[CINN] Fix bug of broadcast shape (#63891)

zyfncg · web-flow · commit 0564b70d3cc7 · 2024-04-28T10:21:55.000+08:00
* fix bug of broadcast shape

* fix bug

* fix unittest
diff --git a/paddle/cinn/hlir/dialect/operator/ir/manual_op.cc b/paddle/cinn/hlir/dialect/operator/ir/manual_op.cc
@@ -103,7 +103,7 @@ void GroupOp::Print(pir::IrPrinter& printer) {
   auto& os = printer.os;
   auto op = operation();
   printer.PrintOpResult(op);
-  os << " = " << name();
+  os << " = " << name() << " [id:" << op->id() << "]";
   printer.PrintOpOperands(op);
   os << " -> ";
   printer.PrintOpReturnType(op);
@@ -181,7 +181,7 @@ void FusionOp::Print(pir::IrPrinter& printer) {
   auto& os = printer.os;
   auto op = operation();
   printer.PrintOpResult(op);
-  os << " = " << name();
+  os << " = " << name() << " [id:" << op->id() << "]";
   printer.PrintOpOperands(op);
   os << " -> ";
   printer.PrintOpReturnType(op);
diff --git a/paddle/cinn/hlir/dialect/operator/transforms/add_broadcast_to_elementwise_pass.cc b/paddle/cinn/hlir/dialect/operator/transforms/add_broadcast_to_elementwise_pass.cc
@@ -262,7 +262,7 @@ class AddBroadcastToElementwisePass : public pir::PatternRewritePass {
   }
 
   bool CanApplyOn(pir::Operation* op) const override {
-    return op->num_regions() > 0;
+    return op->num_regions() > 0 && op->isa<cinn::dialect::GroupOp>();
   }
 };
 
diff --git a/paddle/cinn/hlir/framework/pir/op_lowering_group.cc b/paddle/cinn/hlir/framework/pir/op_lowering_group.cc
@@ -183,7 +183,8 @@ std::ostream& operator<<(std::ostream& os, const OpLoweringGroup& group) {
     os << "}";
   };
   ::pir::IrPrinter printer(os);
-  os << "Group " << group.group_id() << " :\n";
+  os << "Group id: " << group.group_id() << ", func_name: " << group.FuncName()
+     << "\n";
   for (auto* op : group.ops()) {
     printer.PrintOperation(op);
     PrintSymbolDims(*op);
diff --git a/paddle/fluid/pir/dialect/operator/ir/manual_op.cc b/paddle/fluid/pir/dialect/operator/ir/manual_op.cc
@@ -3994,7 +3994,7 @@ void ShapeBroadcastOp::Build(pir::Builder &builder,
 }
 
 void ShapeBroadcastOp::InferMeta(phi::InferMetaContext *infer_meta) {
-  auto fn = PD_INFER_META(phi::ElementwiseInferMeta);
+  auto fn = PD_INFER_META(phi::ShapeBroadcastInferMeta);
   fn(infer_meta);
 }
 
@@ -4051,7 +4051,7 @@ std::vector<pir::Type> ShapeBroadcastOp::InferMeta(
   paddle::dialect::IrTensor dense_out;
   paddle::dialect::IrMetaTensor meta_out(&dense_out);
 
-  phi::ElementwiseInferMeta(meta_x, meta_y, &meta_out);
+  phi::ShapeBroadcastInferMeta(meta_x, meta_y, &meta_out);
 
   std::vector<pir::Type> argument_outputs;
   pir::Type out_dense_tensor_type = paddle::dialect::DenseTensorType::get(
@@ -4069,17 +4069,15 @@ namespace {
 
 symbol::DimExpr GetBroadcastDimExpr(const symbol::DimExpr &lhs,
                                     const symbol::DimExpr &rhs) {
-  if (lhs.isa<std::int64_t>() && rhs.isa<std::int64_t>()) {
-    return std::max(lhs.dyn_cast<std::int64_t>(), rhs.dyn_cast<std::int64_t>());
-  } else if (lhs.isa<std::int64_t>()) {
-    return lhs.dyn_cast<std::int64_t>() == 1 ? rhs : lhs;
-  } else if (rhs.isa<std::int64_t>()) {
-    return rhs.dyn_cast<std::int64_t>() == 1 ? lhs : rhs;
-  } else if (lhs == rhs) {
+  if (lhs == rhs) {
+    return lhs;
+  } else if (lhs == 1) {
+    return rhs;
+  } else if (rhs == 1) {
     return lhs;
   } else {
-    return symbol::Broadcast<symbol::DimExpr>{
-        symbol::List<symbol::DimExpr>{lhs, rhs}};
+    return symbol::SimplifyDimExpr(symbol::Broadcast<symbol::DimExpr>{
+        symbol::List<symbol::DimExpr>{lhs, rhs}});
   }
   PADDLE_THROW(phi::errors::Fatal("Dead code"));
 }
diff --git a/paddle/phi/infermeta/binary.cc b/paddle/phi/infermeta/binary.cc
@@ -3110,6 +3110,30 @@ void SearchsortedInferMeta(const MetaTensor& sorted_sequence,
   }
 }
 
