Pick in XNNPACK

.bazelrc

@@ -137,8 +137,19 @@ build --announce_rc
 # Other build flags.
 build --define=grpc_no_ares=true
 
-# Prevent regression of https://github.com/bazelbuild/bazel/issues/7362
-build --incompatible_remove_legacy_whole_archive
+# See https://github.com/bazelbuild/bazel/issues/7362 for information on what
+# --incompatible_remove_legacy_whole_archive flag does.
+# This flag is set to true in Bazel 1.0 and newer versions. We tried to migrate
+# Tensorflow to the default, however test coverage wasn't enough to catch the
+# errors.
+# There is ongoing work on Bazel team's side to provide support for transitive
+# shared libraries. As part of migrating to transitive shared libraries, we
+# hope to provide a better mechanism for control over symbol exporting, and
+# then tackle this issue again.
+#
+# TODO: Remove this line once TF doesn't depend on Bazel wrapping all library
+# archives in -whole_archive -no_whole_archive.
+build --noincompatible_remove_legacy_whole_archive
 
 # Modular TF build options
 build:dynamic_kernels --define=dynamic_loaded_kernels=true
@@ -149,6 +160,10 @@ build:c++17 --cxxopt=-std=c++1z
 build:c++17 --cxxopt=-stdlib=libc++
 build:c++1z --config=c++17
 
+# Tensorflow uses M_* math constants that only get defined by MSVC headers if
+# _USE_MATH_DEFINES is defined.
+build:windows --copt=/D_USE_MATH_DEFINES
+
 # Default paths for TF_SYSTEM_LIBS
 build --define=PREFIX=/usr
 build --define=LIBDIR=$(PREFIX)/lib

.bazelversion

@@ -1 +1 @@
-0.26.1
+0.29.1

configure.py

@@ -33,7 +33,7 @@
   from distutils.spawn import find_executable as which
 # pylint: enable=g-import-not-at-top
 
-_DEFAULT_CUDA_VERSION = '10'
+_DEFAULT_CUDA_VERSION = '10.1'
 _DEFAULT_CUDNN_VERSION = '7'
 _DEFAULT_TENSORRT_VERSION = '6'
 _DEFAULT_CUDA_COMPUTE_CAPABILITIES = '3.5,7.0'
@@ -817,7 +817,7 @@ def valid_api_level(api_level):
   android_ndk_api_level = prompt_loop_or_load_from_env(
       environ_cp,
       var_name='ANDROID_NDK_API_LEVEL',
-      var_default='18',  # 18 is required for GPU acceleration.
+      var_default='21',  # 21 is required for ARM64 support.
       ask_for_var=('Please specify the (min) Android NDK API level to use. '
                    '[Available levels: %s]') % api_levels,
       check_success=valid_api_level,
@@ -1196,6 +1196,34 @@ def set_system_libs_flag(environ_cp):
     write_to_bazelrc('build --define=INCLUDEDIR=%s' % environ_cp['INCLUDEDIR'])
 
 
+def is_reduced_optimize_huge_functions_available(environ_cp):
+  """Check to see if the system supports /d2ReducedOptimizeHugeFunctions.
+
+  The above compiler flag is a new compiler flag introduced to the Visual Studio
+  compiler in version 16.4 (available in Visual Studio 2019, Preview edition
+  only, as of 2019-11-19). TensorFlow needs this flag to massively reduce
+  compile times, but until 16.4 is officially released, we can't depend on it.
+
+  See also https://groups.google.com/a/tensorflow.org/g/build/c/SsW98Eo7l3o
+
+  Because it's very annoying to check this manually (to check the MSVC installed
+  versions, you need to use the registry, and it's not clear if Bazel will be
+  using that install version anyway), we expect enviroments who know they may
+  use this flag to export TF_VC_VERSION=16.4
+
+  TODO(angerson, gunan): Remove this function when TensorFlow's minimum VS
+  version is upgraded to 16.4.
+
+  Arguments:
+    environ_cp: Environment of the current execution
+
+  Returns:
+    boolean, whether or not /d2ReducedOptimizeHugeFunctions is available on this
+    machine.
+  """
+  return float(environ_cp.get('TF_VC_VERSION', '0')) >= 16.4
+
+
 def set_windows_build_flags(environ_cp):
   """Set Windows specific build options."""
   # The non-monolithic build is not supported yet
@@ -1205,13 +1233,18 @@ def set_windows_build_flags(environ_cp):
   # Fix winsock2.h conflicts
   write_to_bazelrc(
       'build --copt=-DWIN32_LEAN_AND_MEAN --host_copt=-DWIN32_LEAN_AND_MEAN '
-      '--copt=-DNOGDI --host_copt=-DNOGDI')
+      '--copt=-DNOGDI --host_copt=-DNOGDI --copt=-D_USE_MATH_DEFINES')
   # Output more verbose information when something goes wrong
   write_to_bazelrc('build --verbose_failures')
   # The host and target platforms are the same in Windows build. So we don't
   # have to distinct them. This avoids building the same targets twice.
   write_to_bazelrc('build --distinct_host_configuration=false')
 
