diff --git a/paddle/phi/kernels/sparse/cpu/coalesce_kernel.cc b/paddle/phi/kernels/sparse/cpu/coalesce_kernel.cc
index 95d8abd6bcf5c50e829ac2be0c44211a77649907..2906a1f226ce7b96648d42df5e442c39548dbc21 100644
--- a/paddle/phi/kernels/sparse/cpu/coalesce_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/coalesce_kernel.cc
@@ -98,7 +98,7 @@ template <typename T, typename Context>
 void CoalesceKernel(const Context& dev_ctx,
                     const SparseCooTensor& x,
                     SparseCooTensor* out) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "CoalesceCPUKernel", ([&] {
         CoalesceCPUKernel<T, data_t>(dev_ctx, x, out);
       }));
diff --git a/paddle/phi/kernels/sparse/cpu/conv_grad_kernel.cc b/paddle/phi/kernels/sparse/cpu/conv_grad_kernel.cc
index 44ad2fa588b55c4410c17633b0c789bfbac73d5d..753f6f97b5bc8ba9d82f9f39312e6dd1fe3c3f70 100644
--- a/paddle/phi/kernels/sparse/cpu/conv_grad_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/conv_grad_kernel.cc
@@ -196,7 +196,7 @@ void Conv3dCooGradKernel(const Context& dev_ctx,
                          const std::string& key,
                          SparseCooTensor* x_grad,
                          DenseTensor* kernel_grad) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "Conv3dCooGradCPUKernel", ([&] {
         Conv3dCooGradCPUKernel<T, data_t>(dev_ctx,
                                           x,
diff --git a/paddle/phi/kernels/sparse/cpu/conv_kernel.cc b/paddle/phi/kernels/sparse/cpu/conv_kernel.cc
index f15a636f96d45e1868d0db5f43f2c62f66edcbaf..217aca56e6dd3aed03f838e6408677b16d23565d 100644
--- a/paddle/phi/kernels/sparse/cpu/conv_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/conv_kernel.cc
@@ -186,7 +186,7 @@ void Conv3dCooKernel(const Context& dev_ctx,
                      SparseCooTensor* out,
                      DenseTensor* rulebook,
                      DenseTensor* counter) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "Conv3dCooCPUKernel", ([&] {
         Conv3dCooCPUKernel<T, data_t>(dev_ctx,
                                       x,
diff --git a/paddle/phi/kernels/sparse/cpu/elementwise_grad_kernel.cc b/paddle/phi/kernels/sparse/cpu/elementwise_grad_kernel.cc
index 972b4537b95547509eb7e486907306c310f24ab7..2abd4957de49bcb2382070ab80f57348dec7cf24 100644
--- a/paddle/phi/kernels/sparse/cpu/elementwise_grad_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/elementwise_grad_kernel.cc
@@ -236,7 +236,7 @@ void ElementWiseDivideCsrGradKernel(const Context& dev_ctx,
                                     const SparseCsrTensor& dout,
                                     SparseCsrTensor* dx,
                                     SparseCsrTensor* dy) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_crows().dtype(), "ElementWiseDivideCsrGradCPUKernel", ([&] {
         ElementWiseDivideCsrGradCPUKernel<T, data_t>(
             dev_ctx, x, y, out, dout, dx, dy);
@@ -250,7 +250,7 @@ void ElementWiseDivideCooGradKernel(const Context& dev_ctx,
                                     const SparseCooTensor& dout,
                                     SparseCooTensor* dx,
                                     SparseCooTensor* dy) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "ElementWiseDivideCooGradCPUKernel", ([&] {
         ElementWiseDivideCooGradCPUKernel<T, data_t>(
             dev_ctx, x, y, out, dout, dx, dy);
@@ -262,36 +262,38 @@ void ElementWiseDivideCooGradKernel(const Context& dev_ctx,
                                              \
   DEFINE_ELEMENTWISE_GRAD_KERNEL_COO(name)
 
