[TF2XLA] Display Python location information in XLA metadata

As a heuristic, we show the last frame which does not have "tensorflow/python"
substring.

PiperOrigin-RevId: 346892818
Change-Id: I72b5953ba5de9c0a10648805a852fcc033dd960b

tensorflow/compiler/mlir/xla/transforms/legalize_tf_with_tf2xla.cc

@@ -352,7 +352,8 @@ LogicalResult Tf2XlaRewriter::PrepareParams() {
   // XlaCompiler within the context is only used by the functional ops to
   // compile functions. We are not handling those at the moment so XlaCompiler
   // is not required.
-  context_ = new tensorflow::XlaContext(/*compiler=*/nullptr, &hlo_builder_);
+  context_ = new tensorflow::XlaContext(/*compiler=*/nullptr, &hlo_builder_,
+                                        /*graph=*/nullptr);
   context_->Ref();
 
   device_mgr_ = CreateDeviceMgr(device_type_);

tensorflow/compiler/tf2xla/xla_compilation_device.cc

@@ -83,14 +83,31 @@ Allocator* XlaCompilationDevice::GetAllocator(AllocatorAttributes attr) {
   return allocator_.get();
 }
 
+// Attaches location from the node stack trace to metadata. As a heuristic,
+// picks the last frame which does not contain the "tensorflow/python" substring
+// (making exception for frames containing "test" to allow for testing the
+// feature).
+static void AttachLocationToMetadata(xla::OpMetadata& metadata,
+                                     OpKernel* op_kernel, XlaContext& context) {
+  if (const AbstractStackTrace* stack_trace =
+          context.StackTraceForNodeName(op_kernel->def().name())) {
+    if (absl::optional<StackFrame> frame = stack_trace->LastUserFrame()) {
+      metadata.set_source_file(frame->file_name);
+      metadata.set_source_line(frame->line_number);
+    }
+  }
+}
+
 void XlaCompilationDevice::Compute(OpKernel* op_kernel,
                                    OpKernelContext* context) {
   VLOG(4) << "XlaCompilationDevice::Compute "
           << FormatNodeDefForError(op_kernel->def());
-  auto* b = XlaContext::Get(context).builder();
+  XlaContext& xla_context = XlaContext::Get(context);
+  auto* b = xla_context.builder();
   xla::OpMetadata metadata;
   metadata.set_op_type(op_kernel->type_string());
   metadata.set_op_name(op_kernel->name());
+  AttachLocationToMetadata(metadata, op_kernel, xla_context);
   b->SetOpMetadata(metadata);
 
   auto sharding_parse_result = ParseShardingFromDevice(

tensorflow/compiler/tf2xla/xla_compiler.cc

@@ -1309,7 +1309,7 @@ Status XlaCompiler::CompileGraph(
                                    options_.device_type, name));
 
   xla::XlaBuilder builder(name);
-  XlaContext* context = new XlaContext(this, &builder);
+  XlaContext* context = new XlaContext(this, &builder, graph.get());
   core::ScopedUnref context_unref(context);
 
   std::vector<XlaCompiler::Argument> real_args(args.begin(), args.end());

tensorflow/compiler/tf2xla/xla_context.cc

@@ -57,8 +57,15 @@ void XlaContext::set_args(std::vector<XlaExpression> args) {
   args_ = std::move(args);
 }
 
-XlaContext::XlaContext(XlaCompiler* compiler, xla::XlaBuilder* builder)
-    : compiler_(compiler), builder_(builder) {}
+XlaContext::XlaContext(XlaCompiler* compiler, xla::XlaBuilder* builder,
+                       const Graph* graph)
+    : compiler_(compiler), builder_(builder) {
+  if (graph) {
+    for (const Node* node : graph->nodes()) {
+      stack_traces_[node->name()] = node->GetStackTrace();
+    }
+  }
+}
 
 string XlaContext::DebugString() const { return "XLA JIT context"; }
 

tensorflow/compiler/tf2xla/xla_context.h

@@ -27,6 +27,7 @@ limitations under the License.
 #include "tensorflow/compiler/xla/xla_data.pb.h"
 #include "tensorflow/core/framework/op_kernel.h"
 #include "tensorflow/core/framework/resource_mgr.h"
+#include "tensorflow/core/graph/graph.h"
 #include "tensorflow/core/platform/macros.h"
 
 namespace tensorflow {
@@ -44,13 +45,22 @@ class XlaContext : public ResourceBase {
 