+void ShapeBroadcastInferMeta(const MetaTensor& x,
+                             const MetaTensor& y,
+                             MetaTensor* out) {
+  const auto& x_dims = x.dims();
+  const auto& y_dims = y.dims();
+  PADDLE_ENFORCE_EQ(
+      x_dims.size(),
+      1,
+      phi::errors::InvalidArgument("The rank of x must be 1. But received: %d",
+                                   x_dims.size()));
+  PADDLE_ENFORCE_EQ(
+      y_dims.size(),
+      1,
+      phi::errors::InvalidArgument("The rank of y must be 1. But received: %d",
+                                   y_dims.size()));
+
+  if (x_dims[0] <= y_dims[0]) {
+    out->set_dims(y_dims);
+  } else {
+    out->set_dims(x_dims);
+  }
+  out->set_dtype(x.dtype());
+}
+
 void ShuffleBatchInferMeta(const MetaTensor& x,
                            const MetaTensor& seed,
                            int startup_seed,
diff --git a/paddle/phi/infermeta/binary.h b/paddle/phi/infermeta/binary.h
@@ -538,6 +538,10 @@ void SequenceMaskInferMeta(const MetaTensor& x,
                            DataType out_dtype,
                            MetaTensor* y);
 
+void ShapeBroadcastInferMeta(const MetaTensor& x,
+                             const MetaTensor& y,
+                             MetaTensor* out);
+
 void ShuffleBatchInferMeta(const MetaTensor& x,
                            const MetaTensor& seed,
                            int startup_seed,
diff --git a/test/cpp/pir/cinn/add_broadcast_to_elementwise_test.cc b/test/cpp/pir/cinn/add_broadcast_to_elementwise_test.cc
@@ -17,16 +17,32 @@
 #include <memory>
 
 #include "paddle/cinn/hlir/dialect/operator/ir/cinn_op.h"
+#include "paddle/cinn/hlir/dialect/operator/ir/manual_op.h"
 #include "paddle/cinn/hlir/dialect/operator/ir/op_dialect.h"
 #include "paddle/cinn/hlir/dialect/operator/transforms/add_broadcast_to_elementwise_pass.h"
 #include "paddle/fluid/pir/dialect/operator/ir/op_dialect.h"
 #include "paddle/fluid/pir/dialect/operator/ir/pd_op.h"
 #include "paddle/pir/include/core/builtin_dialect.h"
+#include "paddle/pir/include/dialect/control_flow/ir/cf_op.h"
 #include "paddle/pir/include/pass/pass.h"
 #include "paddle/pir/include/pass/pass_manager.h"
 #include "paddle/pir/include/pattern_rewrite/pattern_rewrite_driver.h"
 
+std::vector<pir::Type> CreateDenseTensorTypes(const phi::DDim &dims) {
+  pir::IrContext *ctx = ::pir::IrContext::Instance();
+  pir::Type fp32_dtype = ::pir::Float32Type::get(ctx);
+  phi::DataLayout data_layout = phi::DataLayout::NCHW;
+  phi::LoD lod = {};
+  size_t offset = 0;
+  std::vector<::pir::Type> op_output_types = {::pir::DenseTensorType::get(
+      ctx, fp32_dtype, dims, data_layout, lod, offset)};
+  return op_output_types;
+}
+
 void BuildProgram(pir::Builder &builder) {  // NOLINT
+  auto group_op = builder.Build<cinn::dialect::GroupOp>(
+      CreateDenseTensorTypes(common::make_ddim({4, 3, 16})));
+  builder.SetInsertionPointToBlockEnd(group_op.block());
   paddle::dialect::FullOp full_input_x =
       builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{4, 3, 16},
                                              1.5,
@@ -39,9 +55,13 @@ void BuildProgram(pir::Builder &builder) {  // NOLINT
                                                       full_input_y.result(0));
 
   auto relu_op = builder.Build<paddle::dialect::ReluOp>(add_op.result(0));
+  builder.Build<pir::YieldOp>(std::vector<pir::Value>{relu_op.out()});
 }
 
 void BuildProgramBoth(pir::Builder &builder) {  // NOLINT
+  auto group_op = builder.Build<cinn::dialect::GroupOp>(
+      CreateDenseTensorTypes(common::make_ddim({10, 10})));
+  builder.SetInsertionPointToBlockEnd(group_op.block());
   paddle::dialect::FullOp full_input_x =
       builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{10, 1},
                                              1.5,
@@ -57,9 +77,13 @@ void BuildProgramBoth(pir::Builder &builder) {  // NOLINT
                                                       full_input_y.result(0));
 
   auto relu_op = builder.Build<paddle::dialect::ReluOp>(add_op.result(0));
+  builder.Build<pir::YieldOp>(std::vector<pir::Value>{relu_op.out()});
 }
 
 void BuildProgramSubBoth(pir::Builder &builder) {  // NOLINT
+  auto group_op = builder.Build<cinn::dialect::GroupOp>(
+      CreateDenseTensorTypes(common::make_ddim({10, 10})));
+  builder.SetInsertionPointToBlockEnd(group_op.block());
   paddle::dialect::FullOp full_input_x =
       builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{10, 1},
                                              1.5,
@@ -75,6 +99,7 @@ void BuildProgramSubBoth(pir::Builder &builder) {  // NOLINT
       full_input_x.result(0), full_input_y.result(0));
 
   auto relu_op = builder.Build<paddle::dialect::ReluOp>(sub_op.result(0));
+  builder.Build<pir::YieldOp>(std::vector<pir::Value>{relu_op.out()});
 }
 