+  if is_reduced_optimize_huge_functions_available(environ_cp):
+    write_to_bazelrc(
+        'build --copt=/d2ReducedOptimizeHugeFunctions --host_copt=/d2ReducedOptimizeHugeFunctions'
+    )
+
   if get_var(
       environ_cp, 'TF_OVERRIDE_EIGEN_STRONG_INLINE', 'Eigen strong inline',
       True, ('Would you like to override eigen strong inline for some C++ '

tensorflow/BUILD

@@ -607,6 +607,18 @@ tf_cc_shared_object(
     ] + tf_additional_binary_deps(),
 )
 
+# This is intended to be the same as tf_binary_additional_srcs:
+# https://github.com/tensorflow/tensorflow/blob/cd67f4f3723f9165aabedd0171aaadc6290636e5/tensorflow/tensorflow.bzl#L396-L425
+# And is usable in the "deps" attribute instead of the "srcs" attribute
+# as a workaround for https://github.com/tensorflow/tensorflow/issues/34117
+cc_import(
+    name = "libtensorflow_framework_import_lib",
+    shared_library = select({
+        "//tensorflow:macos": ":libtensorflow_framework.dylib",
+        "//conditions:default": ":libtensorflow_framework.so",
+    }),
+)
+
 # -------------------------------------------
 # New rules should be added above this target.
 # -------------------------------------------

tensorflow/api_template.__init__.py

@@ -119,11 +119,11 @@ def _running_from_pip_package():
       _current_file_location.startswith(dir_) for dir_ in _site_packages_dirs)
 
 if _running_from_pip_package():
-  for s in _site_packages_dirs:
+  for _s in _site_packages_dirs:
     # TODO(gunan): Add sanity checks to loaded modules here.
-    plugin_dir = _os.path.join(s, 'tensorflow-plugins')
-    if _fi.file_exists(plugin_dir):
-      _ll.load_library(plugin_dir)
+    _plugin_dir = _os.path.join(_s, 'tensorflow-plugins')
+    if _fi.file_exists(_plugin_dir):
+      _ll.load_library(_plugin_dir)
 
 # Add module aliases
 if hasattr(_current_module, 'keras'):
@@ -136,3 +136,5 @@ def _running_from_pip_package():
   setattr(_current_module, "optimizers", optimizers)
   setattr(_current_module, "initializers", initializers)
 # pylint: enable=undefined-variable
+
+# __all__ PLACEHOLDER

tensorflow/api_template_v1.__init__.py

@@ -132,9 +132,10 @@ def _running_from_pip_package():
       _current_file_location.startswith(dir_) for dir_ in _site_packages_dirs)
 
 if _running_from_pip_package():
-  for s in _site_packages_dirs:
+  for _s in _site_packages_dirs:
     # TODO(gunan): Add sanity checks to loaded modules here.
-    plugin_dir = _os.path.join(s, 'tensorflow-plugins')
-    if _fi.file_exists(plugin_dir):
-      _ll.load_library(plugin_dir)
+    _plugin_dir = _os.path.join(_s, 'tensorflow-plugins')
+    if _fi.file_exists(_plugin_dir):
+      _ll.load_library(_plugin_dir)
 