   // Creates a new XlaContext. See the documentation on the class data fields
   // for descriptions of the arguments.
-  XlaContext(XlaCompiler* compiler, xla::XlaBuilder* builder);
+  XlaContext(XlaCompiler* compiler, xla::XlaBuilder* builder,
+             const Graph* graph);
 
   // Virtual method defined by ResourceBase.
   string DebugString() const override;
 
   XlaCompiler* compiler() const { return compiler_; }
 
+  const AbstractStackTrace* StackTraceForNodeName(const std::string& name) {
+    const auto& it = stack_traces_.find(name);
+    if (it != stack_traces_.end()) {
+      return it->second.get();
+    }
+    return nullptr;
+  }
+
   // Returns the XlaBuilder that Ops use for compiling new expressions.
   xla::XlaBuilder* builder() { return builder_; }
 
@@ -100,6 +110,9 @@ class XlaContext : public ResourceBase {
   // The XlaBuilder used to construct the subgraph's compiled representation.
   xla::XlaBuilder* builder_;
 
+  // Stack traces for the graph used for compilation.
+  StackTracesMap stack_traces_;
+
   // Arguments to the Tensorflow graph, indexed by _Arg index.
   // Includes both compile-time constant arguments and runtime parameters.
   std::vector<XlaExpression> args_;

tensorflow/core/common_runtime/inline_function_utils.cc

@@ -616,6 +616,7 @@ Status InlineFunctionBody(const FunctionLibraryDefinition& flib_def, Graph* g,
     Node* clone = g->AddNode(ndef, &added_node);
     TF_CHECK_OK(added_node);
     node_map[n->id()] = clone;
+    clone->SetStackTrace(n->GetStackTrace());
 
     // If there is an input control node, and one of:
     // a) the node has no data or control inputs, or

tensorflow/core/graph/graph.cc

@@ -454,6 +454,7 @@ Node* Graph::CopyNode(const Node* node) {
     copy->MaybeCopyOnWrite();
     copy->props_->op_def = op_def;
   }
+  copy->SetStackTrace(node->GetStackTrace());
 
   return copy;
 }

tensorflow/python/eager/def_function_xla_jit_test.py

@@ -170,6 +170,36 @@ class DefFunctionTest(xla_test.XLATestCase):
                                   'not compilable'):
         xla_func(inputs)
 
+  @test_util.disable_mlir_bridge('TODO(b/155782411): MLIR bridge does not'
+                                 'support stack traces')
+  def testPythonLocationInMetadata(self):
+    with ops.device('device:{}:0'.format(self.device)):
+
+      @def_function.function(jit_compile=True)
+      def fn(x, y):
+        return x + y
+
+      inputs = constant_op.constant([1, 2, 2, 3, 3])
+      self.assertIn('def_function_xla_jit_test',
+                    fn.experimental_get_compiler_ir(inputs, inputs)())
+
+  @test_util.disable_mlir_bridge('TODO(b/155782411): MLIR bridge does not'
+                                 'support stack traces')
+  def testPythonLocationNestedInMetadata(self):
+    with ops.device('device:{}:0'.format(self.device)):
+
+      @def_function.function(jit_compile=True)
+      def f(x, y):
+        return x + y
+
+      @def_function.function(jit_compile=True)
+      def g(x, y):
+        return f(x, y)
+
+      inputs = constant_op.constant([1, 2, 2, 3, 3])
+      self.assertIn('def_function_xla_jit_test',
+                    g.experimental_get_compiler_ir(inputs, inputs)())
+
   @test_util.disable_mlir_bridge('TODO(b/155782411): MLIR bridge does not'
                                  'support stack traces')
   def testPythonStackTrace(self):

tensorflow/python/util/tf_stack.cc

@@ -177,11 +177,11 @@ class StackTraceWrapper : public AbstractStackTrace {
   void GenerateCache() const {
     // Grabbing the GIL solves two purposes: 1) makes the class thread-safe, and
     // 2) ToStackFrames and LineContents actually need it.
-    PyGILState_STATE state = PyGILState_Ensure();
     if (stack_frames_cache_) {
       return;
     }
 
+    PyGILState_STATE state = PyGILState_Ensure();
     absl::flat_hash_map<std::pair<std::string, int>, StackFrame> m;
     absl::flat_hash_set<std::string> f;