diff --git a/paddle/fluid/operators/jit/helper.cc b/paddle/fluid/operators/jit/helper.cc
index fe508788efe7e880a2e0b64e3a57351e982de2f0..f868c847bd80e874da2d2babde58129122e0bc70 100644
--- a/paddle/fluid/operators/jit/helper.cc
+++ b/paddle/fluid/operators/jit/helper.cc
@@ -56,7 +56,7 @@ const char* to_string(KernelType kt) {
     ONE_CASE(kMatMul);
     ONE_CASE(kHMax);
     ONE_CASE(kHSum);
-    ONE_CASE(kStrideSum);
+    ONE_CASE(kStrideASum);
     ONE_CASE(kSoftmax);
     ONE_CASE(kEmbSeqPool);
     ONE_CASE(kSgd);
diff --git a/paddle/fluid/operators/jit/kernel_base.h b/paddle/fluid/operators/jit/kernel_base.h
index 6fd8a59d55926e38fc5dd8a538e06c5f3834c34e..fdd41a830a2ca4337091faf2dc412f36c036974c 100644
--- a/paddle/fluid/operators/jit/kernel_base.h
+++ b/paddle/fluid/operators/jit/kernel_base.h
@@ -53,7 +53,7 @@ typedef enum {
   kVSquare,
   kVSub,
   kVTanh,
-  kStrideSum,
+  kStrideASum,
   kStrideScal,
 } KernelType;
 
@@ -132,7 +132,7 @@ DECLARE_KERNELTUPLE(XYNTuple, VCopy);
 DECLARE_KERNELTUPLE(XRNTuple, HMax);
 DECLARE_KERNELTUPLE(XRNTuple, HSum);
 
-DECLARE_KERNELTUPLE(XRNSTuple, StrideSum);
+DECLARE_KERNELTUPLE(XRNSTuple, StrideASum);
 
 typedef struct {
   void* gates;  // gates: x_ch, x_ih, x_fh, x_oh
diff --git a/paddle/fluid/operators/jit/more/mix/mix.cc b/paddle/fluid/operators/jit/more/mix/mix.cc
index 58a44d4b55c30dc98e755466ff7f92c43fa5d206..463e45f6ce1bc8f9085e8efe05ae53486242b181 100644
--- a/paddle/fluid/operators/jit/more/mix/mix.cc
+++ b/paddle/fluid/operators/jit/more/mix/mix.cc
@@ -54,7 +54,7 @@ void Softmax(const T* x, T* y, int n, int bs, int m) {
   auto compute_hmax = KernelFuncs<HMaxTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_hsum = KernelFuncs<HSumTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_vscal = KernelFuncs<VScalTuple<T>, CPUPlace>::Cache().At(n);
-  auto compute_stridesum = KernelFuncs<StrideSumTuple<T>, CPUPlace>::Cache().At(n);
+  auto compute_stridesum = KernelFuncs<StrideASumTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_stridescal = KernelFuncs<StrideScalTuple<T>, CPUPlace>::Cache().At(n);
   auto compute_vaddbias =
       KernelFuncs<VAddBiasTuple<T>, CPUPlace>::Cache().At(n);
diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.cc b/paddle/fluid/operators/jit/more/mkl/mkl.cc
index 2828d758159df1ecc97f9b05f3298cc1bc847842..9e21e2b8d39c4827e78c63f9dcf6a08f9be67cf9 100644
--- a/paddle/fluid/operators/jit/more/mkl/mkl.cc
+++ b/paddle/fluid/operators/jit/more/mkl/mkl.cc
@@ -147,12 +147,12 @@ void ASum<double>(const double* x, double* res, int n) {
 }
 
 template <>
-void StrideSum<float>(const float* x, float* res, int n, int stride) {
+void StrideASum<float>(const float* x, float* res, int n, int stride) {
   res[0] = platform::dynload::cblas_sasum(n, x, stride);
 }
 
 template <>
-void StrideSum<double>(const double* x, double* res, int n, int stride) {
+void StrideASum<double>(const double* x, double* res, int n, int stride) {
   res[0] = platform::dynload::cblas_dasum(n, x, stride);
 }
 
@@ -174,7 +174,7 @@ bool VScalKernel<float>::CanBeUsed(const int& d) const {
 
 template <>
 bool StrideScalKernel<float>::CanBeUsed(const int& d) const {
-  return platform::MayIUse(platform::avx512f) && d > 512;
+  return true;
 }
 
 template <>
diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.h b/paddle/fluid/operators/jit/more/mkl/mkl.h
index 1e974c095fb4e8387ee9e9334e638f1b1d997e92..2f135f9e7a4a1fae903ff239d6328fd5ee717b12 100644
--- a/paddle/fluid/operators/jit/more/mkl/mkl.h
+++ b/paddle/fluid/operators/jit/more/mkl/mkl.h
@@ -129,7 +129,7 @@ template <typename T>
 void ASum(const T* x, T* res, int n);
 
 template <typename T>
-void StrideSum(const T* x, T* res, int n, int stride);
+void StrideASum(const T* x, T* res, int n, int stride);
 
