refactor(mgb/gopt): refactor tensor reformat opt pass

GitOrigin-RevId: a1b1e89b76e4fbdca4f481156bb8af6cae8fe4d8

dnn/src/common/named_tensor.cpp

@@ -120,10 +120,6 @@ Dimension Dimension::operator/(const Dimension& rhs) const {
                   static_cast<char>(m_name), static_cast<char>(rhs.m_name));
     if (operator==(rhs))
         return Dimension(m_name, 1, 1);
-    megdnn_assert(
-            !(*this < rhs),
-            "Divisor must be smaller than dividend(dividend:%s, divisor:%s)",
-            to_string().c_str(), rhs.to_string().c_str());
     if (m_stride == rhs.m_stride) {
         if (m_extent == UNDETERMINED_EXTENT) {
             megdnn_assert(rhs.m_extent != UNDETERMINED_EXTENT,

src/gopt/impl/folding_conv_dimshuffle.cpp

+/**
+ * \file src/gopt/impl/folding_conv_dimshuffle.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megbrain/gopt/inference.h"
+#include "megbrain/opr/basic_arith.h"
+#include "megbrain/opr/dnn/convolution.h"
+#include "megbrain/opr/tensor_manip.h"
+#include "megbrain/opr/utility.h"
+#include "megbrain/serialization/opr_shallow_copy.h"
+
+#include "megdnn/opr_param_defs.h"
+
+#include "megbrain/opr/internal/megdnn_opr_wrapper.h"
+
+#include "megbrain/utils/hash_ct.h"
+
+#include "midout.h"
+
+#include "megbrain/gopt/reformat_manager.h"
+
+#if CUDA_VERSION >= 10020
+MIDOUT_DECL(megbrain_folding_conv_dimshuffle)
+#define MIDOUT_B(tag)                              \
+    MIDOUT_BEGIN(megbrain_folding_conv_dimshuffle, \
+                 midout_iv(MGB_HASH_STR(tag))) {
+#define MIDOUT_E \
+    }            \
+    MIDOUT_END();
+
+using namespace mgb;
+using namespace gopt;
+using ReformatKey = ReformatManager::ReformatKey;
+
+/* ==================== FoldingConvBiasDimshufflePass ================= */
+const char* FoldingConvBiasDimshufflePass::name() const {
+    return mgb_cstr_log("folding conv bias dimshuffle pass");
+}
+
+void FoldingConvBiasDimshufflePass::apply(OptState& opt) const {
+    MIDOUT_B("FoldingConvBiasDimshufflePass::apply");
+    using DepType = cg::OperatorNodeProp::DepType;
+    ThinHashMap<OperatorNodeBase*,
+                SmallVector<std::pair<OperatorNodeBase*, DepType>>>
+            readers;
+    static const ThinHashSet<Typeinfo*> opr_type_list = {
+            opr::TypeCvt::typeinfo(), opr::Dimshuffle::typeinfo(),
+            opr::Reshape::typeinfo(), opr::ConvBias::typeinfo()};
+    opt.graph().iter([&readers](OperatorNodeBase* opr) {
+        for (auto&& i : opr->node_prop().dep_map()) {
+            if (opr_type_list.count(i.first->owner_opr()->dyn_typeinfo())) {
+                readers[i.first->owner_opr()].emplace_back(opr, i.second);
+            }
+        }
+    });
+
+    auto rewriter = opt.graph().make_rewriter();
+    auto try_conv_dimshuffle_reshape_typecvt = [&rewriter, &readers](
+                                                       OperatorNodeBase* opr) {
+        ThinHashSet<OperatorNodeBase*> opr_set;
+        ThinHashSet<OperatorNodeBase*> reader_set;
+        // check typecvt
+        auto typecvt = try_cast_as_op<opr::TypeCvt>(opr);
+        if (typecvt == nullptr)
+            return false;
+        auto inp_dtype = typecvt->input(0)->dtype(),
+             out_dtype = typecvt->output(0)->dtype();
+        bool is_s82f32 = inp_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                         out_dtype.enumv() == DTypeEnum::Float32;
+        if (!is_s82f32)
+            return false;
+        opr_set.insert(opr);
+
+        // check reshape
+        auto reshape =
+                try_cast_as_op<opr::Reshape>(typecvt->input(0)->owner_opr());
+        if (reshape == nullptr)
+            return false;
+        opr_set.insert(reshape);
+        for (auto&& i : readers[reshape]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+
+        // check shuffle
+        auto shuffle =
+                try_cast_as_op<opr::Dimshuffle>(reshape->input(0)->owner_opr());
+        if (shuffle == nullptr)
+            return false;
+        auto&& param = shuffle->param();
+        if (param.pattern_len != 5)
+            return false;
+        bool is_nchw42nchw = param.pattern[0] == 0 && param.pattern[1] == 1 &&
+                             param.pattern[2] == 4 && param.pattern[3] == 2 &&
+                             param.pattern[4] == 3 &&
+                             shuffle->input(0)->shape()[4] == 4;
+        if (!is_nchw42nchw)
