Add fuse_resnet_unit pass (#36818)

* GeneratePass support attr condition and mapping, test=develop

* fix coverage, test=develop

* Add fuse_resnet_unit pass, test=develop

* fix CI errors, test=develop

* fix CI errors, test=develop

* fix unittest error when compiling without CUDA, test=develop

* fix static ci error, test=develop

* limit kernel size must equal 1, test=develop

paddle/fluid/framework/ir/generate_pass.cc

@@ -19,6 +19,32 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+class element_visitor : public boost::static_visitor<Attribute> {
+ public:
+  explicit element_visitor(int index) : index_(index) {}
+
+  template <typename T>
+  Attribute operator()(const T& attr) const {
+    PADDLE_THROW(platform::errors::Unimplemented("Unimplemented operand."));
+  }
+
+  template <typename T>
+  Attribute operator()(const std::vector<T>& attr) const {
+    using ET = std::conditional_t<std::is_same<T, double>::value, float, T>;
+    int index = index_;
+    if (index < 0) {
+      index += attr.size();
+    }
+    if (index >= 0 && static_cast<size_t>(index) < attr.size()) {
+      return static_cast<ET>(attr[index]);
+    }
+    return boost::blank();
+  }
+
+ private:
+  int index_;
+};
+
 class operation_visitor : public boost::static_visitor<Attribute> {
  public:
   explicit operation_visitor(const proto::PassDesc::OperationType& type)
@@ -29,15 +55,17 @@ class operation_visitor : public boost::static_visitor<Attribute> {
     PADDLE_THROW(platform::errors::Unimplemented("Unimplemented operand."));
   }
 
-  template <typename T,
-            std::enable_if_t<std::is_integral<T>::value ||
-                             std::is_floating_point<T>::value>* = nullptr>
+  template <typename T, std::enable_if_t<std::is_integral<T>::value>* = nullptr>
   Attribute operator()(const T& attr, const T& operation) const {
     switch (type_) {
       case proto::PassDesc_OperationType_kSub: {
         return attr - operation;
       }
 
+      case proto::PassDesc_OperationType_kMod: {
+        return attr % operation;
+      }
+
       default:
         PADDLE_THROW(
             platform::errors::Unimplemented("Unimplemented operation type."));
@@ -72,6 +100,15 @@ Attribute GetVarAttrValue(const VarDesc* desc,
   return boost::blank();
 }
 
+Attribute GetOpAttrValue(const OpDesc* desc,
+                         const proto::PassDesc::Attr& attr) {
+  Attribute value = desc->GetAttr(attr.name());
+  if (attr.has_element_index()) {
+    value = boost::apply_visitor(element_visitor(attr.element_index()), value);
+  }
+  return value;
+}
+
 void InitGeneratePattern(const proto::PassDesc& pass_desc, PDPattern* pattern) {
   // Traverse all operators to create subgraph.
   for (int index = 0; index < pass_desc.pattern_size(); ++index) {
@@ -163,6 +200,11 @@ void InitGeneratePattern(const proto::PassDesc& pass_desc, PDPattern* pattern) {
       PDNode* pdnode = pattern->RetrieveNode(condition.attr().var_name());
       pdnode->assert_more([&](Node* x) {
         Attribute attr = GetVarAttrValue(x->Var(), condition.attr());
+        if (condition.has_operation()) {
+          Attribute operation = GetAttrValue(condition.operation().value());
+          attr = boost::apply_visitor(
+              operation_visitor(condition.operation().type()), attr, operation);
+        }
         switch (condition.type()) {
           case proto::PassDesc_ConditionType_kEQ: {
             return attr == GetAttrValue(condition.condition_value());
@@ -305,7 +347,12 @@ GraphPatternDetector::handle_t GetGenerateRewrite(
             node = graph->CreateVarNode(&var_desc);
             var_node_maps.insert({argument, node});
           } else {
-            node = iter->second;
+            if (in_nodes.end() ==
+                std::find(in_nodes.begin(), in_nodes.end(), iter->second)) {
+              node = iter->second;
+            } else {
+              node = graph->CreateVarNode(iter->second->Var());
+            }
           }
           out_nodes.push_back(node);
           arguments.push_back(node->Name());
@@ -329,7 +376,7 @@ GraphPatternDetector::handle_t GetGenerateRewrite(
             Node* condition_node = subgraph.at(pattern.RetrieveNode(
                 std::to_string(attr_map.pattern_attr().op_index())));
             attr =
-                condition_node->Op()->GetAttr(attr_map.pattern_attr().name());
+                GetOpAttrValue(condition_node->Op(), attr_map.pattern_attr());
           }
           if (attr_map.has_operation()) {
             Attribute operation = GetAttrValue(attr_map.operation().value());

