context.get(), components, second_device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  TensorHandlePtr multiply_result(
      Multiply(context.get(), second_combined_value.get(),
               second_negative_one.get(), status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

	//
	//   x = xh*b + x0  (0 <= x0 < b)
	//   y = yh*b + y0  (0 <= y0 < b)
	//   b = 1<<(_W*k)  ("base" of digits xi, yi)
	//
	k := karatsubaLen(n, karatsubaThreshold)
	// k <= n

	// multiply x0 and y0 via Karatsuba
	x0 := x[0:k]              // x0 is not normalized
	y0 := y[0:k]              // y0 is not normalized
	z = z.make(max(6*k, m+n)) // enough space for karatsuba of x0*y0 and full result of x*y

func Sub64(x, y, borrow uint64) (diff, borrowOut uint64) {
	diff = x - y - borrow
	// See Sub32 for the bit logic.
	borrowOut = ((^x & y) | (^(x ^ y) & diff)) >> 63
	return
}

// --- Full-width multiply ---

// Mul returns the full-width product of x and y: (hi, lo) = x * y
// with the product bits' upper half returned in hi and the lower
// half returned in lo.
//

		// Let x = arg0*arg1 (full 32x32->64  unsigned multiply). Returns uint32(x), and flags set to overflow if uint32(x) != x.
		{name: "MULLU", argLength: 2, reg: regInfo{inputs: []regMask{ax, gpsp}, outputs: []regMask{ax, 0}, clobbers: dx}, typ: "(UInt32,Flags)", asm: "MULL", commutative: true, clobberFlags: true},
		// Let x = arg0*arg1 (full 64x64->128 unsigned multiply). Returns uint64(x), and flags set to overflow if uint64(x) != x.

			}
			// Order of operations matters.
			// For very fast benchmarks, prevIters ~= prevns.
			// If you divide first, you get 0 or 1,
			// which can hide an order of magnitude in execution time.
			// So multiply first, then divide.
			n = goalns * prevIters / prevns
			// Run more iterations than we think we'll need (1.2x).
			n += n / 5
			// Don't grow too fast in case we had timing errors previously.
			n = min(n, 100*last)

		{name: "LoweredNilCheck", argLength: 2, reg: regInfo{inputs: []regMask{ptrsp}}, clobberFlags: true, nilCheck: true, faultOnNilArg0: true},
		// Round ops to block fused-multiply-add extraction.
		{name: "LoweredRound32F", argLength: 1, reg: fp11, resultInArg0: true, zeroWidth: true},
		{name: "LoweredRound64F", argLength: 1, reg: fp11, resultInArg0: true, zeroWidth: true},

		{name: "SLTIU", argLength: 1, reg: gp11, asm: "SLTIU", aux: "Int64"}, // arg0 < auxint, unsigned, result is 0 or 1

		// Round ops to block fused-multiply-add extraction.
		{name: "LoweredRound32F", argLength: 1, reg: fp11, resultInArg0: true},
		{name: "LoweredRound64F", argLength: 1, reg: fp11, resultInArg0: true},

		// Calls

* **Deep Learning**: this is a sub-field of Machine Learning, so, the same applies. It's just that there is not a single spreadsheet of numbers to multiply, but a huge set of them, and in many cases, you use a special processor to build and / or use those models.

### Concurrency + Parallelism: Web + Machine Learning

		p.firstProg = prog
	} else {
		p.lastProg.Link = prog
	}
	p.lastProg = prog
	if doLabel {
		p.pc++
		for _, label := range p.pendingLabels {
			if p.labels[label] != nil {
				p.errorf("label %q multiply defined", label)
				return
			}
			p.labels[label] = prog
		}
		p.pendingLabels = p.pendingLabels[0:0]
	}
	prog.Pc = p.pc
	if *flags.Debug {
		fmt.Println(p.lineNum, prog)
	}
	if testOut != nil {

	/* Vector multiply */
	{as: AVMULESB, a1: C_VREG, a2: C_VREG, a6: C_VREG, type_: 82, size: 4},              /* vector multiply, vx-form */
	{as: AVPMSUM, a1: C_VREG, a2: C_VREG, a6: C_VREG, type_: 82, size: 4},               /* vector polynomial multiply & sum, vx-form */
	{as: AVMSUMUDM, a1: C_VREG, a2: C_VREG, a3: C_VREG, a6: C_VREG, type_: 83, size: 4}, /* vector multiply-sum, va-form */

	/* Vector rotate */