Merge branch 'develop' into split_api

Author: Enigmatisms

paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc

@@ -195,7 +195,7 @@ int FCLstmFusePass::BuildFusion(Graph* graph,
                           Node* cell,
                           Node* xx,
                           Node* fc_bias,
-                          const bool use_mkldnn) {
+                          const bool use_onednn) {
     OpDesc op_desc;
     op_desc.SetType("fusion_lstm");
 #define SET_IN(Key, node__) op_desc.SetInput(#Key, {node__->Name()});
@@ -235,7 +235,7 @@ int FCLstmFusePass::BuildFusion(Graph* graph,
     op_desc.SetOutput("XX", {xx->Name()});
     op_desc.SetAttr("is_reverse", lstm->Op()->GetAttr("is_reverse"));
     op_desc.SetAttr("use_peepholes", lstm->Op()->GetAttr("use_peepholes"));
-    op_desc.SetAttr("use_mkldnn", use_mkldnn);
+    op_desc.SetAttr("use_onednn", use_onednn);
     // TODO(TJ): get from attr
     op_desc.SetAttr("use_seq", true);
 
@@ -300,8 +300,9 @@ int FCLstmFusePass::BuildFusion(Graph* graph,
     GET_IR_NODE_FROM_SUBGRAPH(Cell, Cell, lstm_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(w, w, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(mul, mul, fc_pattern);
-    const bool use_mkldnn =
-        (mul->Op()->GetAttrIfExists<bool>("use_mkldnn") &&
+    const bool use_onednn =
+        ((mul->Op()->GetAttrIfExists<bool>("use_mkldnn") ||
+          mul->Op()->GetAttrIfExists<bool>("use_onednn")) &&
          lstm->Op()->GetAttrIfExists<std::string>("gate_activation") ==
              "sigmoid" &&
          lstm->Op()->GetAttrIfExists<std::string>("cell_activation") ==
@@ -323,7 +324,7 @@ int FCLstmFusePass::BuildFusion(Graph* graph,
                    Cell,
                    fc_out,
                    fc_bias,
-                   use_mkldnn);
+                   use_onednn);
       // Remove unneeded nodes.
       std::unordered_set<const Node*> marked_nodes(
           {mul, lstm, elementwise_add, mul_out, BatchGate, BatchCellPreAct});
@@ -339,7 +340,7 @@ int FCLstmFusePass::BuildFusion(Graph* graph,
                    Cell,
                    fc_out,
                    nullptr,
-                   use_mkldnn);
+                   use_onednn);
       // Remove unneeded nodes.
       std::unordered_set<const Node*> marked_nodes(
           {mul, lstm, BatchGate, BatchCellPreAct});

paddle/fluid/framework/ir/fuse_pass_base.cc

@@ -58,10 +58,16 @@ void FusePassBase::AddStatis(int count_of_fused) const {
 FuseOptions FusePassBase::FindFuseOption(const Node& node1,
                                          const Node& node2) const {
 #ifdef PADDLE_WITH_DNNL
-  bool node1_onednn = node1.Op()->HasAttr("use_mkldnn") &&
-                      PADDLE_GET_CONST(bool, node1.Op()->GetAttr("use_mkldnn"));
-  bool node2_onednn = node2.Op()->HasAttr("use_mkldnn") &&
-                      PADDLE_GET_CONST(bool, node2.Op()->GetAttr("use_mkldnn"));
+  bool node1_onednn =
+      (node1.Op()->HasAttr("use_mkldnn") &&
+       PADDLE_GET_CONST(bool, node1.Op()->GetAttr("use_mkldnn"))) ||
+      (node1.Op()->HasAttr("use_onednn") &&
+       PADDLE_GET_CONST(bool, node1.Op()->GetAttr("use_onednn")));
+  bool node2_onednn =
+      (node2.Op()->HasAttr("use_mkldnn") &&
+       PADDLE_GET_CONST(bool, node2.Op()->GetAttr("use_mkldnn"))) ||
+      (node2.Op()->HasAttr("use_onednn") &&
+       PADDLE_GET_CONST(bool, node2.Op()->GetAttr("use_onednn")));
   if (node1_onednn && node2_onednn)
     return FUSE_ONEDNN;
   else if (!node1_onednn && !node2_onednn)

paddle/fluid/framework/ir/gpu_cpu_map_matmul_to_mul_pass.cc

@@ -444,6 +444,9 @@ void GpuCpuMapMatmulV2ToMatmulPass::ApplyImpl(ir::Graph* graph) const {
     if (matmul_v2_op->Op()->HasAttr("use_mkldnn")) {
       desc.SetAttr("use_mkldnn", matmul_v2_op->Op()->GetAttr("use_mkldnn"));
     }
+    if (matmul_v2_op->Op()->HasAttr("use_onednn")) {
+      desc.SetAttr("use_onednn", matmul_v2_op->Op()->GetAttr("use_onednn"));
+    }
     if (matmul_v2_op->Op()->HasAttr("enable_int8")) {
       desc.SetAttr("enable_int8", matmul_v2_op->Op()->GetAttr("enable_int8"));
       desc.SetAttr("Input_scale", matmul_v2_op->Op()->GetAttr("Input_scale"));