paddle/fluid/framework/pass_desc.proto

@@ -26,6 +26,7 @@ message PassDesc {
     kMul = 2;
     kDiv = 3;
     kSize = 4;
+    kMod = 5;
   }
   enum ConditionType {
     kEQ = 0;

paddle/fluid/operators/fused/resnet_unit_op.cc

@@ -109,7 +109,12 @@ class ResNetUnitOp : public framework::OperatorWithKernel {
     // Check dims of inputs
     const auto x_dims = ctx->GetInputDim("X");
     const auto w_dims = ctx->GetInputDim("FilterX");
-    const auto bn_param_dims = ctx->GetInputDim("ScaleX");
+    std::vector<int64_t> bn_param_shape =
+        framework::vectorize(ctx->GetInputDim("ScaleX"));
+    if (1 == bn_param_shape.size()) {
+      bn_param_shape = {1, 1, 1, bn_param_shape[0]};
+    }
+    framework::DDim bn_param_dims = framework::make_ddim(bn_param_shape);
     PADDLE_ENFORCE_EQ(x_dims.size(), 4, platform::errors::InvalidArgument(
                                             "The dimensions of input "
                                             "must equal to 4."

paddle/fluid/operators/fused/resnet_unit_op.cu

@@ -68,8 +68,16 @@ class ResNetUnitKernel : public framework::OpKernel<T> {
 
     auto input_x_shape = framework::vectorize<int>(input_x->dims());
     auto filter_x_shape = framework::vectorize<int>(filter_x->dims());
+    // std::swap used to convert shape of filter from conv2d when kernel size is
+    // 1.
+    if (filter_x_shape[1] != filter_x_shape[2] && 1 == filter_x_shape[2]) {
+      std::swap(filter_x_shape[1], filter_x_shape[3]);
+    }
     auto param_dims = scale_x->dims();
     auto param_shape = framework::vectorize<int>(scale_x->dims());
+    if (1 == param_shape.size()) {
+      param_shape = {1, 1, 1, param_shape[0]};
+    }
     auto output_shape = framework::vectorize<int>(output->dims());
     auto bitmask_shape = framework::vectorize<int>(bitmask->dims());
     int output_channel = filter_x_shape[0];

paddle/fluid/pybind/pybind.cc

@@ -2508,13 +2508,13 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("is_profiler_enabled", platform::IsProfileEnabled);
   m.def("reset_profiler", platform::ResetProfiler);
-  m.def("register_pass", [](const std::string &pass_type,
-                            const py::object &callable) {
+  m.def("register_pass", [](const std::string &pass_type, py::object callable) {
     PADDLE_ENFORCE_EQ(
         framework::ir::PassRegistry::Instance().Has(pass_type), false,
         platform::errors::AlreadyExists(
             "Pass '%s' is registered more than once. Please use another name.",
             pass_type));
+    callable.inc_ref();
     framework::ir::PassRegistry::Instance().Insert(pass_type, [pass_type,
                                                                callable]() {
       py::gil_scoped_acquire guard;

python/paddle/fluid/ir.py

@@ -201,6 +201,12 @@ class RegisterPassHelper(object):
         return vars, program.current_block().ops
 
     def _convert_vars_to_pass_desc(self, patterns, replaces, desc):
+        def _add_element_conditions(conditions, elements):
+            for element in elements:
+                if element._condition:
+                    conditions.append(element._condition)
+                _add_element_conditions(conditions, element._elements)
+
         for (pattern, replace) in zip(patterns, replaces):
             # Convert maps of inputs and outputs.
             var_map = desc.var_maps.add()
@@ -218,10 +224,7 @@ class RegisterPassHelper(object):
             # Convert attr conditions.
             if PassDesc.VarHelper == pattern.__class__:
                 for attr in pattern._attrs.values():
-                    if attr._condition is not None:
-                        conditions.append(attr._condition)
-                    conditions.extend(
-                        [e._condition for e in attr._elements if e._condition])
+                    _add_element_conditions(conditions, [attr])
 
     def _convert_ops_to_pass_desc(self, patterns, replaces, desc):
         for replace in replaces:
@@ -324,6 +327,10 @@ class PassDesc(object):
             return self._clone_with_operation(
                 pass_desc_pb2.PassDesc.OperationType.kAdd, value)
 
