[Inference] Enable infer shape cache. (#48312)

paddle/fluid/framework/ir/runtime_context_cache_pass.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/ir/runtime_context_cache_pass.h"
 
+#include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/framework/operator.h"
 
 namespace paddle {
@@ -21,10 +22,28 @@ namespace framework {
 namespace ir {
 
 void RuntimeContextCachePass::ApplyImpl(ir::Graph* graph) const {
+  static constexpr char kNotAllowInferShapeCahce[] =
+      "@NOT_ALLOW_INFERSHAPE_CACHE@";
   VLOG(3) << "Applies Runtime Context Cache strategy.";
   for (const Node* n : graph->Nodes()) {
     if (n->IsOp() && n->Op()) {
-      n->Op()->SetAttr(kEnableCacheRuntimeContext, true);
+      n->Op()->SetAttr(framework::kEnableCacheRuntimeContext, true);
+    }
+  }
+
+  // if op1 -> var0 and op2 -> var0, then op1 and op2 not support
+  // InferShapeCache.
+  std::unordered_map<std::string, std::vector<Node*>> var2ops;
+  for (auto* op_node : TopologySortOperations(*graph)) {
+    for (auto* var_node : op_node->outputs) {
+      var2ops[var_node->Name()].push_back(op_node);
+    }
+  }
+  for (auto& it : var2ops) {
+    if (it.second.size() > 1) {
+      for (auto op_node : it.second) {
+        op_node->Op()->SetAttr(kNotAllowInferShapeCahce, true);
+      }
     }
   }
 }

paddle/fluid/framework/operator.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 
 #include <sstream>
 #include <string>
+#include <unordered_set>
 
 #include "gflags/gflags.h"
 #include "paddle/fluid/framework/convert_utils.h"
@@ -36,6 +37,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler/supplement_tracing.h"
 #include "paddle/phi/common/int_array.h"
 #include "paddle/phi/common/scalar.h"
+#include "paddle/phi/core/ddim.h"
 #include "paddle/phi/core/kernel_context.h"
 #include "paddle/phi/core/kernel_factory.h"
 #include "paddle/phi/ops/compat/signatures.h"
@@ -562,6 +564,14 @@ phi::DenseTensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var) {
   }
 }
 
+OperatorWithKernel::OperatorWithKernel(const std::string& type,
+                                       const VariableNameMap& inputs,
+                                       const VariableNameMap& outputs,
+                                       const AttributeMap& attrs)
+    : OperatorBase(type, inputs, outputs, attrs) {}
+
+OperatorWithKernel::~OperatorWithKernel() = default;
+
 bool ExecutionContext::HasInput(const std::string& name) const {
   auto* var = InputVar(name);
   return var != nullptr;
@@ -1204,19 +1214,54 @@ class RuntimeInferShapeContext : public InferShapeContext {
 };
 
 struct OperatorWithKernel::CacheImpl {
+  static const char kNotAllowInferShapeCahce[];
   explicit CacheImpl(phi::KernelContext* kernel_ctx,
-                     RuntimeInferShapeContext* infer_shape_ctx)
-      : kernel_ctx_(kernel_ctx), infer_shape_ctx_(infer_shape_ctx) {}
+                     RuntimeInferShapeContext* infer_shape_ctx,
+                     const std::vector<phi::DenseTensor*>& tensors,
+                     bool not_allow_infer_shape_cache)
+      : kernel_ctx_(kernel_ctx),
+        infer_shape_ctx_(infer_shape_ctx),
+        tensors_(tensors),
+        not_allow_infer_shape_cache_(not_allow_infer_shape_cache) {}
 
   phi::KernelContext* getKernelContext() { return kernel_ctx_.get(); }
   RuntimeInferShapeContext* getRuntimeInferShapeContext() {
     return infer_shape_ctx_.get();
   }
 
