Merge remote-tracking branch 'ups/develop' into refine/infershape

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -130,12 +130,13 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler {
 
   std::shared_ptr<mkldnn::memory> AcquireWeightsMemoryFromPrimitive(
       const std::shared_ptr<mkldnn::memory> user_weights_memory_p,
-      std::vector<mkldnn::primitive>& pipeline) {  // NOLINT
+      std::vector<mkldnn::primitive>& pipeline,  // NOLINT
+      bool is_persistent = false) {
     auto user_weights_pd = user_weights_memory_p->get_primitive_desc();
     auto weights_pd = conv_pd_->weights_primitive_desc();
     return this->AcquireMemory(weights_pd, user_weights_pd,
                                user_weights_memory_p, "@weights_mem_p",
-                               pipeline);
+                               pipeline, is_persistent);
   }
 
   std::shared_ptr<mkldnn::memory> AcquireBiasMemoryFromPrimitive(
@@ -266,6 +267,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                    "It must use CPUPlace.");
 
+    const bool is_test = ctx.Attr<bool>("is_test");
+
     auto& dev_ctx =
         ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
     const auto& mkldnn_engine = dev_ctx.GetEngine();
@@ -296,6 +299,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
     std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
+    bool fuse_relu = ctx.Attr<bool>("fuse_relu");
     int groups = ctx.Attr<int>("groups");
 
     // TODO(pzelazko-intel) add support for group convolution and dilation
@@ -348,11 +352,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       bias_tz = paddle::framework::vectorize2int(bias->dims());
       auto bias_md = platform::MKLDNNMemDesc(
           bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
-      conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md,
-                                     strides, paddings, mkldnn_engine);
+      conv_pd =
+          ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md, strides,
+                               paddings, mkldnn_engine, fuse_relu);
     } else {
       conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides,
-                                     paddings, mkldnn_engine);
+                                     paddings, mkldnn_engine, fuse_relu);
     }
     // Save conv_pd/src_memory/weights_memory for backward pass
     dev_ctx.SetBlob(key_conv_pd, conv_pd);
@@ -371,7 +376,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto src_memory_p =
         handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
     auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive(
-        user_weights_memory_p, pipeline);
+        user_weights_memory_p, pipeline, is_test);
     auto dst_memory_p =
         handler.AcquireDstMemoryFromPrimitive(to_void_cast<T>(output_data));
 
@@ -402,11 +407,26 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   }
 
  private:
+  mkldnn::primitive_attr AddRelu() const {
+    // Fusion with ReLU layer is executed through the PostOps feature. Create a
+    // PostOps object and configure it to execute an eltwise relu operation.
+    mkldnn::primitive_attr conv_attr;
+    constexpr float scale = 1.0f;
+    constexpr float negative_slope = 0.0f;
+    constexpr float placeholder = 0.0f;
+    mkldnn::post_ops post_operations;
+    post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
+                                   negative_slope, placeholder);
+    conv_attr.set_post_ops(post_operations);
+    return conv_attr;
+  }
+
   std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
   ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
                        const memory::desc& dst, const std::vector<int>& strides,
                        const std::vector<int>& paddings,
-                       const mkldnn::engine& engine) const {
+                       const mkldnn::engine& engine,
+                       const bool fuse_relu) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -415,8 +435,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         dst, stride_dims, padding_dims, padding_dims,
         mkldnn::padding_kind::zero);
 
-    auto p_conv_pd =
-        new mkldnn::convolution_forward::primitive_desc(conv_desc, engine);
+    mkldnn::primitive_attr conv_attr;
+    if (fuse_relu) {
+      conv_attr = AddRelu();
+    }
+
+    auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
+        conv_desc, conv_attr, engine);
 
     return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
         p_conv_pd);
@@ -427,7 +452,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const memory::desc& bias, const memory::desc& dst,
                        const std::vector<int>& strides,
                        const std::vector<int>& paddings,
-                       const mkldnn::engine& engine) const {
+                       const mkldnn::engine& engine,
+                       const bool fuse_relu) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -436,8 +462,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         bias, dst, stride_dims, padding_dims, padding_dims,
         mkldnn::padding_kind::zero);
 