+# __all__ PLACEHOLDER

tensorflow/c/kernels/BUILD

@@ -1,8 +1,8 @@
 load(
     "//tensorflow:tensorflow.bzl",
     "tf_cc_test",
-    "tf_kernel_library",
     "tf_gen_op_libs",
+    "tf_kernel_library",
 )
 
 package(

tensorflow/cc/gradients/math_grad.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#define _USE_MATH_DEFINES
 #include <cmath>
 
 #include "tensorflow/cc/ops/array_ops_internal.h"

tensorflow/compiler/jit/BUILD

@@ -1,4 +1,4 @@
-load("//tensorflow:tensorflow.bzl", "if_mlir", "tf_cc_test", "cc_header_only_library")
+load("//tensorflow:tensorflow.bzl", "cc_header_only_library", "if_mlir", "tf_cc_test")
 load("//tensorflow/stream_executor:build_defs.bzl", "if_cuda_or_rocm")
 load("//tensorflow:tensorflow.bzl", "tf_custom_op_py_library", "tf_jit_compilation_passes_extra_deps")
 load("//tensorflow/core/platform:default/build_config.bzl", "tf_additional_all_protos", "tf_proto_library")

tensorflow/compiler/xla/BUILD

@@ -1,4 +1,4 @@
-load("//tensorflow:tensorflow.bzl", "tf_cc_test", "cc_header_only_library")
+load("//tensorflow:tensorflow.bzl", "cc_header_only_library", "tf_cc_test")
 load("//tensorflow/compiler/xla:xla.bzl", "xla_proto_library")
 load(
     "//tensorflow/core/platform:default/build_config.bzl",

tensorflow/compiler/xla/client/lib/math.cc

@@ -15,9 +15,7 @@ limitations under the License.
 
 #include "tensorflow/compiler/xla/client/lib/math.h"
 
-// This macro is required to make MSVC defines math constants in math.h
-#define _USE_MATH_DEFINES
-#include <math.h>
+#include <cmath>
 
 #include "tensorflow/compiler/xla/client/lib/arithmetic.h"
 #include "tensorflow/compiler/xla/client/lib/constants.h"

tensorflow/compiler/xla/client/lib/tridiagonal.cc

@@ -36,6 +36,8 @@ namespace {
 struct TridiagonalSystemShape {
   const int64 rank;
   const int64 num_equations;
+  TridiagonalSystemShape(int64 rk, int64 num_eqs)
+      : rank(rk), num_equations(num_eqs) {}
 };
 
 Status CheckSecondToLastDimension(const Shape& op_shape, int64 rank,
@@ -109,9 +111,7 @@ StatusOr<TridiagonalSystemShape> CheckSystemAndReturnShape(XlaOp lower_diagonal,
   TF_RETURN_IF_ERROR(CheckSecondToLastDimension(upper_diagonal_shape, rank, 1,
                                                 "upper diagonal"));
 
-  TridiagonalSystemShape result = {.rank = rank,
-                                   .num_equations = num_equations};
-  return result;
+  return TridiagonalSystemShape(rank, num_equations);
 }
 
 XlaOp Coefficient(XlaOp operand, int64 i) {

tensorflow/compiler/xla/service/dynamic_dimension_inference.cc

@@ -469,7 +469,7 @@ Status DynamicDimensionInferenceVisitor::HandleSetDimensionSize(
     // Propagate dynamic dimension indicated by this set dimension size
     // instruction.
     parent_->SetDynamicSize(hlo, {}, hlo->dimension(), hlo->mutable_operand(1),
-                            {.stride = 1, .multiple_of = 1});
+                            DimensionConstraint(1, 1));
   }
 
