feat(dnn/fallback): fix fallback interface of weight preprocess

GitOrigin-RevId: ca860f487e2c1d1264ab2b3ce0a5515c383b8dcb

dnn/include/megdnn/oprs/nn.h

@@ -234,10 +234,10 @@ public:
             const TensorLayout& dst) = 0;
 
 protected:
-    CanonizedFilterMeta check_exec(const TensorLayout& src,
-                                   const TensorLayout& filter,
-                                   const TensorLayout& dst,
-                                   size_t workspace_in_bytes);
+    CanonizedFilterMeta check_exec(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& dst, size_t workspace_in_bytes,
+            const PreprocessedFilter* preprocessed_filter);
 };
 using Convolution = ConvolutionForward;
 
@@ -408,12 +408,11 @@ public:
     static WinogradParam parse_winograd_name(const std::string& algo_name);
 
 protected:
-    CanonizedFilterMeta check_exec(const TensorLayout& src,
-                                   const TensorLayout& filter,
-                                   const TensorLayout& bias,
-                                   const TensorLayout& z,
-                                   const TensorLayout& dst,
-                                   size_t workspace_in_bytes);
+    CanonizedFilterMeta check_exec(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& bias, const TensorLayout& z,
+            const TensorLayout& dst, size_t workspace_in_bytes,
+            const PreprocessedFilter* preprocessed_filter);
 };
 using ConvBias = ConvBiasForward;
 

dnn/src/common/conv_bias.cpp

@@ -32,7 +32,8 @@ void ConvBiasForward::deduce_layout(const TensorLayout& src,
 ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& bias, const TensorLayout& z,
-        const TensorLayout& dst, size_t workspace_in_bytes) {
+        const TensorLayout& dst, size_t workspace_in_bytes,
+        const PreprocessedFilter* preprocessed_filter) {
     if ((param().format == param::ConvBias::Format::NCHW_WINOGRAD ||
          param().format == param::ConvBias::Format::NCHW88_WINOGRAD ||
          param().format == param::ConvBias::Format::NCHW44_WINOGRAD) &&
@@ -82,9 +83,11 @@ ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
 
     auto ret = check_layout_fwd(src, filter, dst);
     megdnn_assert_contiguous(bias);
-    auto required_workspace_in_bytes =
-            get_workspace_in_bytes(src, filter, bias, z, dst, nullptr);
-    megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
+    auto required_workspace_in_bytes = get_workspace_in_bytes(
+            src, filter, bias, z, dst, preprocessed_filter);
+    megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes,
+                  "worksapce have size of %zu, but need %zu",
+                  workspace_in_bytes, required_workspace_in_bytes);
     if (bias.ndim != 0) {
         //! bias.layout == dst.layout failed, no assert information
         auto check_eq = [](const TensorLayout& bias, const TensorLayout& dst) {

dnn/src/common/convolution.cpp

@@ -1028,10 +1028,11 @@ void ConvolutionForward::deduce_layout(const TensorLayout& src,
 
 ConvolutionForward::CanonizedFilterMeta ConvolutionForward::check_exec(
         const TensorLayout& src, const TensorLayout& filter,
-        const TensorLayout& dst, size_t workspace_in_bytes) {
+        const TensorLayout& dst, size_t workspace_in_bytes,
+        const PreprocessedFilter* preprocessed_filter) {
     auto ret = check_layout_fwd(src, filter, dst);
     auto required_workspace_in_bytes =
-            get_workspace_in_bytes(src, filter, dst, nullptr);
+            get_workspace_in_bytes(src, filter, dst, preprocessed_filter);
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
     return ret;
 }

dnn/src/cuda/conv_bias/opr_impl.cpp

@@ -25,10 +25,10 @@ namespace cuda {
 void ConvBiasForwardImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                                _megdnn_tensor_in bias, _megdnn_tensor_in z,
                                _megdnn_tensor_out dst,
-                               const PreprocessedFilter*,
+                               const PreprocessedFilter* preprocessed_filter,
                                _megdnn_workspace workspace) {
     check_exec(src.layout, filter.layout, bias.layout, z.layout, dst.layout,
-               workspace.size);
+               workspace.size, preprocessed_filter);
     AlgoBase::ExecArgs args(this, src, filter, bias, z, dst, workspace);
     auto algo = get_algorithm(this, src.layout, filter.layout, bias.layout,
                               z.layout, dst.layout);

dnn/src/cuda/conv_bias/opr_impl.h

@@ -52,13 +52,10 @@ public:
             const TensorLayout&, const TensorLayout&) override {
         return {};
     }
-    void exec_preprocess(const TensorLayout& ,
-                         _megdnn_tensor_in ,
-                         const TensorLayout& ,
-                         const TensorLayout& ,
-                         const TensorLayout& ,
-                         PreprocessedFilter* ,
-                         _megdnn_workspace ) override {
+    void exec_preprocess(const TensorLayout&, _megdnn_tensor_in,
+                         const TensorLayout&, const TensorLayout&,
+                         const TensorLayout&, PreprocessedFilter*,
+                         _megdnn_workspace) override {
         megdnn_throw("cuda conv_bias exec_preprocess has not implemeted yet");
     }
 

dnn/src/fallback/conv_bias/opr_impl.cpp

@@ -119,17 +119,22 @@ SmallVector<ConvBiasImpl::AlgoBase*> ConvBiasImpl::algo_pack() {
 bool ConvBiasImpl::is_naive_algo(ConvBiasImpl::Algorithm* algo) {
     return algo == nullptr || strcmp(algo->name(), "DEFAULT") == 0;
 }
+
+#define NCB_ALGO_FUNC(name, algo, param) \
+    static_cast<AlgoBase*>(algo)->name(this, param)
+
 void ConvBiasImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                         _megdnn_tensor_in bias, _megdnn_tensor_in z,
                         _megdnn_tensor_out dst,
                         const PreprocessedFilter* preprocessed_filter,
                         _megdnn_workspace workspace) {
     check_exec(src.layout, filter.layout, bias.layout, z.layout, dst.layout,
-               workspace.size);
-    auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace);
+               workspace.size, preprocessed_filter);
+    auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace,
+                                      preprocessed_filter);
     ConvBiasImpl::Algorithm* algo = get_algorithm(fparam, workspace.size);
     if (!is_naive_algo(algo) &&
-        ncb_algo_get_workspace(algo, fparam) <= workspace.size) {
+        NCB_ALGO_FUNC(get_workspace, algo, fparam) <= workspace.size) {
         exec_with_ncb_kern(fparam, algo);
     } else {
         naive::ConvBiasForwardImpl::exec(src, filter, bias, z, dst,
@@ -137,18 +142,71 @@ void ConvBiasImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
     }
 }
 