+            return false;
+        opr_set.insert(shuffle);
+        for (auto&& i : readers[shuffle]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+
+        // check conv bias
+        auto conv_bias =
+                try_cast_as_op<opr::ConvBias>(shuffle->input(0)->owner_opr());
+        if (conv_bias == nullptr)
+            return false;
+        inp_dtype = conv_bias->input(0)->dtype();
+        bool is_s8nchw4 = inp_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                          conv_bias->param().format ==
+                                  megdnn::param::ConvBias::Format::NCHW4;
+        if (!is_s8nchw4)
+            return false;
+        if (conv_bias->input().size() != 3)
+            return false;
+        opr_set.insert(conv_bias);
+        for (auto&& i : readers[conv_bias]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        for (auto reader : reader_set) {
+            if (opr_set.count(reader) <= 0) {
+                return false;
+            }
+        }
+        auto src = rewriter.get_var(conv_bias->input(0)),
+             filter = rewriter.get_var(conv_bias->input(1)),
+             bias = rewriter.get_var(conv_bias->input(2));
+        auto new_bias = ReformatManager::instance().get(ReformatKey{
+                TensorFormats::NCHWc4, TensorFormats::NCHW})({bias});
+        new_bias = opr::TypeCvt::make(new_bias, dtype::Float32()).node();
+        auto new_param = conv_bias->param();
+        new_param.format = megdnn::param::ConvBias::Format::NCHW4_NCHW;
+        auto conv_bias_shuffle = opr::ConvBias::make(
+                src, filter, new_bias, new_param, conv_bias->execution_policy(),
+                OperatorNodeConfig{dtype::Float32()});
+        rewriter.replace_var(opr->output(0), conv_bias_shuffle.node(),
+                             mgb_cstr_log("replace conv_bias + typecvt + "
+                                          "dimshuffle + "
+                                          "reshape to conv_bias(NCHW4_NCHW)"));
+        return true;
+    };
+
+    auto try_conv_reformat_nchw42nchw32 = [&rewriter,
+                                           &readers](OperatorNodeBase* opr) {
+        ThinHashSet<OperatorNodeBase*> opr_set;
+        ThinHashSet<OperatorNodeBase*> reader_set;
+        // check reshape
+        auto reshape1 = try_cast_as_op<opr::Reshape>(opr);
+        if (reshape1 == nullptr)
+            return false;
+        opr_set.insert(opr);
+        // check dimshuffle
+        auto shuffle = try_cast_as_op<opr::Dimshuffle>(
+                reshape1->input(0)->owner_opr());
+        if (shuffle == nullptr)
+            return false;
+        auto&& param = shuffle->param();
+        if (param.pattern_len != 6)
+            return false;
+        bool is_nchw42nchw32 = param.pattern[0] == 0 && param.pattern[1] == 1 &&
+                               param.pattern[2] == 3 && param.pattern[3] == 4 &&
+                               param.pattern[4] == 2 && param.pattern[5] == 5 &&
+                               shuffle->output(0)->shape()[5] == 4 &&
+                               shuffle->output(0)->shape()[4] == 8;
+        if (!is_nchw42nchw32)
+            return false;
+        opr_set.insert(shuffle);
+        for (auto&& i : readers[shuffle]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        // check reshape
+        auto reshape2 =
+                try_cast_as_op<opr::Reshape>(shuffle->input(0)->owner_opr());
+        if (reshape2 == nullptr)
+            return false;
+        opr_set.insert(reshape2);
+        for (auto&& i : readers[reshape2]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        // check conv bias
+        auto conv_bias =
+                try_cast_as_op<opr::ConvBias>(reshape2->input(0)->owner_opr());
+        if (conv_bias == nullptr)
+            return false;
+        auto inp_dtype = conv_bias->input(0)->dtype();
+        bool is_s8nchw4 = inp_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                          conv_bias->param().format ==
+                                  megdnn::param::ConvBias::Format::NCHW4;
+        if (!is_s8nchw4)
+            return false;
+        if (conv_bias->input().size() != 3)
+            return false;
+        opr_set.insert(conv_bias);
+        for (auto&& i : readers[conv_bias]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        for (auto reader : reader_set) {