   // Also Propagate dynamic dimension already set by operands.
@@ -813,7 +813,7 @@ Status DynamicDimensionInferenceVisitor::HandleReshape(HloInstruction* hlo) {
 
           parent_->SetDynamicSize(
               reshape, {}, output_dynamic_dimension, new_dynamic_size,
-              {.stride = 1, .multiple_of = constraint.multiple_of / divisor});
+              DimensionConstraint(1, constraint.multiple_of / divisor));
         }
 
         if (input_dim_size < output_dim_size) {
@@ -850,12 +850,12 @@ Status DynamicDimensionInferenceVisitor::HandleReshape(HloInstruction* hlo) {
               hlo->parent()->AddInstruction(HloInstruction::CreateBinary(
                   output_dynamic_size->shape(), HloOpcode::kMultiply,
                   new_dynamic_size, operand_dynamic_size));
+          int64 new_multiple_of_constraint =
+              constraint.multiple_of * output_dim_size /
+              operand->shape().dimensions(input_dynamic_dimension);
           parent_->SetDynamicSize(
               reshape, {}, output_dynamic_dimension, new_dynamic_size,
-              {.stride = 1,
-               .multiple_of =
-                   constraint.multiple_of * output_dim_size /
-                   operand->shape().dimensions(input_dynamic_dimension)});
+              DimensionConstraint(1, new_multiple_of_constraint));
         }
 
         return Status::OK();
@@ -1227,7 +1227,7 @@ Status DynamicDimensionInferenceVisitor::HandleParameter(HloInstruction* hlo) {
         parent_->SetDynamicSize(target_parameter,
                                 dynamic_dimension.parameter_index,
                                 dynamic_dimension.dimension, dynamic_size,
-                                {.stride = 1, .multiple_of = 1});
+                                DimensionConstraint(1, 1));
         return Status::OK();
       });
 }

tensorflow/compiler/xla/service/dynamic_dimension_inference.h

@@ -149,6 +149,9 @@ class DynamicDimensionInference {
   //
   //
   struct DimensionConstraint {
+    explicit DimensionConstraint(int64 s, int64 m)
+        : stride(s), multiple_of(m) {}
+    DimensionConstraint() : stride(1), multiple_of(1) {}
     // Stride represents the distance of a newly placed element and the previous
     // placed element on this dynamic dimension.
     int64 stride;

tensorflow/compiler/xla/service/hlo_parser.cc

@@ -2611,18 +2611,37 @@ struct MinMaxFiniteValue<bfloat16> {
   static double min() { return -max(); }
 };
 
+// MSVC's standard C++ library does not define isnan/isfinite for integer types.
+// To work around that we will need to provide our own.
+template <typename T>
+std::enable_if_t<std::is_floating_point<T>::value, bool> IsFinite(T val) {
+  return std::isfinite(val);
+}
+template <typename T>
+std::enable_if_t<std::is_floating_point<T>::value, bool> IsNaN(T val) {
+  return std::isnan(val);
+}
+template <typename T>
+std::enable_if_t<std::is_integral<T>::value, bool> IsFinite(T val) {
+  return std::isfinite(static_cast<double>(val));
+}
+template <typename T>
+std::enable_if_t<std::is_integral<T>::value, bool> IsNaN(T val) {
+  return std::isnan(static_cast<double>(val));
+}
+
 template <typename LiteralNativeT, typename ParsedElemT>
 bool HloParserImpl::CheckParsedValueIsInRange(LocTy loc, ParsedElemT value) {
   if (std::is_floating_point<ParsedElemT>::value) {
     auto value_as_native_t = static_cast<LiteralNativeT>(value);
     auto value_double_converted = static_cast<ParsedElemT>(value_as_native_t);
-    if (!std::isfinite(value) || std::isfinite(value_double_converted)) {
+    if (!IsFinite(value) || IsFinite(value_double_converted)) {
       value = value_double_converted;
     }
   }
   PrimitiveType literal_ty =
       primitive_util::NativeToPrimitiveType<LiteralNativeT>();
-  if (std::isnan(value) ||
+  if (IsNaN(value) ||
       (std::numeric_limits<ParsedElemT>::has_infinity &&
        (std::numeric_limits<ParsedElemT>::infinity() == value ||
         -std::numeric_limits<ParsedElemT>::infinity() == value))) {

tensorflow/core/api_def/python_api/api_def_Erfinv.pbtxt

@@ -0,0 +1,4 @@
+op {
+  graph_op_name: "Erfinv"
+  visibility: HIDDEN
+}

tensorflow/core/api_def/python_api/api_def_Ndtri.pbtxt

@@ -0,0 +1,4 @@
+op {
+  graph_op_name: "Ndtri"
+  visibility: HIDDEN
+}

tensorflow/core/common_runtime/eager/mkl_eager_op_rewrite.cc

@@ -87,7 +87,7 @@ REGISTER_REWRITE(EagerOpRewriteRegistry::PRE_EXECUTION, MklEagerOpRewrite);
 