+        def Mod(self, value):
+            return self._clone_with_operation(
+                pass_desc_pb2.PassDesc.OperationType.kMod, value)
+
         def Size(self):
             return self._clone_with_operation(
                 pass_desc_pb2.PassDesc.OperationType.kSize)
@@ -336,13 +343,20 @@ class PassDesc(object):
                 value._to_pass_desc_attr(condition.condition_attr)
             else:
                 self._to_op_desc_attr(value, condition.condition_value)
+            if self._operation:
+                condition.operation.CopyFrom(self._operation)
             self._condition = condition
 
         def EQ(self, value):
             self._set_with_condition(pass_desc_pb2.PassDesc.ConditionType.kEQ,
                                      value)
 
-        def MappedPattern(self, var=None, op=None, index=0, name=None):
+        def MappedPattern(self,
+                          var=None,
+                          op=None,
+                          index=0,
+                          name=None,
+                          element_index=None):
             if all([var, op]):
                 raise ValueError("Only mapped one of which var or op.")
 
@@ -356,7 +370,8 @@ class PassDesc(object):
                     raise ValueError(
                         "Index '{}' of operator '{}' is incorrect.".format(
                             index, op))
-                return PassDesc.AttrHelper(ops[index], name)
+                return PassDesc.AttrHelper(
+                    ops[index], name, element_index=element_index)
 
             self._mapped = mapped_op if var is None else mapped_var
 
@@ -460,6 +475,13 @@ class PassDesc(object):
         def Outputs(self):
             return self._outputs
 
+        def SetOutputs(self, **kwargs):
+            for param, arg in kwargs.items():
+                if arg is None:
+                    self._desc.remove_output(param)
+                else:
+                    self._desc.set_output(param, [arg.name])
+
     OP = OpHelper()
 
 

python/paddle/fluid/tests/unittests/ir/CMakeLists.txt

@@ -10,3 +10,5 @@ foreach(target ${TEST_IR_PASSES})
 endforeach()
 
 add_subdirectory(inference)
+
+set_tests_properties(test_fuse_resnet_unit PROPERTIES TIMEOUT 120)

python/paddle/fluid/tests/unittests/ir/test_fuse_resnet_unit.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+import paddle
+import paddle.incubate
+from paddle.fluid import core
+from paddle.vision.models import ResNet
+from paddle.vision.models.resnet import BottleneckBlock, BasicBlock
+
+paddle.enable_static()
+np.random.seed(0)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_cuda() or
+                 paddle.get_cudnn_version() < 8000,
+                 "only support with cuda and cudnn version is at least 8.0.")
+class TestFuseResNetUnit(unittest.TestCase):
+    def test_fuse_resenet_unit(self):
+        place = paddle.CUDAPlace(0)
+        program = paddle.static.Program()
+        startup_program = paddle.static.Program()
+        with paddle.static.amp.fp16_guard():
+            with paddle.static.program_guard(program, startup_program):
+                x = paddle.static.data("x", [1, 64, 64, 8])
+                conv2d = paddle.nn.Conv2D(
+                    8, 32, 1, bias_attr=False, data_format='NHWC')
+                batch_norm = paddle.nn.BatchNorm(
+                    32, act='relu', data_layout='NHWC')
+                out = batch_norm(conv2d(x))
+        graph = core.Graph(program.desc)
+        core.get_pass("fuse_resnet_unit").apply(graph)
+        after_program = paddle.fluid.framework.IrGraph(graph).to_program()
+        params = paddle.static.amp.cast_model_to_fp16(program)
+        after_params = paddle.static.amp.cast_model_to_fp16(after_program)
+        exe = paddle.static.Executor(place)
+        exe.run(startup_program)
+        paddle.static.amp.cast_parameters_to_fp16(
+            place, program, to_fp16_var_names=params)
+        paddle.static.amp.cast_parameters_to_fp16(
+            place, after_program, to_fp16_var_names=after_params)
+        feed = {"x": np.random.randn(1, 64, 64, 8).astype("float16")}
+        before_out = exe.run(program, feed=feed, fetch_list=[out.name])
+        after_out = exe.run(after_program, feed=feed, fetch_list=[out.name])
+        self.assertTrue(np.allclose(before_out[0], after_out[0], atol=5e-3))

python/paddle/incubate/__init__.py

@@ -23,6 +23,7 @@ from .tensor import segment_sum
 from .tensor import segment_mean
 from .tensor import segment_max
 from .tensor import segment_min
+from .passes import fuse_resnet_unit_pass
 
 from . import nn  #noqa: F401
 

python/paddle/incubate/passes/__init__.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