+            if (opr_set.count(reader) <= 0) {
+                return false;
+            }
+        }
+        auto src = rewriter.get_var(conv_bias->input(0)),
+             filter = rewriter.get_var(conv_bias->input(1)),
+             bias = rewriter.get_var(conv_bias->input(2));
+        auto new_bias = ReformatManager::instance().get(ReformatKey{
+                TensorFormats::NCHWc4, TensorFormats::NCHWc32})({bias});
+        auto new_param = conv_bias->param();
+        new_param.format = megdnn::param::ConvBias::Format::NCHW4_NCHW32;
+        auto conv_bias_shuffle = opr::ConvBias::make(
+                src, filter, new_bias, new_param, conv_bias->execution_policy(),
+                conv_bias->config());
+        rewriter.replace_var(
+                opr->output(0), conv_bias_shuffle.node(),
+                mgb_cstr_log("replace conv_bias + "
+                             "reformat to conv_bias(NCHW4_NCHW32)"));
+        return true;
+    };
+
+    auto try_conv_reformat_nchw42nhwc = [&rewriter,
+                                         &readers](OperatorNodeBase* opr) {
+        ThinHashSet<OperatorNodeBase*> opr_set;
+        ThinHashSet<OperatorNodeBase*> reader_set;
+        // check reshape
+        auto reshape = try_cast_as_op<opr::Reshape>(opr);
+        if (reshape == nullptr)
+            return false;
+        opr_set.insert(opr);
+
+        // check dimshuffle
+        auto shuffle =
+                try_cast_as_op<opr::Dimshuffle>(reshape->input(0)->owner_opr());
+        if (shuffle == nullptr)
+            return false;
+        auto&& param = shuffle->param();
+        if (param.pattern_len != 5)
+            return false;
+        bool is_nchw42nhwc = param.pattern[0] == 0 && param.pattern[1] == 2 &&
+                             param.pattern[2] == 3 && param.pattern[3] == 1 &&
+                             param.pattern[4] == 4 &&
+                             shuffle->output(0)->shape()[4] == 4;
+        if (!is_nchw42nhwc)
+            return false;
+        opr_set.insert(shuffle);
+        for (auto&& i : readers[shuffle]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+
+        auto typecvt =
+                try_cast_as_op<opr::TypeCvt>(shuffle->input(0)->owner_opr());
+        if (typecvt == nullptr)
+            return false;
+        auto in_dtype = typecvt->input(0)->dtype(),
+             out_dtype = typecvt->output(0)->dtype();
+        bool is_s82s4 = in_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                        (out_dtype.enumv() == DTypeEnum::QuantizedS4 ||
+                         out_dtype.enumv() == DTypeEnum::Quantized4Asymm);
+        if (!is_s82s4)
+            return false;
+        opr_set.insert(typecvt);
+        for (auto&& i : readers[typecvt]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+
+        // check conv bias
+        auto conv_bias =
+                try_cast_as_op<opr::ConvBias>(typecvt->input(0)->owner_opr());
+        if (conv_bias == nullptr)
+            return false;
+        auto inp_dtype = conv_bias->input(0)->dtype();
+        bool is_s8nchw4 = inp_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                          conv_bias->param().format ==
+                                  megdnn::param::ConvBias::Format::NCHW4;
+        if (!is_s8nchw4)
+            return false;
+        if (conv_bias->input().size() != 3)
+            return false;
+        opr_set.insert(conv_bias);
+        for (auto&& i : readers[conv_bias]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        for (auto reader : reader_set) {
+            if (opr_set.count(reader) <= 0) {
+                return false;
+            }
+        }
+        auto src = rewriter.get_var(conv_bias->input(0)),
+             filter = rewriter.get_var(conv_bias->input(1)),
+             bias = rewriter.get_var(conv_bias->input(2));
+        auto new_bias = ReformatManager::instance().get(ReformatKey{
+                TensorFormats::NCHWc4, TensorFormats::NHWC})({bias});
+        auto new_param = conv_bias->param();
+        new_param.format = megdnn::param::ConvBias::Format::NCHW4_NHWC;
+        auto conv_bias_shuffle = opr::ConvBias::make(
+                src, filter, new_bias, new_param, conv_bias->execution_policy(),
+                OperatorNodeConfig{out_dtype});
+        rewriter.replace_var(opr->output(0), conv_bias_shuffle.node(),