+void ConvBiasImpl::exec_preprocess(const TensorLayout& src_layout,
+                                   _megdnn_tensor_in filter,
+                                   const TensorLayout& bias_layout,
+                                   const TensorLayout& z_layout,
+                                   const TensorLayout& dst_layout,
+                                   PreprocessedFilter* preprocessed_filter,
+                                   _megdnn_workspace workspace) {
+    //! exec_preprocess currently only support preprocess weights before exec,
+    //! src/dst/bias/z will be ignored, just set to nullptr
+    TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout},
+            bias{nullptr, bias_layout};
+    auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace,
+                                      preprocessed_filter);
+    ConvolutionImpl::Algorithm* algo = get_algorithm(fparam, workspace.size);
+    if (!is_naive_algo(algo) && NCB_ALGO_FUNC(get_preprocess_workspace, algo,
+                                              fparam) <= workspace.size) {
+        exec_preprocess_with_ncb_kern(fparam, algo);
+    } else {
+        naive::ConvBiasForwardImpl::exec_preprocess(
+                src_layout, filter, bias_layout, z_layout, dst_layout,
+                preprocessed_filter, workspace);
+    }
+}
+
 size_t ConvBiasImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter) {
-    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst);
+    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst,
+                                           preprocessed_filter);
     ConvBiasImpl::Algorithm* algo = get_algorithm(fparam);
     if (is_naive_algo(algo)) {
         return naive::ConvBiasForwardImpl::get_workspace_in_bytes(
                 src, filter, bias, z, dst, preprocessed_filter);
     } else {
-        return ncb_algo_get_workspace(algo, fparam);
+        return NCB_ALGO_FUNC(get_workspace, algo, fparam);
+    }
+}
+
+size_t ConvBiasImpl::get_preprocess_workspace_in_bytes(
+        const TensorLayout& src, const TensorLayout& filter,
+        const TensorLayout& bias, const TensorLayout& z,
+        const TensorLayout& dst) {
+    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
+    Algorithm* algo = get_algorithm(fparam);
+    if (is_naive_algo(algo)) {
+        return naive::ConvBiasForwardImpl::get_preprocess_workspace_in_bytes(
+                src, filter, bias, z, dst);
+    } else {
+        return NCB_ALGO_FUNC(get_preprocess_workspace, algo, fparam);
+    }
+}
+
+SmallVector<TensorLayout> ConvBiasImpl::deduce_preprocessed_filter_layout(
+        const TensorLayout& src, const TensorLayout& filter,
+        const TensorLayout& bias, const TensorLayout& z,
+        const TensorLayout& dst) {
+    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
+    Algorithm* algo = get_algorithm(fparam);
+    if (is_naive_algo(algo)) {
+        return naive::ConvBiasForwardImpl::deduce_preprocessed_filter_layout(
+                src, filter, bias, z, dst);
+    } else {
+        return NCB_ALGO_FUNC(deduce_preprocessed_filter_layout, algo, fparam);
     }
 }
 
@@ -156,7 +214,7 @@ std::vector<ConvBiasImpl::Algorithm*> ConvBiasImpl::get_all_algorithms(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst) {
-    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst);
+    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
     auto ret = get_all_algorithms_with_ncb(fparam);
     if (ret.empty()) {
         return naive::ConvBiasForwardImpl::get_all_algorithms(src, filter, bias,
@@ -170,7 +228,7 @@ ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic(
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst, size_t workspace_limit_in_bytes,
         bool reproducible) {
-    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst);
+    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
     auto result = get_algorithm_heuristic_with_ncb(
             fparam, workspace_limit_in_bytes, reproducible);
     if (result == nullptr) {
@@ -181,9 +239,25 @@ ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic(
     return result;
 }
 
+ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic_with_ncb(
+        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
+        bool reproducible) {
+    for (auto i : get_all_algorithms_with_ncb(param)) {
+        size_t need_workspace = NCB_ALGO_FUNC(get_workspace, i, param);
+        if (static_cast<AlgoBase*>(i)->usable_reproducible(
+                    this, param, AlgoSelectionStrategy::HEURISTIC,
+                    reproducible) &&
+            need_workspace <= workspace_limit_in_bytes) {
+            return i;
+        }
+    }
+    return nullptr;
+}
+
 ConvBiasImpl::NCBKernSizeParam ConvBiasImpl::make_ncb_kern_size_param(
         const TensorLayout& src, const TensorLayout& filter,
-        const TensorLayout& bias, const TensorLayout& dst) {
+        const TensorLayout& bias, const TensorLayout& dst,
+        const PreprocessedFilter* preprocessed_filter) {
     auto safe_u32 = [](size_t v) -> uint32_t {
         megdnn_assert(v <= std::numeric_limits<uint32_t>::max(),
                       "value too large: %zu", v);
@@ -258,7 +332,9 @@ ConvBiasImpl::NCBKernSizeParam ConvBiasImpl::make_ncb_kern_size_param(
              {src.stride[0], src.stride[1], src.stride[2], src.stride[3]},
              {dst.stride[0], dst.stride[1], dst.stride[2], dst.stride[3]},
              param().compute_mode,
-             nr_threads},
+             nr_threads,
+             reinterpret_cast<const ConvolutionForward::PreprocessedFilter*>(
+                     preprocessed_filter)},
             param().output_block_size,
             format,
             bias.dtype,
@@ -269,10 +345,12 @@ ConvBiasImpl::NCBKernSizeParam ConvBiasImpl::make_ncb_kern_size_param(
 
 ConvBiasImpl::NCBKernParam ConvBiasImpl::make_ncb_kern_param(
         _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_in bias,
-        _megdnn_tensor_out dst, _megdnn_workspace workspace) {
+        _megdnn_tensor_out dst, _megdnn_workspace workspace,
+        const PreprocessedFilter* preprocessed_filter) {
     NCBKernParam ret;
-    static_cast<NCBKernSizeParam&>(ret) = make_ncb_kern_size_param(
-            src.layout, filter.layout, bias.layout, dst.layout);
+    static_cast<NCBKernSizeParam&>(ret) =
+            make_ncb_kern_size_param(src.layout, filter.layout, bias.layout,
+                                     dst.layout, preprocessed_filter);
     ret.src_ptr = src.raw_ptr;
     ret.filter_ptr = filter.raw_ptr;
     ret.bias_ptr = bias.raw_ptr;
@@ -284,7 +362,7 @@ ConvBiasImpl::NCBKernParam ConvBiasImpl::make_ncb_kern_param(
 
 void ConvBiasImpl::exec_with_ncb_kern(const NCBKernParam& param,
                                       ConvBiasImpl::Algorithm* algo) {
-    auto ncb_kerns = ncb_algo_dispatch_kerns(algo, param);
+    auto ncb_kerns = NCB_ALGO_FUNC(dispatch_kerns, algo, param);
     for (auto&& kernel : ncb_kerns) {
         auto run = [kernel, param](size_t index, size_t thread_id) {
             CpuNDRange ndrange_id(kernel.global_size, index);
@@ -295,21 +373,17 @@ void ConvBiasImpl::exec_with_ncb_kern(const NCBKernParam& param,
     }
 }
 
-ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic_with_ncb(
-        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
-        bool reproducible) {
-    return ncb_algo_get_algorithm_heuristic(param, workspace_limit_in_bytes,
-                                            reproducible);
-}
-
-size_t ConvBiasImpl::ncb_algo_get_workspace(Algorithm* algo,
-                                            const NCBKernSizeParam& param) {
-    return static_cast<AlgoBase*>(algo)->get_workspace(this, param);
-}
-
-SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::ncb_algo_dispatch_kerns(
-        Algorithm* algo, const NCBKernSizeParam& param) {
-    return static_cast<AlgoBase*>(algo)->dispatch_kerns(this, param);
+void ConvBiasImpl::exec_preprocess_with_ncb_kern(
+        const NCBKernParam& param, ConvBiasImpl::Algorithm* algo) {
+    auto ncb_kerns = NCB_ALGO_FUNC(dispatch_preprocess_kerns, algo, param);
+    for (auto&& kernel : ncb_kerns) {
+        auto run = [kernel, param](size_t index, size_t thread_id) {
+            CpuNDRange ndrange_id(kernel.global_size, index);
+            kernel.kern(param, {thread_id, ndrange_id});
+        };
+        static_cast<naive::HandleImpl*>(handle())->dispatch_kern(
+                run, kernel.global_size.total_size());
+    }
 }
 
 std::vector<ConvBiasImpl::Algorithm*> ConvBiasImpl::get_all_algorithms_with_ncb(
@@ -332,20 +406,6 @@ std::vector<ConvBiasImpl::Algorithm*> ConvBiasImpl::get_all_algorithms_with_ncb(
     return algos;
 }
 
-ConvBiasImpl::Algorithm* ConvBiasImpl::ncb_algo_get_algorithm_heuristic(
-        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
-        bool reproducible) {
-    for (auto i : get_all_algorithms_with_ncb(param)) {
-        if (static_cast<AlgoBase*>(i)->usable_reproducible(
-                    this, param, AlgoSelectionStrategy::HEURISTIC,
-                    reproducible) &&
-            ncb_algo_get_workspace(i, param) <= workspace_limit_in_bytes) {
-            return i;
-        }
-    }
-    return nullptr;
-}
-
 ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm(
         const NCBKernSizeParam& param, size_t workspace_size) {
     if (auto set = execution_policy().algorithm) {

dnn/src/fallback/conv_bias/opr_impl.h

@@ -51,6 +51,25 @@ public:
               _megdnn_tensor_out dst, const PreprocessedFilter*,
               _megdnn_workspace workspace) override;
 
+    void exec_preprocess(const TensorLayout& src_layout,
+                         _megdnn_tensor_in filter,
+                         const TensorLayout& bias_layout,
+                         const TensorLayout& z_layout,
+                         const TensorLayout& dst_layout,
+                         PreprocessedFilter* preprocessed_filter,
+                         _megdnn_workspace workspace) override;
+
+    SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& bias, const TensorLayout& z,
+            const TensorLayout& dst) override;
+
+    size_t get_preprocess_workspace_in_bytes(const TensorLayout& src,
+                                             const TensorLayout& filter,
+                                             const TensorLayout& bias,
+                                             const TensorLayout& z,
+                                             const TensorLayout& dst) override;
+
     //! implemented by get_workspace_with_ncb()
     size_t get_workspace_in_bytes(const TensorLayout& src,
                                   const TensorLayout& filter,
@@ -198,6 +217,23 @@ public:
         virtual SmallVector<NCBKern> dispatch_kerns(
                 ConvBiasImpl* opr, const NCBKernSizeParam& param) const = 0;
 
+        virtual SmallVector<NCBKern> dispatch_preprocess_kerns(
+                ConvBiasImpl*, const NCBKernSizeParam&) const {
+            return {};
+        };
+
+        //! get the layouts of weight_prerocess dst
+        virtual SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
+                ConvBiasImpl*, const NCBKernSizeParam&) const {
+            return {};
+        };
+
+        //! get the workspace when weight_prerocess
+        virtual size_t get_preprocess_workspace(ConvBiasImpl*,
+                                                const NCBKernSizeParam&) const {
+            return 0_z;
+        };
+
         //! Temporarily used to identify whether the matmul algorithm is
         //! is_preferred.
         virtual bool is_preferred(ConvBiasImpl*,
@@ -219,40 +255,19 @@ public:
     virtual SmallVector<AlgoBase*> algo_pack();
 
 protected:
-    //! default impl calls ncb_algo_dispatch_kern()
     virtual void exec_with_ncb_kern(const NCBKernParam& param,
                                     ConvBiasImpl::Algorithm* algo);
 
-    //! default impl calls ncb_algo_get_all_algorithms()
+    virtual void exec_preprocess_with_ncb_kern(const NCBKernParam& param,
+                                               Algorithm* algo);
+
     virtual std::vector<Algorithm*> get_all_algorithms_with_ncb(
             const NCBKernSizeParam& param);
 
-    //! default impl calls ncb_algo_get_algorithm_heuristic()
     virtual Algorithm* get_algorithm_heuristic_with_ncb(
             const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
             bool reproducible = false);
 
-    /**
-     * \brief get kernel pointer for non-contiguous batch  kernel or
-     *     simply conv bias kernel.
-     *
-     *  whether the kernel processing batch 1-group is decided by the
-     *  algo.
-     */
-
-    virtual SmallVector<NCBKern> ncb_algo_dispatch_kerns(
-            Algorithm* algo, const NCBKernSizeParam& param);
-
-    virtual size_t ncb_algo_get_workspace(Algorithm* algo,
-                                          const NCBKernSizeParam& param);
-    /*!
-     * the default impl iterates over all ncb_algo_get_all_algorithms()
-     * and return the first one whose workspace does not exceed the limit.
-     */
-    virtual Algorithm* ncb_algo_get_algorithm_heuristic(
-            const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
-            bool reproducible = false);
-
     const char* get_algorithm_set_name() const override;
 
 private:
@@ -276,16 +291,16 @@ private:
             const NCBKernSizeParam& param,
             size_t workspace_size = std::numeric_limits<size_t>::max());
 
-    NCBKernSizeParam make_ncb_kern_size_param(const TensorLayout& src,
-                                              const TensorLayout& filter,
-                                              const TensorLayout& bias,
-                                              const TensorLayout& dst);
-
-    NCBKernParam make_ncb_kern_param(_megdnn_tensor_in src,
-                                     _megdnn_tensor_in filter,
-                                     _megdnn_tensor_in bias,
-                                     _megdnn_tensor_out dst,
-                                     _megdnn_workspace workspace);
+    NCBKernSizeParam make_ncb_kern_size_param(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& bias, const TensorLayout& dst,
+            const PreprocessedFilter* preprocessed_filter);
+
+    NCBKernParam make_ncb_kern_param(
+            _megdnn_tensor_in src, _megdnn_tensor_in filter,
+            _megdnn_tensor_in bias, _megdnn_tensor_out dst,
+            _megdnn_workspace workspace,
+            const PreprocessedFilter* preprocessed_filter);
 };
 
 }  // namespace fallback

dnn/src/fallback/convolution/algos.cpp

@@ -376,7 +376,67 @@ size_t ConvolutionImpl::AlgoDefault::get_workspace(
     return get_bundle(param).total_size_in_bytes();
 }
 
