[PIR+CINN]Polish CompilationCache for Parsing GroupInputDimExprs in c…

…ase of BroadcastTree (#63750)
PaddlePaddle · Apr 23, 2024 · 66818fb · 66818fb
1 parent 20481c7
commit 66818fb
Show file tree

Hide file tree

Showing 6 changed files with 36 additions and 29 deletions.
diff --git a/paddle/cinn/hlir/dialect/operator/transforms/lowering_pass/lower_cinn_fusion_op_pass.cc b/paddle/cinn/hlir/dialect/operator/transforms/lowering_pass/lower_cinn_fusion_op_pass.cc
@@ -61,26 +61,8 @@ pir::Operation* ProcessDyShapeGroup(
   } else {  // no condition block
     // compile group to jit_kernel_op
     std::vector<pir::Type> output_types;
-    const auto& group_output_values = group->output_values();
-    for (size_t i = 0; i < group_output_values.size(); ++i) {
-      auto base_type =
-          group_output_values[i].type().dyn_cast<::pir::DenseTensorType>();
-      auto dim_info = base_type.dims();
-      if (shape_analysis.HasShapeOrDataForValue(group_output_values[i])) {
-        auto shape = group->GetShapeOrDataExprs(group_output_values[i]).shape();
-        for (size_t k = 0; k < shape.size(); ++k) {
-          if (shape[k].isa<int64_t>()) {
-            dim_info[k] = shape[k].Get<int64_t>();
-          }
-        }
-      }
-      auto new_type = ::pir::DenseTensorType::get(pir::IrContext::Instance(),
-                                                  base_type.dtype(),
-                                                  dim_info,
-                                                  base_type.data_layout(),
-                                                  base_type.lod(),
-                                                  base_type.offset());
-      output_types.push_back(new_type);
+    for (const auto& value : group->output_values()) {
+      output_types.push_back(value.type());
     }
     auto jit_kernel_op = rewriter.Build<cinn::dialect::JitKernelOp>(
         group_inputs, GetJitKernelAttr(group), output_types);

diff --git a/paddle/cinn/hlir/framework/pir/fusion_info.cc b/paddle/cinn/hlir/framework/pir/fusion_info.cc
@@ -112,6 +112,11 @@ std::ostream& operator<<(std::ostream& os, const FusionOpInfo& info) {
 }
 
 FusionInfo::FusionInfo(const OpLoweringGroup& group) {
+  ParseOpInfos(group);
+  ParseInputDimExprs(group);
+}
+
+void FusionInfo::ParseOpInfos(const OpLoweringGroup& group) {
   std::unordered_map<const ::pir::Operation*, size_t> op_mapper;
   unique_fn_name_ = group.FuncName();
 
@@ -141,15 +146,37 @@ FusionInfo::FusionInfo(const OpLoweringGroup& group) {
     op_infos_.emplace_back(*op, GetInnerUpstreamOps(op));
     op_mapper.insert({op, i});
   }
-  auto& shape_analysis =
-      ::pir::ShapeAnalysisManager::Instance().Get(group.GetParentProgram());
-  for (const auto& value : group.GetInputOpValues()) {
+}
+
+void FusionInfo::ParseInputDimExprs(const OpLoweringGroup& group) {
+  // NOTE(Aurelius84): [Why try get DimExpr from Group firstly? ]
+  // In case of BroadcastTree, we will clone many Groups containing same ops.
+  // But its input valus is defining outside and will have same DimExprs in
+  // global ShapeAnalysis, which leading hash conflict unexpected.
+  const auto TryGetDimExprsFromGroup = [&](const ::pir::Value& value) -> bool {
+    if (!group.HasShapeOrDataExprs(value)) return false;
+    input_dim_exprs_.push_back(group.GetShapeOrDataExprs(value));
+    return true;
+  };
+  // NOTE(Aurelius84): If we can't get DimExpr from Group, we will find them
+  // from global ShapeAnalysis.
+  const auto TryeGetDimExprsFromGlobal =
+      [&](const ::pir::Value& value) -> bool {
+    auto& shape_analysis =
+        ::pir::ShapeAnalysisManager::Instance().Get(group.GetParentProgram());
     if (!shape_analysis.HasShapeOrDataForValue(value)) {
       VLOG(4) << "FusionInfo: input value doesn't have shape or data, skip it."
               << value.impl();
-      continue;
+      return false;
     }
     input_dim_exprs_.push_back(shape_analysis.GetShapeOrDataForValue(value));
+    return true;
+  };
+
+  for (const auto& value : group.GetInputOpValues()) {
+    if (!TryGetDimExprsFromGroup(value)) {
+      TryGetDimExprsFromGroup(value);
+    }
   }
 }
 

diff --git a/paddle/cinn/hlir/framework/pir/fusion_info.h b/paddle/cinn/hlir/framework/pir/fusion_info.h
@@ -90,6 +90,9 @@ class FusionInfo {
   friend std::ostream &operator<<(std::ostream &os, const FusionInfo &info);
 
  private:
+  void ParseOpInfos(const OpLoweringGroup &group);
+  void ParseInputDimExprs(const OpLoweringGroup &group);
+
   std::vector<FusionOpInfo> op_infos_;
   std::vector<::symbol::ShapeOrDataDimExprs> input_dim_exprs_;
   std::size_t cached_hash_value_{0};

diff --git a/test/ir/pir/cinn/inference/test_llama_forward.py b/test/ir/pir/cinn/inference/test_llama_forward.py
@@ -87,8 +87,6 @@ def eval(self, use_cinn):
         return out
 
     def test_eval(self):
-        # TODO(Aurelius84):disable compilation cache
-        paddle.set_flags({"FLAGS_enable_cinn_compile_cache": False})
         dy_out = self.eval(use_cinn=False)
         cinn_out = self.eval(use_cinn=True)
         np.testing.assert_allclose(

diff --git a/test/ir/pir/cinn/inference/test_llama_inference.py b/test/ir/pir/cinn/inference/test_llama_inference.py
@@ -190,7 +190,6 @@ def test_eval(self):
         paddle.set_flags(
             {
                 "FLAGS_prim_forward_blacklist": "pd_op.embedding;pd_op.softmax",
-                "FLAGS_enable_cinn_compile_cache": False,
             }
         )
         cinn_out = self.eval(use_cinn=True)

diff --git a/test/ir/pir/cinn/inference/test_llama_postprocess.py b/test/ir/pir/cinn/inference/test_llama_postprocess.py
@@ -109,8 +109,6 @@ def eval(self, use_cinn):
         return out
 
     def test_eval(self):
-        # TODO(Aurelius84):disable compilation cache
-        paddle.set_flags({"FLAGS_enable_cinn_compile_cache": False})
         dy_out = self.eval(use_cinn=False)
         cinn_out = self.eval(use_cinn=True)
         # TODO(Aurelius84): fix the precision with inf