-#define DEFINE_ELEMENTWISE_GRAD_KERNEL_CSR(name)                              \
-  template <typename T, typename Context>                                     \
-  void ElementWise##name##CsrGradKernel(const Context& dev_ctx,               \
-                                        const SparseCsrTensor& x,             \
-                                        const SparseCsrTensor& y,             \
-                                        const SparseCsrTensor& dout,          \
-                                        SparseCsrTensor* dx,                  \
-                                        SparseCsrTensor* dy) {                \
-    PD_VISIT_INTEGRAL_TYPES(x.non_zero_crows().dtype(),                       \
-                            "ElementWise##name##CsrGradCPUKernel",            \
-                            ([&] {                                            \
-                              ElementWise##name##CsrGradCPUKernel<T, data_t>( \
-                                  dev_ctx, x, y, dout, dx, dy);               \
-                            }));                                              \
+#define DEFINE_ELEMENTWISE_GRAD_KERNEL_CSR(name)                     \
+  template <typename T, typename Context>                            \
+  void ElementWise##name##CsrGradKernel(const Context& dev_ctx,      \
+                                        const SparseCsrTensor& x,    \
+                                        const SparseCsrTensor& y,    \
+                                        const SparseCsrTensor& dout, \
+                                        SparseCsrTensor* dx,         \
+                                        SparseCsrTensor* dy) {       \
+    PD_VISIT_BASE_INTEGRAL_TYPES(                                    \
+        x.non_zero_crows().dtype(),                                  \
+        "ElementWise##name##CsrGradCPUKernel",                       \
+        ([&] {                                                       \
+          ElementWise##name##CsrGradCPUKernel<T, data_t>(            \
+              dev_ctx, x, y, dout, dx, dy);                          \
+        }));                                                         \
   }
 
-#define DEFINE_ELEMENTWISE_GRAD_KERNEL_COO(name)                              \
-  template <typename T, typename Context>                                     \
-  void ElementWise##name##CooGradKernel(const Context& dev_ctx,               \
-                                        const SparseCooTensor& x,             \
-                                        const SparseCooTensor& y,             \
-                                        const SparseCooTensor& dout,          \
-                                        SparseCooTensor* dx,                  \
-                                        SparseCooTensor* dy) {                \
-    PD_VISIT_INTEGRAL_TYPES(x.non_zero_indices().dtype(),                     \
-                            "ElementWise##name##CooGradCPUKernel",            \
-                            ([&] {                                            \
-                              ElementWise##name##CooGradCPUKernel<T, data_t>( \
-                                  dev_ctx, x, y, dout, dx, dy);               \
-                            }));                                              \
+#define DEFINE_ELEMENTWISE_GRAD_KERNEL_COO(name)                     \
+  template <typename T, typename Context>                            \
+  void ElementWise##name##CooGradKernel(const Context& dev_ctx,      \
+                                        const SparseCooTensor& x,    \
+                                        const SparseCooTensor& y,    \
+                                        const SparseCooTensor& dout, \
+                                        SparseCooTensor* dx,         \
+                                        SparseCooTensor* dy) {       \
+    PD_VISIT_BASE_INTEGRAL_TYPES(                                    \
+        x.non_zero_indices().dtype(),                                \
+        "ElementWise##name##CooGradCPUKernel",                       \
+        ([&] {                                                       \
+          ElementWise##name##CooGradCPUKernel<T, data_t>(            \
+              dev_ctx, x, y, dout, dx, dy);                          \
+        }));                                                         \
   }
 
 DEFINE_ELEMENTWISE_GRAD_KERNEL(Add)
diff --git a/paddle/phi/kernels/sparse/cpu/elementwise_kernel.cc b/paddle/phi/kernels/sparse/cpu/elementwise_kernel.cc
index fc8592cbc9d4de51d64baaf41e833e166d60fd1d..3f7b26ee92e1e0337dabfb2e184e1574e1804cd8 100644
--- a/paddle/phi/kernels/sparse/cpu/elementwise_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/elementwise_kernel.cc
@@ -57,11 +57,12 @@ void Merge(const IntT el_len,
            const IntT len_b_max,
            IntT* c_index,
            T* c_values,
-           IntT& nnz,
+           IntT* out_nnz,
            const Functor& functor_org,
            const bool is_divide) {
   IntT a = 0;
   IntT b = 0;
+  IntT& nnz = (*out_nnz);
   nnz = 0;
   const IntT* b_index = nullptr;
   std::vector<IntT> b_full_index;
@@ -94,9 +95,7 @@ void Merge(const IntT el_len,
       }
       ++a;
       ++b;
-    }
-    // coordinate x[a] < coordinate y[b]
-    else if (a_index[a] < b_index[b]) {
+    } else if (a_index[a] < b_index[b]) {  // coordinate x[a] < coordinate y[b]
       if (!functor(a_values + a * el_len,
                    zero.data(),
                    c_values + nnz * el_len,
@@ -105,9 +104,7 @@ void Merge(const IntT el_len,
         ++nnz;
       }
       ++a;
-    }
-    // coordinate x[a] > coordinate y[b]
-    else if (a_index[a] > b_index[b]) {
+    } else if (a_index[a] > b_index[b]) {  // coordinate x[a] > coordinate y[b]
       if (!functor(zero.data(),
                    b_values[b_index[b]],
                    c_values + nnz * el_len,
@@ -215,7 +212,7 @@ void ElementWiseCooKernelImpl(const Context& dev_ctx,
                           max_len,
                           out_indexs.data(),
                           out_values_vec.data(),
-                          nnz,
+                          &nnz,
                           functor,
                           is_divide);
 