-//! Return the implment kernel
+size_t ConvolutionImpl::AlgoDefault::get_preprocess_workspace(
+        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+    ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
+            init_convbias_opr_and_param(m_conv_bias_opr, param);
+    m_conv_bias_opr->execution_policy() = {m_algorithm};
+    return m_algorithm->get_preprocess_workspace(m_conv_bias_opr,
+                                                 conv_bias_param);
+}
+
+SmallVector<TensorLayout>
+ConvolutionImpl::AlgoDefault::deduce_preprocessed_filter_layout(
+        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+    ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
+            init_convbias_opr_and_param(m_conv_bias_opr, param);
+    m_conv_bias_opr->execution_policy() = {m_algorithm};
+    return m_algorithm->deduce_preprocessed_filter_layout(m_conv_bias_opr,
+                                                          conv_bias_param);
+}
+
+//! Return the implement preprocess kernel
+SmallVector<ConvolutionImpl::NCBKern>
+ConvolutionImpl::AlgoDefault::get_preprocess_kimpl(
+        ::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase* algo,
+        const NCBKernSizeParam& param) {
+    MIDOUT_BEGIN(megdnn_fallback_conv, midout_iv("get_preprocess_kimpl"_hash)) {
+        // construct the conv_bias kern param
+        ::ConvBiasImpl::NCBKernParam conv_bias_param;
+        ::ConvBiasImpl::NCBKernSizeParam conv_bias_size_param =
+                init_convbias_opr_and_param(conv_bias_opr, param);
+        static_cast<::ConvBiasImpl::NCBKernSizeParam&>(conv_bias_param) =
+                conv_bias_size_param;
+        auto conv_bias_preprocess_kerns =
+                algo->dispatch_preprocess_kerns(conv_bias_opr, conv_bias_param);
+        SmallVector<ConvolutionImpl::NCBKern> convolution_preprocess_kerns;
+
+        //! Set the conv_bias param using convolution param
+        auto set_copy_param_filter_workspace_ptr =
+                [](const NCBKernParam& conv_param,
+                   ::ConvBiasImpl::NCBKernParam& copied_param) {
+                    copied_param.filter_ptr = conv_param.filter_ptr;
+                    copied_param.workspace_ptr = conv_param.workspace_ptr;
+                    copied_param.workspace_size = conv_param.workspace_size;
+                };
+        for (size_t i = 0; i < conv_bias_preprocess_kerns.size(); i++) {
+            auto kernel = conv_bias_preprocess_kerns[i];
+            //! If the kerenl batch parallel
+            auto run = [=](const NCBKernParam& p,
+                           const NCBKernIndex& ncb_index) {
+                auto copy_param = conv_bias_param;
+                set_copy_param_filter_workspace_ptr(p, copy_param);
+                kernel.kern(copy_param,
+                            {ncb_index.thread_id, ncb_index.ndrange_id});
+            };
+            convolution_preprocess_kerns.push_back({run, kernel.global_size});
+        }
+        return convolution_preprocess_kerns;
+    }
+    MIDOUT_END();
+}
+
+//! Return the implement kernel
 SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoDefault::get_kimpl(
         ::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase* algo,
         const NCBKernSizeParam& param) {
@@ -392,7 +452,7 @@ SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoDefault::get_kimpl(
         SmallVector<ConvolutionImpl::NCBKern> convolution_kerns;
 
         //! Set the conv_bias param using convolution param
-        auto set_copy_param_run_time_address =
+        auto set_copy_param_compute_address =
                 [](const NCBKernParam& conv_param,
                    ::ConvBiasImpl::NCBKernParam& copied_param) {
                     copied_param.src_ptr = conv_param.src_ptr;
@@ -407,7 +467,7 @@ SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoDefault::get_kimpl(
             auto run = [=](const NCBKernParam& p,
                            const NCBKernIndex& ncb_index) {
                 auto copy_param = conv_bias_param;
-                set_copy_param_run_time_address(p, copy_param);
+                set_copy_param_compute_address(p, copy_param);
                 kernel.kern(copy_param,
                             {ncb_index.thread_id, ncb_index.ndrange_id});
             };

dnn/src/fallback/convolution/algos.h

@@ -110,6 +110,9 @@ class ConvolutionImpl::AlgoDefault final : public AlgoBase {
     static SmallVector<NCBKern> get_kimpl(ConvBiasImpl* conv_bias_opr,
                                           ConvBiasImpl::AlgoBase* algo,
                                           const NCBKernSizeParam& param);
+    static SmallVector<NCBKern> get_preprocess_kimpl(
+            ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase* algo,
+            const NCBKernSizeParam& param);
 
 public:
     AlgoDefault(fallback::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase*);
@@ -121,6 +124,17 @@ public:
     size_t get_workspace(ConvolutionImpl* opr,
                          const NCBKernSizeParam& param) const override;
 
+    size_t get_preprocess_workspace(ConvolutionImpl*,
+                                    const NCBKernSizeParam&) const override;
+
+    SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
+            ConvolutionImpl*, const NCBKernSizeParam&) const override;
+
+    SmallVector<NCBKern> dispatch_preprocess_kern(
+            ConvolutionImpl*, const NCBKernSizeParam& param) const override {
+        return get_preprocess_kimpl(m_conv_bias_opr, m_algorithm, param);
+    }
+
     SmallVector<NCBKern> dispatch_kern(
             ConvolutionImpl* /*opr*/,
             const NCBKernSizeParam& param) const override {

dnn/src/fallback/convolution/opr_impl.cpp

@@ -80,14 +80,19 @@ SmallVector<ConvolutionImpl::AlgoBase*> ConvolutionImpl::algo_pack() {
 bool ConvolutionImpl::is_naive_algo(ConvolutionImpl::Algorithm* algo) {
     return algo == nullptr || strcmp(algo->name(), "DEFAULT") == 0;
 }
+
+#define NCB_ALGO_FUNC(name, algo, param) \
+    static_cast<AlgoBase*>(algo)->name(this, fparam)
+
 void ConvolutionImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                            _megdnn_tensor_out dst,
                            const PreprocessedFilter* preprocessed_filter,
                            _megdnn_workspace workspace) {
-    auto fparam = make_ncb_kern_param(src, filter, dst, workspace);
+    auto fparam = make_ncb_kern_param(src, filter, dst, preprocessed_filter,
+                                      workspace);
     ConvolutionImpl::Algorithm* algo = get_algorithm(fparam, workspace.size);
     if (!is_naive_algo(algo) &&
-        ncb_algo_get_workspace(algo, fparam) <= workspace.size) {
+        NCB_ALGO_FUNC(get_workspace, algo, fparam) <= workspace.size) {
         exec_with_ncb_kern(fparam, algo);
     } else {
         naive::ConvolutionForwardImpl::exec(src, filter, dst,
@@ -95,24 +100,73 @@ void ConvolutionImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
     }
 }
 
