feat(gopt): add channel padding pass when enable nchw44/nchw88/nchw44-dot

GitOrigin-RevId: fbe0516ffc

src/gopt/impl/framework.cpp

@@ -743,20 +743,28 @@ const GraphOptimizer& GraphOptimizer::add_passes_for_optimize_options(
     });
     cb(nhwcd4, {
         add_pass<FuseConvBiasNonlinPass>();
+        add_pass(PaddingChannelPass::make(
+                cg::GraphCommonOptimizeOptions::LayoutTransform::NHWCD4, true));
         add_pass(ConvertFormatPass::make_nhwcd4_converter());
     });
     cb(nchw88, {
         add_pass<FuseConvBiasNonlinPass>();
+        add_pass(PaddingChannelPass::make(
+                cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW88, true));
         add_pass(EnableNchwxxPass::make_nchwxx_converter(8));
         add_pass<ShuffleShuffleRemovePass>();
     });
     cb(nchw44, {
         add_pass<FuseConvBiasNonlinPass>();
+        add_pass(PaddingChannelPass::make(
+                cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44, true));
         add_pass(EnableNchwxxPass::make_nchwxx_converter(4));
         add_pass<ShuffleShuffleRemovePass>();
     });
     cb(nchw44_dot, {
         add_pass<FuseConvBiasNonlinPass>();
+        add_pass(PaddingChannelPass::make(
+                cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44_DOT, true));
         add_pass(EnableNchw44DotPass::make_nchw44_dot_converter());
         add_pass<ShuffleShuffleRemovePass>();
     });
@@ -784,7 +792,7 @@ const GraphOptimizer& GraphOptimizer::add_passes_for_optimize_options(
     cb(nchw64, {
         add_pass<FuseConvBiasNonlinPass>();
         add_pass(PaddingChannelPass::make(
-                cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW64));
+                cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW64, false));
         add_pass<FuseConvBiasZPass>();
         add_pass(EnableNCHW64Pass::make_nchw64_converter());
         add_pass<ShuffleShuffleRemovePass>();

src/gopt/impl/padding_channel.cpp

@@ -1,5 +1,6 @@
 #include "megbrain/gopt/inference.h"
 #include "megbrain/opr/basic_arith.h"
+#include "megbrain/opr/dnn/adaptive_pooling.h"
 #include "megbrain/opr/dnn/convolution.h"
 #include "megbrain/opr/dnn/pooling.h"
 #include "megbrain/opr/imgproc.h"
@@ -34,8 +35,8 @@ using ReformatKey = ReformatManager::ReformatKey;
 
 /* ==================== PaddingChannelPass ================= */
 namespace {
-size_t padding_int4(size_t in_channel, bool flag) {
-    static_cast<void>(flag);
+
+size_t padding_int4(size_t in_channel, bool) {
     if (in_channel <= 32) {
         return (8 - (in_channel % 8)) % 8;
     } else {
@@ -43,6 +44,8 @@ size_t padding_int4(size_t in_channel, bool flag) {
     }
 }
 
+//! flag is used by user to identify some case, such as in nchw64, flag is used
+//! to identify the convbias and convolution backward
 size_t padding_int8(size_t in_channel, bool flag) {
     if (flag) {
         if (in_channel <= 16) {
@@ -58,24 +61,41 @@ size_t padding_4(size_t in_channel, bool) {
     return (4 - (in_channel % 4)) % 4;
 };
 
+size_t padding_8(size_t in_channel, bool) {
+    return (8 - (in_channel % 8)) % 8;
+};
+
 }  // namespace
 
 std::unique_ptr<PaddingChannelPass> PaddingChannelPass::make(
-        cg::GraphCommonOptimizeOptions::LayoutTransform layout_transform) {
+        cg::GraphCommonOptimizeOptions::LayoutTransform layout_transform,
+        bool only_padding_weights) {
     MIDOUT_B("PaddingChannelPass::make")
     using LayoutTrans = cg::GraphCommonOptimizeOptions::LayoutTransform;
-    auto ret = std::make_unique<PaddingChannelPass>();
+    auto ret = std::unique_ptr<PaddingChannelPass>(
+            new PaddingChannelPass(only_padding_weights));
     auto& alignment_map = ret->m_alignment_map;
     if (layout_transform == LayoutTrans::NCHW64) {
         alignment_map[DTypeEnum::QuantizedS4] = padding_int4;
         alignment_map[DTypeEnum::Quantized4Asymm] = padding_int4;
         alignment_map[DTypeEnum::QuantizedS8] = padding_int8;
     } else if (
+            layout_transform == LayoutTrans::NHWCD4 ||
             layout_transform == LayoutTrans::NCHW44 ||
             layout_transform == LayoutTrans::NCHW44_DOT) {
         alignment_map[DTypeEnum::QuantizedS8] = padding_4;
         alignment_map[DTypeEnum::Quantized8Asymm] = padding_4;
         alignment_map[DTypeEnum::Float32] = padding_4;
+#if !MEGDNN_DISABLE_FLOAT16
+        alignment_map[DTypeEnum::Float16] = padding_4;
+#endif
+    } else if (layout_transform == LayoutTrans::NCHW88) {
+        alignment_map[DTypeEnum::QuantizedS8] = padding_8;
+        alignment_map[DTypeEnum::Quantized8Asymm] = padding_8;
+        alignment_map[DTypeEnum::Float32] = padding_8;
+#if !MEGDNN_DISABLE_FLOAT16
+        alignment_map[DTypeEnum::Float16] = padding_8;
+#endif
     }
     ret->fill_opr_convert_fun(layout_transform);
     return ret;
@@ -138,6 +158,10 @@ VarNode* PaddingChannelPass::extract_subtensor(
     mgb_assert(inp->shape()[2] == orig_shape[2]);
     mgb_assert(inp->shape()[3] == orig_shape[3]);
     size_t orig_channels = orig_shape[1];
+    //! if channel is not padding, do nothing
+    if (orig_channels == inp->shape()[1]) {
+        return inp;
+    }
     auto x = SymbolVar(inp);
     auto cv = [&x](int v) { return x.make_scalar(v); };
     using AIdx = opr::Subtensor::AxisIndexer;
@@ -150,8 +174,25 @@ VarNode* PaddingChannelPass::extract_subtensor(
 };
 
