[Dynamic Shape] Add FullyInsertBroadcastPass and Broadcast Op (#60511)

jiahy0825 · zhangboSJTU · web-flow · commit a05f19533d6a · 2024-01-04T10:49:28.000+08:00
* add ShapeBroadcastOp

* add pass FullyInsertBroadcastPass

* InferSymbolicShape of BroadcastShape Op

* Delete unit test

* Fix return error

* Code format

* Fix error message

* Update paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.cc

Co-authored-by: Bo Zhang &lt;105368690+zhangbopd@users.noreply.github.com&gt;

---------

Co-authored-by: Bo Zhang &lt;105368690+zhangbopd@users.noreply.github.com&gt;
diff --git a/paddle/cinn/hlir/dialect/operator/transforms/CMakeLists.txt b/paddle/cinn/hlir/dialect/operator/transforms/CMakeLists.txt
@@ -29,4 +29,13 @@ if(NOT CINN_ONLY)
     cinn_op_dialect
     op_dialect_vjp)
 
+  cinn_cc_library(
+    fully_insert_broadcast_pass
+    SRCS
+    fully_insert_broadcast_pass.cc
+    DEPS
+    pir
+    cinn_op_dialect
+    op_dialect_vjp)
+
 endif()
diff --git a/paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.cc b/paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.cc
@@ -0,0 +1,112 @@
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// 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/ir/cinn_op.h"
+#include "paddle/cinn/hlir/framework/pir/utils.h"
+#include "paddle/common/ddim.h"
+#include "paddle/fluid/pir/dialect/operator/ir/manual_op.h"
+#include "paddle/fluid/pir/dialect/operator/ir/op_attribute.h"
+#include "paddle/fluid/pir/dialect/operator/ir/op_type.h"
+#include "paddle/fluid/pir/dialect/operator/ir/pd_op.h"
+#include "paddle/fluid/pir/drr/api/match_context.h"
+#include "paddle/pir/core/builtin_dialect.h"
+#include "paddle/pir/pass/pass.h"
+#include "paddle/pir/pattern_rewrite/pattern_applicator.h"
+#include "paddle/pir/pattern_rewrite/pattern_match.h"
+#include "paddle/pir/pattern_rewrite/pattern_rewrite_driver.h"
+
+namespace cinn {
+namespace dialect {
+namespace ir {
+
+pir::Value GetOutputDimTensor(pir::PatternRewriter* rewriter,
+                              pir::Value x,
+                              pir::Value y) {
+  pir::Value x_shape = rewriter->Build<paddle::dialect::ShapeOp>(x).out();
+  pir::Value y_shape = rewriter->Build<paddle::dialect::ShapeOp>(y).out();
+  return rewriter->Build<paddle::dialect::ShapeBroadcastOp>(x_shape, y_shape)
+      .out();
+}
+
+bool ProcessOp(pir::Operation* op, pir::PatternRewriter* rewriter) {
+  pir::Value x = op->operand_source(0);
+  pir::Value y = op->operand_source(1);
+  pir::Value output_dim_tensor = GetOutputDimTensor(rewriter, x, y);
+  {
+    pir::Value broadcasted_x =
+        rewriter->Build<paddle::dialect::ExpandOp>(x, output_dim_tensor).out();
+    op->operand(0).set_source(broadcasted_x);
+  }
+  {
+    pir::Value broadcasted_y =
+        rewriter->Build<paddle::dialect::ExpandOp>(y, output_dim_tensor).out();
+    op->operand(1).set_source(broadcasted_y);
+  }
+  return true;
+}
+
+template <typename OPTYPE>
+class FullyInsertBroadcastPattern : public pir::OpRewritePattern<OPTYPE> {
+ public:
+  using pir::OpRewritePattern<OPTYPE>::OpRewritePattern;
+
+  bool MatchAndRewrite(OPTYPE op,
+                       pir::PatternRewriter& rewriter) const override {
+    return ProcessOp(op, &rewriter);
+  }
+};
+
+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;
+}
+
+bool FullyInsertBroadcastPass::CanApplyOn(pir::Operation* op) const {
+  return op->isa<pir::ModuleOp>() && op->num_regions() > 0;
+}
+
+}  // namespace ir
+}  // namespace dialect
+}  // namespace cinn
diff --git a/paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.h b/paddle/cinn/hlir/dialect/operator/transforms/fully_insert_broadcast_pass.h
@@ -0,0 +1,35 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/pir/pass/pass.h"
+#include "paddle/pir/pattern_rewrite/frozen_rewrite_pattern_set.h"
+
+namespace cinn {
+namespace dialect {
+namespace ir {
+
+class FullyInsertBroadcastPass : public pir::PatternRewritePass {
+ public:
+  FullyInsertBroadcastPass();
+
+  pir::RewritePatternSet InitializePatterns(pir::IrContext *context) override;
+
+  bool CanApplyOn(pir::Operation *op) const override;
+};
+
+}  // namespace ir
+}  // namespace dialect
+}  // namespace cinn
diff --git a/paddle/fluid/pir/dialect/operator/ir/manual_op.cc b/paddle/fluid/pir/dialect/operator/ir/manual_op.cc
@@ -23,7 +23,7 @@ paddle::dialect::AddNOp, paddle::dialect::AddN_Op,
     paddle::dialect::SliceArrayOp, paddle::dialect::SliceArrayDenseOp,
     paddle::dialect::AssignArray_Op, paddle::dialect::ArrayToTensorOp,
     paddle::dialect::SelectInputOp, paddle::dialect::IncrementOp,
-    paddle::dialect::Increment_Op
+    paddle::dialect::Increment_Op, paddle::dialect::ShapeBroadcastOp
 #else
 
