feat(mbg/core): support bias preprocess in conv_bias

GitOrigin-RevId: d2e1e14d414a9dcb9a6816515100b854eedb8a0e

dnn/include/megdnn/oprs/nn.h

@@ -378,15 +378,15 @@ public:
                       _megdnn_workspace workspace) = 0;
 
     /**
-     * \brief execute weight preprocessing, read weights form filter and write
-     * to preprocessed_filter after preprocessed.
+     * \brief execute weight preprocessing, read weights form filter and bias,
+     * write to preprocessed_filter after preprocessed.
      *
      * \praram[in] workspace the needed tmp workspace when exec_preprocess
      * running, the size is got by get_preprocess_workspace_in_bytes
      */
     virtual void exec_preprocess(const TensorLayout& src_layout,
                                  _megdnn_tensor_in filter,
-                                 const TensorLayout& bias_layout,
+                                 _megdnn_tensor_in bias,
                                  const TensorLayout& z_layout,
                                  const TensorLayout& dst_layout,
                                  PreprocessedFilter* preprocessed_filter,

dnn/src/cuda/conv_bias/opr_impl.cpp

@@ -238,11 +238,11 @@ ConvBiasForwardImpl::deduce_preprocessed_filter_layout(
 
 void ConvBiasForwardImpl::exec_preprocess(
         const TensorLayout& src_layout, _megdnn_tensor_in filter,
-        const TensorLayout& bias_layout, const TensorLayout& z_layout,
+        _megdnn_tensor_in bias, const TensorLayout& z_layout,
         const TensorLayout& dst_layout, PreprocessedFilter* preprocessed_filter,
         _megdnn_workspace workspace) {
     TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout},
-            z{nullptr, z_layout}, bias{nullptr, bias_layout};
+            z{nullptr, z_layout};
     AlgoBase::ExecArgs args(this, src, filter, bias, z, dst, workspace,
                             preprocessed_filter);
     auto algo = get_algorithm(this, src.layout, filter.layout, bias.layout,

dnn/src/cuda/conv_bias/opr_impl.h

@@ -49,7 +49,7 @@ public:
             const TensorLayout&, const TensorLayout&, const TensorLayout&,
             const TensorLayout&, const TensorLayout&) override;
     void exec_preprocess(const TensorLayout&, _megdnn_tensor_in,
-                         const TensorLayout&, const TensorLayout&,
+                         _megdnn_tensor_in, const TensorLayout&,
                          const TensorLayout&, PreprocessedFilter*,
                          _megdnn_workspace) override;
     const char* get_algorithm_set_name() const override;

dnn/src/fallback/conv_bias/opr_impl.cpp

@@ -178,15 +178,14 @@ 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,
+                                   _megdnn_tensor_in bias,
                                    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};
+    //! exec_preprocess currently only support preprocess weights and bias
+    //! before exec, src/dst/z will be ignored, just set to nullptr
+    TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout};
     auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace,
                                       preprocessed_filter);
     //! should not pass workspace_size limit otherwise can not find match algo
@@ -196,7 +195,7 @@ void ConvBiasImpl::exec_preprocess(const TensorLayout& src_layout,
         exec_preprocess_with_ncb_kern(fparam, algo);
     } else {
         naive::ConvBiasForwardImpl::exec_preprocess(
-                src_layout, filter, bias_layout, z_layout, dst_layout,
+                src_layout, filter, bias, z_layout, dst_layout,
                 preprocessed_filter, workspace);
     }
 }

dnn/src/fallback/conv_bias/opr_impl.h

@@ -57,7 +57,7 @@ public:
 
     void exec_preprocess(const TensorLayout& src_layout,
                          _megdnn_tensor_in filter,
-                         const TensorLayout& bias_layout,
+                         _megdnn_tensor_in bias,
                          const TensorLayout& z_layout,
                          const TensorLayout& dst_layout,
                          PreprocessedFilter* preprocessed_filter,

dnn/src/naive/conv_bias/opr_impl.h

@@ -59,7 +59,7 @@ public:
     }
 
     void exec_preprocess(const TensorLayout&, _megdnn_tensor_in,
-                         const TensorLayout&, const TensorLayout&,
+                         _megdnn_tensor_in, const TensorLayout&,
                          const TensorLayout&, PreprocessedFilter*,
                          _megdnn_workspace) override {}
 

dnn/test/common/opr_proxy.h

@@ -601,7 +601,7 @@ struct OprWeightPreprocessProxy<ConvBiasForward>
                         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,
+        opr->exec_preprocess(tensors[0].layout, tensors[1], tensors[2],
                              tensors[3].layout, tensors[4].layout,
                              &preprocessed_filter,
                              preprocess_workspace.workspace());

src/core/test/graph/misc.cpp

@@ -1955,6 +1955,39 @@ typename DnnOp::Algorithm* try_find_any_weight_preprocess_algo(
     return nullptr;
 }
 