 VarNode* PaddingChannelPass::pad_in_channels(VarNode* inp, size_t pad_channels) {
-    mgb_assert(inp->shape().ndim == 4);
-    TensorShape shape{inp->shape()[0], pad_channels, inp->shape()[2], inp->shape()[3]};
+    TensorShape shape;
+    size_t axis = 0;
+    if (inp->shape().ndim == 4) {
+        shape = TensorShape{
+                inp->shape()[0], pad_channels, inp->shape()[2], inp->shape()[3]};
+        axis = 1;
+    } else {
+        mgb_assert(inp->shape().ndim == 5);
+        //! the channel wise convolution
+        if (inp->shape()[1] == 1 && inp->shape()[2] == 1) {
+            shape = TensorShape{
+                    pad_channels, inp->shape()[1], inp->shape()[2], inp->shape()[3],
+                    inp->shape()[4]};
+            axis = 0;
+        } else {
+            //! the group convolution
+            mgb_assert(0, "group convolution can't padding cahnnel\n");
+        }
+    }
     std::shared_ptr<HostTensorND> host_val =
             std::make_shared<HostTensorND>(inp->comp_node(), inp->dtype());
     host_val->resize(shape);
@@ -159,13 +200,30 @@ VarNode* PaddingChannelPass::pad_in_channels(VarNode* inp, size_t pad_channels)
     size_t size_bytes = TensorLayout{shape, inp->dtype()}.span().dist_byte();
     std::memset(ptr, 0, size_bytes);
     auto padding = opr::ImmutableTensor::make(*inp->owner_graph(), *host_val);
-    auto out = opr::Concat::make({inp, padding}, 1);
+    auto out = opr::Concat::make({inp, padding}, axis);
     return out.node();
 };
 
 VarNode* PaddingChannelPass::pad_out_channels(VarNode* inp, size_t pad_channels) {
-    mgb_assert(inp->shape().ndim == 4);
-    TensorShape shape{pad_channels, inp->shape()[1], inp->shape()[2], inp->shape()[3]};
+    TensorShape shape;
+    size_t axis = 0;
+    if (inp->shape().ndim == 4) {
+        shape = TensorShape{
+                pad_channels, inp->shape()[1], inp->shape()[2], inp->shape()[3]};
+        axis = 0;
+    } else {
+        mgb_assert(inp->shape().ndim == 5);
+        //! the channel wise convolution
+        if (inp->shape()[1] == 1 && inp->shape()[2] == 1) {
+            shape = TensorShape{
+                    pad_channels, inp->shape()[1], inp->shape()[2], inp->shape()[3],
+                    inp->shape()[4]};
+            axis = 0;
+        } else {
+            //! the group convolution
+            mgb_assert(0, "group convolution can't padding cahnnel\n");
+        }
+    }
     std::shared_ptr<HostTensorND> host_val =
             std::make_shared<HostTensorND>(inp->comp_node(), inp->dtype());
     host_val->resize(shape);
@@ -173,15 +231,15 @@ VarNode* PaddingChannelPass::pad_out_channels(VarNode* inp, size_t pad_channels)
     size_t size_bytes = TensorLayout{shape, inp->dtype()}.span().dist_byte();
     std::memset(ptr, 0, size_bytes);
     auto padding = opr::ImmutableTensor::make(*inp->owner_graph(), *host_val);
-    auto out = opr::Concat::make({inp, padding}, 0);
+    auto out = opr::Concat::make({inp, padding}, axis);
     return out.node();
 };
 
-// padding policy for conv bias with data type qint8
-OperatorNodeBase* PaddingChannelPass::padding_policy(
+// padding policy for dense convolution
+OperatorNodeBase* PaddingChannelPass::padding_conv_policy(
         OperatorNodeBase* opr, const VarNodeArray& new_inp) {
     mgb_assert(opr->input().size() == new_inp.size());
-    mgb_assert(new_inp.size() == 3);
+    mgb_assert(new_inp.size() >= 2);
     //! new weights and old weights are same shape
     mgb_assert(opr->input(1)->shape().eq_shape(new_inp[1]->shape()));
     auto inps = new_inp;
@@ -198,7 +256,8 @@ OperatorNodeBase* PaddingChannelPass::padding_policy(
     if (m_padding_oprs.count(opr->input(0)->owner_opr())) {
         //! as the opr of input var is padding, but the dtype of input and output of
         //! the input opr maybe different, so the alignment is not the same
-        size_t pad_channels_0 = it->second(new_in_channels, true);
+        size_t pad_channels_0 =
+                m_only_padding_weights ? 0 : it->second(new_in_channels, true);
         size_t pad_channels_1 = it->second(in_channels, true);
         if (pad_channels_0) {
             inps[0] = pad_in_channels(new_inp[0], pad_channels_0);
@@ -211,7 +270,7 @@ OperatorNodeBase* PaddingChannelPass::padding_policy(
     } else {
         mgb_assert(new_in_channels == in_channels);
         size_t pad_channels = it->second(in_channels, true);
-        if (pad_channels > 0) {
+        if (pad_channels > 0 && !m_only_padding_weights) {
             inps[0] = pad_in_channels(new_inp[0], pad_channels);
             inps[1] = pad_in_channels(new_inp[1], pad_channels);
         }
@@ -220,31 +279,63 @@ OperatorNodeBase* PaddingChannelPass::padding_policy(
     size_t pad_channels = it->second(out_channels, true);
     if (pad_channels > 0) {
         inps[1] = pad_out_channels(inps[1], pad_channels);
-        inps[2] = pad_in_channels(inps[2], pad_channels);
+        if (inps.size() >= 3) {
+            inps[2] = pad_in_channels(inps[2], pad_channels);
+        }
         m_padding_oprs.insert(opr);
     }
     return serialization::copy_opr_shallow(*opr, inps, opr->config());
 };
 
+//! padding policy for channel wise convolution
+OperatorNodeBase* PaddingChannelPass::padding_channel_wise_conv_policy(
+        OperatorNodeBase* opr, const VarNodeArray& new_inp) {
+    mgb_assert(opr->input().size() == new_inp.size());
+    mgb_assert(opr->input()[1]->shape().ndim == 5);
+    mgb_assert(new_inp.size() >= 2);
+    //! new weights and old weights are same shape
+    mgb_assert(opr->input(1)->shape().eq_shape(new_inp[1]->shape()));
+    auto inps = new_inp;
+    size_t group = opr->input(1)->shape()[0];
+    size_t new_in_channels = new_inp[0]->shape()[1];
+    auto it = m_alignment_map.find(opr->input(0)->dtype().enumv());
+    if (it != m_alignment_map.end()) {
+        mgb_assert(it->second);
+    } else {
+        return serialization::copy_opr_shallow(*opr, inps, opr->config());
+    }
+    // pad input channels
+    if (m_padding_oprs.count(opr->input(0)->owner_opr())) {
+        size_t pad_channels_1 = new_in_channels - group;
+        if (pad_channels_1) {
+            inps[1] = pad_in_channels(new_inp[1], pad_channels_1);
+            m_padding_oprs.insert(opr);
+        }
+    }
+    return serialization::copy_opr_shallow(*opr, inps, opr->config());
+};
+
 void PaddingChannelPass::fill_opr_convert_fun(LayoutTrans layout_trans) {
-    add_convbias_replace_func(layout_trans);
+    add_conv_replace_func(layout_trans);
     add_conv_backward_data_replace_func(layout_trans);
     add_format_aware_opr_replace_func(layout_trans);
     add_elemwise_like_opr_replace_func(layout_trans);
+    add_condition_padding_oprs_replace_func(layout_trans);
     add_nonpadding_oprs_replace_func(layout_trans);
 }
 