+                             mgb_cstr_log("replace conv_bias + "
+                                          "reformat to conv_bias(NCHW4_NHWC)"));
+        return true;
+    };
+
+    auto try_conv_reformat_nchw322nchw4 = [&rewriter,
+                                           &readers](OperatorNodeBase* opr) {
+        ThinHashSet<OperatorNodeBase*> opr_set;
+        ThinHashSet<OperatorNodeBase*> reader_set;
+        // check reshape
+        auto reshape1 = try_cast_as_op<opr::Reshape>(opr);
+        if (reshape1 == nullptr)
+            return false;
+        opr_set.insert(opr);
+        // check dimshuffle
+        auto shuffle = try_cast_as_op<opr::Dimshuffle>(
+                reshape1->input(0)->owner_opr());
+        if (shuffle == nullptr)
+            return false;
+        auto&& param = shuffle->param();
+        if (param.pattern_len != 6)
+            return false;
+        bool is_nchw322nchw4 = param.pattern[0] == 0 && param.pattern[1] == 1 &&
+                               param.pattern[2] == 4 && param.pattern[3] == 2 &&
+                               param.pattern[4] == 3 && param.pattern[5] == 5 &&
+                               shuffle->input(0)->shape()[5] == 4 &&
+                               shuffle->input(0)->shape()[4] == 8;
+        if (!is_nchw322nchw4)
+            return false;
+        opr_set.insert(shuffle);
+        for (auto&& i : readers[shuffle]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        // check reshape
+        auto reshape2 =
+                try_cast_as_op<opr::Reshape>(shuffle->input(0)->owner_opr());
+        if (reshape2 == nullptr)
+            return false;
+        opr_set.insert(reshape2);
+        for (auto&& i : readers[reshape2]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        // check conv bias
+        auto conv_bias =
+                try_cast_as_op<opr::ConvBias>(reshape2->input(0)->owner_opr());
+        if (conv_bias == nullptr)
+            return false;
+        auto inp_dtype = conv_bias->input(0)->dtype();
+        bool is_s8nchw32 = inp_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                           conv_bias->param().format ==
+                                   megdnn::param::ConvBias::Format::NCHW32;
+        if (!is_s8nchw32)
+            return false;
+        if (conv_bias->input().size() != 3)
+            return false;
+        opr_set.insert(conv_bias);
+        for (auto&& i : readers[conv_bias]) {
+            if (i.second & DepType::DEV_VALUE) {
+                reader_set.insert(i.first);
+            }
+        }
+        for (auto reader : reader_set) {
+            if (opr_set.count(reader) <= 0) {
+                return false;
+            }
+        }
+        auto src = rewriter.get_var(conv_bias->input(0)),
+             filter = rewriter.get_var(conv_bias->input(1)),
+             bias = rewriter.get_var(conv_bias->input(2));
+        auto new_bias = ReformatManager::instance().get(ReformatKey{
+                TensorFormats::NCHWc32, TensorFormats::NCHWc4})({bias});
+        auto new_param = conv_bias->param();
+        new_param.format = megdnn::param::ConvBias::Format::NCHW32_NCHW4;
+        auto conv_bias_shuffle = opr::ConvBias::make(
+                src, filter, new_bias, new_param, conv_bias->execution_policy(),
+                conv_bias->config());
+        rewriter.replace_var(
+                opr->output(0), conv_bias_shuffle.node(),
+                mgb_cstr_log("replace conv_bias + "
+                             "reformat to conv_bias(NCHW32_NCHW4)"));
+        return true;
+    };
+    MGB_MARK_USED_VAR(try_conv_reformat_nchw322nchw4);
+    MGB_MARK_USED_VAR(try_conv_reformat_nchw42nchw32);
+
+    auto on_opr = [&try_conv_dimshuffle_reshape_typecvt,
+                   &try_conv_reformat_nchw42nchw32,
+                   &try_conv_reformat_nchw42nhwc,
+                   &try_conv_reformat_nchw322nchw4,
+                   &rewriter](OperatorNodeBase* opr) {
+        if (!try_conv_dimshuffle_reshape_typecvt(opr) &&
+            !try_conv_reformat_nchw42nchw32(opr) &&
+            !try_conv_reformat_nchw42nhwc(opr) &&
+            !try_conv_reformat_nchw322nchw4(opr)) {
+            rewriter.auto_replace_outputs(opr);
+        }
+    };
+    opt.graph().iter(on_opr);
+    rewriter.apply_inplace();
+
+    MIDOUT_E
+}
+#endif
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/gopt/impl/reformat_manager.cpp