+template <typename DnnOp, typename... Args>
+typename DnnOp::Algorithm* try_find_any_bias_preprocess_algo(
+        DnnOp* dnn_op, const char* mgb_info, Maybe<bool>& found,
+        Args&& ...args) {
+    if (found.valid()) {
+        if (found.val()) {
+            return dnn_op->execution_policy().algorithm;
+        } else {
+            return nullptr;
+        }
+    }
+    for (auto&& algo : dnn_op->get_all_algorithms(
+            std::forward<Args>(args)...)) {
+        dnn_op->execution_policy().algorithm = algo;
+        auto layouts = dnn_op->deduce_preprocessed_filter_layout(
+                std::forward<Args>(args)...);
+        if (layouts.size() <= 1)
+            continue;
+        bool valid = false;
+        if (!layouts[1].is_empty()) {
+            valid = true;
+            break;
+        }
+        if (valid) {
+            found.emplace(true);
+            return algo;
+        }
+    }
+    found.emplace(false);
+    mgb_log_warn("Can't find bias preprocess algo for op %s", mgb_info);
+    return nullptr;
+}
+
 void test_free_memory_in_weight_preprocess(int record_level, CompNode cn) {
     HostTensorGenerator<> gen;
     auto graph = ComputingGraph::make();
@@ -2152,4 +2185,54 @@ TEST(TestGraph, FreeMemoryInWeightPreprocessWithMultiReader) {
                         .empty());
 }
 
+TEST(TestGraph, FreeBias) {
+    HostTensorGenerator<> gen;
+    auto graph = ComputingGraph::make();
+    auto cn = CompNode::load("xpu0");
+    graph->options().graph_opt.weight_preprocess = true;
+    auto mkvar = [&](const char* name, const TensorShape& shp) {
+        return opr::Host2DeviceCopy::make(*graph, gen(shp, cn)).rename(name);
+    };
+    auto mkcvar = [&](const char* name, const TensorShape& shp) {
+        return opr::SharedDeviceTensor::make_const(*graph, *gen(shp, cn))
+                .rename(name);
+    };
+    auto x = mkvar("x", {1, 32, 16, 16});
+    // ConvBias test dense
+    opr::ConvBias::Param param_conv_bias;
+    param_conv_bias.pad_h = param_conv_bias.pad_w = 0;
+    param_conv_bias.sparse = opr::ConvBias::Param::Sparse::DENSE;
+    auto w1 = mkcvar("w1", {32, 32, 1, 1}), b1 = mkcvar("b1", {1, 32, 1, 1});
+    auto conv1 = opr::ConvBias::make(x, w1, b1, param_conv_bias);
+    Maybe<bool> wp1;
+    conv1.node()->owner_opr()->cast_final_safe<opr::ConvBias>()
+        .setup_algo_chooser([&](const cg::OperatorNodeBase* opr) {
+            return try_find_any_bias_preprocess_algo(
+                opr->cast_final_safe<opr::ConvBias>().megdnn_opr(),
+                opr->cname(), wp1,
+                opr->input(0)->layout(), opr->input(1)->layout(),
+                opr->input(2)->layout(), TensorLayout{},
+                opr->output(0)->layout());
+    });
+
+    HostTensorND host_y;
+    auto func =graph->compile({make_callback_copy(conv1, host_y)});
+    //!flag the no need memory of var
+    func->execute();
+    //!free the no need memory of var
+    func->execute();
+    auto check = [&](SymbolVar v) {
+        ASSERT_TRUE(v.node()->contain_flag(VarNode::Flag::MEMORY_NO_NEED));
+        ASSERT_TRUE(v.node()->dev_tensor().empty());
+        ASSERT_TRUE(v.node()->owner_opr()
+                            ->cast_final_safe<opr::SharedDeviceTensor>()
+                            .get_dev_tensor()
+                            .empty());
+    };
+    ASSERT_TRUE(wp1.valid());
+    if (wp1.val()) {
+        check(b1);
+    }
+}
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/dnn/convolution.cpp

