update

core/predictor/framework/infer.h

@@ -528,7 +528,7 @@ class CloneDBReloadableInferEngine
 };
 
 template <typename FluidFamilyCore>
-#ifdef WITH_TRT
+#ifdef WITH_TRT | WITH_XPU
 class FluidInferEngine : public DBReloadableInferEngine<FluidFamilyCore> {
 #else
 class FluidInferEngine : public CloneDBReloadableInferEngine<FluidFamilyCore> {

paddle_inference/inferencer-fluid-arm/include/fluid_arm_engine.h

@@ -128,14 +128,6 @@ class FluidArmAnalysisCore : public FluidFamilyCore {
     config.DisableGpu();
     config.SetCpuMathLibraryNumThreads(1);
 
-    if (params.enable_memory_optimization()) {
-      config.EnableMemoryOptim();
-    }
-
-    if (params.enable_memory_optimization()) {
-      config.EnableMemoryOptim();
-    }
-
     if (params.use_lite()) {
       config.EnableLiteEngine(PrecisionType::kFloat32, true);
     }
@@ -144,6 +136,16 @@ class FluidArmAnalysisCore : public FluidFamilyCore {
       config.EnableXpu(2 * 1024 * 1024);
     }
 
+    if (params.enable_memory_optimization()) {
+      config.EnableMemoryOptim();
+    }
+
+    if (params.enable_ir_optimization()) {
+      config.SwitchIrOptim(true);
+    } else {
+      config.SwitchIrOptim(false);
+    }
+
     config.SwitchSpecifyInputNames(true);
     AutoLock lock(GlobalPaddleCreateMutex::instance());
     _core = CreatePredictor(config);
@@ -173,6 +175,14 @@ class FluidArmAnalysisDirCore : public FluidFamilyCore {
     config.SwitchSpecifyInputNames(true);
     config.SetCpuMathLibraryNumThreads(1);
 
+    if (params.use_lite()) {
+      config.EnableLiteEngine(PrecisionType::kFloat32, true);
+    }
+
+    if (params.use_xpu()) {
+      config.EnableXpu(2 * 1024 * 1024);
+    }
+
     if (params.enable_memory_optimization()) {
       config.EnableMemoryOptim();
     }
@@ -183,14 +193,6 @@ class FluidArmAnalysisDirCore : public FluidFamilyCore {
       config.SwitchIrOptim(false);
     }
 
-    if (params.use_lite()) {
-      config.EnableLiteEngine(PrecisionType::kFloat32, true);
-    }
-
-    if (params.use_xpu()) {
-      config.EnableXpu(1 * 1024 * 1024);
-    }
-
     AutoLock lock(GlobalPaddleCreateMutex::instance());
     _core = CreatePredictor(config);
     if (NULL == _core.get()) {