@@ -11,7 +11,6 @@
  */
 
 #include "megbrain/gopt/reformat_manager.h"
-#include <numeric>
 #include "megbrain/opr/tensor_manip.h"
 
 using namespace mgb;
@@ -65,6 +64,10 @@ NamedTensorShape tensor_formats_to_named_tensor_shape(TensorFormats format) {
             return {{"C//8"}, {"C%1"}, {"C%1"}, {"R"}, {"S"}, {"C%8"}};
         case TensorFormats::KRSCk8:
             return {{"K//8"}, {"R"}, {"S"}, {"C"}, {"K%8"}};
+        case TensorFormats::KCRSc4:
+            return {{"K"}, {"C//4"}, {"R"}, {"S"}, {"C%4"}};
+        case TensorFormats::GKCRSc4:
+            return {{"G"}, {"K"}, {"C//4"}, {"R"}, {"S"}, {"C%4"}};
         case TensorFormats::KCRS:
             return {{"K"}, {"C"}, {"R"}, {"S"}};
         case TensorFormats::GKCRS:
@@ -130,70 +133,40 @@ bool ReformatManager::ReformatKey::Equal::operator()(
            lhs.attribute == rhs.attribute;
 }
 
+ReformatManager::ReformatKey&
+ReformatManager::ReformatKey::deduce_reformat_dtype_enum(const DType& dt) {
+    static const ThinHashSet<std::pair<TensorFormats, TensorFormats>> set = {
+            {TensorFormats::NCHW, TensorFormats::NCHWc64},
+            {TensorFormats::NCHWc64, TensorFormats::NCHW},
+            {TensorFormats::NCHW, TensorFormats::NHWC},
+            {TensorFormats::NHWC, TensorFormats::NCHW}};
+    if (set.count({input_format, output_format}) > 0 &&
+        (dt.enumv() == DTypeEnum::QuantizedS4 ||
+         dt.enumv() == DTypeEnum::Quantized4Asymm)) {
+        input_dtype = output_dtype = dt.enumv();
+    }
+    return *this;
+}
+
 // =================== ReformatManager ====================*/
-#define FOREACH_FEATURE_TENSOR_FORMATS(cb)                                     \
-    cb(NCHW) cb(NHWC) cb(NCHWc4) cb(NCHWc8) cb(NCHWc32) cb(NCHWc64) cb(CHWNc4) \
-            cb(NHCWc4)
-#define FOREACH_WEIGHT_TENSOR_FORMATS(cb)                                     \
-    cb(KRSCk4) cb(KRSCk4c4) cb(KCRSk4c4) cb(KCRSc4k4) cb(KCRSc8k8) cb(KRSCk8) \
-            cb(GKRSCk4) cb(GKRSCk4c4) cb(GKCRSc4k4) cb(GKCRSk4c4)             \
-                    cb(GKCRSc8k8) cb(C11RSc4) cb(C11RSc8)
 ReformatManager::ReformatManager() {
-    static constexpr TensorFormats feature_tensor_formats[] = {
-#define cb(_fmt) TensorFormats::_fmt,
-            FOREACH_FEATURE_TENSOR_FORMATS(cb)
-#undef cb
+    using Attribute = ReformatKey::Attribute;
+    {
+        auto i = TensorFormats::NCHWc4, o = TensorFormats::CHWNc4;
+        auto&& impl1 = [](const VarNodeArray& vars) {
+            return opr::RelayoutFormat::make(
+                           vars[0],
+                           megdnn::param::RelayoutFormat::Mode::NCHW4_CHWN4)
+                    .node();
         };
-    static constexpr int nr_feature_tensor_formats =
-            sizeof(feature_tensor_formats) / sizeof(TensorFormats);
-    for (int i = 0; i < nr_feature_tensor_formats; ++i) {
-        for (int o = 0; o < nr_feature_tensor_formats; ++o) {
-            if (i == o)
-                continue;
-            NamedTensorShape input_shape = tensor_formats_to_named_tensor_shape(
-                    feature_tensor_formats[i]);
-            NamedTensorShape output_shape =
-                    tensor_formats_to_named_tensor_shape(
-                            feature_tensor_formats[o]);
-            auto impl = std::get<0>(
-                    ReformatEmitter{input_shape, output_shape}.emit());
-            m_cache.emplace(ReformatKey{feature_tensor_formats[i],
-                                        feature_tensor_formats[o]},
-                            impl);
-        }
-    }
-    static constexpr TensorFormats default_weight_tensor_formats =
-            TensorFormats::KCRS;
-    static constexpr TensorFormats default_group_conv_weight_tensor_formats =
-            TensorFormats::GKCRS;
-    static constexpr TensorFormats default_chan_conv_weight_tensor_formats =
-            TensorFormats::C11RS;
-    static constexpr TensorFormats weight_tensor_formats[] = {
-#define cb(_fmt) TensorFormats::_fmt,
-            FOREACH_WEIGHT_TENSOR_FORMATS(cb)
-#undef cb
+        m_cache.emplace(ReformatKey{i, o}, impl1);
+        auto&& impl2 = [](const VarNodeArray& vars) {
+            return opr::RelayoutFormat::make(
+                           vars[0],
+                           megdnn::param::RelayoutFormat::Mode::CHWN4_NCHW4)
+                    .node();
         };
-    static constexpr int nr_weight_tensor_formats =
-            sizeof(weight_tensor_formats) / sizeof(TensorFormats);
-    using Name = megdnn::Dimension::Name;
-    for (int o = 0; o < nr_weight_tensor_formats; ++o) {
-        NamedTensorShape output_shape =
-                tensor_formats_to_named_tensor_shape(weight_tensor_formats[o]);
-        TensorFormats input_format;
-        if (output_shape[0].name() == Name::G) {
-            input_format = default_group_conv_weight_tensor_formats;
-        } else if (output_shape[0].name() == Name::C) {
-            input_format = default_chan_conv_weight_tensor_formats;
-        } else {
-            mgb_assert(output_shape[0].name() == Name::K);
-            input_format = default_weight_tensor_formats;
-        }
-        NamedTensorShape input_shape =
-                tensor_formats_to_named_tensor_shape(input_format);
-        auto impl =
-                std::get<0>(ReformatEmitter{input_shape, output_shape}.emit());
-        m_cache.emplace(ReformatKey{input_format, weight_tensor_formats[o]},
-                        impl);
+        m_cache.emplace(ReformatKey{o, i}, impl2);
     }
     {
         auto i = TensorFormats::NCHW, o = TensorFormats::NCHWc4;
@@ -206,7 +179,7 @@ ReformatManager::ReformatManager() {
         m_cache.emplace(ReformatKey{i, o, Attribute::IC_SMALL}, impl);
     }
     {
-        auto i = TensorFormats::KCRS, o = TensorFormats::KCRSc4k4;
+        auto i = TensorFormats::KCRS, o = TensorFormats::KCRSc4;
         auto&& impl = [](const VarNodeArray& vars) {
             return opr::RelayoutFormat::make(
                            vars[0],
@@ -238,7 +211,7 @@ ReformatManager::ReformatManager() {
         auto&& impl = [](const VarNodeArray& vars) {
             return opr::RelayoutFormat::make(
                            vars[0],
-                           megdnn::param::RelayoutFormat::Mode::NCHW_NCHW64)
+                           megdnn::param::RelayoutFormat::Mode::NCHW64_NCHW)
                     .node();
         };
         m_cache.emplace(
@@ -272,7 +245,7 @@ ReformatManager::ReformatManager() {
         auto&& impl = [](const VarNodeArray& vars) {
             return opr::RelayoutFormat::make(
                            vars[0],
-                           megdnn::param::RelayoutFormat::Mode::NCHW_NHWC)
+                           megdnn::param::RelayoutFormat::Mode::NHWC_NCHW)
                     .node();
         };
         m_cache.emplace(
@@ -371,14 +344,23 @@ ReformatManager::ReformatManager() {
                         impl);
     }
 }
-#undef FOREACH_FEATURE_TENSOR_FORMATS
-#undef FOREACH_WEIGHT_TENSOR_FORMATS
 
