Method Compute

tensorflow/core/kernels/sparse_matmul_op.cc:968–1079

  }

  void Compute(OpKernelContext* ctx) override {
    const Tensor& a = ctx->input(0);
    const Tensor& b = ctx->input(1);
    OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(a.shape()),
                errors::InvalidArgument("a is not a matrix"));
    OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(b.shape()),
                errors::InvalidArgument("b is not a matrix"));

    const int m = transpose_a_ ? a.dim_size(1) : a.dim_size(0);
    const int k = transpose_a_ ? a.dim_size(0) : a.dim_size(1);
    const int n = transpose_b_ ? b.dim_size(0) : b.dim_size(1);
    const int k2 = transpose_b_ ? b.dim_size(1) : b.dim_size(0);

    OP_REQUIRES(ctx, k == k2,
                errors::InvalidArgument(
                    "Matrix size incompatible: a: ", a.shape().DebugString(),
                    ", b: ", b.shape().DebugString()));
    OP_REQUIRES(ctx, m >= 0 && n >= 0 && k >= 0,
                errors::InvalidArgument(
                    "Matrix dimensions cannot be negative: a: ",
                    a.shape().DebugString(), ", b: ", b.shape().DebugString()));
    Tensor* output = nullptr;
    OP_REQUIRES_OK(ctx, ctx->allocate_output(0, TensorShape({m, n}), &output));

    // Return early if at least one of the output dimension size is 0.
    if (m == 0 || n == 0) {
      return;
    }

    if (k == 0) {
      // If the inner dimension k in the matrix multiplication is zero, we fill
      // the output with zeros.
      functor::SetZeroFunctor<CPUDevice, float> f;
      f(ctx->eigen_device<CPUDevice>(), output->flat<float>());
      return;
    }

    auto out = output->matrix<float>();

    std::unique_ptr<Tensor> a_float;
    std::unique_ptr<Tensor> b_float;
    if (!a_is_sparse_ && !b_is_sparse_) {
      auto left = &a;
      auto right = &b;
      // TODO(agarwal): multi-thread the conversions from bfloat16 to float.
      if (std::is_same<TL, bfloat16>::value) {
        a_float.reset(new Tensor(DT_FLOAT, a.shape()));
        BFloat16ToFloat(a.flat<bfloat16>().data(),
                        a_float->flat<float>().data(), a.NumElements());
        left = a_float.get();
      }
      if (std::is_same<TR, bfloat16>::value) {
        b_float.reset(new Tensor(DT_FLOAT, b.shape()));
        BFloat16ToFloat(b.flat<bfloat16>().data(),
                        b_float->flat<float>().data(), b.NumElements());
        right = b_float.get();
      }
      Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair;