diff --git a/include/ideep/operators/matmul.hpp b/include/ideep/operators/matmul.hpp
index 36b15ede..a1bb4e7e 100644
--- a/include/ideep/operators/matmul.hpp
+++ b/include/ideep/operators/matmul.hpp
@@ -335,11 +335,11 @@ struct matmul_forward : public dnnl::matmul,
       if (bias.is_empty()) {
         do_prepare_dynamic_quant</*with_bias=*/false>(
             param, src, weights, bias, dst, weights_scales,
-            sum_coeff, attr, data_type::f32, aengine);
+            sum_coeff, attr, data_type::f32, alowp_kind, aengine);
       } else {
         do_prepare_dynamic_quant</*with_bias=*/true>(
             param, src, weights, bias, dst, weights_scales,
-            sum_coeff, attr, data_type::f32, aengine);
+            sum_coeff, attr, data_type::f32, alowp_kind, aengine);
       }
     } else {
       if (bias.is_empty()) {
@@ -379,7 +379,7 @@ struct matmul_forward : public dnnl::matmul,
     if (is_dynamic) {
       do_prepare_dynamic_quant</*with_bias=*/false>(
           param, src, weights, dummy_bias, dst, weights_scales,
-          sum_coeff, attr, data_type::f32, aengine);
+          sum_coeff, attr, data_type::f32, alowp_kind, aengine);
     } else {
       do_prepare_static_quant</*with_bias=*/false>(
           param, src, weights, dummy_bias, dst,
@@ -606,11 +606,11 @@ struct matmul_forward : public dnnl::matmul,
       if (bias.is_empty()) {
         do_prepare_dynamic_quant</*with_bias=*/false>(
             param, src, weights, bias, dst, weights_scales,
-            sum_coeff, attr, data_type::f32, aengine);
+            sum_coeff, attr, data_type::f32, alowp_kind, aengine);
       } else {
         do_prepare_dynamic_quant</*with_bias=*/true>(
             param, src, weights, bias, dst, weights_scales,
-            sum_coeff, attr, data_type::f32, aengine);
+            sum_coeff, attr, data_type::f32, alowp_kind, aengine);
       }
     } else {
       if (bias.is_empty()) {
@@ -710,7 +710,7 @@ struct matmul_forward : public dnnl::matmul,
         // prepare
         do_prepare_dynamic_quant<with_bias>(param, src, weights,
                               bias, dst, weights_scales, sum_coeff,
-                              attr, data_type::f32, aengine);
+                              attr, data_type::f32, alowp_kind, aengine);
         // compute
         if (bias.is_empty()) {
           do_compute_dynamic_quant</*with_bias=*/false, reorder_weight>(
@@ -1126,6 +1126,7 @@ struct matmul_forward : public dnnl::matmul,
       const float sum_coeff = 1.0f, // for post-op sum
       const attr_t& attr = attr_t(),
       const data_type dst_type = data_type::f32,
+      const lowp_kind alowp_kind = u8s8,
       const engine& aengine = engine::cpu_engine()) {
     /* This function does the following things:
      * - Determine expected descs of src/weight/dst
@@ -1153,7 +1154,7 @@ struct matmul_forward : public dnnl::matmul,
     auto& weights_scales_in =
         weights.has_scale() ? weights.get_scale() : weights_scales;
 
-    auto src_data_type = data_type::u8;
+    auto src_data_type = alowp_kind == s8s8 ? data_type::s8 : data_type::u8;
     std::vector<int64_t> src_strides = (ndims == 3) ?
         std::vector<int64_t>({src_dims[1] * src_dims[2], src_dims[1], 1}) :
         std::vector<int64_t>({src_dims[1], 1});
@@ -1373,11 +1374,11 @@ struct matmul_forward : public dnnl::matmul,
     auto& expected_src = reorder_src ?
                          src.reorder_if_differ_in(expected_src_desc) :
                          src;
-    
+
     auto& expected_other = reorder_src ?
                            other.reorder_if_differ_in(expected_dst_desc) :
                            other;
-    
+
     auto& expected_weights = reorder_weight ?
                              weights.reorder_if_differ_in(expected_wei_desc) :
                              weights;