-const ReformatManager::ReformatImpl& ReformatManager::get(
+ReformatManager::ReformatImpl ReformatManager::get(
         const ReformatKey& key) const {
+    using Attribute = ReformatKey::Attribute;
     MGB_TRY {
-        auto&& impl = m_cache.at(key);
-        return impl;
+        auto find = m_cache.find(key);
+        if (find != m_cache.end()) {
+            auto rst = find->second;
+            return rst;
+        }
+        mgb_assert(key.attribute == Attribute::DEFAULT);
+        auto&& i = key.input_format;
+        auto&& o = key.output_format;
+        auto ishp = tensor_formats_to_named_tensor_shape(i);
+        auto oshp = tensor_formats_to_named_tensor_shape(o);
+        auto builder = std::get<0>(ReformatEmitter{ishp, oshp}.emit());
+        return builder;
     }
     MGB_CATCH(std::exception & exc, {
         mgb_log_error(
@@ -390,10 +372,7 @@ const ReformatManager::ReformatImpl& ReformatManager::get(
 }
 
 const ReformatManager& ReformatManager::instance() {
-    static ReformatManager* inst = nullptr;
-    if (inst == nullptr) {
-        inst = new ReformatManager();
-    }
-    return *inst;
+    static ReformatManager inst;
+    return inst;
 }
 // vim: syntax=cpp.doxygen

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

@@ -227,6 +227,7 @@ namespace gopt {
         VarReplaceCheckFlag m_var_replace_check_flag =
                 VarReplaceCheckFlag::CHECK_ALL;
         class RelayoutPlaceholder;
+        friend class ShuffleShuffleRemovePass;
 
     public:
         TensorReformatPass& set_var_replace_check_flag(VarReplaceCheckFlag flag) {

src/gopt/include/megbrain/gopt/reformat_manager.h

@@ -49,10 +49,14 @@ enum class TensorFormats : uint32_t {
 
     KRSCk8 = 21,  ///< [K/8, R, S, C, K%8]
 