 TEST(PatternRewrite, broadcast_elementwise) {
@@ -91,7 +116,11 @@ TEST(PatternRewrite, broadcast_elementwise) {
 
   pm.Run(&program);
 
-  auto it = program.block()->begin();
+  auto it = program.block()
+                ->begin()
+                ->dyn_cast<cinn::dialect::GroupOp>()
+                .block()
+                ->begin();
 
   CHECK_EQ(it->isa<paddle::dialect::FullOp>(), true);
   it++;
@@ -116,7 +145,11 @@ TEST(PatternRewrite, broadcast_elementwise_both) {
 
   pm.Run(&program);
 
-  auto it = program.block()->begin();
+  auto it = program.block()
+                ->begin()
+                ->dyn_cast<cinn::dialect::GroupOp>()
+                .block()
+                ->begin();
 
   CHECK_EQ(it->isa<paddle::dialect::FullOp>(), true);
   it++;
@@ -143,7 +176,11 @@ TEST(PatternRewrite, broadcast_elementwise_sub_both) {
 
   pm.Run(&program);
 
-  auto it = program.block()->begin();
+  auto it = program.block()
+                ->begin()
+                ->dyn_cast<cinn::dialect::GroupOp>()
+                .block()
+                ->begin();
 
   CHECK_EQ(it->isa<paddle::dialect::FullOp>(), true);
   it++;
diff --git a/test/cpp/pir/cinn/pir_all_path_test.cc b/test/cpp/pir/cinn/pir_all_path_test.cc
@@ -62,20 +62,24 @@ static void RunAndCheckResult(::pir::Program* program,
   ctx->GetOrRegisterDialect<paddle::dialect::OperatorDialect>();
   ctx->GetOrRegisterDialect<cinn::dialect::OperatorDialect>();
 
-  pir::PassManager pm(ctx);
-  pm.AddPass(cinn::dialect::ir::CreatePdOpToCinnOpPass());
-  pm.AddPass(cinn::dialect::ir::CreateAddBroadcastToElementwisePass());
-  pm.AddPass(
+  pir::PassManager stage_1_pm(ctx);
+  stage_1_pm.AddPass(cinn::dialect::ir::CreatePdOpToCinnOpPass());
+  stage_1_pm.AddPass(
       std::make_unique<cinn::dialect::ir::MergeReshapeWithBroadcastPass>());
 
-  pm.AddPass(pir::CreateDeadCodeEliminationPass());
-  pm.AddPass(pir::CreateBuildCinnPass());
-  pm.AddPass(cinn::dialect::ir::CreateCinnGroupClusterPass());
-  pm.AddPass(cinn::dialect::ir::CreateAddStoreInFusionOpPass());
-  pm.AddPass(pir::CreateDeadCodeEliminationPass());
-  pm.AddPass(cinn::dialect::ir::CreateLowerCinnFusionOpPass());
-  pm.EnableIRPrinting();
-  CHECK_EQ(pm.Run(program), true);
+  stage_1_pm.AddPass(pir::CreateDeadCodeEliminationPass());
+  stage_1_pm.AddPass(pir::CreateBuildCinnPass());
+  stage_1_pm.AddPass(cinn::dialect::ir::CreateAddBroadcastToElementwisePass());
+  stage_1_pm.EnableIRPrinting();
+  CHECK_EQ(stage_1_pm.Run(program), true);
+
+  pir::PassManager stage_2_pm(ctx);
+  stage_2_pm.AddPass(cinn::dialect::ir::CreateCinnGroupClusterPass());
+  stage_2_pm.AddPass(cinn::dialect::ir::CreateAddStoreInFusionOpPass());
+  stage_2_pm.AddPass(pir::CreateDeadCodeEliminationPass());
+  stage_2_pm.AddPass(cinn::dialect::ir::CreateLowerCinnFusionOpPass());
+  stage_2_pm.EnableIRPrinting();
+  CHECK_EQ(stage_2_pm.Run(program), true);
 
   paddle::platform::Place place = paddle::platform::CUDAPlace(0);
 

Original file line number	Diff line number	Diff line change
`@@ -262,7 +262,7 @@ class AddBroadcastToElementwisePass : public pir::PatternRewritePass {`
`262`	`262`	`}`
`263`	`263`
`264`	`264`	`bool CanApplyOn(pir::Operation* op) const override {`
`265`		`- return op->num_regions() > 0;`
	`265`	`+ return op->num_regions() > 0 && op->isa<cinn::dialect::GroupOp>();`
`266`	`266`	`}`
`267`	`267`	`};`
`268`	`268`