diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
index 35c4268a7f9242c9582deca947203ed118626cc5..07931de4ffda808fc4cc895572453908f01d2e07 100644
--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -49,7 +49,7 @@ if (WITH_DISTRIBUTE)
 endif()
 
 SET(OP_ONLY_MKL "")
-if (NOT WITH_MKL)
+if (NOT WITH_MKL OR NOT WITH_AVX)
     SET(OP_ONLY_MKL ${OP_ONLY_MKL} match_matrix_tensor_op)
     SET(OP_ONLY_MKL ${OP_ONLY_MKL} var_conv_2d_op)
 endif()
diff --git a/paddle/fluid/operators/match_matrix_tensor_op.cc b/paddle/fluid/operators/match_matrix_tensor_op.cc
index e9a645d2e0b7b6ac2f4b204f5150161cda9d7d39..d30f396f9c1e8cf1c480c159ec9943b9542ca546 100644
--- a/paddle/fluid/operators/match_matrix_tensor_op.cc
+++ b/paddle/fluid/operators/match_matrix_tensor_op.cc
@@ -286,8 +286,8 @@ class CPUMatchMatrixTensorOPGradKernel : public framework::OpKernel<T> {
             auto* r_data = bottom_r_data + (offset_r[b] + j) * dim_in;
             auto* r_diff = bottom_r_diff + (offset_r[b] + j) * dim_in;
             if (diff != 0.0) {
-              sse_axpy(r_data, l_trans_diff, dim_in, diff);
-              sse_axpy(l_trans_data, r_diff, dim_in, diff);
+              avx_axpy(r_data, l_trans_diff, dim_in, diff);
+              avx_axpy(l_trans_data, r_diff, dim_in, diff);
             }
           }
         }
diff --git a/paddle/fluid/operators/search_compute.h b/paddle/fluid/operators/search_compute.h
index c795f1e390b8a38407b856321eeda75c9ff57895..995c85d9ff4d9de0a8eb804d9caf52f470056283 100644
--- a/paddle/fluid/operators/search_compute.h
+++ b/paddle/fluid/operators/search_compute.h
@@ -73,22 +73,10 @@ void call_gemm_batched(const framework::ExecutionContext& ctx,
   }
 }
 
-#ifndef TYPE_USE_FLOAT
-#define TYPE_USE_FLOAT
-#endif
-#ifndef USE_SSE
-#define USE_SSE
-#endif
-
-#if defined(TYPE_USE_FLOAT)
-
 #define __m256x __m256
-#define __m128x __m128
 
 static const unsigned int AVX_STEP_SIZE = 8;
-static const unsigned int SSE_STEP_SIZE = 4;
 static const unsigned int AVX_CUT_LEN_MASK = 7U;
-static const unsigned int SSE_CUT_LEN_MASK = 3U;
 
 #define _mm256_mul_px _mm256_mul_ps
 #define _mm256_add_px _mm256_add_ps
@@ -96,20 +84,11 @@ static const unsigned int SSE_CUT_LEN_MASK = 3U;
 #define _mm256_store_px _mm256_storeu_ps
 #define _mm256_broadcast_sx _mm256_broadcast_ss
 
-#define _mm_add_px _mm_add_ps
-#define _mm_mul_px _mm_mul_ps
-#define _mm_load_px _mm_loadu_ps
-#define _mm_store_px _mm_storeu_ps
-#define _mm_load1_px _mm_load1_ps
-
-#endif
-
 template <typename T>
-inline void sse_axpy(const T* x, T* y, size_t len, const T alpha) {
+inline void avx_axpy(const T* x, T* y, size_t len, const T alpha) {
   unsigned int jjj, lll;
   jjj = lll = 0;
 
-#if defined(USE_AVX)
   lll = len & ~AVX_CUT_LEN_MASK;
   __m256x mm_alpha = _mm256_broadcast_sx(&alpha);
   for (jjj = 0; jjj < lll; jjj += AVX_STEP_SIZE) {
@@ -119,16 +98,6 @@ inline void sse_axpy(const T* x, T* y, size_t len, const T alpha) {
                       _mm256_mul_px(mm_alpha, _mm256_load_px(x + jjj))));
   }
 
-#elif defined(USE_SSE)
-  lll = len & ~SSE_CUT_LEN_MASK;
-  __m128x mm_alpha = _mm_load1_px(&alpha);
-  for (jjj = 0; jjj < lll; jjj += SSE_STEP_SIZE) {
-    _mm_store_px(y + jjj,
-                 _mm_add_px(_mm_load_px(y + jjj),
-                            _mm_mul_px(mm_alpha, _mm_load_px(x + jjj))));
-  }
-
-#endif
   for (; jjj < len; jjj++) {
     y[jjj] += alpha * x[jjj];
   }
diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt
index 7215e3dde4726bf178ac00e625ce582113792c35..f77f018f65ba8d07485a4c4e0a77544188f24d00 100644
--- a/python/paddle/fluid/tests/unittests/CMakeLists.txt
+++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt
@@ -70,10 +70,10 @@ if(NOT WITH_MKLML)
     list(REMOVE_ITEM TEST_OPS test_fusion_seqexpand_concat_fc_op)
 endif()
 
-if(NOT WITH_MKL)
+if(NOT WITH_MKL OR NOT WITH_AVX)
   list(REMOVE_ITEM TEST_OPS test_match_matrix_tensor_op)
   list(REMOVE_ITEM TEST_OPS test_var_conv_2d)
-endif(NOT WITH_MKL)
+endif()
 
 if(WITH_GPU OR NOT WITH_MKLML)
     # matmul with multiple heads need MKL support