Fix clang compilation errors in TF-MKL variant.

PiperOrigin-RevId: 257656880

tensorflow/core/common_runtime/mkl_cpu_allocator.h

@@ -253,6 +253,7 @@ class MklCPUAllocator : public Allocator {
     auto l_stats = large_size_allocator_->GetStats();
 
     // Combine statistics from small-size and large-size allocator.
+    mutex_lock l(mutex_);
     stats_.num_allocs = l_stats->num_allocs + s_stats->num_allocs;
     stats_.bytes_in_use = l_stats->bytes_in_use + s_stats->bytes_in_use;
     stats_.peak_bytes_in_use =

tensorflow/core/kernels/BUILD

@@ -34,7 +34,6 @@ load(
 )
 load(
     "//third_party/mkl:build_defs.bzl",
-    "if_mkl",
     "if_mkl_ml",
     "mkl_deps",
 )
@@ -7465,8 +7464,10 @@ tf_mkl_kernel_library(
     ],
     deps = [
         ":bounds_check",
+        ":fill_functor",
         ":matmul_op",
         ":ops_util",
+        "//third_party/eigen3",
         "//tensorflow/core:core_cpu",
         "//tensorflow/core:framework",
         "//tensorflow/core:lib",

tensorflow/core/kernels/gather_nd_op_cpu_impl.h

@@ -107,13 +107,16 @@ struct GatherNdSlice<CPUDevice, T, Index, IXDIM> {
     Eigen::Tensor<Eigen::DenseIndex, 1>::Dimensions reshape_dims{{ 1 }};
     Eigen::array<Eigen::DenseIndex, 1> broadcast_dims{{ batch_size }};
 #else
+#if !defined(INTEL_MKL) || !defined(ENABLE_MKL)
     Eigen::IndexList<Eigen::type2index<1> > reshape_dims;
+#endif  // defined(INTEL_MKL) && defined(ENABLE_MKL)
     Eigen::IndexList<Eigen::DenseIndex> broadcast_dims;
     broadcast_dims.set(0, batch_size);
-#endif
+#endif  // !defined(EIGEN_HAS_INDEX_LIST)
     generator::GatherNdSliceGenerator<T, Index, IXDIM> gather_nd_generator(
         slice_size, Tindices, Tparams, Tout, &error_loc);
 
+// TODO(b/137289929): Parallelize this with ParallelFor and remove OpenMP call.
 #if defined(INTEL_MKL) && defined(ENABLE_MKL)
 // Eigen implementation below is not highly performant. gather_nd_generator
 // does not seem to be called in parallel, leading to very poor performance.

tensorflow/core/kernels/mkl_conv_ops.h

@@ -562,8 +562,6 @@ class MklConvBackpropCommonOp : public OpKernel {
     OP_REQUIRES_OK(context, context->GetAttr("strides", &strides_));
     int stride_n = GetTensorDim(strides_, data_format_, 'N');
     int stride_c = GetTensorDim(strides_, data_format_, 'C');
-    const int64 stride_h = GetTensorDim(strides_, data_format_, 'H');
-    const int64 stride_w = GetTensorDim(strides_, data_format_, 'W');
     OP_REQUIRES(
         context, (stride_n == 1 && stride_c == 1),
         errors::InvalidArgument("Current implementation does not yet support "

tensorflow/core/kernels/mkl_pooling_ops_common.h

@@ -464,7 +464,8 @@ class MklPoolingOpBase : public OpKernel {
     // We may not get this attribute for this node if it does not go through
     // graph rewrite pass. So we do not check for error while retrieving this
     // attribute value.
-    context->GetAttr("workspace_enabled", &this->workspace_enabled_);
+    TF_CHECK_OK(
+        context->GetAttr("workspace_enabled", &this->workspace_enabled_));
   }
   void Compute(OpKernelContext* context) override = 0;
 

tensorflow/core/kernels/mkl_qmatmul_op.cc

@@ -222,10 +222,10 @@ class MklDnnMatMulFwdPrimitive : public MklPrimitive {
           bias_mem(nullptr),
           dst_mem(nullptr),
           fwd_desc(nullptr),
+          fwd_pd(nullptr),
           src_md(nullptr),
           weight_md(nullptr),
           bias_md(nullptr),
-          fwd_pd(nullptr),
           matmul_fwd(nullptr),
           fwd_stream(nullptr) {}
   };
@@ -719,6 +719,7 @@ class MklDnnQuantizedMatMulOp : public OpKernel {
         context->CtxFailure(
             errors::InvalidArgument("Quantization mode must be"
                                     "either MIN_FIRST or SCALED."));
+        return nullptr;
       }
     }
   }

tensorflow/core/util/mkl_util.h

@@ -525,8 +525,8 @@ inline Tensor ConvertMklToTF(OpKernelContext* context, const Tensor& mkl_tensor,
     TensorShape output_shape = mkl_shape.GetTfShape();
 
     // Allocate output tensor.
-    context->allocate_temp(DataTypeToEnum<T>::v(), output_shape,
-                           &output_tensor);
+    TF_CHECK_OK(context->allocate_temp(DataTypeToEnum<T>::v(), output_shape,
+                                       &output_tensor));
 
     auto cpu_engine = engine(engine::cpu, 0);
     MklDnnData<T> input(&cpu_engine);
@@ -576,7 +576,7 @@ inline const Tensor& MklGetInput(OpKernelContext* ctext, int n) {
 inline void GetMklInputList(OpKernelContext* ctext, StringPiece name,
                             OpInputList* input_tensors) {
   CHECK_NOTNULL(input_tensors);
-  ctext->input_list(name, input_tensors);
+  TF_CHECK_OK(ctext->input_list(name, input_tensors));
 }
 
 inline void GetMklShapeList(OpKernelContext* ctext, StringPiece name,