"TestFn", {"a:float", "b:float", "c:float"}, {"m:float"}, {},
      {{{"d"}, "Add", {"a", "b"}, {{"T", DT_FLOAT}}},
       {{"e"}, "Mul", {"d", "c"}, {{"T", DT_FLOAT}}},
       {{"f"}, "Add", {"e", "a"}, {{"T", DT_FLOAT}}},
       {{"g"}, "Mul", {"f", "b"}, {{"T", DT_FLOAT}}},
       // Force two clusters by excluding this node explicitly.
       {{"h"}, "Add", {"g", "f"}, {{"T", DT_FLOAT}, {"_XlaCompile", false}}},

            VAR name:a type:kotlin.Int [val]
              CONST Int type=kotlin.Int value=5
            CALL 'public final fun times (other: kotlin.Int): kotlin.Int [operator] declared in kotlin.Int' type=kotlin.Int origin=MUL
              $this: GET_VAR 'val a: kotlin.Int [val] declared in <root>.CodeFragment.run' type=kotlin.Int origin=null

  %0 = tfl.add %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

func.func @cond_false(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
  %0 = tfl.mul %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<*xf32>
  func.return %0 : tensor<*xf32>

  %1 = "tfl.mul"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<256x8x7x3xf32>, tensor<3xf32>) -> tensor<256x8x7x3xf32>
  func.return %1 : tensor<256x8x7x3xf32>

  AbstractTensorHandle* expand_dims_outputs;
  TF_RETURN_IF_ERROR(
      ops::ExpandDims(ctx, vec, dim.get(), &expand_dims_outputs, "ExpandDims"));
  TF_RETURN_IF_ERROR(
      ops::Mul(ctx, expand_dims_outputs, mat, &outputs[0], "Mul"));
  expand_dims_outputs->Unref();
  return absl::OkStatus();
}

class SparseSoftmaxCrossEntropyWithLogitsGradientFunction
    : public GradientFunction {
 public:

	var p Point
	p.ScalarMult(dalekScalar, B)
	if dalekScalarBasepoint.Equal(&p) != 1 {
		t.Error("Scalar mul does not match dalek")
	}
	checkOnCurve(t, &p)
}

func TestBaseMultVsDalek(t *testing.T) {
	var p Point
	p.ScalarBaseMult(dalekScalar)
	if dalekScalarBasepoint.Equal(&p) != 1 {
		t.Error("Scalar mul does not match dalek")
	}
	checkOnCurve(t, &p)
}

func TestVarTimeDoubleBaseMultVsDalek(t *testing.T) {

  %0 = "tf.Conv2D"(%arg0, %cst0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<1x112x112x3xf32>, tensor<1x3x3x2xf32>) -> tensor<1x28x23x2xf32>
  %1 = "tf.Mul"(%0, %cst2) : (tensor<1x28x23x2xf32>, tensor<2xf32>) -> tensor<1x28x23x2xf32>

  func.return %1 : tensor<1x28x23x2xf32>
  // CHECK: %[[CST:.*]] = "tf.Const{{.*}} dense<

  %9 = "tf.AddV2"(%8, %cst_6) : (tensor<1x98x98x1xf32>, tensor<1xf32>) -> tensor<1x98x98x1xf32>
  %10 = "tf.Mul"(%9, %cst_5) : (tensor<1x98x98x1xf32>, tensor<1xf32>) -> tensor<1x98x98x1xf32>
  %11 = "tf.AddV2"(%10, %cst_8) : (tensor<1x98x98x1xf32>, tensor<f32>) -> tensor<1x98x98x1xf32>

// Calibration used to determine thresholds for using
// different algorithms.  Ideally, this would be converted
// to go generate to create thresholds.go

// This file prints execution times for the Mul benchmark
// given different Karatsuba thresholds. The result may be
// used to manually fine-tune the threshold constant. The
// results are somewhat fragile; use repeated runs to get
// a clear picture.

	"LL",
	"LLV",
	"LUI",
	"MOVB",
	"MOVBU",
	"MOVD",
	"MOVDF",
	"MOVDW",
	"MOVF",
	"MOVFD",
	"MOVFW",
	"MOVH",
	"MOVHU",
	"MOVW",
	"MOVWD",
	"MOVWF",
	"MOVWL",
	"MOVWR",
	"MUL",
	"MULD",
	"MULF",
	"MULU",
	"MULH",
	"MULHU",
	"MULW",
	"NEGD",
	"NEGF",
	"NEGW",
	"NEGV",
	"NOOP",
	"NOR",
	"OR",
	"REM",
	"REMU",
	"RFE",
	"SC",
	"SCV",
	"SGT",