-void PaddingChannelPass::add_convbias_replace_func(LayoutTrans layout_trans) {
+void PaddingChannelPass::add_conv_replace_func(LayoutTrans layout_trans) {
     if (layout_trans == LayoutTrans::NCHW64) {
         m_opr_replace_funcs[opr::ConvBiasForward::typeinfo()] =
                 [this](OperatorNodeBase* opr, const VarNodeArray& new_inp) {
-                    if (opr->input(0)->dtype().enumv() == DTypeEnum::QuantizedS8) {
-                        return padding_policy(opr, new_inp);
-                    } else if (
-                            opr->input(0)->dtype().enumv() == DTypeEnum::QuantizedS4 ||
-                            opr->input(0)->dtype().enumv() ==
-                                    DTypeEnum::Quantized4Asymm) {
-                        return padding_policy(opr, new_inp);
+                    mgb_assert(
+                            opr->input()[1]->shape().ndim == 4,
+                            "nchw64 format only support padding channel in dense "
+                            "convolution\n");
+                    if (opr->input(0)->dtype().enumv() == DTypeEnum::QuantizedS8 ||
+                        opr->input(0)->dtype().enumv() == DTypeEnum::QuantizedS4 ||
+                        opr->input(0)->dtype().enumv() == DTypeEnum::Quantized4Asymm) {
+                        return padding_conv_policy(opr, new_inp);
                     } else {
                         mgb_assert(
                                 m_padding_oprs.count(opr->input(0)->owner_opr()) == 0,
@@ -257,11 +348,36 @@ void PaddingChannelPass::add_convbias_replace_func(LayoutTrans layout_trans) {
                                 *opr, new_inp, opr->config());
                     }
                 };
-    } else if (layout_trans == LayoutTrans::NCHW44) {
-        m_opr_replace_funcs[opr::ConvBiasForward::typeinfo()] =
-                [this](OperatorNodeBase* opr, const VarNodeArray& new_inp) {
-                    return padding_policy(opr, new_inp);
-                };
+    } else if (
+            layout_trans == LayoutTrans::NCHW44 ||
+            layout_trans == LayoutTrans::NCHW44_DOT ||
+            layout_trans == LayoutTrans::NCHW88) {
+        auto padding_conv = [this](OperatorNodeBase* opr, const VarNodeArray& new_inp) {
+            if (opr->input()[1]->shape().ndim == 4) {
+                return padding_conv_policy(opr, new_inp);
+            } else {
+                mgb_assert(opr->input()[1]->shape().ndim == 5);
+                if (opr->input()[1]->shape()[1] == 1 &&
+                    opr->input()[1]->shape()[2] == 1) {
+                    return padding_channel_wise_conv_policy(opr, new_inp);
+                } else {
+                    //! group convolution can't padding channel
+                    mgb_assert(opr->input().size() == new_inp.size());
+                    auto inps = new_inp;
+                    for (size_t i = 0; i < new_inp.size(); ++i) {
+                        auto cur_inp = opr->input(i);
+                        bool padding_cur_inp =
+                                m_padding_oprs.count(cur_inp->owner_opr()) > 0;
+                        if (padding_cur_inp) {
+                            inps[i] = extract_subtensor(inps[i], cur_inp->shape());
+                        }
+                    }
+                    return serialization::copy_opr_shallow(*opr, inps, opr->config());
+                }
+            }
+        };
+        m_opr_replace_funcs[opr::ConvBiasForward::typeinfo()] = padding_conv;
+        m_opr_replace_funcs[opr::Convolution::typeinfo()] = padding_conv;
     }
 }
 
@@ -298,7 +414,9 @@ void PaddingChannelPass::add_conv_backward_data_replace_func(LayoutTrans layout_
                         size_t pad_channels = new_out_channels - out_channels;
                         inps[0] = pad_out_channels(new_inp[0], pad_channels);
                     } else {
-                        size_t pad_channels = it->second(out_channels, false);
+                        size_t pad_channels = m_only_padding_weights
+                                                    ? 0
+                                                    : it->second(out_channels, false);
                         if (pad_channels > 0) {
                             inps[0] = pad_out_channels(new_inp[0], pad_channels);
                             inps[1] = pad_in_channels(new_inp[1], pad_channels);
@@ -311,26 +429,45 @@ void PaddingChannelPass::add_conv_backward_data_replace_func(LayoutTrans layout_
                         inps[0] = pad_in_channels(inps[0], pad_channels);
                         m_padding_oprs.insert(opr);
                     }
+                    return serialization::copy_opr_shallow(*opr, inps, opr->config());
+                };
+    } else {
+        m_opr_replace_funcs[opr::ConvolutionBackwardData::typeinfo()] =
+                [this](OperatorNodeBase* opr, const VarNodeArray& new_inp) {
+                    mgb_assert(opr->input(0)->shape().eq_shape(new_inp[0]->shape()));
+                    auto inps = new_inp;
+                    size_t out_channels = opr->input(0)->shape()[0];
+                    size_t new_out_channels = new_inp[1]->shape()[1];
+                    // pad output channels
+                    if (m_padding_oprs.count(opr->input(1)->owner_opr())) {
+                        size_t pad_channels = new_out_channels - out_channels;
+                        inps[0] = pad_out_channels(new_inp[0], pad_channels);
+                    }
+                    out_channels = inps[0]->shape()[0];
+
                     return serialization::copy_opr_shallow(*opr, inps, opr->config());
                 };
     }
 }
 
-void PaddingChannelPass::add_format_aware_opr_replace_func(LayoutTrans) {
-    auto replace_format_aware_opr = [this](OperatorNodeBase* opr,
-                                           const VarNodeArray& new_inp) {
-        if (opr->input(0)->dtype().enumv() != DTypeEnum::QuantizedS8 &&
-            opr->input(0)->dtype().enumv() != DTypeEnum::QuantizedS4 &&
-            opr->input(0)->dtype().enumv() != DTypeEnum::Quantized4Asymm) {
-            mgb_assert(
-                    m_padding_oprs.count(opr->input(0)->owner_opr()) == 0,
-                    "operator(type:%s,name:%s) for data type(%s) cannot be "
-                    "padded channel. extra info:"
-                    "consumer(%s), producer(%s)",
-                    opr->dyn_typeinfo()->name, opr->cname(),
-                    opr->input(0)->dtype().name(), opr->cname(),
-                    opr->input(0)->owner_opr()->cname());
-            return serialization::copy_opr_shallow(*opr, new_inp, opr->config());
+void PaddingChannelPass::add_format_aware_opr_replace_func(LayoutTrans layout_trans) {
+    auto replace_format_aware_opr = [this, layout_trans](
+                                            OperatorNodeBase* opr,
+                                            const VarNodeArray& new_inp) {
+        if (layout_trans == LayoutTrans::NCHW64) {
+            if (opr->input(0)->dtype().enumv() != DTypeEnum::QuantizedS8 &&
+                opr->input(0)->dtype().enumv() != DTypeEnum::QuantizedS4 &&
+                opr->input(0)->dtype().enumv() != DTypeEnum::Quantized4Asymm) {
+                mgb_assert(
+                        m_padding_oprs.count(opr->input(0)->owner_opr()) == 0,
+                        "operator(type:%s,name:%s) for data type(%s) cannot be "
+                        "padded channel. extra info:"
+                        "consumer(%s), producer(%s)",
+                        opr->dyn_typeinfo()->name, opr->cname(),
+                        opr->input(0)->dtype().name(), opr->cname(),
+                        opr->input(0)->owner_opr()->cname());
+                return serialization::copy_opr_shallow(*opr, new_inp, opr->config());
+            }
         }
         mgb_assert(opr->input().size() == new_inp.size());
         if (m_padding_oprs.count(opr->input(0)->owner_opr())) {
@@ -341,6 +478,9 @@ void PaddingChannelPass::add_format_aware_opr_replace_func(LayoutTrans) {
     m_opr_replace_funcs[opr::PoolingForward::typeinfo()] = replace_format_aware_opr;
     m_opr_replace_funcs[opr::WarpPerspectiveForward::typeinfo()] =
             replace_format_aware_opr;
+    m_opr_replace_funcs[opr::WarpAffine::typeinfo()] = replace_format_aware_opr;
+    m_opr_replace_funcs[opr::AdaptivePooling::typeinfo()] = replace_format_aware_opr;
+    m_opr_replace_funcs[opr::ResizeForward::typeinfo()] = replace_format_aware_opr;
 }
 
 void PaddingChannelPass::add_elemwise_like_opr_replace_func(LayoutTrans) {
@@ -353,6 +493,10 @@ void PaddingChannelPass::add_elemwise_like_opr_replace_func(LayoutTrans) {
         size_t channels_after_padding = 0;
         size_t i = 0;
         for (auto&& cur_inp : opr->input()) {
+            if (cur_inp->shape().is_scalar()) {
+                ++i;
+                continue;
+            }
             bool padding_cur_inp = m_padding_oprs.count(cur_inp->owner_opr()) > 0;
             if (padding_cur_inp) {
                 if (!have_padding_inp)
@@ -363,8 +507,9 @@ void PaddingChannelPass::add_elemwise_like_opr_replace_func(LayoutTrans) {
                     same_padding = channels_after_padding == new_inp[i]->shape()[1];
                 }
             }
-            if (padding_all_inps && (!padding_cur_inp || !same_padding))
+            if (padding_all_inps && (!padding_cur_inp || !same_padding)) {
                 padding_all_inps = false;
+            }
             ++i;
         }
         if (have_padding_inp && !padding_all_inps) {
@@ -378,7 +523,7 @@ void PaddingChannelPass::add_elemwise_like_opr_replace_func(LayoutTrans) {
             }
             return serialization::copy_opr_shallow(*opr, inps, opr->config());
         }
-        if (padding_all_inps) {
+        if (padding_all_inps && have_padding_inp) {
             m_padding_oprs.insert(opr);
         }
         return serialization::copy_opr_shallow(*opr, new_inp, opr->config());
@@ -386,6 +531,53 @@ void PaddingChannelPass::add_elemwise_like_opr_replace_func(LayoutTrans) {
     m_opr_replace_funcs[opr::ElemwiseMultiType::typeinfo()] = replace_elemwise_like_opr;
     m_opr_replace_funcs[opr::Elemwise::typeinfo()] = replace_elemwise_like_opr;
     m_opr_replace_funcs[opr::TypeCvt::typeinfo()] = replace_elemwise_like_opr;
+    m_opr_replace_funcs[opr::PowC::typeinfo()] = replace_elemwise_like_opr;
+}
+
+void PaddingChannelPass::add_condition_padding_oprs_replace_func(LayoutTrans) {
+    auto replace_condition_oprs = [this](OperatorNodeBase* opr,
+                                         const VarNodeArray& new_inp) {
+        mgb_assert(opr->input().size() == new_inp.size());
+        bool can_forward_padding = true;
+        if (auto reduce = opr->try_cast_final<opr::Reduce>()) {
+            auto axis = reduce->param().axis;
+            if (axis < 0) {
+                axis += reduce->input(0)->layout().ndim;
+            }
+            //! don't reduce in channel
+            if (reduce->input().size() > 1) {
+                can_forward_padding = false;
+            } else {
+                can_forward_padding = reduce->param().axis != 1;
+            }
+        } else if (auto subtensor = opr->try_cast_final<opr::Subtensor>()) {
+            auto indexs = subtensor->index_desc();
+            size_t input_dim = subtensor->input(0)->shape().ndim;
+            for (size_t id = 0; id < indexs.size(); id++) {
+                if (indexs[id].axis.get(input_dim) == 1) {
+                    //! when subtensor perform on channel dim, if is idx mode or
+                    //! end is valid, it can forward without add subtensor
+                    can_forward_padding &=
+                            indexs[id].idx.node() || indexs[id].end.node();
+                }
+            }
+        }
+        auto inps = new_inp;
+        for (size_t i = 0; i < new_inp.size(); ++i) {
+            auto cur_inp = opr->input(i);
+            bool padding_cur_inp = m_padding_oprs.count(cur_inp->owner_opr()) > 0;
+            if (padding_cur_inp) {
+                if (can_forward_padding) {
+                    m_padding_oprs.insert(opr);
+                } else {
+                    inps[i] = extract_subtensor(inps[i], cur_inp->shape());
+                }
+            }
+        }
+        return serialization::copy_opr_shallow(*opr, inps, opr->config());
+    };
+    m_opr_replace_funcs[opr::Reduce::typeinfo()] = replace_condition_oprs;
+    m_opr_replace_funcs[opr::Subtensor::typeinfo()] = replace_condition_oprs;
 }
 
 void PaddingChannelPass::add_nonpadding_oprs_replace_func(LayoutTrans) {
@@ -405,8 +597,11 @@ void PaddingChannelPass::add_nonpadding_oprs_replace_func(LayoutTrans) {
     m_opr_replace_funcs[opr::Reshape::typeinfo()] = replace_nonpadding_oprs;
     m_opr_replace_funcs[opr::GetVarShape::typeinfo()] = replace_nonpadding_oprs;
     m_opr_replace_funcs[opr::Concat::typeinfo()] = replace_nonpadding_oprs;
-    m_opr_replace_funcs[opr::Reduce::typeinfo()] = replace_nonpadding_oprs;
-    m_opr_replace_funcs[opr::Subtensor::typeinfo()] = replace_nonpadding_oprs;
+    m_opr_replace_funcs[opr::Dimshuffle::typeinfo()] = replace_nonpadding_oprs;
+    m_opr_replace_funcs[opr::Argmax::typeinfo()] = replace_nonpadding_oprs;
+    m_opr_replace_funcs[opr::Argmin::typeinfo()] = replace_nonpadding_oprs;
+    m_opr_replace_funcs[opr::IncrSubtensor::typeinfo()] = replace_nonpadding_oprs;
+    m_opr_replace_funcs[opr::AssertEqual::typeinfo()] = replace_nonpadding_oprs;
 }
 
