptr<pir::Pass> CreateInsertBroadcastPass() {
+  return std::make_unique<InsertBroadcastPass>();
 }
 
 }  // namespace ir

…ransforms/fully_insert_broadcast_pass.cc …ator/transforms/insert_broadcast_pass.ccpaddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.cc renamed to paddle/cinn/hlir/dialect/operator/transforms/insert_broadcast_pass.cc paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.cc renamed to paddle/cinn/hlir/dialect/operator/transforms/insert_broadcast_pass.cc

@@ -14,13 +14,14 @@
 
 #pragma once
 
-#include <memory>
-#include "paddle/pir/core/dll_decl.h"
+#include "paddle/pir/pass/pass.h"
 
-namespace pir {
+namespace cinn {
+namespace dialect {
+namespace ir {
 
-class Pass;
+IR_API std::unique_ptr<pir::Pass> CreateInsertBroadcastPass();
 
-IR_API std::unique_ptr<Pass> CreateFcWithSpecialOpFusePass();
-
-}  // namespace pir
+}  // namespace ir
+}  // namespace dialect
+}  // namespace cinn