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -3289,7 +3289,7 @@ PDNode *patterns::UnsupportedBfloat16::operator()() {
   return op;
 }
 
-PDNode *patterns::Bloat16Ops::operator()() {
+PDNode *patterns::Bfloat16Ops::operator()() {
   auto op = pattern->NewNode(op_repr())->assert_is_op();
   op->assert_more([&](Node *node) {
     return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") ==

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -1774,8 +1774,8 @@ struct UnsupportedBfloat16 : public PatternBase {
   PATTERN_DECL_NODE(op);
 };
 
-struct Bloat16Ops : public PatternBase {
-  Bloat16Ops(PDPattern* pattern, const std::string& name_scope)
+struct Bfloat16Ops : public PatternBase {
+  Bfloat16Ops(PDPattern* pattern, const std::string& name_scope)
       : PatternBase(pattern, name_scope, "many_bfloat16_ops") {}
 
   PDNode* operator()();

paddle/fluid/framework/ir/onednn/cpu_bfloat16_pass.cc

@@ -249,11 +249,11 @@ void CPUBFloat16Pass::ApplyImpl(ir::Graph* graph) const {
   int dequantize_counter = 0;
 
   GraphPatternDetector gpd;
-  patterns::Bloat16Ops Bloat16Ops{gpd.mutable_pattern(), "Bloat16Ops"};
-  Bloat16Ops();
+  patterns::Bfloat16Ops Bfloat16Ops{gpd.mutable_pattern(), "Bfloat16Ops"};
+  Bfloat16Ops();
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* graph) {
-    GET_IR_NODE_FROM_SUBGRAPH(op, op, Bloat16Ops);
+    GET_IR_NODE_FROM_SUBGRAPH(op, op, Bfloat16Ops);
 
     Quantizer quantizer(graph, op);
     quantizer.AddQuantOps();

paddle/fluid/framework/ir/onednn/cpu_bfloat16_pass_tester.cc

@@ -28,7 +28,7 @@ void SetOp(ProgramDesc* prog,
            const std::string& onednn_data_type = "float32") {
   auto* op = prog->MutableBlock(0)->AppendOp();
   op->SetType(type);
-  op->SetAttr("use_mkldnn", use_onednn);
+  op->SetAttr("use_onednn", use_onednn);
   op->SetAttr("name", name);
 
   if (type == "conv2d") {

paddle/fluid/framework/ir/onednn/interpolate_onednn_pass.cc

@@ -31,7 +31,8 @@ void InterpolateOneDNNPass::ApplyImpl(ir::Graph* graph) const {
   PADDLE_ENFORCE_NOT_NULL(graph,
                           common::errors::InvalidArgument(
                               "Pointer to graph argument should not be NULL."));
-  if (!(graph->Has("use_mkldnn") && graph->Get<bool>("use_mkldnn"))) {
+  if (!(graph->Has("use_mkldnn") && graph->Get<bool>("use_mkldnn")) &&
+      !(graph->Has("use_onednn") && graph->Get<bool>("use_onednn"))) {
     VLOG(3) << "Do not handle interpolate_onednn_pass";
     return;
   }
@@ -53,7 +54,7 @@ void InterpolateOneDNNPass::ApplyImpl(ir::Graph* graph) const {
                                   interpolate_op_types.end(),
                                   node->Name()) != interpolate_op_types.end()) {
       auto* op_desc = node->Op();
-      op_desc->SetAttr("use_mkldnn", true);
+      op_desc->SetAttr("use_onednn", true);
       ++found_count;
     }
   }

paddle/fluid/framework/ir/onednn/multi_gru_fuse_pass.cc

@@ -186,7 +186,7 @@ MultiGRUFusePass::MultiGRUFusePass() {
       .AddAttr("origin_mode")
       .IsType<bool>()
       .End()
-      .AddAttr("use_mkldnn")
+      .AddAttr("use_onednn")
       .IsType<bool>()
       .End()
       .AddAttr("mkldnn_data_type")

paddle/fluid/framework/op_desc.h

@@ -235,7 +235,7 @@ class TEST_API OpDesc {
   // attribute name => all original attrs
   AttributeMap attrs_;
   // runtime_attrs_ contains the attributes which used for dispatching kernel
-  // (use_mkldnn, use_cudnn, ...) or passing additional configuration for
+  // (use_onednn, use_cudnn, ...) or passing additional configuration for
   // special heterogeneous kernel (workspace_size_MB, ...).
   // The attributes in runtime_attrs_ are set by framework (such as PASS),
   // and not in the python api.