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/gopt/include/megbrain/gopt/inference.h

@@ -508,39 +508,48 @@ public:
  * assume input network is built in NCHW tensor format
  */
 class PaddingChannelPass final : public Pass {
-public:
+private:
     using ChannelAlignmentMap =
             ThinHashMap<DTypeEnum, std::function<size_t(size_t, bool)>>;
     using LayoutTrans = cg::GraphCommonOptimizeOptions::LayoutTransform;
+    using ReplaceFuncs = ThinHashMap<
+            Typeinfo*,
+            thin_function<OperatorNodeBase*(OperatorNodeBase*, const VarNodeArray&)>>;
 
+public:
     const char* name() const override;
     void apply(OptState& opt) const override;
 
     void fill_opr_convert_fun(LayoutTrans layout_trans);
 
-    using ReplaceFuncs = ThinHashMap<
-            Typeinfo*,
-            thin_function<OperatorNodeBase*(OperatorNodeBase*, const VarNodeArray&)>>;
-
     //! make channel padding opt pass with given tensor format
-    static std::unique_ptr<PaddingChannelPass> make(LayoutTrans layout_transform);
+    static std::unique_ptr<PaddingChannelPass> make(
+            LayoutTrans layout_transform, bool only_padding_weights = false);
 
 private:
+    PaddingChannelPass(bool only_padding_weights = false)
+            : m_only_padding_weights(only_padding_weights) {}
+
     VarNode* extract_subtensor(VarNode* inp, const TensorShape& orig_shape) const;
     VarNode* pad_in_channels(VarNode* inp, size_t pad_channels);
     VarNode* pad_out_channels(VarNode* inp, size_t pad_channels);
-    OperatorNodeBase* padding_policy(
+    OperatorNodeBase* padding_conv_policy(
+            OperatorNodeBase* opr, const VarNodeArray& new_inp);
+
+    OperatorNodeBase* padding_channel_wise_conv_policy(
             OperatorNodeBase* opr, const VarNodeArray& new_inp);
 
-    void add_convbias_replace_func(LayoutTrans layout_transform);
+    void add_conv_replace_func(LayoutTrans layout_transform);
     void add_conv_backward_data_replace_func(LayoutTrans layout_transform);
     void add_format_aware_opr_replace_func(LayoutTrans layout_transform);
     void add_elemwise_like_opr_replace_func(LayoutTrans layout_transform);
+    void add_condition_padding_oprs_replace_func(LayoutTrans layout_transform);
     void add_nonpadding_oprs_replace_func(LayoutTrans layout_transform);
 
     ChannelAlignmentMap m_alignment_map;
     ReplaceFuncs m_opr_replace_funcs;
     mutable ThinHashSet<OperatorNodeBase*> m_padding_oprs;
+    bool m_only_padding_weights;
 };
 
 /*!

src/gopt/test/inference.cpp

@@ -4004,8 +4004,11 @@ TEST(TestGoptInference, ConvertFormatNCHW44) {
     ASSERT_EQ(
             opr::Convolution::Param::Format::NCHW44,
             find_opr<opr::ConvBias>(y_opt, "conv4").param().format);
+
+    //! nchw44 add the PaddingChannel pass, it will padding conv5 output
+    //! channel, so it will convert to nchw44 format
     ASSERT_EQ(
-            opr::Convolution::Param::Format::NCHW,
+            opr::Convolution::Param::Format::NCHW44,
             find_opr<opr::ConvBias>(y_opt, "conv5").param().format);
 
     graph->compile({{y_opt, {}}})
@@ -4206,7 +4209,6 @@ TEST(TestGoptInference, ConvertFormatNCHW44_DOT) {
     auto w1 = mkcvar("w1", {8, 3, 3, 3}),
          conv1 = opr::Convolution::make(
                  x, w1, param_conv, {}, OperatorNodeConfig("conv1"));
-    printf("create conv1 %s\n", conv1.node()->owner_opr()->dyn_typeinfo()->name);
     param_conv.pad_h = param_conv.pad_w = 1;
     //! no supported hybrid nchw44
     opr::ConvBias::Param param_conv_bias_pad0;
@@ -4313,8 +4315,10 @@ TEST(TestGoptInference, ConvertFormatNCHW44_DOT) {
     ASSERT_EQ(
             opr::Convolution::Param::Format::NCHW44,
             find_opr<opr::ConvBias>(y_opt, "conv4").param().format);
+    //! nchw44-dot default add PaddingChannel pass, so it output channel will be
+    //! padding to times of 4, so it can't be used as nchw44 format.
     ASSERT_EQ(
-            opr::Convolution::Param::Format::NCHW,
+            opr::Convolution::Param::Format::NCHW44,
             find_opr<opr::ConvBias>(y_opt, "conv5").param().format);
 
     graph->compile({{y_opt, {}}})