+void ConvolutionImpl::exec_preprocess(const TensorLayout& src_layout,
+                         _megdnn_tensor_in filter,
+                         const TensorLayout& dst_layout,
+                         PreprocessedFilter* preprocessed_filter,
+                         _megdnn_workspace workspace) {
+    //! exec_preprocess currently only support preprocess weights before exec,
+    //! src/dst will be ignored, just set to nullptr
+    TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout};
+    auto fparam = make_ncb_kern_param(src, filter, dst, preprocessed_filter,
+                                      workspace);
+    ConvolutionImpl::Algorithm* algo = get_algorithm(fparam, workspace.size);
+    if (!is_naive_algo(algo) && NCB_ALGO_FUNC(get_preprocess_workspace, algo,
+                                              fparam) <= workspace.size) {
+        exec_preprocess_with_ncb_kern(fparam, algo);
+    } else {
+        naive::ConvolutionForwardImpl::exec_preprocess(
+                src_layout, filter, dst_layout, preprocessed_filter, workspace);
+    }
+}
+
 size_t ConvolutionImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter) {
-    auto fparam = make_ncb_kern_size_param(src, filter, dst);
+    auto fparam =
+            make_ncb_kern_size_param(src, filter, dst, preprocessed_filter);
     Algorithm* algo = get_algorithm(fparam);
     if (is_naive_algo(algo)) {
         return naive::ConvolutionForwardImpl::get_workspace_in_bytes(
                 src, filter, dst, preprocessed_filter);
     } else {
-        return ncb_algo_get_workspace(algo, fparam);
+        return static_cast<AlgoBase*>(algo)->get_workspace(this, fparam);
+    }
+}
+
+size_t ConvolutionImpl::get_preprocess_workspace_in_bytes(
+        const TensorLayout& src, const TensorLayout& filter,
+        const TensorLayout& dst) {
+    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
+    Algorithm* algo = get_algorithm(fparam);
+    if (is_naive_algo(algo)) {
+        return naive::ConvolutionForwardImpl::get_preprocess_workspace_in_bytes(
+                src, filter, dst);
+    } else {
+        return static_cast<AlgoBase*>(algo)->get_preprocess_workspace(this,
+                                                                      fparam);
+    }
+}
+
+SmallVector<TensorLayout> ConvolutionImpl::deduce_preprocessed_filter_layout(
+        const TensorLayout& src, const TensorLayout& filter,
+        const TensorLayout& dst){
+    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
+    Algorithm* algo = get_algorithm(fparam);
+    if (is_naive_algo(algo)) {
+        return naive::ConvolutionForwardImpl::deduce_preprocessed_filter_layout(
+                src, filter, dst);
+    } else {
+        return static_cast<AlgoBase*>(algo)->deduce_preprocessed_filter_layout(
+                this, fparam);
     }
 }
 
 std::vector<ConvolutionImpl::Algorithm*> ConvolutionImpl::get_all_algorithms(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst) {
-    auto fparam = make_ncb_kern_size_param(src, filter, dst);
+    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
     auto ret = get_all_algorithms_with_ncb(fparam);
     if (ret.empty()) {
         return naive::ConvolutionForwardImpl::get_all_algorithms(src, filter,
@@ -125,7 +179,7 @@ ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_heuristic(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst, size_t workspace_limit_in_bytes,
         bool reproducible) {
-    auto fparam = make_ncb_kern_size_param(src, filter, dst);
+    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
     auto result = get_algorithm_heuristic_with_ncb(
             fparam, workspace_limit_in_bytes, reproducible);
     if (result == nullptr) {
@@ -137,7 +191,8 @@ ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_heuristic(
 
 ConvolutionImpl::NCBKernSizeParam ConvolutionImpl::make_ncb_kern_size_param(
         const TensorLayout& src, const TensorLayout& filter,
-        const TensorLayout& dst) {
+        const TensorLayout& dst,
+        const PreprocessedFilter* preprocessed_filter) {
     auto safe_u32 = [](size_t v) -> uint32_t {
         megdnn_assert(v <= std::numeric_limits<uint32_t>::max(),
                       "value too large: %zu", v);
@@ -175,15 +230,17 @@ ConvolutionImpl::NCBKernSizeParam ConvolutionImpl::make_ncb_kern_size_param(
             {src.stride[0], src.stride[1], src.stride[2], src.stride[3]},
             {dst.stride[0], dst.stride[1], dst.stride[2], dst.stride[3]},
             param().compute_mode,
-            nr_threads};
+            nr_threads,
+            preprocessed_filter};
 }
 
 ConvolutionImpl::NCBKernParam ConvolutionImpl::make_ncb_kern_param(
         _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_out dst,
+        const PreprocessedFilter* preprocessed_filter,
         _megdnn_workspace workspace) {
     NCBKernParam ret;
-    static_cast<NCBKernSizeParam&>(ret) =
-            make_ncb_kern_size_param(src.layout, filter.layout, dst.layout);
+    static_cast<NCBKernSizeParam&>(ret) = make_ncb_kern_size_param(
+            src.layout, filter.layout, dst.layout, preprocessed_filter);
     ret.src_ptr = src.raw_ptr;
     ret.filter_ptr = filter.raw_ptr;
     ret.dst_ptr = dst.raw_ptr;
@@ -192,9 +249,30 @@ ConvolutionImpl::NCBKernParam ConvolutionImpl::make_ncb_kern_param(
     return ret;
 }
 
+void ConvolutionImpl::exec_preprocess_with_ncb_kern(const NCBKernParam& param,
+                                                    Algorithm* algo) {
+    auto kerns =
+            static_cast<AlgoBase*>(algo)->dispatch_preprocess_kern(this, param);
+    auto fallback_handle = handle();
+    for (auto kernel : kerns) {
+        megdnn_assert(
+                param.filter_meta.format == Param::Format::NCHW ||
+                        param.filter_meta.format == Param::Format::NHWC ||
+                        param.filter_meta.format == Param::Format::NCHW88 ||
+                        param.filter_meta.format == Param::Format::NCHW44,
+                "invalid conv format");
+        auto run = [param, kernel](size_t index, size_t thread_id) {
+            CpuNDRange ndrange_id(kernel.global_size, index);
+            kernel.kern(param, {thread_id, ndrange_id});
+        };
+        static_cast<naive::HandleImpl*>(fallback_handle)
+                ->dispatch_kern(run, kernel.global_size.total_size());
+    }
+}
+
 void ConvolutionImpl::exec_with_ncb_kern(const NCBKernParam& param,
                                          Algorithm* algo) {
-    auto kerns = ncb_algo_dispatch_kern(algo, param);
+    auto kerns = static_cast<AlgoBase*>(algo)->dispatch_kern(this, param);
     auto fallback_handle = handle();
     for (auto kernel : kerns) {
         megdnn_assert(param.filter_meta.format == Param::Format::NCHW ||
@@ -215,10 +293,13 @@ ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_heuristic_with_ncb(
         const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
         bool reproducible) {
     for (auto i : get_all_algorithms_with_ncb(param)) {
-        if (static_cast<AlgoBase*>(i)->usable_reproducible(
-                    this, param, AlgoSelectionStrategy::HEURISTIC,
-                    reproducible) &&
-            ncb_algo_get_workspace(i, param) <= workspace_limit_in_bytes) {
+        size_t need_workspace =
+                static_cast<AlgoBase*>(i)->get_workspace(this, param);
+        bool usable_reproducible =
+                static_cast<AlgoBase*>(i)->usable_reproducible(
+                        this, param, AlgoSelectionStrategy::HEURISTIC,
+                        reproducible);
+        if (usable_reproducible && need_workspace <= workspace_limit_in_bytes) {
             return i;
         }
     }

dnn/src/fallback/convolution/opr_impl.h

@@ -39,12 +39,26 @@ public:
               _megdnn_tensor_out dst, const PreprocessedFilter*,
               _megdnn_workspace workspace) override;
 