@@ -292,7 +289,7 @@ void ElementWiseCooKernelImpl(const Context& dev_ctx,
                                     const SparseCsrTensor& x,                 \
                                     const SparseCsrTensor& y,                 \
                                     SparseCsrTensor* out) {                   \
-    PD_VISIT_INTEGRAL_TYPES(                                                  \
+    PD_VISIT_BASE_INTEGRAL_TYPES(                                             \
         x.non_zero_crows().dtype(), "ElementWise##name##CsrCPUKernel", ([&] { \
           ElementWise##name##CsrCPUKernel<T, data_t>(dev_ctx, x, y, out);     \
         }));                                                                  \
@@ -309,18 +306,18 @@ void ElementWiseCooKernelImpl(const Context& dev_ctx,
         dev_ctx, x, y, out, functor);                                    \
   }
 
-#define DEFINE_COO_ELEMENTWISE_KERNEL(name)                               \
-  template <typename T, typename Context>                                 \
-  void ElementWise##name##CooKernel(const Context& dev_ctx,               \
-                                    const SparseCooTensor& x,             \
-                                    const SparseCooTensor& y,             \
-                                    SparseCooTensor* out) {               \
-    PD_VISIT_INTEGRAL_TYPES(x.non_zero_indices().dtype(),                 \
-                            "ElementWise##name##CooCPUKernel",            \
-                            ([&] {                                        \
-                              ElementWise##name##CooCPUKernel<T, data_t>( \
-                                  dev_ctx, x, y, out);                    \
-                            }));                                          \
+#define DEFINE_COO_ELEMENTWISE_KERNEL(name)                                    \
+  template <typename T, typename Context>                                      \
+  void ElementWise##name##CooKernel(const Context& dev_ctx,                    \
+                                    const SparseCooTensor& x,                  \
+                                    const SparseCooTensor& y,                  \
+                                    SparseCooTensor* out) {                    \
+    PD_VISIT_BASE_INTEGRAL_TYPES(x.non_zero_indices().dtype(),                 \
+                                 "ElementWise##name##CooCPUKernel",            \
+                                 ([&] {                                        \
+                                   ElementWise##name##CooCPUKernel<T, data_t>( \
+                                       dev_ctx, x, y, out);                    \
+                                 }));                                          \
   }
 
 DEFINE_CSR_ELEMENTWISE_CPU_KERNEL(Add)
diff --git a/paddle/phi/kernels/sparse/cpu/mask_kernel.cc b/paddle/phi/kernels/sparse/cpu/mask_kernel.cc
index 92c015101264c0fc259434a06ea8a59355ac381f..8dbf7a84c19388e0874748c7172fab7726450d30 100644
--- a/paddle/phi/kernels/sparse/cpu/mask_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/mask_kernel.cc
@@ -79,7 +79,7 @@ void SparseMaskKernel(const Context& dev_ctx,
                       const DenseTensor& x,
                       const SparseCooTensor& mask,
                       SparseCooTensor* out) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       mask.non_zero_indices().dtype(), "SparseMaskCPUKernel", ([&] {
         SparseMaskCPUKernel<T, data_t>(dev_ctx, x, mask, out);
       }));
@@ -146,7 +146,7 @@ void SparseMaskHelperKernel(const Context& dev_ctx,
                             const SparseCooTensor& x,
                             const DenseTensor& mask_indices,
                             DenseTensor* out) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "SparseMaskHelperCPUKernel", ([&] {
         SparseMaskHelperCPUKernel<T, data_t>(dev_ctx, x, mask_indices, out);
       }));
diff --git a/paddle/phi/kernels/sparse/cpu/pool_grad_kernel.cc b/paddle/phi/kernels/sparse/cpu/pool_grad_kernel.cc
index 077ac07e8d38edde198cc1c8a88b5967f013454d..0d68e712cdd540e001790b3c02de6f9c4bf56809 100644
--- a/paddle/phi/kernels/sparse/cpu/pool_grad_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/pool_grad_kernel.cc
@@ -83,7 +83,7 @@ void MaxPoolCooGradKernel(const Context& dev_ctx,
                           const SparseCooTensor& out_grad,
                           const std::vector<int>& kernel_sizes,
                           SparseCooTensor* x_grad) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "MaxPoolCooGradCPUKernel", ([&] {
         MaxPoolCooGradCPUKernel<T, data_t>(
             dev_ctx, x, rulebook, counter, out, out_grad, kernel_sizes, x_grad);
diff --git a/paddle/phi/kernels/sparse/cpu/pool_kernel.cc b/paddle/phi/kernels/sparse/cpu/pool_kernel.cc
index f01017bba56f5d01dee00284f927c92ef6338de0..8479036789766c92efddef6ad2136a915539a2a7 100644
--- a/paddle/phi/kernels/sparse/cpu/pool_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/pool_kernel.cc
@@ -109,7 +109,7 @@ void MaxPoolCooKernel(const Context& dev_ctx,
                       SparseCooTensor* out,
                       DenseTensor* rulebook,
                       DenseTensor* counter) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "MaxPoolCooCPUKernel", ([&] {
         MaxPoolCooCPUKernel<T, data_t>(dev_ctx,
                                        x,
diff --git a/paddle/phi/kernels/sparse/cpu/softmax_grad_kernel.cc b/paddle/phi/kernels/sparse/cpu/softmax_grad_kernel.cc
index 18d4f4a9c246b8ad37f48decbf30ee5ffdd958ae..e56fe869705aae8cf1caa516aae5dd000a7c6f9c 100644
--- a/paddle/phi/kernels/sparse/cpu/softmax_grad_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/softmax_grad_kernel.cc
@@ -62,7 +62,7 @@ void SoftmaxCsrGradKernel(const Context& dev_ctx,
   T* dx_data = dx_values->data<T>();
 