+    // NCHW4
+    KCRSc4 = 22,   ///< [K, C/4, R, S, C%4]
+    GKCRSc4 = 23,  ///< [G, K, C/4, R, S, C%4]
+
     // default weight format
-    KCRS = 22,   ///< [K, C, R, S]
-    GKCRS = 23,  ///< [G, K, C, R, S]
-    C11RS = 24,  ///< [C, 1, 1, R, S]
+    KCRS = 24,   ///< [K, C, R, S]
+    GKCRS = 25,  ///< [G, K, C, R, S]
+    C11RS = 26,  ///< [C, 1, 1, R, S]
 };
 
 class ReformatManager : public NonCopyableObj {
@@ -60,16 +64,20 @@ class ReformatManager : public NonCopyableObj {
 
 public:
     using ReformatImpl = thin_function<VarNode*(const VarNodeArray&)>;
+    struct ReformatKey {
         enum class Attribute : uint32_t {
             DEFAULT = 0,
             IMAGE2D = 1 << 0,
             IC_SMALL = 1 << 1,
         };
-    struct ReformatKey {
         TensorFormats input_format, output_format;
         DTypeEnum input_dtype, output_dtype;
         Attribute attribute;
         std::string to_string() const;
+        ReformatKey()
+                : input_dtype{DTypeEnum::Float32},
+                  output_dtype{DTypeEnum::Float32},
+                  attribute{Attribute::DEFAULT} {}
         ReformatKey(TensorFormats input_format_, TensorFormats output_format_,
                     Attribute attribute_ = Attribute::DEFAULT,
                     DTypeEnum input_dtype_ = DTypeEnum::Float32,
@@ -86,11 +94,13 @@ public:
             bool operator()(const ReformatKey& lhs,
                             const ReformatKey& rhs) const;
         };
+        ReformatKey& deduce_reformat_dtype_enum(const DType& dt);
     };
     using ReformatCache =
             std::unordered_map<ReformatKey, ReformatImpl, ReformatKey::Hash,
                                ReformatKey::Equal>;
-    const ReformatImpl& get(const ReformatKey& key) const;
+    ReformatImpl get(const ReformatKey& key) const;
+    ReformatImpl get(ReformatKey&& key) const { return get(key); }
     static const ReformatManager& instance();
 
 private:

