Multiplier multiplier = new Multiplier();
                        Assert.assertEquals(1, multiplier.multiply(1, 1));
                        Assert.assertEquals(4, multiplier.multiply(2, 2));
                        Assert.assertEquals(2, multiplier.multiply(1, 2));
                    }
                }
            """
            file("transitive/build.gradle") << """

	// add 2¹²⁸ to each message block value
	VLEIB $4, $1, M_4
	VLEIB $12, $1, M_4

multiply:
	// accumulate the incoming message
	VAG H_0, M_0, M_0
	VAG H_3, M_3, M_3
	VAG H_1, M_1, M_1
	VAG H_4, M_4, M_4
	VAG H_2, M_2, M_2

	// multiply the accumulator by the key coefficient
	MULTIPLY(M_0, M_1, M_2, M_3, M_4, R_0, R_1, R_2, R_3, R_4, R5_1, R5_2, R5_3, R5_4, H_0, H_1, H_2, H_3, H_4)

					}
					atomic.AddInt64(&ops, int64(n))
					atomic.AddInt64(&lat, latency)
				})
				spent := time.Since(t)
				if spent > 0 && n > 0 {
					// Since we are benchmarking n parallel servers we need to multiply by n.
					// This will give an estimate of the total ops/s.
					latency := float64(atomic.LoadInt64(&lat)) / float64(time.Millisecond)
					b.ReportMetric(float64(n)*float64(ops)/spent.Seconds(), "vops/s")

//   ⎡x / c⎤ = ⎣x * m / 2^(n+s)⎦ + 1
// (TODO: derivation?)
//
// The multiply is a bit odd, as it is a signed n-bit value
// times an unsigned n-bit value.  For n smaller than the
// word size, we can extend x and m appropriately and use the
// signed multiply instruction.  For n == word size,
// we must use the signed multiply high and correct
// the result by adding x*2^n.
//

    %16 = stablehlo.subtract %10, %15 : tensor<1x3x3x4xf32>
    %17 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2, 3] : (tensor<1x1x1x4xf32>) -> tensor<1x3x3x4xf32>
    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>
    %19 = call @uniform_quantize_1(%18, %5, %6) : (tensor<1x3x3x4xf32>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xi8>) -> tensor<1x3x3x4xi8>

	HasAVX512IFMA       bool // Advanced vector extension 512 Integer Fused Multiply Add
	HasAVX512VBMI       bool // Advanced vector extension 512 Vector Byte Manipulation Instructions
	HasAVX5124VNNIW     bool // Advanced vector extension 512 Vector Neural Network Instructions Word variable precision
	HasAVX5124FMAPS     bool // Advanced vector extension 512 Fused Multiply Accumulation Packed Single precision

                        Multiplier multiplier = new Multiplier();
                        Assert.assertEquals(1, multiplier.multiply(1, 1));
                        Assert.assertEquals(4, multiplier.multiply(2, 2));
                        Assert.assertEquals(2, multiplier.multiply(1, 2));
                    }
                }
            """
            file("transitive/build.gradle") << """

    auto subtraction_result = Subtract(scope, grad_grad, squeeze_result);
    auto multiply_result = Multiply(scope, subtraction_result, logits_softmax);
    grad = Add(scope, grad, multiply_result);
  }
  auto minus_log_softmax = Multiply(scope, LogSoftmax(scope, logits), -1.0f);
  grad_outputs->push_back(grad);
  grad_outputs->push_back(BroadcastMul(scope, grad_loss, minus_log_softmax));

  %10 = mhlo.add %arg0, %5 : tensor<2xf32>
  %11 = mhlo.clamp %7, %10, %3 : tensor<2xf32>
  %12 = mhlo.clamp %9, %11, %1 : tensor<2xf32>
  %13 = mhlo.multiply %arg0, %12 : tensor<2xf32>
  %14 = mhlo.divide %13, %1 : tensor<2xf32>
  return %14 : tensor<2xf32>
}

// CHECK-LABEL:   func.func @hardswish(

	}

	if y32+v32_n < 0 {
		t.Errorf("'%#x + %#x < 0' failed", y32, v32_n)
	}

	if y32+v32_n <= 0 {
		t.Errorf("'%#x + %#x <= 0' failed", y32, v32_n)
	}
}

// multiply-add
func testMAddVar64(t *testing.T) {
	if x64+v64*one64 < 0 {
	} else {
		t.Errorf("'%#x + %#x*1 < 0' failed", x64, v64)
	}

	if x64+v64*one64 <= 0 {
	} else {