src/gopt/test/padding_channel.cpp

@@ -0,0 +1,446 @@
+#include "megbrain/graph/cg.h"
+#include "megbrain/opr/dnn/local.h"
+
+#include "megbrain/gopt/basic_arith.h"
+#include "megbrain/gopt/gtrans.h"
+#include "megbrain/gopt/inference.h"
+
+#include "megbrain/opr/basic_arith_wrapper.h"
+#include "megbrain/opr/blas.h"
+#include "megbrain/opr/dnn/adaptive_pooling.h"
+#include "megbrain/opr/dnn/batch_norm.h"
+#include "megbrain/opr/dnn/convolution.h"
+#include "megbrain/opr/dnn/pooling.h"
+#include "megbrain/opr/imgproc.h"
+#include "megbrain/opr/io.h"
+#include "megbrain/opr/nn_int.h"
+#include "megbrain/opr/tensor_gen.h"
+#include "megbrain/opr/tensor_manip.h"
+#include "megbrain/opr/utility.h"
+
+#include "helper.h"
+#include "megbrain/comp_node_env.h"
+#include "megbrain/test/helper.h"
+
+#include "megdnn/tensor_format.h"
+
+#include <random>
+#include <vector>
+
+using namespace mgb;
+
+namespace {
+//! find first the operator of specific type; raise exception if not found
+template <typename T>
+T* find_opr(SymbolVar endpoint) {
+    T* found = nullptr;
+    auto cb = [&found](cg::OperatorNodeBase* opr) {
+        if (!found && opr->same_type<T>()) {
+            found = &opr->cast_final_safe<T>();
+        }
+    };
+    cg::DepOprIter{cb}.add(endpoint.node()->owner_opr());
+    mgb_assert(found, "not found opr from %s", endpoint.node()->name().c_str());
+    return found;
+}
+
+template <typename T>
+T* find_opr(SymbolVar endpoint, const std::string& node_name) {
+    T* found = nullptr;
+    auto cb = [&found, &node_name](cg::OperatorNodeBase* opr) {
+        if (!found && opr->same_type<T>() && opr->name() == node_name) {
+            found = &opr->cast_final_safe<T>();
+        }
+    };
+    cg::DepOprIter{cb}.add(endpoint.node()->owner_opr());
+    mgb_assert(
+            found, "not found opr %s from %s", node_name.c_str(),
+            endpoint.node()->name().c_str());
+    return found;
+}
+}  // namespace
+
+TEST(TestGoptInference, ChannelPaddingNCHW44) {
+    HostTensorGenerator<> gen;
+    auto cn = CompNode::load("cpu0");
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+    auto mkcvar = [&](const char* name, const TensorShape& shp) {
+        return opr::SharedDeviceTensor::make(*graph, *gen(shp, cn)).rename(name);
+    };
+
+    auto host_x = gen({1, 3, 8, 8}, cn);
+    auto x = opr::Host2DeviceCopy::make(*graph, host_x);
+    //! Hybrid nchw44 mode
+    opr::ConvBias::Param param_conv;
+    param_conv.pad_h = param_conv.pad_w = 1;
+    auto w1 = mkcvar("w1", {8, 3, 3, 3}), b1 = mkcvar("w1", {1, 8, 1, 1}),
+         conv1 = opr::ConvBias::make(
+                 x, w1, b1, param_conv, {}, OperatorNodeConfig("conv1"));
+
+    auto w2 = mkcvar("w2", {6, 8, 3, 3}), b2 = mkcvar("b2", {1, 6, 1, 1}),
+         conv2 = opr::ConvBias::make(
+                 conv1, w2, b2, param_conv, {}, OperatorNodeConfig("conv2"));
+    auto w3 = mkcvar("w3", {3, 6, 3, 3}), b3 = mkcvar("b3", {1, 3, 1, 1}),
+         conv3 = opr::ConvBias::make(
+                 conv2, w3, b3, param_conv, {}, OperatorNodeConfig("conv3"));
+
+    opr::Convolution::Param param_convolution;
+    param_convolution.sparse = opr::Convolution::Param::Sparse::GROUP;
+    //! channel wise convolution
+    auto w4 = mkcvar("w4", {3, 1, 1, 1, 1}),
+         conv4 = opr::Convolution::make(
+                 conv3, w4, param_convolution, {}, OperatorNodeConfig("conv4"));
+
+    param_convolution.sparse = opr::Convolution::Param::Sparse::DENSE;
+    auto w5 = mkcvar("w5", {6, 3, 1, 1}),
+         conv5 = opr::Convolution::make(
+                 conv4, w5, param_convolution, {}, OperatorNodeConfig("conv5"));
+
+    //! group convolution
+    param_convolution.sparse = opr::Convolution::Param::Sparse::GROUP;
+    auto w6 = mkcvar("w6", {2, 4, 3, 1, 1}),
+         conv6 = opr::Convolution::make(
+                 conv5, w6, param_convolution, {}, OperatorNodeConfig("conv6"));
+
+    param_convolution.sparse = opr::Convolution::Param::Sparse::DENSE;
+    auto w7 = mkcvar("w7", {3, 8, 1, 1}),
+         y = opr::Convolution::make(
+                 conv6, w7, param_convolution, {}, OperatorNodeConfig("conv7"));
+
+    SymbolVar y_opt;
+    auto options = gopt::OptimizeForInferenceOptions{};
+    options.enable_fuse_conv_bias_nonlinearity();
+    options.enable_nchw44();
+    unpack_vector(gopt::optimize_for_inference({y}, options), y_opt);
+    auto conv1_opt = find_opr<opr::ConvBias>(y_opt, "conv1");
+    auto conv2_opt = find_opr<opr::ConvBias>(y_opt, "conv2");
+    auto conv3_opt = find_opr<opr::ConvBias>(y_opt, "conv3");
+    auto conv4_opt = find_opr<opr::Convolution>(y_opt, "conv4");
+    auto conv6_opt = find_opr<opr::Convolution>(y_opt, "conv6");
+    //! do not padding input tensor
+    ASSERT_EQ(conv1_opt->input(0)->shape()[1], 3);
+    ASSERT_EQ(opr::Convolution::Param::Format::NCHW44, conv1_opt->param().format);
+    //! output tensor padding input tensor
+    ASSERT_EQ(conv2_opt->input(1)->shape()[0], 2);
+    ASSERT_EQ(opr::Convolution::Param::Format::NCHW44, conv2_opt->param().format);
+    ASSERT_EQ(conv3_opt->input(1)->shape()[0], 1);
+    ASSERT_EQ(opr::Convolution::Param::Format::NCHW44, conv3_opt->param().format);
+
+    ASSERT_EQ(conv4_opt->input(1)->shape()[0], 1);
+    ASSERT_EQ(opr::Convolution::Param::Format::NCHW44, conv4_opt->param().format);
+    ASSERT_EQ(conv6_opt->input(0)->shape()[1], 6);
+    ASSERT_EQ(opr::Convolution::Param::Format::NCHW, conv6_opt->param().format);
+
+    //! the dst tensor channel must stay unchange
+    ASSERT_EQ(y_opt.node()->shape()[1], 3);
+    graph->compile({{y_opt, {}}})
+            ->to_json()
+            ->writeto_fpath(output_file("TestGoptInference.ChannelPaddingNCHW44.json"));
+
+    HostTensorND host_y_opt, host_y;
+    auto func = graph->compile(
+            {make_callback_copy(y, host_y), make_callback_copy(y_opt, host_y_opt)});
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+
+    //! test change the input shape
+    *host_x = *gen({2, 3, 32, 32}, cn);
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+}