 #include "paddle/fluid/pir/dialect/operator/ir/manual_op.h"
@@ -35,6 +35,7 @@ paddle::dialect::AddNOp, paddle::dialect::AddN_Op,
 #include "paddle/fluid/pir/dialect/operator/ir/op_type.h"
 #include "paddle/fluid/pir/dialect/operator/ir/pd_op.h"
 #include "paddle/fluid/primitive/rule/vjp/vjp.h"
+#include "paddle/phi/api/lib/data_type_set.h"
 #include "paddle/phi/api/lib/utils/allocator.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/enforce.h"
@@ -2925,6 +2926,158 @@ phi::DataType Increment_Op::GetKernelTypeForVar(
   return expected_kernel_dtype;
 }
 
+void ShapeBroadcastOp::Build(pir::Builder &builder,
+                             pir::OperationArgument &argument,
+                             pir::Value x_,
+                             pir::Value y_) {
+  VLOG(4) << "Start build ShapeBroadcastOp";
+
+  VLOG(4) << "Builder construction inputs";
+  std::vector<pir::Value> argument_inputs = {x_, y_};
+  argument.AddInputs(argument_inputs);
+
+  VLOG(4) << "Builder construction attributes";
+
+  VLOG(4) << "Builder construction outputs";
+  paddle::dialect::DenseTensorType x =
+      x_.type().dyn_cast<paddle::dialect::DenseTensorType>();
+  paddle::dialect::DenseTensorType y =
+      y_.type().dyn_cast<paddle::dialect::DenseTensorType>();
+
+  VLOG(4) << "Builder construction  dense_x";
+  paddle::dialect::IrTensor ir_tensor_x(
+      paddle::dialect::TransToPhiDataType(x.dtype()),
+      x.dims(),
+      x.data_layout(),
+      x.lod(),
+      x.offset());
+  VLOG(4) << "Builder construction  meta_x";
+  paddle::dialect::IrMetaTensor meta_x(&ir_tensor_x);
+
+  VLOG(4) << "Builder construction  dense_y";
+  paddle::dialect::IrTensor ir_tensor_y(
+      paddle::dialect::TransToPhiDataType(y.dtype()),
+      y.dims(),
+      y.data_layout(),
+      y.lod(),
+      y.offset());
+  VLOG(4) << "Builder construction  meta_y";
+  paddle::dialect::IrMetaTensor meta_y(&ir_tensor_y);
+  paddle::dialect::IrTensor dense_out;
+  paddle::dialect::IrMetaTensor meta_out(&dense_out);
+
+  phi::ElementwiseInferMeta(meta_x, meta_y, &meta_out);
+
+  std::vector<pir::Type> argument_outputs;
+  pir::Type out_dense_tensor_type = paddle::dialect::DenseTensorType::get(
+      pir::IrContext::Instance(),
+      paddle::dialect::TransToIrDataType(dense_out.dtype()),
+      dense_out.dims(),
+      dense_out.layout(),
+      dense_out.lod(),
+      dense_out.offset());
+  argument_outputs.push_back(out_dense_tensor_type);
+  argument.AddOutputs(argument_outputs.begin(), argument_outputs.end());
+  ::pir::PassStopGradientsDefaultly(argument);
+}
+
+namespace {
+
+void ShapeBroadcastOpInferMeta(const phi::MetaTensor &x,
+                               const phi::MetaTensor &y,
+                               phi::MetaTensor *out) {
+  PADDLE_ENFORCE_EQ(
+      x.dims().size(),
+      1,
+      phi::errors::PreconditionNotMet(
+          "The size %d of x.dims() must be equal to 1.", x.dims().size()));