@@ -961,17 +961,37 @@ void ConvBiasForward::scn_do_execute_preprocess() {
         z_layout = input(3)->layout();
     }
     megdnn_opr()->exec_preprocess(
-            input(0)->layout(), input(1)->dev_tensor().as_megdnn(), bias_layout,
-            z_layout, output(0)->layout(), preprocessed_filter(),
+            input(0)->layout(), input(1)->dev_tensor().as_megdnn(),
+            input(2)->dev_tensor().as_megdnn(), z_layout, output(0)->layout(),
+            preprocessed_filter(),
             intl::get_megdnn_workspace_from_var(output().back()));
-    //! Flag the input(1) no use later, which can be freed when no other
+    //! Flag the weight and bias no use later, which can be freed when no other
     //! var depend on its dev_value, host_value and shape.
-    auto receiver_info =
+    auto receiver_info_weight =
             input(1)->owner_graph()->var_receiver_in_current_comp_seq(input(1));
-    if (receiver_info.dev_value == 1 && receiver_info.host_value == 0 &&
-        receiver_info.shape == 0) {
+    if (receiver_info_weight.dev_value == 1 &&
+        receiver_info_weight.host_value == 0 &&
+        receiver_info_weight.shape == 0) {
         input(1)->add_flag(VarNode::Flag::MEMORY_NO_NEED);
     }
+    //! if bias is preprocessd
+    if (input().size() > 3) {
+        auto preprocessed_layouts =
+                megdnn_opr()->deduce_preprocessed_filter_layout(
+                        input(0)->layout(), input(1)->layout(), bias_layout,
+                        z_layout, output(0)->layout());
+        if (preprocessed_layouts.size() > 1 &&
+            !preprocessed_layouts[1].is_empty()) {
+            auto receiver_info_bias =
+                    input(2)->owner_graph()->var_receiver_in_current_comp_seq(
+                            input(2));
+            if (receiver_info_bias.dev_value == 1 &&
+                receiver_info_bias.host_value == 0 &&
+                receiver_info_bias.shape == 0) {
+                input(2)->add_flag(VarNode::Flag::MEMORY_NO_NEED);
+            }
+        }
+    }
 }
 
 /* ===================== LocalShareForward ==================== */

src/opr/impl/search_policy/profiler.cpp

@@ -178,9 +178,26 @@ typename TimedProfiler<Opr>::TResult TimedProfiler<Opr>::prof_impl(
             for (size_t i = 0; i < flt_val.size(); i++) {
                 pf.tensors[i] = flt_val[i].as_megdnn();
             }
-            APPLY(_(megdnn_opr)->exec_preprocess(args..., &pf, mdn_workspace),
-                  std::forward_as_tuple(layouts[0], inp_val[1].as_megdnn()),
-                  array_skip<2>(layouts));
+            if_constexpr<opr_contain_bias<Opr>()>(
+                    //! convbias
+                    [&](auto __) {
+                        APPLY(__(megdnn_opr)
+                                      ->exec_preprocess(args..., &pf,
+                                                        mdn_workspace),
+                              std::forward_as_tuple(layouts[0],
+                                                    inp_val[1].as_megdnn(),
+                                                    inp_val[2].as_megdnn()),
+                              array_skip<arity_in - 1>(layouts));
+                    },
+                    //! Convolution
+                    [&](auto __) {
+                        APPLY(__(megdnn_opr)
+                                      ->exec_preprocess(args..., &pf,
+                                                        mdn_workspace),
+                              std::forward_as_tuple(layouts[0],
+                                                    inp_val[1].as_megdnn()),
+                              array_skip<2>(layouts));
+                    });
         }
     });
 

src/opr/include/megbrain/opr/search_policy/profiler.h

@@ -75,6 +75,11 @@ constexpr bool opr_supports_preprocess() {
            std::is_same<Opr, megdnn::ConvBias>::value;
 }
 
+template <typename Opr>
+constexpr bool opr_contain_bias() {
+    return std::is_same<Opr, megdnn::ConvBias>::value;
+}
+
 template <typename Opr, bool has_prep>
 struct PreprocessFilterImpl {
     using T = union {};