 // Constructor
 MklEagerOpRewrite::MklEagerOpRewrite(string name, string file, string line)
-    : EagerOpRewrite(name, file, line) {
+    : EagerOpRewrite(name, file, line), registered_kernels_map_() {
   InsertMKLEagerOps({"BatchMatMul", AlwaysRewrite, CreateGenericMklOp});
   InsertMKLEagerOps({"BatchMatMulV2", AlwaysRewrite, CreateGenericMklOp});
   InsertMKLEagerOps({"Conv2D", RewriteConv2D, CreateMklConv2DOp});

tensorflow/core/common_runtime/eager/mkl_eager_op_rewrite_test.cc

@@ -33,14 +33,15 @@ class EagerOpRewriteTest {
         absl::make_unique<StaticDeviceMgr>(DeviceFactory::NewDevice(
             "CPU", {}, "/job:localhost/replica:0/task:0/device:CPU:0"));
     bool async = false;
+    bool lazy_remote_tensor_copy = false;
     tensorflow::Rendezvous* rendezvous =
         new tensorflow::IntraProcessRendezvous(device_mgr.get());
     std::unique_ptr<tensorflow::EagerContext> eager_ctx =
         std::unique_ptr<tensorflow::EagerContext>(new tensorflow::EagerContext(
             SessionOptions(),
             tensorflow::ContextDevicePlacementPolicy::DEVICE_PLACEMENT_SILENT,
             tensorflow::ContextMirroringPolicy::MIRRORING_NONE, async,
-            device_mgr.get(), false, rendezvous,
+            lazy_remote_tensor_copy, device_mgr.get(), false, rendezvous,
             GetDefaultCustomKernelCreator()));
 
     EagerExecutor executor_(false);

tensorflow/core/debug/BUILD

@@ -15,10 +15,10 @@
 load(
     "//tensorflow:tensorflow.bzl",
     "check_deps",
+    "tf_cc_binary",
     "tf_cc_test",
     "tf_copts",
     "tf_cuda_library",
-    "tf_cc_binary",
 )
 
 # For platform specific build config

tensorflow/core/framework/dataset.cc

@@ -403,6 +403,7 @@ Status DatasetBaseIterator::GetNext(IteratorContext* ctx,
                                     bool* end_of_sequence) {
   profiler::TraceMe activity([&] { return BuildTraceMeName(); },
                              profiler::TraceMeLevel::kInfo);
+  DVLOG(3) << prefix() << " GetNext enter";
   RecordStart(ctx, /*stop_output=*/true);
   Status s = GetNextInternal(ctx, out_tensors, end_of_sequence);
   if (s.ok() && !*end_of_sequence) RecordElement(ctx);
@@ -415,6 +416,7 @@ Status DatasetBaseIterator::GetNext(IteratorContext* ctx,
                          s.error_message());
     LOG(ERROR) << s;
   }
+  DVLOG(3) << prefix() << " GetNext exit";
   return s;
 }
 

tensorflow/core/framework/dataset.h

@@ -815,9 +815,13 @@ class DatasetBaseIterator : public IteratorBase {
 
   explicit DatasetBaseIterator(const BaseParams& params) : params_(params) {
     params_.dataset->Ref();
+    VLOG(2) << prefix() << " constructor";
   }
 
-  ~DatasetBaseIterator() override { params_.dataset->Unref(); }
+  ~DatasetBaseIterator() override {
+    VLOG(2) << prefix() << " destructor";
+    params_.dataset->Unref();
+  }
 
   const DataTypeVector& output_dtypes() const override {
     return params_.dataset->output_dtypes();