+  PADDLE_ENFORCE_EQ(
+      y.dims().size(),
+      1,
+      phi::errors::PreconditionNotMet(
+          "The size %d of y.dims() must be equal to 1.", y.dims().size()));
+  out->set_dims({std::max<int64_t>(x.dims().at(0), y.dims().at(0))});
+  // dtype need promote when meet input dtype with more precision
+  paddle::experimental::DataTypeSet dtype_set{x.dtype()};
+  dtype_set = dtype_set | paddle::experimental::DataTypeSet(y.dtype());
+  DataType promote_result = PromoteTypes(dtype_set);
+  if (promote_result == DataType::UNDEFINED) {
+    promote_result = x.dtype();
+  }
+  out->set_dtype(promote_result);
+  out->set_layout(x.layout());
+  out->share_lod(x);
+}
+
+}  // namespace
+
+void ShapeBroadcastOp::InferMeta(phi::InferMetaContext *infer_meta) {
+  auto fn = PD_INFER_META(ShapeBroadcastOpInferMeta);
+  fn(infer_meta);
+}
+
+phi::DataType ShapeBroadcastOp::GetKernelTypeForVar(
+    const std::string &var_name,
+    const phi::DataType &tensor_dtype,
+    const phi::DataType &expected_kernel_dtype) {
+  VLOG(4) << "Get KernelType for Var of op: ShapeBroadcastOp";
+
+  return expected_kernel_dtype;
+}
+
+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 {
+    return symbol::Broadcast<symbol::DimExpr>{
+        symbol::List<symbol::DimExpr>{lhs, rhs}};
+  }
+  LOG(FATAL) << "Dead code";
+}
+
+}  // namespace
+
+bool ShapeBroadcastOp::InferSymbolicShape(
+    pir::ShapeConstraintIRAnalysis *shape_analysis) {
+  pir::Value x = operand_source(0);
+  pir::Value y = operand_source(1);
+  std::string x_id = pir::GetValueId(&x);
+  std::string y_id = pir::GetValueId(&y);
+
+  IR_ENFORCE(shape_analysis->value_id_to_shapeordata_.count(x_id) > 0,
+             "x_id does not exist.");
+  IR_ENFORCE(shape_analysis->value_id_to_shapeordata_.count(y_id) > 0,
+             "y_id does not exist.");
+  const auto &x_data_shape = shape_analysis->value_id_to_shapeordata_.at(x_id);
+  const auto &y_data_shape = shape_analysis->value_id_to_shapeordata_.at(y_id);
+  IR_ENFORCE(x_data_shape.data().has_value(),
+             "Value x comes from ShapeOp, it must have data");
+  IR_ENFORCE(y_data_shape.data().has_value(),
+             "Value y comes from ShapeOp, it must have data");
+  const auto &x_data = x_data_shape.data().value();
+  const auto &y_data = y_data_shape.data().value();
+  IR_ENFORCE(x_data.size() == y_data.size(), "Support same rank temporarily");
+
+  std::vector<symbol::DimExpr> output_data;
+  for (std::size_t i = 0; i < x_data.size(); ++i) {
+    output_data.emplace_back(GetBroadcastDimExpr(x_data.at(i), y_data.at(i)));
+  }
+
+  pir::OpResult res = result(0);
+  std::string res_id = pir::GetValueId(&res);
+  symbol::ShapeOrDataDimExprs output_data_shape =
+      symbol::ShapeOrDataDimExprs::MakeConsistentShapeOrData(output_data);
+  shape_analysis->value_id_to_shapeordata_[res_id] = output_data_shape;
+  return true;
+}
+
 }  // namespace dialect
 }  // namespace paddle
 