+    void exec_preprocess(const TensorLayout& src_layout,
+                         _megdnn_tensor_in filter,
+                         const TensorLayout& dst_layout,
+                         PreprocessedFilter* preprocessed_filter,
+                         _megdnn_workspace workspace) override;
+
     //! implemented by get_workspace_with_ncb()
     size_t get_workspace_in_bytes(const TensorLayout& src,
                                   const TensorLayout& filter,
                                   const TensorLayout& dst,
                                   const PreprocessedFilter*) override;
 
+    SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& dst) override;
+
+    size_t get_preprocess_workspace_in_bytes(const TensorLayout& src,
+                                             const TensorLayout& filter,
+                                             const TensorLayout& dst) override;
+
     //! implemented by get_all_algorithms_with_ncb()
     std::vector<Algorithm*> get_all_algorithms(
             const TensorLayout& src, const TensorLayout& filter,
@@ -70,6 +84,8 @@ public:
         ptrdiff_t inp_s[4], out_s[4];
         Param::ComputeMode compute_mode;
         size_t nr_threads;
+        //! weight_preprocess info
+        const PreprocessedFilter* preprocessed_filter;
     };
 
     //! memory param for kernels with non-contiguous batch
@@ -169,6 +185,23 @@ public:
         virtual SmallVector<NCBKern> dispatch_kern(
                 ConvolutionImpl* opr, const NCBKernSizeParam& param) const = 0;
 
+        virtual SmallVector<NCBKern> dispatch_preprocess_kern(
+                ConvolutionImpl*, const NCBKernSizeParam&) const {
+            return {};
+        };
+
+        //! get the layouts of weight_prerocess dst
+        virtual SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
+                ConvolutionImpl*, const NCBKernSizeParam&) const {
+            return {};
+        };
+
+        //! get the workspace when weight_prerocess
+        virtual size_t get_preprocess_workspace(ConvolutionImpl*,
+                                                const NCBKernSizeParam&) const {
+            return 0_z;
+        };
+
         //! Temporarily used to identify whether the matmul algorithm is
         //! is_preferred.
         virtual bool is_preferred(ConvolutionImpl*,
@@ -192,6 +225,9 @@ public:
 protected:
     virtual void exec_with_ncb_kern(const NCBKernParam& param, Algorithm* algo);
 
+    virtual void exec_preprocess_with_ncb_kern(const NCBKernParam& param,
+                                               Algorithm* algo);
+
     virtual std::vector<Algorithm*> get_all_algorithms_with_ncb(
             const NCBKernSizeParam& param);
 
@@ -199,21 +235,6 @@ protected:
             const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
             bool reproducible = false);
 
-    //! get kernel pointer
-    virtual SmallVector<NCBKern> ncb_algo_dispatch_kern(
-            Algorithm* algo, const NCBKernSizeParam& param) {
-        return static_cast<AlgoBase*>(algo)->dispatch_kern(this, param);
-    }
-    //! get algo workspace
-    virtual size_t ncb_algo_get_workspace(Algorithm* algo,
-                                          const NCBKernSizeParam& param) {
-        return static_cast<AlgoBase*>(algo)->get_workspace(this, param);
-    }
-    /*!
-     * the default impl iterates over all ncb_1g_get_all_algorithms()
-     * and return the first one whose workspace does not exceed the limit.
-     */
-
     const char* get_algorithm_set_name() const override;
 
     class AlgoFallback;
@@ -231,14 +252,16 @@ private:
             const NCBKernSizeParam& param,
             size_t workspace_size = std::numeric_limits<size_t>::max());
 
-    NCBKernSizeParam make_ncb_kern_size_param(const TensorLayout& src,
-                                              const TensorLayout& filter,
-                                              const TensorLayout& dst);
-
-    NCBKernParam make_ncb_kern_param(_megdnn_tensor_in src,
-                                     _megdnn_tensor_in filter,
-                                     _megdnn_tensor_out dst,
-                                     _megdnn_workspace workspace);
+    NCBKernSizeParam make_ncb_kern_size_param(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& dst,
+            const PreprocessedFilter* preprocessed_filter);
+
+    NCBKernParam make_ncb_kern_param(
+            _megdnn_tensor_in src, _megdnn_tensor_in filter,
+            _megdnn_tensor_out dst,
+            const PreprocessedFilter* preprocessed_filter,
+            _megdnn_workspace workspace);
 };
 
 class ConvolutionBackwardDataImpl : public naive::ConvolutionBackwardDataImpl {

dnn/src/naive/conv_bias/opr_impl.cpp

@@ -80,14 +80,15 @@ size_t ConvBiasForwardImpl::get_workspace_in_bytes(const TensorLayout& src,
 void ConvBiasForwardImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                                _megdnn_tensor_in bias, _megdnn_tensor_in z,
                                _megdnn_tensor_out dst,
-                               const PreprocessedFilter*,
+                               const PreprocessedFilter* preprocessed_filter,
                                _megdnn_workspace workspace) {
     MIDOUT_BEGIN(megdnn_naive_conv_bias_fwd) {
         dt_byte *workspace_ptr = workspace.raw_ptr;
         // ============================w * f + b================================
 