+  bool NeedInferShape() {
+    if (not_allow_infer_shape_cache_) return true;
+
+    bool ret{false};
+    if (last_ddims_.empty() || tensors_.empty()) ret = true;
+    if (!ret) {
+      CHECK_EQ(last_ddims_.size(), tensors_.size());
+      for (size_t i = 0; i < last_ddims_.size(); ++i) {
+        if (tensors_[i]->dims() != last_ddims_[i]) {
+          ret = true;
+          break;
+        }
+      }
+    }
+    if (ret) {
+      last_ddims_.resize(tensors_.size());
+      for (size_t i = 0; i < last_ddims_.size(); ++i) {
+        last_ddims_[i] = tensors_[i]->dims();
+      }
+    }
+    VLOG(3) << "need infer shape is " << ret;
+    return ret;
+  }
+
  private:
   std::unique_ptr<phi::KernelContext> kernel_ctx_;
   std::unique_ptr<RuntimeInferShapeContext> infer_shape_ctx_;
+  std::vector<phi::DenseTensor*> tensors_;
+  bool not_allow_infer_shape_cache_;
+  std::vector<phi::DDim> last_ddims_;
 };
+const char OperatorWithKernel::CacheImpl::kNotAllowInferShapeCahce[] =
+    "@NOT_ALLOW_INFERSHAPE_CACHE@";
 
 static void CheckTensorNANOrInf(const std::string& op_type,
                                 const std::string& name,
@@ -1524,8 +1569,9 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
     pre_scope_ = cur_scope;
   } else if (run_phi_kernel_ && impl_ != nullptr && !need_prepare_data_ &&
              !need_prepare_phi_data_) {
-    if (!all_kernels_must_compute_runtime_shape_)
+    if (!all_kernels_must_compute_runtime_shape_ && impl_->NeedInferShape()) {
       this->Info().infer_shape_(impl_->getRuntimeInferShapeContext());
+    }
     (*phi_kernel_)(impl_->getKernelContext());
   } else {
     if (runtime_ctx_.get() == nullptr || pre_scope_ != cur_scope) {
@@ -1828,9 +1874,31 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
       phi::KernelContext phi_kernel_context;
       if (enable_cache_runtime_context_ && !need_prepare_phi_data_ &&
           !need_prepare_data_) {
-        impl_ =
+        // TODO(inference): Now we only suppor dense_tensor cache, we may be
+        // support ScalarTensor, SparseTensor in future.
+        bool all_dense_tensor_input_{true};
+        for (auto& iter : Inputs()) {
+          for (auto& name : iter.second) {
+            all_dense_tensor_input_ &=
+                scope.FindVar(name)->IsType<phi::DenseTensor>();
+          }
+        }
+
+        std::vector<phi::DenseTensor*> tensors;
+        if (all_dense_tensor_input_) {
+          for (auto& iter : Inputs()) {
+            for (auto& name : iter.second) {
+              auto* t = scope.FindVar(name)->GetMutable<phi::DenseTensor>();
+              tensors.push_back(t);
+            }
+          }
+        }
+
+        impl_.reset(
             new CacheImpl(new phi::KernelContext(),
-                          new RuntimeInferShapeContext(*this, *runtime_ctx));
+                          new RuntimeInferShapeContext(*this, *runtime_ctx),
+                          tensors,
+                          HasAttr(CacheImpl::kNotAllowInferShapeCahce)));
         BuildPhiKernelContext(*runtime_ctx, dev_ctx, impl_->getKernelContext());
         (*phi_kernel_)(impl_->getKernelContext());
       } else {
@@ -3246,6 +3314,7 @@ void OperatorWithKernel::BuildPhiKernelContext(
   if (phi::OneDNNContext::classof(dev_ctx)) {
     phi::OneDNNContext* one_dnn_ctx = static_cast<phi::OneDNNContext*>(dev_ctx);
     one_dnn_ctx->ClearDnnAttr();
+    if (!RuntimeAttrs().empty()) need_prepare_phi_data_ = true;
   }
 #endif
 
