diff --git a/dnn/test/arm_common/conv_bias.cpp b/dnn/test/arm_common/conv_bias.cpp
index 620e12a882f4c2968f7f9c34b0998857da2f3ddb..fe7f01104127c2c0fecfe1c48c85855272672c3a 100644
--- a/dnn/test/arm_common/conv_bias.cpp
+++ b/dnn/test/arm_common/conv_bias.cpp
@@ -1745,6 +1745,62 @@ TEST_F(ARM_COMMON, BENCHMARK_CONV_BIAS_1X1_S1_NCHW_VS_NCHW44_INT8x8x32) {
 }
 #endif
 
+TEST_F(ARM_COMMON, BENCHMARK_CONV_BIAS_WINOGRAD_VS_IM2COL_INT8) {
+    auto&& args = get_winograd_benchmark_args(3, 8);
+    using namespace conv_bias;
+    constexpr size_t RUN = 10;
+
+    Benchmarker<ConvBias> benchmark_im2col(handle());
+    benchmark_im2col.set_display(false);
+    benchmark_im2col.set_times(RUN);
+    benchmark_im2col.set_dtype(0, dtype::QuantizedS8(2.5f))
+            .set_dtype(1, dtype::QuantizedS8(2.5f))
+            .set_dtype(2, dtype::QuantizedS32(6.25f))
+            .set_dtype(4, dtype::QuantizedS8(60.25f));
+
+    Benchmarker<ConvBias> benchmark_winograd(handle());
+    benchmark_winograd.set_display(false);
+    benchmark_winograd.set_times(RUN);
+    benchmark_winograd.set_dtype(0, dtype::QuantizedS8(2.5f))
+            .set_dtype(1, dtype::QuantizedS8(2.5f))
+            .set_dtype(2, dtype::QuantizedS32(6.25f))
+            .set_dtype(4, dtype::QuantizedS8(60.25f));
+
+    for (auto&& arg : args) {
+        TensorLayout dst_layout;
+        auto opr = handle()->create_operator<ConvBias>();
+        opr->param() = arg.param;
+        opr->deduce_layout({arg.src, dtype::Float32()},
+                        {arg.filter, dtype::Float32()},
+                        {arg.bias, dtype::Float32()}, {}, dst_layout);
+        //! dst.nr_elems * IC * FH * FW * 2
+        float computations = dst_layout.total_nr_elems() * arg.filter[1] *
+                            arg.filter[2] * arg.filter[3] * 2.0 /
+                            (1024 * 1024 * 1024) * 1e3;
+
+        benchmark_im2col.set_param(arg.param);
+        auto im2col_used =
+                algo_benchmark<ConvBias>(
+                        benchmark_im2col, {arg.src, arg.filter, {}, {}, {}},
+                        "IM2COLMATMUL:AARCH64_INT8X8X32_K4X4X16") /
+                RUN;
+
+        benchmark_winograd.set_param(arg.param);
+        auto winograd_used =
+                algo_benchmark<ConvBias>(
+                        benchmark_winograd, {arg.src, arg.filter, {}, {}, {}},
+                        "WINOGRAD:AARCH64_INT16X16X32_MK8_8X8:8:2") /
+                RUN;
+
+        printf("%s %s: im2col: %f ms %f Gflops winograd: %f ms %f GFlops "
+            "speedup: "
+            "%f\n",
+            arg.src.to_string().c_str(), arg.filter.to_string().c_str(),
+            im2col_used, computations / im2col_used, winograd_used,
+            computations / winograd_used, im2col_used / winograd_used);
+    }
+}
+
 #endif
 
 // vim: syntax=cpp.doxygen
diff --git a/dnn/test/common/conv_bias.cpp b/dnn/test/common/conv_bias.cpp
index c95c340c1bb5b885388b0d22e3c60bedbb7a028b..e4b87c601b9ebc19bde0f612baf81b16d9adad37 100644
--- a/dnn/test/common/conv_bias.cpp
+++ b/dnn/test/common/conv_bias.cpp
@@ -736,6 +736,12 @@ std::vector<conv_bias::TestArg> get_winograd_benchmark_args(size_t kernel,
     pack(64, 64, 123, 123, kernel, kernel / 2);
     pack(64, 24, 123, 123, kernel, kernel / 2);
     pack(24, 24, 224, 224, kernel, kernel / 2);
+
+    //! conv in resnet18
+    pack(64, 64, 56, 56, kernel, kernel / 2);
+    pack(128, 128, 28, 28, kernel, kernel / 2);
+    pack(256, 256, 14, 14, kernel, kernel / 2);
+    pack(512, 512, 7, 7, kernel, kernel / 2);
     return args;
 }
 
diff --git a/src/opr/impl/dnn/convolution.cpp b/src/opr/impl/dnn/convolution.cpp
index d8a365087c3722be8ff3b6ffaf626fce66cd8f24..a2f01451e499bd6a2f598ba925092af793c3f1e6 100644
--- a/src/opr/impl/dnn/convolution.cpp
+++ b/src/opr/impl/dnn/convolution.cpp
@@ -309,6 +309,7 @@ typename TimedProfiler<Opr>::TResult TimedProfiler<Opr>::prof_impl(
         return _dt(1.0f)
 
             cb(dtype::QuantizedS8);
+            cb(dtype::QuantizedS16);
             cb(dtype::QuantizedS32);
             default:
                 return DType::from_enum(enumv);
diff --git a/src/opr/impl/tensor_manip.cpp b/src/opr/impl/tensor_manip.cpp
index ba93d3d3eb5c499d838af8bf7218ac4a7d4e3576..2abdd581f8409db3d7c746e56aec9a80f5e166c0 100644
--- a/src/opr/impl/tensor_manip.cpp
+++ b/src/opr/impl/tensor_manip.cpp
@@ -1549,6 +1549,12 @@ void RelayoutFormat::init_output_format() {
 /* f{{{ ===================== WinogradFilterPreprocess ===================== */
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(WinogradFilterPreprocess);
 MEGDNN_OPR_INIT1(WinogradFilterPreprocess, "winograd_filter_preprocess")
+void WinogradFilterPreprocess::init_output_dtype() {
+    TensorLayout dst;
+    TensorLayout src{input(0)->shape(), input(0)->dtype(), input(0)->format()};
+    megdnn_opr()->deduce_layout(src, dst);
+    output(0)->dtype(dst.dtype);
+}
 // f}}}
 
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
diff --git a/src/opr/include/megbrain/opr/tensor_manip.h b/src/opr/include/megbrain/opr/tensor_manip.h
index 4d80558a176abbe6faa619ca3d54cf5c3ea127ff..f02c9a38fdaaacb1c3919082d9de9ad7b20a85c8 100644
--- a/src/opr/include/megbrain/opr/tensor_manip.h
+++ b/src/opr/include/megbrain/opr/tensor_manip.h
@@ -637,7 +637,16 @@ MGB_DEFINE_OPR_CLASS(RelayoutFormat,
  *
  * See docs of megdnn params for more details
  */
-MGB_DEFINE_MEGDNN_OPR_WRAPPER_FWD1(WinogradFilterPreprocess);
+MGB_DEFINE_OPR_CLASS(WinogradFilterPreprocess,
+                     intl::MegDNNOprWrapperFwd<megdnn::WinogradFilterPreprocess>)
+    public:
+        WinogradFilterPreprocess(VarNode* p0, const Param& param,
+                const OperatorNodeConfig& config);
+        static SymbolVar make(SymbolVar p0, const Param& param = {},
+                const OperatorNodeConfig& config = {});
+        void init_output_dtype() override final;
+};
+
 } // opr
 } // mgb