-    auto p_conv_pd =
-        new mkldnn::convolution_forward::primitive_desc(conv_desc, engine);
+    mkldnn::primitive_attr conv_attr;
+    if (fuse_relu) {
+      conv_attr = AddRelu();
+    }
+
+    auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
+        conv_desc, conv_attr, engine);
 
     return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
         p_conv_pd);

paddle/fluid/operators/conv_op.cc

@@ -109,6 +109,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
 }
 
 void Conv2DOpMaker::Make() {
+  AddAttr<bool>("is_test", "").SetDefault(false);
   AddInput(
       "Input",
       "(Tensor) The input tensor of convolution operator. "
@@ -161,6 +162,8 @@ void Conv2DOpMaker::Make() {
   AddAttr<bool>("use_mkldnn",
                 "(bool, default false) Only used in mkldnn kernel")
       .SetDefault(false);
+  AddAttr<bool>("fuse_relu", "(bool, default false) Only used in mkldnn kernel")
+      .SetDefault(false);
   AddAttr<std::string>(
       "data_format",
       "(string, default NCHW) Only used in "

paddle/fluid/platform/mkldnn_helper.h

@@ -192,7 +192,8 @@ class MKLDNNHandler {
       mkldnn::memory::primitive_desc& user_mpd,  // NOLINT
       const std::shared_ptr<mkldnn::memory> user_memory_p,
       const std::string& suffix,
-      std::vector<mkldnn::primitive>& pipeline) {  // NOLINT
+      std::vector<mkldnn::primitive>& pipeline,  // NOLINT
+      bool is_persistent = false) {
     // create reorder primitive if the input format is not the preferred one
     auto local_key = key_ + suffix;
     auto key_reorder_p = key_ + suffix + "reorder_p";
@@ -213,7 +214,7 @@ class MKLDNNHandler {
         pipeline.push_back(*reorder_p);
       }
       dev_ctx_.SetBlob(local_key, target_memory_p);
-    } else {
+    } else if (!is_persistent) {
       // Make reorder if needed
       auto reorder_p = std::static_pointer_cast<mkldnn::reorder>(
           dev_ctx_.GetBlob(key_reorder_p));

python/paddle/fluid/transpiler/inference_transpiler.py

@@ -60,12 +60,46 @@ class InferenceTranspiler(object):
         if not isinstance(scope, core.Scope):
             raise TypeError("scope should be as Scope type or None")
         use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False))
+
         self._fuse_batch_norm(program, place, scope)
         if use_mkldnn:
-            self._fuse_relu_mkldnn(program)
             self._fuse_conv_bias_mkldnn(program)
+            self._fuse_conv_relu_mkldnn(program)
+            self._fuse_bn_relu_mkldnn(program)
+
+    def _fuse_conv_relu_mkldnn(self, program):
+        '''
+        Transpile the program by fused relu activation for MKLDNN program.
+        Relu activation following convolution OP can be fused by adding
+        'fuse_relu' attribute to convolution OP.
+        The result of fuse is:
+            - before:
+                - conv->relu->any_other_op
+            - after:
+                - conv->any_other_op
+        :param program: program to transpile
+        :type program: Program
+        '''
+        self.block = program.block(0)
+
+        i = 0
+        while i < len(self.block.ops):
+            current_op = self.block.ops[i]
+            if current_op.type in ['conv2d']:
+                next_op = self.block.ops[i + 1]
+                if next_op.type == 'relu':
+                    # modify conv OP to include relu
+                    current_op.set_attr("fuse_relu", True)
+                    # remove conv OP
+                    self.block._remove_op(i + 1)
+            i = i + 1
+
+        # TODO(luotao): use clone() method to flush the program.desc in force,
+        # since some large program.desc will not be flushed immediately.
+        # And a better solution will be considered later.
+        program = program.clone()
 
-    def _fuse_relu_mkldnn(self, program):
+    def _fuse_bn_relu_mkldnn(self, program):
         '''
         Transpile the program by fused relu activation for MKLDNN program.
 
@@ -159,7 +193,6 @@ class InferenceTranspiler(object):
                 self._fuse_conv_bias(i, current_op, next_op)
                 self.block._remove_op(i + 1)  # Remove old conv
                 self.block._remove_op(i + 1)  # Remove elementwise_add
-                i = i + 1
             i = i + 1
 
         self._remove_unused_var()