diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc
index 1b9360afcecf63ff0c3e306cdf303cc426e80f1e..97ddf223aefcdfaf8a488f93a152336c1ed458f4 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/helper.h b/paddle/fluid/operators/jit/helper.h
index bba3a13619619b6de3f797a4efc4a0d09c3b281f..d5773d65940127ea0a9b77ed2760bd371b778f4c 100644
--- a/paddle/fluid/operators/jit/helper.h
+++ b/paddle/fluid/operators/jit/helper.h
@@ -118,26 +118,33 @@ typename KernelTuples::func_type Get(
   return GetRefer<KT, KernelTuples>();
 }
 
-template <KernelType KT, typename KernelTuples>
-class KernelFuncsCache {
+template <KernelType KT, typename KernelTuples, typename PlaceType>
+class KernelFuncs {
  public:
-  KernelFuncsCache() = default;
-  static KernelFuncsCache& Instance() {
-    static thread_local KernelFuncsCache<KT, KernelTuples> g_func_cache;
+  KernelFuncs() = default;
+  static KernelFuncs& Cache() {
+    static thread_local KernelFuncs<KT, KernelTuples, PlaceType> g_func_cache;
     return g_func_cache;
   }
 
   bool Has(int key) const { return funcs_.find(key) != funcs_.end(); }
 
-  typename KernelTuples::func_type At(int key) { return funcs_.at(key); }
-
   void Insert(int key, typename KernelTuples::func_type func) {
     funcs_.emplace(key, func);
   }
 
+  typename KernelTuples::func_type At(int key) {
+    if (Has(key)) {
+      return funcs_.at(key);
+    }
+    auto func = Get<KT, KernelTuples, PlaceType>(key);
+    Insert(key, func);
+    return func;
+  }
+
  private:
   std::unordered_map<int, typename KernelTuples::func_type> funcs_;
-  DISABLE_COPY_AND_ASSIGN(KernelFuncsCache);
+  DISABLE_COPY_AND_ASSIGN(KernelFuncs);
 };
 
 const char* to_string(KernelType kt);
diff --git a/paddle/fluid/operators/jit/more/mix/mix.cc b/paddle/fluid/operators/jit/more/mix/mix.cc
index 0f42ac158ca7926981df55936cb903d5f4ae4806..0036d1c238b17768c4df61af22a85588990e1815 100644
--- a/paddle/fluid/operators/jit/more/mix/mix.cc
+++ b/paddle/fluid/operators/jit/more/mix/mix.cc
@@ -49,49 +49,16 @@ void VTanh(const T* x, T* y, int n) {
 }
 
 void Softmax(const T* x, T* y, int n, int bs) {
-  typename XRNTuples<T>::func_type compute_hmax{nullptr};
-  typename XRNTuples<T>::func_type compute_hsum{nullptr};
-  typename AXYNTuples<T>::func_type compute_vscal{nullptr};
-  typename AXYNTuples<T>::func_type compute_vaddbias{nullptr};
-  typename XYNTuples<T>::func_type compute_vexp{nullptr};
-
-  if (!KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Has(n)) {
-    compute_hmax = Get<kHMax, XRNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Insert(n, compute_hmax);
-  } else {
-    compute_hmax = KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Has(n)) {
-    compute_hsum = Get<kHSum, XRNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Insert(n, compute_hsum);
-  } else {
-    compute_hsum = KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Has(n)) {
-    compute_vscal = Get<kVScal, AXYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Insert(n,
-                                                               compute_vscal);
-  } else {
-    compute_vscal = KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Has(n)) {
-    compute_vaddbias = Get<kVAddBias, AXYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Insert(
-        n, compute_vaddbias);
-  } else {
-    compute_vaddbias =
-        KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().At(n);
-  }
-
-  if (!KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Has(n)) {
-    compute_vexp = Get<KernelType::kVExp, XYNTuples<T>, platform::CPUPlace>(n);
-    KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Insert(n, compute_vexp);
-  } else {
-    compute_vexp = KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().At(n);
-  }
+  auto compute_hmax =
+      KernelFuncs<kHMax, XRNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_hsum =
+      KernelFuncs<kHSum, XRNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vscal =
+      KernelFuncs<kVScal, AXYNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vaddbias =
+      KernelFuncs<kVAddBias, AXYNTuples<T>, platform::CPUPlace>::Cache().At(n);
+  auto compute_vexp =
+      KernelFuncs<kVExp, XYNTuples<T>, platform::CPUPlace>::Cache().At(n);
 
   for (int i = 0; i < bs; ++i) {
     T scalar;
diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.cc b/paddle/fluid/operators/jit/more/mkl/mkl.cc
index c7d0215eda9d1e14fcad16da7b70f45824789266..4c999131ab116ebe3484355158993558b02cc4b2 100644
--- a/paddle/fluid/operators/jit/more/mkl/mkl.cc
+++ b/paddle/fluid/operators/jit/more/mkl/mkl.cc
@@ -136,7 +136,7 @@ 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 <>
diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h
index cddd0a18db53a7ddf9ca14d5f373180586ef6a31..0ad57c51be79cd3577b43c9af777bff710308fac 100644
--- a/paddle/fluid/operators/math/fc_compute.h
+++ b/paddle/fluid/operators/math/fc_compute.h
@@ -30,15 +30,17 @@ inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
     return;
   }
   if (relu) {
-    auto compute =
-        jit::Get<jit::kVAddRelu, jit::XYZNTuples<T>, platform::CPUPlace>(N);
+    auto compute = jit::KernelFuncs<jit::kVAddRelu, jit::XYZNTuples<T>,
+                                    platform::CPUPlace>::Cache()
+                       .At(N);
     for (int i = 0; i < M; i++) {
       T* dst = Y + i * N;
       compute(B, dst, dst, N);
     }
   } else {
-    auto compute =
-        jit::Get<jit::kVAdd, jit::XYZNTuples<T>, platform::CPUPlace>(N);
+    auto compute = jit::KernelFuncs<jit::kVAdd, jit::XYZNTuples<T>,
+                                    platform::CPUPlace>::Cache()
+                       .At(N);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif
diff --git a/paddle/fluid/operators/math/softmax_impl.h b/paddle/fluid/operators/math/softmax_impl.h
index 1ff9ff684fc8001afb0f768a033b4c5bd1592702..a1cb3f972826a67721b00ce6df0ec48cc34d6e03 100644
--- a/paddle/fluid/operators/math/softmax_impl.h
+++ b/paddle/fluid/operators/math/softmax_impl.h
@@ -82,8 +82,9 @@ class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
     const int kClassDim = 1;
     // 2D data. Batch x C
     auto compute_softmax =
-        jit::Get<jit::kSoftmax, jit::SoftmaxTuples<float>, platform::CPUPlace>(
-            in_dims[kClassDim]);
+        jit::KernelFuncs<jit::kSoftmax, jit::SoftmaxTuples<float>,
+                         platform::CPUPlace>::Cache()
+            .At(in_dims[kClassDim]);
     compute_softmax(in_data, out_data, in_dims[kClassDim], in_dims[kBatchDim]);
   }
 };