VPCOMPRESSB X13, K1, X28                           // 62127d0963ec
	VPCOMPRESSB X8, K1, X28                            // 62127d0963c4
	VPCOMPRESSB X7, K1, -7(CX)(DX*1)                   // 62f27d09637c11f9
	VPCOMPRESSB X13, K1, -7(CX)(DX*1)                  // 62727d09636c11f9
	VPCOMPRESSB X8, K1, -7(CX)(DX*1)                   // 62727d09634411f9
	VPCOMPRESSB X7, K1, -15(R14)(R15*4)                // 62927d09637cbef1

	VADDPD 2032(DX), X29, X0        // 62f1950058427f
	VADDPD 2032(DX), X1, X29        // 6261f508586a7f
	VADDPD 2032(DX), X29, X28       // 6261950058627f
	VADDPD 2032(DX)(AX*2), X29, X0  // 62f195005844427f
	VADDPD 2032(DX)(AX*2), X1, X29  // 6261f508586c427f
	VADDPD 2032(DX)(AX*2), X29, X28 // 626195005864427f
	VADDPD 4064(DX), Y0, Y29        // 6261fd28586a7f
	VADDPD 4064(DX), Y29, Y1        // 62f19520584a7f

// i.e., d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...
//
// `dx` are used as initial gradients (which represent the symbolic partial
// derivatives of some loss function `L` w.r.t. `y`).
// `dx` must be nullptr or have size `ny`.
// If `dx` is nullptr, the implementation will use dx of `OnesLike` for all
// shapes in `y`.

  TF_Operation* y = Placeholder(graph_, s_, "y", TF_FLOAT);
  TF_Operation* xy = Mul(x, y, graph_, s_, "xy");
  TF_Output dxy_dx, dxy_dy;

  TF_Output outputs[1] = {{xy, 0}};
  TF_Output inputs[1] = {{x, 0}};
  TF_AddGradients(graph_, outputs, 1, inputs, 1, nullptr, s_, &dxy_dx);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

  inputs[0] = {y, 0};