python/paddle/incubate/passes/fuse_resnet_unit_pass.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.fluid.ir as ir
+
+
+def set_resnet_unit_attrs(resnet_unit, has_shortcut):
+    resnet_unit.SetAttr("fuse_add", False)
+    resnet_unit.SetAttr("act_type", "relu")
+    resnet_unit.SetAttr("has_shortcut", has_shortcut)
+    resnet_unit.SetAttr("data_format", 'NHWC')
+    resnet_unit.SetAttr("dilation", 1)
+    resnet_unit.Attr("stride").MappedPattern(
+        op="conv2d", name="strides", element_index=0)
+    resnet_unit.Attr("padding").MappedPattern(
+        op="conv2d", name="paddings", element_index=0)
+    resnet_unit.Attr("group").MappedPattern(op="conv2d", name="groups")
+    resnet_unit.Attr("op_device").MappedPattern(op="conv2d", name="op_device")
+    resnet_unit.Attr("op_namescope").MappedPattern(
+        op="conv2d", name="op_namescope")
+    resnet_unit.Attr("momentum").MappedPattern(op="batch_norm", name="momentum")
+    resnet_unit.Attr("epsilon").MappedPattern(op="batch_norm", name="epsilon")
+    resnet_unit.Attr("use_global_stats").MappedPattern(
+        op="batch_norm", name="use_global_stats")
+
+
+def set_resnet_unit_outputs(resnet_unit, meanX, varX, meanZ=None, varZ=None):
+    resnet_unit.SetOutputs(
+        RunningMeanX=meanX,
+        RunningVarX=varX,
+        RunningMeanZ=meanZ,
+        RunningVarZ=varZ)
+
+
+@ir.RegisterPass
+def fuse_resnet_unit():
+    def pattern_conv_bn(x, filter, scale, bias, mean, var):
+        filter.Attr("shape")[0].Mod(32).EQ(0)
+        filter.Attr("shape")[1].Mod(8).EQ(0)
+        filter.Attr("shape")[2].EQ(1)
+        filter.Attr("shape")[3].EQ(1)
+        conv2d = ir.PassDesc.OP.conv2d(Input=x, Filter=filter)
+        conv2d.SetAttr("data_format", 'NHWC')
+        bn = ir.PassDesc.OP.batch_norm(
+            X=conv2d, Bias=bias, Mean=mean, Scale=scale, Variance=var)
+        return bn
+
+    def pattern_one_input(x, filter, scale, bias, mean, var):
+        bn = pattern_conv_bn(x, filter, scale, bias, mean, var)
+        relu = ir.PassDesc.OP.relu(X=bn.Output("Y"))
+        return relu
+
+    def replace_one_input(x, filter, scale, bias, mean, var):
+        resnet_unit = ir.PassDesc.OP.resnet_unit(
+            X=x, FilterX=filter, ScaleX=scale, BiasX=bias, MeanX=mean, VarX=var)
+        set_resnet_unit_attrs(resnet_unit, False)
+        set_resnet_unit_outputs(resnet_unit, mean, var)
+        return resnet_unit.Output("Y")
+
+    def pattern_two_input(x, filterX, scaleX, biasX, meanX, varX, z, filterZ,
+                          scaleZ, biasZ, meanZ, varZ):
+        bnX = pattern_conv_bn(x, filterX, scaleX, biasX, meanX, varX)
+        bnZ = pattern_conv_bn(x, filterZ, scaleZ, biasZ, meanZ, varZ)
+        ewadd = ir.PassDesc.OP.elementwise_add(
+            X=bnX.Output("Y"), Y=bnZ.Output("Y"))
+        relu = ir.PassDesc.OP.relu(X=ewadd)
+        return relu
+
+    def replace_two_input(x, filterX, scaleX, biasX, meanX, varX, z, filterZ,
+                          scaleZ, biasZ, meanZ, varZ):
+        resnet_unit = ir.PassDesc.OP.resnet_unit(
+            X=x,
+            FilterX=filterX,
+            ScaleX=scaleX,
+            BiasX=biasX,
+            MeanX=meanX,
+            VarX=varX,
+            Z=z,
+            FilterZ=filterZ,
+            ScaleZ=scaleZ,
+            BiasZ=biasZ,
+            MeanZ=meanZ,
+            VarZ=varZ)
+        set_resnet_unit_attrs(resnet_unit, True)
+        set_resnet_unit_outputs(resnet_unit, meanX, varX, meanZ, varZ)
+        return resnet_unit.Output("Y")
+
+    return (pattern_one_input, replace_one_input), (pattern_two_input,
+                                                    replace_two_input)

python/setup.py.in

@@ -275,6 +275,7 @@ packages=['paddle',
           'paddle.incubate.operators',
           'paddle.incubate.tensor',
           'paddle.incubate.nn',
+          'paddle.incubate.passes',
           'paddle.distributed.fleet',
           'paddle.distributed.fleet.base',
           'paddle.distributed.fleet.elastic',