-        auto filter_meta = check_exec(src.layout, filter.layout, bias.layout,
-                                      z.layout, dst.layout, workspace.size);
+        auto filter_meta =
+                check_exec(src.layout, filter.layout, bias.layout, z.layout,
+                           dst.layout, workspace.size, preprocessed_filter);
         auto sfb = dst;
         if (bias.layout.dtype.enumv() != dst.layout.dtype.enumv()) {
             // intermediate result

dnn/src/naive/conv_bias/opr_impl.h

@@ -61,9 +61,7 @@ public:
     void exec_preprocess(const TensorLayout&, _megdnn_tensor_in,
                          const TensorLayout&, const TensorLayout&,
                          const TensorLayout&, PreprocessedFilter*,
-                         _megdnn_workspace) override{
-        megdnn_throw("conv_bias exec_preprocess is not impl yet");
-    }
+                         _megdnn_workspace) override {}
 
     const char* get_algorithm_set_name() const override;
 };

dnn/src/naive/convolution/convolution.cpp

@@ -28,11 +28,11 @@ using namespace naive;
 void ConvolutionForwardImpl::exec(_megdnn_tensor_in src,
                                   _megdnn_tensor_in filter,
                                   _megdnn_tensor_out dst,
-                                  const PreprocessedFilter*,
+                                  const PreprocessedFilter* preprocessed_filter,
                                   _megdnn_workspace workspace) {
     MIDOUT_BEGIN(megdnn_naive_conv_fwd) {
         auto filter_meta = check_exec(src.layout, filter.layout, dst.layout,
-                                      workspace.size);
+                                      workspace.size, preprocessed_filter);
         using ComputeMode = Param::ComputeMode;
 #define DISPATCH_CMODE(in_dt, out_dt, in_ct, out_ct, comp_ct, cmode)      \
     do {                                                                  \

dnn/src/naive/convolution/opr_impl.h

@@ -44,9 +44,7 @@ class ConvolutionForwardImpl: public ConvolutionForward {
 
         void exec_preprocess(const TensorLayout&, _megdnn_tensor_in,
                              const TensorLayout&, PreprocessedFilter*,
-                             _megdnn_workspace) override {
-            megdnn_throw("convolution exec_preprocess in not impl yet");
-        }
+                             _megdnn_workspace) override {}
 
         SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
                 const TensorLayout& , const TensorLayout& ,

dnn/test/common/opr_proxy.h

@@ -18,6 +18,9 @@
 #include "test/common/workspace_wrapper.h"
 
 #include <algorithm>
+#include <memory>
+
+
 
 namespace megdnn {
 namespace test {
@@ -32,6 +35,9 @@ struct OprProxyDefaultImpl
 template <typename Opr>
 struct OprProxy : public OprProxyDefaultImpl<Opr> {};
 
+template <typename Opr>
+struct OprWeightPreprocessProxy : public OprProxyDefaultImpl<Opr> {};
+
 template <typename Opr>
 struct OprProxyVectorToSingle {};
 
@@ -139,6 +145,28 @@ struct OprProxyProfilingBase
     typename Opr::Algorithm* target_algo = nullptr;
 
     OprProxyProfilingBase(bool profile = false) { m_profiling = profile; }
+
+    //! used for alloc tensor for weight preprocess
+    static std::shared_ptr<TensorNDArray> alloc_tensors(
+            Handle* handle, const TensorLayoutArray& layouts) {
+        auto deleter = [handle](TensorNDArray* ptr) {
+            for (auto&& i : *ptr) {
+                auto pdata = static_cast<dt_byte*>(i.raw_ptr) +
+                             i.layout.span().low_byte;
+                megdnn_free(handle, pdata);
+            }
+            delete ptr;
+        };
+        std::shared_ptr<TensorNDArray> ret{new TensorNDArray, deleter};
+        for (size_t i = 0; i < layouts.size(); ++i) {
+            auto span = layouts[i].span();
+            ret->emplace_back(static_cast<dt_byte*>(
+                                      megdnn_malloc(handle, span.dist_byte())) -
+                                      span.low_byte,
+                              layouts[i]);
+        }
+        return ret;
+    }
 };
 
 template <class Opr>
@@ -207,7 +235,6 @@ DEF_PROF3(LocalShareBackwardData);
 DEF_PROF3(LocalShareBackwardFilter);
 #undef DEF_PROF3
 
-//! TODO: it should adapt weight preprocess later
 template <>
 struct OprProxy<ConvolutionForward>
         : public OprProxyProfilingTernary<ConvolutionForward> {
@@ -263,6 +290,100 @@ struct OprProxy<ConvolutionForward>
     }
 };
 
+template <>
+struct OprWeightPreprocessProxy<ConvolutionForward>
+        : public OprProxyProfilingTernary<ConvolutionForward> {
+    using OprProxyProfilingTernary<ConvolutionForward>::OprProxyProfilingTernary;
+    void exec(ConvolutionForward* opr, const TensorNDArray& tensors) {
+        megdnn_assert(tensors.size() == 3);
+        if (!Base::W.valid()) {
+            Base::W = WorkspaceWrapper(opr->handle(), 0);
+        }
+        if (Base::m_profiling && !Base::target_algo) {
+            size_t min_time = std::numeric_limits<size_t>::max();
+            for (auto algo :
+                 opr->get_all_algorithms(tensors[0].layout, tensors[1].layout,
+                                         tensors[2].layout)) {
+                opr->execution_policy().algorithm = algo;
+
+                auto preprocess_tensors = weight_prerocess(opr, tensors, algo);
+                megcoreSynchronize(opr->handle()->megcore_computing_handle());
+                ConvolutionForward::PreprocessedFilter preprocessed_filter{
+                        algo, *preprocess_tensors};
+
+                auto workspace_size = opr->get_workspace_in_bytes(
+                        tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                        &preprocessed_filter);
+                Base::W.update(workspace_size);
+
+                for (size_t times = 0; times < Base::warmup_times; ++times)
+                    opr->exec(tensors[0], tensors[1], tensors[2],
+                              &preprocessed_filter, Base::W.workspace());
+                megcoreSynchronize(opr->handle()->megcore_computing_handle());
+                Timer timer;
+                timer.start();
+                for (size_t times = 0; times < Base::exec_times; ++times) {
+                    opr->exec(tensors[0], tensors[1], tensors[2],
+                              &preprocessed_filter, Base::W.workspace());
+                }
+                megcoreSynchronize(opr->handle()->megcore_computing_handle());
+                timer.stop();
+                printf("%.3fms %s\n", timer.get_time_in_us() / 1e3,
+                       algo->name());
+                if (min_time > timer.get_time_in_us()) {
+                    min_time = timer.get_time_in_us();
+                    Base::target_algo = algo;
+                }
+            }
+            opr->execution_policy().algorithm = Base::target_algo;
+            auto preprocess_tensors =
+                    weight_prerocess(opr, tensors, Base::target_algo);
+            megcoreSynchronize(opr->handle()->megcore_computing_handle());
+            ConvolutionForward::PreprocessedFilter preprocessed_filter{
+                    Base::target_algo, *preprocess_tensors};
+            auto workspace_size = opr->get_workspace_in_bytes(
+                    tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                    &preprocessed_filter);
+            Base::W.update(workspace_size);
+        }
+        auto preprocess_tensors =
+                weight_prerocess(opr, tensors, Base::target_algo);
+        megcoreSynchronize(opr->handle()->megcore_computing_handle());
+        ConvolutionForward::PreprocessedFilter preprocessed_filter{
+                Base::target_algo, *preprocess_tensors};
+        if (!Base::target_algo) {
+            auto workspace_size = opr->get_workspace_in_bytes(
+                    tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                    &preprocessed_filter);
+            Base::W.update(workspace_size);
+        }
+        opr->exec(tensors[0], tensors[1], tensors[2], &preprocessed_filter,
+                  Base::W.workspace());
+    }
+
+    //! handle weight preprocess
+    std::shared_ptr<TensorNDArray> weight_prerocess(
+            ConvolutionForward* opr, const TensorNDArray& tensors,
+            ConvolutionForward::Algorithm* algo) {
+        auto weight_perprocess_layouts = opr->deduce_preprocessed_filter_layout(
+                tensors[0].layout, tensors[1].layout, tensors[2].layout);
+        auto preprocessed_filter_tensors_ptr =
+                alloc_tensors(opr->handle(), weight_perprocess_layouts);
+        ConvolutionForward::PreprocessedFilter preprocessed_filter{
+                algo, *preprocessed_filter_tensors_ptr};
+        size_t preprocess_workspace_size =
+                opr->get_preprocess_workspace_in_bytes(tensors[0].layout,
+                                                       tensors[1].layout,
+                                                       tensors[2].layout);
+        WorkspaceWrapper preprocess_workspace(opr->handle(),
+                                              preprocess_workspace_size);
+        opr->exec_preprocess(tensors[0].layout, tensors[1], tensors[2].layout,
+                             &preprocessed_filter,
+                             preprocess_workspace.workspace());
+        return preprocessed_filter_tensors_ptr;
+    }
+};
+
 