@@ -2948,4 +3101,5 @@ IR_DEFINE_EXPLICIT_TYPE_ID(paddle::dialect::ExpandOp)
 IR_DEFINE_EXPLICIT_TYPE_ID(paddle::dialect::SelectInputOp)
 IR_DEFINE_EXPLICIT_TYPE_ID(paddle::dialect::IncrementOp)
 IR_DEFINE_EXPLICIT_TYPE_ID(paddle::dialect::Increment_Op)
+IR_DEFINE_EXPLICIT_TYPE_ID(paddle::dialect::ShapeBroadcastOp)
 #endif
diff --git a/paddle/fluid/pir/dialect/operator/ir/manual_op.h b/paddle/fluid/pir/dialect/operator/ir/manual_op.h
@@ -18,6 +18,7 @@
 #include "paddle/fluid/framework/infershape_utils.h"
 #include "paddle/fluid/pir/dialect/operator/interface/decomp.h"
 #include "paddle/fluid/pir/dialect/operator/interface/get_kernel_type_for_var.h"
+#include "paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape.h"
 #include "paddle/fluid/pir/dialect/operator/interface/infermeta.h"
 #include "paddle/fluid/pir/dialect/operator/interface/op_yaml_info.h"
 #include "paddle/fluid/pir/dialect/operator/interface/vjp.h"
@@ -554,6 +555,37 @@ class Increment_Op
       const std::vector<std::vector<bool>> &stop_gradients);
 };
 
+class IR_API ShapeBroadcastOp
+    : public pir::Op<ShapeBroadcastOp,
+                     paddle::dialect::InferSymbolicShapeInterface,
+                     paddle::dialect::InferMetaInterface,
+                     paddle::dialect::GetKernelTypeForVarInterface> {
+ public:
+  using Op::Op;
+  static const char *name() { return "pd_op.shape_broadcast"; }
+  static constexpr const char **attributes_name = nullptr;
+  static constexpr uint32_t attributes_num = 0;
+  static void Build(pir::Builder &builder,             // NOLINT
+                    pir::OperationArgument &argument,  // NOLINT
+                    pir::Value x_,
+                    pir::Value y_);
+
+  void VerifySig() {}
+
+  pir::Value x() { return operand_source(0); }
+  pir::Value y() { return operand_source(1); }
+  pir::OpResult out() { return result(0); }
+
+  static void InferMeta(phi::InferMetaContext *infer_meta);
+
+  static phi::DataType GetKernelTypeForVar(
+      const std::string &var_name,
+      const phi::DataType &tensor_dtype,
+      const phi::DataType &expected_kernel_dtype);
+
+  bool InferSymbolicShape(pir::ShapeConstraintIRAnalysis *shape_analysis);
+};
+
 }  // namespace dialect
 }  // namespace paddle
 
@@ -577,3 +609,4 @@ IR_DECLARE_EXPLICIT_TYPE_ID(paddle::dialect::ExpandOp)
 IR_DECLARE_EXPLICIT_TYPE_ID(paddle::dialect::SelectInputOp)
 IR_DECLARE_EXPLICIT_TYPE_ID(paddle::dialect::IncrementOp)
 IR_DECLARE_EXPLICIT_TYPE_ID(paddle::dialect::Increment_Op)
+IR_DECLARE_EXPLICIT_TYPE_ID(paddle::dialect::ShapeBroadcastOp)