src/gopt/test/reformat_manager.cpp

+/**
+ * \file src/gopt/test/reformat_manager.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "./helper.h"
+
+#include "megbrain/gopt/reformat_manager.h"
+#include "megbrain/opr/tensor_manip.h"
+
+using namespace mgb;
+using namespace gopt;
+
+TEST(TestReformatManager, Feature) {
+    constexpr size_t N = 16, C = 128, H = 7, W = 7;
+    HostTensorGenerator<> gen;
+    using ReformatKey = ReformatManager::ReformatKey;
+    auto src_format = TensorFormats::NHWC, dst_format = TensorFormats::NCHWc64;
+    ReformatKey key{src_format, dst_format};
+    auto reformat = ReformatManager::instance().get(key);
+
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+
+    auto r = [](VarNode* inp) {
+        auto x = SymbolVar(inp);
+        auto xshp = opr::GetVarShape::make(x);
+        auto cv = [&x](int v) { return x.make_scalar(v); };
+        auto sub = [&xshp, &cv](int idx) {
+            return opr::IndexAt::make(xshp, {{0, cv(idx)}});
+        };
+        auto tshp0 = opr::Concat::make(
+                {sub(0), sub(1), sub(2), sub(3) / 64, cv(64)}, 0);
+        auto y0 = opr::Reshape::make(x, tshp0);
+        auto y1 = opr::Dimshuffle::make(y0, {0, 3, 1, 2, 4});
+        return y1;
+    };
+
+    auto mkvar = [&](const char* name, const TensorShape& shp) {
+        return opr::Host2DeviceCopy::make(*graph, gen(shp)).rename(name);
+    };
+    auto x = mkvar("x", {N, H, W, C});
+    auto y1 = SymbolVar(reformat({x.node()}));
+    auto y2 = r(x.node());
+    size_t nr_shapeof = 0;
+    size_t nr_reshape = 0;
+    cg::DepOprIter{[&nr_shapeof, &nr_reshape](cg::OperatorNodeBase* o) {
+        if (o->same_type<opr::GetVarShape>())
+            nr_shapeof++;
+        if (o->same_type<opr::Reshape>())
+            nr_reshape++;
+    }}
+            .add(y1.node()->owner_opr());
+    ASSERT_EQ(nr_shapeof, 1);
+    ASSERT_EQ(nr_reshape, 1);
+    HostTensorND t1, t2;
+    auto func1 = graph->compile({make_callback_copy(y1, t1)});
+    func1->execute();
+    auto func2 = graph->compile({make_callback_copy(y2, t2)});
+    func2->execute();
+    MGB_ASSERT_TENSOR_EQ(t1, t2);
+}
+
+TEST(TestReformatManager, Weight) {
+    constexpr size_t G = 8, K = 128, C = 128, R = 3, S = 3;
+    HostTensorGenerator<> gen;
+    using ReformatKey = ReformatManager::ReformatKey;
+    auto src_format = TensorFormats::GKCRS,
+         dst_format = TensorFormats::GKCRSk4c4;
+    ReformatKey key{src_format, dst_format};
+    auto reformat = ReformatManager::instance().get(key);
+
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+
+    auto r = [](VarNode* inp) {
+        auto x = SymbolVar(inp);
+        auto xshp = opr::GetVarShape::make(x);
+        auto cv = [&x](int v) { return x.make_scalar(v); };
+        auto sub = [&xshp, &cv](int idx) {
+            return opr::IndexAt::make(xshp, {{0, cv(idx)}});
+        };
+        auto tshp0 = opr::Concat::make({sub(0), sub(1) / 4, cv(4), sub(2) / 4,
+                                        cv(4), sub(3), sub(4)},
+                                       0),
+             tshp1 = opr::Concat::make({sub(0), sub(1) / 4, sub(2) / 4, sub(3),
+                                        sub(4), cv(4), cv(4)},
+                                       0);
+        auto y0 = opr::Reshape::make(x, tshp0);
+        auto y1 = opr::Dimshuffle::make(y0, {0, 1, 3, 5, 6, 2, 4});
+        auto y2 = opr::Reshape::make(y1, tshp1);
+        return y2;
+    };
+
+    auto mkvar = [&](const char* name, const TensorShape& shp) {
+        return opr::Host2DeviceCopy::make(*graph, gen(shp)).rename(name);
+    };
+    auto w = mkvar("w", {G, K / G, C / G, R, S});
+    auto y1 = SymbolVar(reformat({w.node()}));
+    auto y2 = r(w.node());
+    size_t nr_shapeof = 0;
+    size_t nr_reshape = 0;
+    cg::DepOprIter{[&nr_shapeof, &nr_reshape](cg::OperatorNodeBase* o) {
+        if (o->same_type<opr::GetVarShape>())
+            nr_shapeof++;
+        if (o->same_type<opr::Reshape>())
+            nr_reshape++;
+    }}
+            .add(y1.node()->owner_opr());
+    ASSERT_EQ(nr_shapeof, 1);
+    ASSERT_EQ(nr_reshape, 1);
+    HostTensorND t1, t2;
+    auto func1 = graph->compile({make_callback_copy(y1, t1)});
+    func1->execute();
+    auto func2 = graph->compile({make_callback_copy(y2, t2)});
+    func2->execute();
+    MGB_ASSERT_TENSOR_EQ(t1, t2);
+}
+
+TEST(TestReformatManager, InvalidKey) {
+    using ReformatKey = ReformatManager::ReformatKey;
+    using Attribute = ReformatKey::Attribute;
+    auto src_format = TensorFormats::GKCRS,
+         dst_format = TensorFormats::GKCRSk4c4;
+    Attribute attribute = Attribute::IMAGE2D;
+    ReformatKey key{src_format, dst_format, attribute};
+    ASSERT_THROW(ReformatManager::instance().get(key), AssertionError);
+}
+
+TEST(TestReformatManager, InputChannelSmall) {
+    constexpr size_t N = 16, C = 3, H = 224, W = 224;
+    auto cn = CompNode::load("cpux");
+    HostTensorGenerator<> gen;
+    using ReformatKey = ReformatManager::ReformatKey;
+    using Attribute = ReformatKey::Attribute;
+    auto src_format = TensorFormats::NCHW, dst_format = TensorFormats::NCHWc4;
+    ReformatKey key{src_format, dst_format, Attribute::IC_SMALL};
+    auto reformat = ReformatManager::instance().get(key);
+
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+
+    auto r = [](VarNode* inp) {
+        auto x = SymbolVar(inp);
+        auto y = opr::RelayoutFormat::make(
+                x, megdnn::param::RelayoutFormat::Mode::NCHW_NCHW4_IC_SMALL);
+        return y;
+    };
+
+    auto mkvar = [&](const char* name, const TensorShape& shp) {
+        return opr::Host2DeviceCopy::make(*graph, gen(shp, cn)).rename(name);
+    };
+    auto x = mkvar("x", {N, C, H, W});
+    auto y1 = SymbolVar(reformat({x.node()}));
+    auto y2 = r(x.node());
+    HostTensorND t1, t2;
+    auto func1 = graph->compile({make_callback_copy(y1, t1)});
+    func1->execute();
+    auto func2 = graph->compile({make_callback_copy(y2, t2)});
+    func2->execute();
+    MGB_ASSERT_TENSOR_EQ(t1, t2);
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}