 template <class Opr>
 struct OprProxyProfiling5 : public OprProxyProfilingBase<Opr, 5> {
@@ -329,11 +450,9 @@ struct OprProxyProfiling5 : public OprProxyProfilingBase<Opr, 5> {
 
 DEF_PROF5(DeformableConvForward);
 DEF_PROF5(DeformableConvBackwardFilter);
-//DEF_PROF5(ConvBiasForward);
 DEF_PROF5(BatchConvBiasForward);
 #undef DEF_PROF5
 
-//! TODO: it should adapt weight preprocess later
 template <>
 struct OprProxy<ConvBiasForward> : public OprProxyProfiling5<ConvBiasForward> {
     using OprProxyProfiling5<ConvBiasForward>::OprProxyProfiling5;
@@ -390,6 +509,106 @@ struct OprProxy<ConvBiasForward> : public OprProxyProfiling5<ConvBiasForward> {
     }
 };
 
+template <>
+struct OprWeightPreprocessProxy<ConvBiasForward>
+        : public OprProxyProfiling5<ConvBiasForward> {
+    using OprProxyProfiling5<ConvBiasForward>::OprProxyProfiling5;
+    void exec(ConvBiasForward* opr, const TensorNDArray& tensors) {
+        megdnn_assert(tensors.size() == 5);
+        if (!Base::W.valid()) {
+            Base::W = WorkspaceWrapper(opr->handle(), 0);
+        }
+        if (Base::m_profiling && !Base::target_algo) {
+            size_t min_time = std::numeric_limits<size_t>::max();
+            for (auto algo :
+                 opr->get_all_algorithms(tensors[0].layout, tensors[1].layout,
+                                         tensors[2].layout, tensors[3].layout,
+                                         tensors[4].layout)) {
+                opr->execution_policy().algorithm = algo;
+
+                auto preprocess_tensors = weight_prerocess(opr, tensors, algo);
+                megcoreSynchronize(opr->handle()->megcore_computing_handle());
+                ConvBiasForward::PreprocessedFilter preprocessed_filter{
+                        algo, *preprocess_tensors};
+
+                auto workspace_size = opr->get_workspace_in_bytes(
+                        tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                        tensors[3].layout, tensors[4].layout,
+                        &preprocessed_filter);
+                Base::W.update(workspace_size);
+
+                for (size_t times = 0; times < Base::warmup_times; ++times)
+                    opr->exec(tensors[0], tensors[1], tensors[2], tensors[3],
+                              tensors[4], &preprocessed_filter,
+                              Base::W.workspace());
+                megcoreSynchronize(opr->handle()->megcore_computing_handle());
+                Timer timer;
+                timer.start();
+                for (size_t times = 0; times < Base::exec_times; ++times) {
+                    opr->exec(tensors[0], tensors[1], tensors[2], tensors[3],
+                              tensors[4], &preprocessed_filter,
+                              Base::W.workspace());
+                }
+                megcoreSynchronize(opr->handle()->megcore_computing_handle());
+                timer.stop();
+                printf("%.3fms %s\n", timer.get_time_in_us() / 1e3,
+                       algo->name());
+                if (min_time > timer.get_time_in_us()) {
+                    min_time = timer.get_time_in_us();
+                    Base::target_algo = algo;
+                }
+            }
+            opr->execution_policy().algorithm = Base::target_algo;
+            auto preprocess_tensors =
+                    weight_prerocess(opr, tensors, Base::target_algo);
+            megcoreSynchronize(opr->handle()->megcore_computing_handle());
+            ConvBiasForward::PreprocessedFilter preprocessed_filter{
+                    Base::target_algo, *preprocess_tensors};
+            auto workspace_size = opr->get_workspace_in_bytes(
+                    tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                    tensors[3].layout, tensors[4].layout, &preprocessed_filter);
+            Base::W.update(workspace_size);
+        }
+        auto preprocess_tensors =
+                    weight_prerocess(opr, tensors, Base::target_algo);
+        megcoreSynchronize(opr->handle()->megcore_computing_handle());
+        ConvBiasForward::PreprocessedFilter preprocessed_filter{
+                Base::target_algo, *preprocess_tensors};
+        if (!Base::target_algo) {
+            auto workspace_size = opr->get_workspace_in_bytes(
+                    tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                    tensors[3].layout, tensors[4].layout, &preprocessed_filter);
+            Base::W.update(workspace_size);
+        }
+        opr->exec(tensors[0], tensors[1], tensors[2], tensors[3], tensors[4],
+                  &preprocessed_filter, Base::W.workspace());
+    }
+
+    //! handle weight preprocess
+    std::shared_ptr<TensorNDArray> weight_prerocess(
+            ConvBiasForward* opr, const TensorNDArray& tensors,
+            ConvBiasForward::Algorithm* algo) {
+        auto weight_perprocess_layouts = opr->deduce_preprocessed_filter_layout(
+                tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                tensors[3].layout, tensors[4].layout);
+        auto preprocessed_filter_tensors_ptr =
+                alloc_tensors(opr->handle(), weight_perprocess_layouts);
+        ConvBiasForward::PreprocessedFilter preprocessed_filter{
+                algo, *preprocessed_filter_tensors_ptr};
+        size_t preprocess_workspace_size =
+                opr->get_preprocess_workspace_in_bytes(
+                        tensors[0].layout, tensors[1].layout, tensors[2].layout,
+                        tensors[3].layout, tensors[4].layout);
+        WorkspaceWrapper preprocess_workspace(opr->handle(),
+                                              preprocess_workspace_size);
+        opr->exec_preprocess(tensors[0].layout, tensors[1], tensors[2].layout,
+                             tensors[3].layout, tensors[4].layout,
+                             &preprocessed_filter,
+                             preprocess_workspace.workspace());
+        return preprocessed_filter_tensors_ptr;
+    }
+};
+
 template <class Opr>
 struct OprProxyProfiling8 : public OprProxyProfilingBase<Opr, 8> {
     using Base = OprProxyProfilingBase<Opr, 8>;