tensorflow--tensorflow/tensorflow/compiler/tf2xla/kernels/const_op.cc

/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include <cstdint>
#include <type_traits>

#include "absl/status/status.h"
#include "absl/strings/str_cat.h"
#include "tensorflow/compiler/tf2xla/type_util.h"
#include "tensorflow/compiler/tf2xla/xla_compiler.h"
#include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
#include "tensorflow/compiler/tf2xla/xla_op_registry.h"
#include "xla/hlo/builder/xla_builder.h"
#include "tensorflow/core/framework/kernel_def_builder.h"
#include "tensorflow/core/framework/tensor.pb.h"
#include "tensorflow/core/framework/types.pb.h"

namespace tensorflow {
namespace {

template <typename DstT, typename SrcT>
DstT CastTo(SrcT src) {
  return static_cast<DstT>(src);
}

template <typename DstT,
          typename std::enable_if<std::is_same<DstT, Eigen::half>::value ||
                                  std::is_same<DstT, bfloat16>::value>::type* =
              nullptr>
DstT CastTo(int32_t src) {
  return absl::bit_cast<DstT>(static_cast<uint16_t>(src));
}

// Returns scalar constant with the value in the tensor, if the given proto has
// exactly one value but more than one elements. This encoding is used to
// efficiently serialize tensors that have one value repeated for all the
// indices.
xla::XlaOp GetScalarConst(const TensorProto& proto, xla::XlaBuilder* b) {
  if (!proto.tensor_content().empty()) return xla::XlaOp();
  TensorShape shape(proto.tensor_shape());
  if (shape.num_elements() > 1) {
    switch (proto.dtype()) {
#define HANDLE_SPLAT(DTYPE, field_name, xla_type)                             \
  case DTYPE:                                                                 \
    if (proto.field_name##_val_size() == 0) {                                 \
      return xla::ConstantR0(b, CastTo<xla_type>(0));                         \
    } else if (proto.field_name##_val_size() == 1) {                          \
      return xla::ConstantR0(b, CastTo<xla_type>(proto.field_name##_val(0))); \
    }                                                                         \
    break;

      HANDLE_SPLAT(DT_BOOL, bool, bool);

      HANDLE_SPLAT(DT_INT8, int, int8_t);
      HANDLE_SPLAT(DT_INT16, int, int16_t);
      HANDLE_SPLAT(DT_INT32, int, int32_t);
      HANDLE_SPLAT(DT_INT64, int64, int64_t);

      HANDLE_SPLAT(DT_UINT8, int, uint8_t);
      HANDLE_SPLAT(DT_UINT16, int, uint16_t);
      HANDLE_SPLAT(DT_UINT32, uint32, uint32_t);
      HANDLE_SPLAT(DT_UINT64, uint64, uint64_t);

      HANDLE_SPLAT(DT_FLOAT, float, float);
      HANDLE_SPLAT(DT_DOUBLE, double, double);

      HANDLE_SPLAT(DT_BFLOAT16, half, bfloat16);
      HANDLE_SPLAT(DT_HALF, half, Eigen::half);

#undef HANDLE_SPLAT

#define HANDLE_COMPLEX_SPLAT(DTYPE, field_name, xla_type)                     \
  case DTYPE:                                                                 \
    if (proto.field_name##_val_size() == 2) {                                 \
      return xla::ConstantR0<xla_type>(                                       \
          b, xla_type(proto.field_name##_val(0), proto.field_name##_val(1))); \
    }                                                                         \
    break;

      HANDLE_COMPLEX_SPLAT(DT_COMPLEX64, scomplex, xla::complex64);
      HANDLE_COMPLEX_SPLAT(DT_COMPLEX128, dcomplex, xla::complex128);

#undef HANDLE_COMPLEXSPLAT

      default:
        break;
    }
  }

  return xla::XlaOp();
}

class ConstOp : public XlaOpKernel {
 public:
  explicit ConstOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {
    const TensorProto* proto = nullptr;
    OP_REQUIRES_OK(ctx, ctx->GetAttr("value", &proto));
    proto_ = *proto;
    OP_REQUIRES(
        ctx, ctx->output_type(0) == proto_.dtype(),
        absl::InvalidArgumentError(absl::StrCat(
            "Type mismatch between value (", DataTypeString(proto_.dtype()),
            ") and dtype (", DataTypeString(ctx->output_type(0)), ")")));
    OP_REQUIRES_OK(ctx, TensorShape::IsValidShape(proto_.tensor_shape()));
  }

  void Compile(XlaOpKernelContext* ctx) override {
    xla::XlaBuilder* b = ctx->builder();

    // To avoid blowups for large constants filled with the same value,
    // recognize that case and emit a scalar broadcast instead.
    TensorShape shape(proto_.tensor_shape());
    if (shape.num_elements() > 1) {
      xla::XlaOp value = GetScalarConst(proto_, b);
      if (value.valid()) {
        ctx->SetOutput(0, xla::Broadcast(value, shape.dim_sizes()));
        return;
      }
    }

    Tensor tensor(proto_.dtype());
    OP_REQUIRES(ctx, tensor.FromProto(cpu_allocator(), proto_),
                absl::InvalidArgumentError(absl::StrCat(
                    "Cannot parse tensor from proto: ", proto_.DebugString())));
    ctx->SetConstantOutput(0, tensor);
  }

 private:
  TensorProto proto_;
  ConstOp(const ConstOp&) = delete;
  void operator=(const ConstOp&) = delete;
};

// XLA_* devices also register a "real" Const operator so we suppress the
// dummy operator using CompilationOnly().
REGISTER_XLA_OP(Name("Const").CompilationOnly(), ConstOp);

}  // namespace
}  // namespace tensorflow