@@ -3267,7 +3336,6 @@ void OperatorWithKernel::BuildPhiKernelContext(
 #if defined(PADDLE_WITH_MKLDNN) || defined(PADDLE_WITH_CUDA)
   auto& runtime_attrs = RuntimeAttrs();
   for (const auto& attr_iter : runtime_attrs) {
-    need_prepare_phi_data_ = true;
     auto& attr_name = attr_iter.first;
     auto& attr = attr_iter.second;
     auto attr_propertys = paddle::operators::GetExtraAttrProperties(attr_name);

paddle/fluid/framework/operator.h

@@ -612,8 +612,9 @@ class OperatorWithKernel : public OperatorBase {
   OperatorWithKernel(const std::string& type,
                      const VariableNameMap& inputs,
                      const VariableNameMap& outputs,
-                     const AttributeMap& attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {}
+                     const AttributeMap& attrs);
+
+  virtual ~OperatorWithKernel();
 
   static paddle::flat_hash_map<std::string /* op_type */, OpKernelMap>&
   AllOpKernels() {
@@ -785,8 +786,9 @@ class OperatorWithKernel : public OperatorBase {
   mutable std::unique_ptr<phi::Kernel> phi_kernel_;
   mutable std::unique_ptr<phi::ArgumentMappingFn> arg_map_fn_;
 
+ private:
   struct CacheImpl;
-  mutable CacheImpl* impl_{nullptr};
+  mutable std::unique_ptr<CacheImpl> impl_;
 };
 
 extern bool OpSupportGPU(const std::string& op_type);

paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.cc

@@ -23,6 +23,8 @@ namespace inference {
 namespace analysis {
 
 void IrGraphToProgramPass::RunImpl(Argument *argument) {
+  auto cache_pass =
+      framework::ir::PassRegistry::Instance().Get("runtime_context_cache_pass");
   auto pass =
       framework::ir::PassRegistry::Instance().Get("graph_to_program_pass");
 
@@ -31,14 +33,12 @@ void IrGraphToProgramPass::RunImpl(Argument *argument) {
               new int(argument->memory_optim_sort_kind()));
   }
 
-  std::unique_ptr<framework::ir::Graph> graph(argument->main_graph_ptr());
-
   // Direct using ProgramDesc desc(argument->main_program()) may cause
   // incomplete copies of information.
   framework::ProgramDesc desc;
   desc.CopyFrom(*argument->main_program().Proto());
   pass->SetNotOwned("program", &desc);
-  pass->Apply(graph.release());  // the argument still own the graph.
+  pass->Apply(cache_pass->Apply(argument->main_graph_ptr()));
 
   argument->SetIrAnalyzedProgram(
       new framework::proto::ProgramDesc(*desc.Proto()));

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -188,7 +188,6 @@ const std::vector<std::string> kGpuLowerPrecisionPasses{
     "fc_fuse_pass",
     "fc_elementwise_layernorm_fuse_pass",
     "embedding_eltwise_layernorm_fuse_pass",
-    "runtime_context_cache_pass",
 };
 
 const std::vector<std::string> kTrtLowerPrecisionPasses{
@@ -254,10 +253,7 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
 #endif                                         //
         "transpose_flatten_concat_fuse_pass",  //
         "constant_folding_pass",               //
-        // following pass should be located in the last, since it will
-        // work on all fused ops.
-        "float_to_half_pass",  //
-        "runtime_context_cache_pass"
+        "float_to_half_pass",                  //
   });
 
   use_gpu_ = true;
@@ -322,10 +318,7 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
                   "conv_transpose_bn_fuse_pass",             //
                   "conv_transpose_eltwiseadd_bn_fuse_pass",  //
                   "is_test_pass",                            //
-                  "constant_folding_pass",
-                  // following pass should be located in the last, since
-                  // it will work on all fused ops.
-                  "runtime_context_cache_pass"});
+                  "constant_folding_pass"});
 
   use_gpu_ = false;
 }
@@ -475,7 +468,6 @@ void CpuPassStrategy::EnableMkldnnInt8() {
     passes_.push_back("int8_scale_calculation_mkldnn_pass");
     passes_.push_back("params_quantization_mkldnn_pass");
     passes_.push_back("mkldnn_inplace_pass");
-    passes_.push_back("runtime_context_cache_pass");
   }
   use_mkldnn_int8_ = true;
 #else