diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc
index 5c5a61f64093802697eb21452267471129c7fcf3..9d2ec5f91a9d2e3a59bec7b00e1663c5548296fd 100644
--- a/paddle/fluid/operators/jit/benchmark.cc
+++ b/paddle/fluid/operators/jit/benchmark.cc
@@ -93,6 +93,7 @@ std::vector<int> TestSizes() {
 template <typename KernelTuples, typename... Args>
 struct BenchFunc {
   // return this function avg time
+  // TODO(TJ): clear cache every time
   double operator()(const typename KernelTuples::func_type tgt, Args... args) {
     for (int i = 0; i < FLAGS_burning; ++i) {
       tgt(args...);
@@ -172,6 +173,9 @@ void BenchXYZNKernel() {
     RandomVec<T>(d, y_data);
     BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data<T>(),
                                                      y.data<T>(), z_data, d);
+    // test inplace
+    BenchAllImpls<KT, jit::XYZNTuples<T>, PlaceType>(d, x.data<T>(), z_data,
+                                                     z_data, d);
   }
 }
 
diff --git a/paddle/fluid/operators/jit/gen/blas.cc b/paddle/fluid/operators/jit/gen/blas.cc
index dee6c7b9d3ee9756c1b11d10d55fdca341cbee85..5da24c359edd2df93333fe0ca8a18cdc7385aadb 100644
--- a/paddle/fluid/operators/jit/gen/blas.cc
+++ b/paddle/fluid/operators/jit/gen/blas.cc
@@ -155,7 +155,7 @@ class NCHW16CMulNCCreator : public JitCodeCreator<int> {
   class name##Creator : public JitCodeCreator<int> {                         \
    public:                                                                   \
     bool UseMe(const int& attr) const override {                             \
-      return platform::MayIUse(platform::avx);                               \
+      return platform::MayIUse(platform::avx) && attr <= 1024;               \
     }                                                                        \
     size_t CodeSize(const int& d) const override {                           \
       return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \
diff --git a/paddle/fluid/operators/jit/gen/blas.h b/paddle/fluid/operators/jit/gen/blas.h
index de6b33f467279124d7acd97709516c31706ec4f9..66a97c1be503b0fa983f9a7ec3b61c986774f16b 100644
--- a/paddle/fluid/operators/jit/gen/blas.h
+++ b/paddle/fluid/operators/jit/gen/blas.h
@@ -61,6 +61,7 @@ class VXXJitCode : public JitCode {
       base += "_Vec";
     }
     base += (with_relu_ ? "_Relu" : "");
+    base += "_D" + std::to_string(num_);
     return base.c_str();
   }
   void genCode() override;
diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.cc b/paddle/fluid/operators/jit/more/mkl/mkl.cc
index 28a37198dae19a57509934ec784746bc23436e7a..3f6814d6c6be1eb76a6b3bfd18adb65991fe3bbb 100644
--- a/paddle/fluid/operators/jit/more/mkl/mkl.cc
+++ b/paddle/fluid/operators/jit/more/mkl/mkl.cc
@@ -139,7 +139,7 @@ bool VMulKernel<float>::UseMe(const int& d) const {
 
 template <>
 bool VAddKernel<float>::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx512f) && d > 512;
+  return platform::MayIUse(platform::avx) && d > 512;
 }
 
 template <>