+
+TEST(TestGoptInference, ChannelPaddingSubtensor) {
+    HostTensorGenerator<> gen;
+    auto cn = CompNode::load("cpu0");
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+    auto mkcvar = [&](const char* name, const TensorShape& shp) {
+        return opr::SharedDeviceTensor::make(*graph, *gen(shp, cn)).rename(name);
+    };
+
+    auto host_x = gen({1, 3, 8, 8}, cn);
+    auto x = opr::Host2DeviceCopy::make(*graph, host_x);
+    //! Hybrid nchw44 mode
+    opr::ConvBias::Param param_conv;
+    param_conv.pad_h = param_conv.pad_w = 1;
+    auto w1 = mkcvar("w1", {8, 3, 3, 3}), b1 = mkcvar("w1", {1, 8, 1, 1}),
+         conv1 = opr::ConvBias::make(
+                 x, w1, b1, param_conv, {}, OperatorNodeConfig("conv1"));
+
+    auto w2 = mkcvar("w2", {6, 8, 1, 1}),
+         conv2 = opr::Convolution::make(conv1, w2, {}, {}, OperatorNodeConfig("conv2"));
+    using AIdx = opr::indexing::AxisIndexer;
+    auto sub0 = opr::Subtensor::make(
+            conv2,
+            {AIdx::make_interval(
+                    2, conv2.make_scalar(1), conv2.make_scalar(4),
+                    conv2.make_scalar(1))},
+            OperatorNodeConfig("sub0"));
+    auto sub1 = opr::Subtensor::make(
+            conv2,
+            {AIdx::make_interval(
+                     1, conv2.make_scalar(1), conv2.make_scalar(2),
+                     conv2.make_scalar(1)),
+             AIdx::make_interval(
+                     2, conv2.make_scalar(1), conv2.make_scalar(4),
+                     conv2.make_scalar(1))},
+            OperatorNodeConfig("sub1"));
+    auto sub2 = opr::Subtensor::make(
+            conv2,
+            {AIdx::make_interval(1, conv2.make_scalar(5), {}, {}),
+             AIdx::make_interval(
+                     2, conv2.make_scalar(1), conv2.make_scalar(4),
+                     conv2.make_scalar(1))},
+            OperatorNodeConfig("sub2"));
+    auto y_sub = sub0 + sub1 + sub2;
+
+    SymbolVar y_pad;
+    unpack_vector(
+            gopt::GraphOptimizer{}
+                    .add_pass(gopt::PaddingChannelPass::make(
+                            cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44,
+                            true))
+                    .apply({{y_sub}})
+                    .endpoint_vars(),
+            y_pad);
+    auto conv1_opt = find_opr<opr::ConvBias>(y_pad, "conv1");
+    auto conv2_opt = find_opr<opr::Convolution>(y_pad, "conv2");
+    auto sub0_opt = find_opr<opr::Subtensor>(y_pad, "sub0");
+    auto sub1_opt = find_opr<opr::Subtensor>(y_pad, "sub1");
+    auto sub2_opt = find_opr<opr::Subtensor>(y_pad, "sub2");
+    //! do not padding input tensor
+    ASSERT_EQ(conv1_opt->input(0)->shape()[1], 3);
+    //! output tensor padding input tensor
+    ASSERT_EQ(conv2_opt->input(1)->shape()[0], 8);
+    ASSERT_EQ(conv2_opt->output(0)->shape()[1], 8);
+
+    //! sub0 do not perform on channel dim, so no need to add subtensor
+    ASSERT_EQ(sub0_opt->input(0)->shape()[1], 8);
+    //! sub1 perform on channel dim, but end is specific, so no need to add subtensor
+    ASSERT_EQ(sub1_opt->input(0)->shape()[1], 8);
+    //! sub1 perform on channel dim, and end is default, so need to add subtensor
+    ASSERT_EQ(sub2_opt->input(0)->shape()[1], 6);
+
+    //! the dst tensor channel must stay unchange
+    ASSERT_EQ(y_pad.node()->shape()[1], 6);
+    graph->compile({{y_pad, {}}})
+            ->to_json()
+            ->writeto_fpath(
+                    output_file("TestGoptInference.ChannelPaddingSubtensor.json"));
+
+    HostTensorND host_y_opt, host_y;
+    auto func = graph->compile(
+            {make_callback_copy(y_sub, host_y), make_callback_copy(y_pad, host_y_opt)});
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+
+    //! test change the input shape
+    *host_x = *gen({2, 3, 32, 32}, cn);
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+}
+
+TEST(TestGoptInference, ChannelPaddingReduce) {
+    HostTensorGenerator<> gen;
+    auto cn = CompNode::load("cpu0");
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+    auto mkcvar = [&](const char* name, const TensorShape& shp) {
+        return opr::SharedDeviceTensor::make(*graph, *gen(shp, cn)).rename(name);
+    };
+
+    auto host_x = gen({1, 3, 8, 8}, cn);
+    auto x = opr::Host2DeviceCopy::make(*graph, host_x);
+    //! Hybrid nchw44 mode
+    opr::ConvBias::Param param_conv;
+    param_conv.pad_h = param_conv.pad_w = 1;
+    auto w1 = mkcvar("w1", {8, 3, 3, 3}), b1 = mkcvar("w1", {1, 8, 1, 1}),
+         conv1 = opr::ConvBias::make(
+                 x, w1, b1, param_conv, {}, OperatorNodeConfig("conv1"));
+
+    auto w2 = mkcvar("w2", {6, 8, 1, 1}),
+         conv2 = opr::Convolution::make(conv1, w2, {}, {}, OperatorNodeConfig("conv2"));
+    auto reduce0 = opr::Reduce::make(
+            conv2, {opr::Reduce::Mode::MAX, 1}, {}, OperatorNodeConfig("reduce0"));
+    auto reduce1 = opr::Reduce::make(
+            conv2, {opr::Reduce::Mode::MAX, 2}, {}, OperatorNodeConfig("reduce1"));
+    auto y_reduce = reduce0 + reduce1;
+
+    SymbolVar y_pad;
+    unpack_vector(
+            gopt::GraphOptimizer{}
+                    .add_pass(gopt::PaddingChannelPass::make(
+                            cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44,
+                            true))
+                    .apply({{y_reduce}})
+                    .endpoint_vars(),
+            y_pad);
+    auto conv1_opt = find_opr<opr::ConvBias>(y_pad, "conv1");
+    auto conv2_opt = find_opr<opr::Convolution>(y_pad, "conv2");
+    auto reduce0_opt = find_opr<opr::Reduce>(y_pad, "reduce0");
+    auto reduce1_opt = find_opr<opr::Reduce>(y_pad, "reduce1");
+    //! do not padding input tensor
+    ASSERT_EQ(conv1_opt->input(0)->shape()[1], 3);
+    //! output tensor padding input tensor
+    ASSERT_EQ(conv2_opt->input(1)->shape()[0], 8);
+    ASSERT_EQ(conv2_opt->output(0)->shape()[1], 8);
+
+    //! reduce0 perform on channel dim, so need to add subtensor
+    ASSERT_EQ(reduce0_opt->input(0)->shape()[1], 6);