   // dx = (dout - sum(dout * out)) * out
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       out.non_zero_crows().dtype(), "SoftmaxCsrGradKernel", ([&] {
         const data_t* out_crows_data = out_crows.data<data_t>();
         for (int i = 0; i < batch_size; ++i) {
diff --git a/paddle/phi/kernels/sparse/cpu/softmax_kernel.cc b/paddle/phi/kernels/sparse/cpu/softmax_kernel.cc
index 5f7342b52addacb7223cd2b410f23a17060dfd94..96b6470e0f39e966fbe5b8c3a2dc582e08fe2548 100644
--- a/paddle/phi/kernels/sparse/cpu/softmax_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/softmax_kernel.cc
@@ -60,7 +60,7 @@ void SoftmaxCsrKernel(const Context& dev_ctx,
   T* out_data = out_values->data<T>();
 
   // out = exp(x-x_max) / sum( exp(x-x_max ))
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_crows().dtype(), "CsrSoftmaxKernel", ([&] {
         const data_t* x_crows_data = x_crows.data<data_t>();
         for (int i = 0; i < batch_size; ++i) {
diff --git a/paddle/phi/kernels/sparse/cpu/sparse_utils_kernel.cc b/paddle/phi/kernels/sparse/cpu/sparse_utils_kernel.cc
index 8bf0104ef0baf755794fa6962e3ca80385380098..85227b7cf64fcb9e36eecf8b1d435c9a0bd530e2 100644
--- a/paddle/phi/kernels/sparse/cpu/sparse_utils_kernel.cc
+++ b/paddle/phi/kernels/sparse/cpu/sparse_utils_kernel.cc
@@ -160,7 +160,7 @@ template <typename T, typename Context>
 void SparseCsrToCooKernel(const Context& dev_ctx,
                           const SparseCsrTensor& x,
                           SparseCooTensor* out) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_crows().dtype(), "SparseCsrToCooCPUKernel", ([&] {
         SparseCsrToCooCPUKernel<T, data_t>(dev_ctx, x, out);
       }));
@@ -250,7 +250,7 @@ template <typename T, typename Context>
 void SparseCooToCsrKernel(const Context& dev_ctx,
                           const SparseCooTensor& x,
                           SparseCsrTensor* out) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "SparseCooToCsrCPUKernel", ([&] {
         SparseCooToCsrCPUKernel<T, data_t>(dev_ctx, x, out);
       }));
@@ -304,7 +304,7 @@ template <typename T, typename Context>
 void SparseCooToDenseKernel(const Context& dev_ctx,
                             const SparseCooTensor& x,
                             DenseTensor* out) {
-  PD_VISIT_INTEGRAL_TYPES(
+  PD_VISIT_BASE_INTEGRAL_TYPES(
       x.non_zero_indices().dtype(), "SparseCooToDenseCPUKernel", ([&] {
         SparseCooToDenseCPUKernel<T, data_t>(dev_ctx, x, out);
       }));