 template <typename T>
 void StrideScal(const T* a, const T* x, T* y, int n, int stride);
@@ -155,7 +155,7 @@ void Softmax(const T* x, T* y, int n, int bs, int m=1) {
       VScal(&sum, &y[i * n], &y[i * n], n);
     } else {
       for (int j = 0; j < m; ++j) {
-        StrideSum(&y[i * n + j], &sum, n/m, m);
+        StrideASum(&y[i * n + j], &sum, n/m, m);
         sum = static_cast<T>(1) / sum;
         StrideScal(&sum, &y[i * n + j], &y[i * n + j], n/m, m);
       }
diff --git a/paddle/fluid/operators/jit/refer/CMakeLists.txt b/paddle/fluid/operators/jit/refer/CMakeLists.txt
index 9a39809c9356cb9a68a5b8113f7d66074dd2e882..7133f596620410d37ffe52a2ee92b7a9974bf1cc 100644
--- a/paddle/fluid/operators/jit/refer/CMakeLists.txt
+++ b/paddle/fluid/operators/jit/refer/CMakeLists.txt
@@ -33,7 +33,7 @@ USE_JITKERNEL_REFER(kMatMul)
 USE_JITKERNEL_REFER(kVSquare)
 USE_JITKERNEL_REFER(kHSum)
 USE_JITKERNEL_REFER(kHMax)
-USE_JITKERNEL_REFER(kStrideSum)
+USE_JITKERNEL_REFER(kStrideASum)
 USE_JITKERNEL_REFER(kSoftmax)
 USE_JITKERNEL_REFER(kEmbSeqPool)
 USE_JITKERNEL_REFER(kSgd)
diff --git a/paddle/fluid/operators/jit/refer/refer.cc b/paddle/fluid/operators/jit/refer/refer.cc
index 704124e805a0efc62fc6d63a22bb31fa8e75cf0b..460cb6c58076d7f6c49b60fed45584bd9b506c63 100644
--- a/paddle/fluid/operators/jit/refer/refer.cc
+++ b/paddle/fluid/operators/jit/refer/refer.cc
@@ -52,7 +52,7 @@ REGISTER_REFER_KERNEL(SeqPool);
 REGISTER_REFER_KERNEL(MatMul);
 REGISTER_REFER_KERNEL(HMax);
 REGISTER_REFER_KERNEL(HSum);
-REGISTER_REFER_KERNEL(StrideSum);
+REGISTER_REFER_KERNEL(StrideASum);
 REGISTER_REFER_KERNEL(Softmax);
 REGISTER_REFER_KERNEL(EmbSeqPool);
 REGISTER_REFER_KERNEL(Sgd);
diff --git a/paddle/fluid/operators/jit/refer/refer.h b/paddle/fluid/operators/jit/refer/refer.h
index dee9245524b492b422805177189dc810cfd58b67..e3387f60a6ec798130a894751cfbec0526243b64 100644
--- a/paddle/fluid/operators/jit/refer/refer.h
+++ b/paddle/fluid/operators/jit/refer/refer.h
@@ -412,10 +412,10 @@ void HSum(const T* x, T* res, int n) {
 }
 
 template <typename T>
-void StrideSum(const T* x, T* res, int n, int stride) {
+void StrideASum(const T* x, T* res, int n, int stride) {
   res[0] = x[0];
   for (int i = stride; i < n; i+=stride) {
-    res[0] += x[i];
+    res[0] += std::abs(x[i]);
   }
 }
 
@@ -442,7 +442,7 @@ void Softmax(const T* x, T* y, int n, int bs = 1, int m = 1) {
       VScal(&scalar, y, y, n);
     } else {
       for (int j = 0; j < m; j++) {
-        StrideSum(&y[j], &scalar, n, m);
+        StrideASum(&y[j], &scalar, n, m);
         scalar = static_cast<T>(1) / scalar;
         StrideScal(&scalar, &y[j], &y[j], n, m);
       }
@@ -554,7 +554,7 @@ DECLARE_REFER_KERNEL(GRUHtPart2);
 DECLARE_REFER_KERNEL(HMax);
 DECLARE_REFER_KERNEL(HSum);
 
-DECLARE_REFER_KERNEL(StrideSum);
+DECLARE_REFER_KERNEL(StrideASum);
 
 // others
 DECLARE_REFER_KERNEL(CRFDecoding);
diff --git a/paddle/fluid/operators/jit/test.cc b/paddle/fluid/operators/jit/test.cc
index 93a448166fa932878b36967419121e6c38d7c2cd..c47ec01d3e01d2a348a226798d0985b3c193a99a 100644
--- a/paddle/fluid/operators/jit/test.cc
+++ b/paddle/fluid/operators/jit/test.cc
@@ -727,6 +727,7 @@ void TestKernelSoftmax() {
         if (m > n || n % m != 0) {
           continue;
         }
+        VLOG(10) << "Softmax: " << bs <<  ", " << n << ", " << m;
         auto ref = jit::GetReferFunc<KernelTuple>();
         EXPECT_TRUE(ref != nullptr);
         std::vector<T> x(bs * n), y(bs * n);