+    //! reduce1 don't perform on channel dim, so no need to add subtensor
+    ASSERT_EQ(reduce1_opt->input(0)->shape()[1], 8);
+
+    //! the dst tensor channel must stay unchange
+    ASSERT_EQ(y_pad.node()->shape()[1], 6);
+    graph->compile({{y_pad, {}}})
+            ->to_json()
+            ->writeto_fpath(output_file("TestGoptInference.ChannelPaddingReduce.json"));
+
+    HostTensorND host_y_opt, host_y;
+    auto func = graph->compile(
+            {make_callback_copy(y_reduce, host_y),
+             make_callback_copy(y_pad, host_y_opt)});
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+
+    //! test change the input shape
+    *host_x = *gen({2, 3, 32, 32}, cn);
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+}
+
+TEST(TestGoptInference, ChannelPaddingMisc) {
+    HostTensorGenerator<> gen;
+    auto cn = CompNode::load("cpu0");
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+    auto mkcvar = [&](const char* name, const TensorShape& shp) {
+        return opr::SharedDeviceTensor::make(*graph, *gen(shp, cn)).rename(name);
+    };
+
+    auto host_x = gen({1, 3, 8, 8}, cn);
+    auto x = opr::Host2DeviceCopy::make(*graph, host_x);
+    //! Hybrid nchw44 mode
+    opr::ConvBias::Param param_conv;
+    param_conv.pad_h = param_conv.pad_w = 1;
+    auto w1 = mkcvar("w1", {8, 3, 3, 3}), b1 = mkcvar("w1", {1, 8, 1, 1}),
+         conv1 = opr::ConvBias::make(
+                 x, w1, b1, param_conv, {}, OperatorNodeConfig("conv1"));
+
+    auto w2 = mkcvar("w2", {6, 8, 1, 1}),
+         conv2 = opr::Convolution::make(conv1, w2, {}, {}, OperatorNodeConfig("conv2"));
+    auto elem0 = conv2 + 1;
+    auto concat = opr::Concat::make({elem0, conv2}, 1, OperatorNodeConfig("concat"));
+
+    SymbolVar y_pad;
+    unpack_vector(
+            gopt::GraphOptimizer{}
+                    .add_pass(gopt::PaddingChannelPass::make(
+                            cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44,
+                            true))
+                    .apply({{concat}})
+                    .endpoint_vars(),
+            y_pad);
+    auto conv1_opt = find_opr<opr::ConvBias>(y_pad, "conv1");
+    auto conv2_opt = find_opr<opr::Convolution>(y_pad, "conv2");
+    auto elemwise0_opt = find_opr<opr::Elemwise>(y_pad);
+    auto concat_opt = find_opr<opr::Concat>(y_pad, "concat");
+    //! do not padding input tensor
+    ASSERT_EQ(conv1_opt->input(0)->shape()[1], 3);
+    //! output tensor padding input tensor
+    ASSERT_EQ(conv2_opt->input(1)->shape()[0], 8);
+    ASSERT_EQ(conv2_opt->output(0)->shape()[1], 8);
+
+    ASSERT_EQ(elemwise0_opt->output(0)->shape()[1], 8);
+    ASSERT_EQ(concat_opt->input(0)->shape()[1], 6);
+    ASSERT_EQ(concat_opt->input(1)->shape()[1], 6);
+
+    //! the dst tensor channel must stay unchange
+    ASSERT_EQ(y_pad.node()->shape()[1], 12);
+    graph->compile({{y_pad, {}}})
+            ->to_json()
+            ->writeto_fpath(output_file("TestGoptInference.ChannelPaddingMisc.json"));
+
+    HostTensorND host_y_opt, host_y;
+    auto func = graph->compile(
+            {make_callback_copy(concat, host_y),
+             make_callback_copy(y_pad, host_y_opt)});
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+
+    //! test change the input shape
+    *host_x = *gen({2, 3, 32, 32}, cn);
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+}
+
+TEST(TestGoptInference, ChannelPaddingMoreOp) {
+    HostTensorGenerator<> gen;
+    auto cn = CompNode::load("cpu0");
+    auto graph = ComputingGraph::make();
+    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(*graph, *gen(shp, cn)).rename(name);
+    };
+
+    auto host_x = gen({2, 3, 8, 8}, cn);
+    auto x = opr::Host2DeviceCopy::make(*graph, host_x);
+
+    opr::Convolution::Param param;
+    param.pad_h = param.pad_w = 1;
+    auto w1 = mkcvar("w1", {6, 3, 3, 3}), conv = opr::Convolution::make(x, w1, param);
+    auto shape_of = opr::GetVarShape::make(conv);
+    auto subtensor = opr::Subtensor::make(
+            shape_of, {opr::Subtensor::AxisIndexer::make_interval(
+                              0, x.make_scalar(2), None, x.make_scalar(1))});
+
+    opr::Resize::Param param_resize;
+    param_resize.format = opr::Resize::Param::Format::NCHW;
+    auto resize = opr::ResizeForward::make(conv, subtensor * 2, param_resize);
+    auto mat = mkcvar("mat", {2, 3, 3}),
+         warp = opr::WarpPerspectiveForward::make(
+                 resize, mat, nullptr, cg::var_from_tensor_shape(x, {4, 4}));
+
+    auto b = mkvar("b", {1, 6, 1, 1}),
+         elem = opr::Elemwise::make({warp + b}, opr::Elemwise::Param::Mode::RELU);
+    param.pad_h = param.pad_w = 1;
+    auto w2 = mkcvar("w2", {7, 6, 3, 3}), y = opr::Convolution::make(elem, w2, param),
+         z = opr::AxisAddRemove::make(y, {opr::AxisAddRemove::AxisDesc::make_add(0)});
+
+    SymbolVar y_pad, z_pad;
+    unpack_vector(
+            gopt::GraphOptimizer{}
+                    .add_pass(gopt::PaddingChannelPass::make(
+                            cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44,
+                            true))
+                    .apply({{y}})
+                    .endpoint_vars(),
+            y_pad);
+    unpack_vector(
+            gopt::GraphOptimizer{}
+                    .add_pass(gopt::PaddingChannelPass::make(
+                            cg::GraphCommonOptimizeOptions::LayoutTransform::NCHW44,
+                            true))
+                    .apply({{z}})
+                    .endpoint_vars(),
+            z_pad);
+
+    graph->compile({{y_pad, {}}})
+            ->to_json()
+            ->writeto_fpath(output_file("TestGoptInference.ChannelPaddingMoreOp.json"));
+
+    HostTensorND host_y_opt, host_y;
+    auto func = graph->compile(
+            {make_callback_copy(y, host_y), make_callback_copy(y_pad, host_y_opt)});
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-3);
+
+    *host_x = *gen({2, 3, 16, 16}, cn);
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-3);
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}