WEBVTT

00:00:01.079 --> 00:00:06.700
One of the biggest and slowest circuits in
an arithmetic and logic unit is the multiplier.

00:00:06.700 --> 00:00:11.280
We'll start by developing a straightforward
implementation and then, in the next section,

00:00:11.280 --> 00:00:16.509
look into tradeoffs to make it either smaller
or faster.

00:00:16.509 --> 00:00:21.510
Here's the multiplication operation for two
unsigned 4-bit operands broken down into its

00:00:21.510 --> 00:00:24.090
component operations.

00:00:24.090 --> 00:00:27.529
This is exactly how we learned to do it in
primary school.

00:00:27.529 --> 00:00:33.330
We take each digit of the multiplier (the
B operand) and use our memorized multiplication

00:00:33.330 --> 00:00:38.540
tables to multiply it with each digit of the
multiplicand (the A operand),

00:00:38.540 --> 00:00:44.640
dealing with any carries into the next column
as we process the multiplicand right-to-left.

00:00:44.640 --> 00:00:49.250
The output from this step is called a partial
product and then we repeat the step for the

00:00:49.250 --> 00:00:51.640
remaining bits of the multiplier.

00:00:51.640 --> 00:00:56.150
Each partial product is shifted one digit
to the left, reflecting the increasing weight

00:00:56.150 --> 00:00:59.119
of the multiplier digits.

00:00:59.119 --> 00:01:05.120
In our case the digits are just single bits,
i.e., they're 0 or 1 and the multiplication

00:01:05.120 --> 00:01:06.610
table is pretty simple!

00:01:06.610 --> 00:01:13.010
In fact, the 1-bit-by-1-bit binary multiplication
circuit is just a 2-input AND gate.

00:01:13.010 --> 00:01:16.420
And look Mom, no carries!

00:01:16.420 --> 00:01:20.580
The partial products are N bits wide since
there are no carries.

00:01:20.580 --> 00:01:25.450
If the multiplier has M bits, there will be
M partial products.

00:01:25.450 --> 00:01:30.390
And when we add the partial products together,
we'll get an N+M bit result if we account

00:01:30.390 --> 00:01:34.220
for the possible carry-out from the high-order
bit.

00:01:34.220 --> 00:01:38.640
The easy part of the multiplication is forming
the partial products - it just requires some

00:01:38.640 --> 00:01:40.030
AND gates.

00:01:40.030 --> 00:01:46.610
The more expensive operation is adding together
the M N-bit partial products.

00:01:46.610 --> 00:01:52.320
Here's the schematic for the combinational
logic needed to implement the 4x4 multiplication,

00:01:52.320 --> 00:01:58.080
which would be easy to extend for larger multipliers
(we'd need more rows) or larger multiplicands

00:01:58.080 --> 00:01:59.930
(we'd need more columns).

00:01:59.930 --> 00:02:06.710
The M*N 2-input AND gates compute the bits
of the M partial products.

00:02:06.710 --> 00:02:11.020
The adder modules add the current row's partial
product with the sum of the partial products

00:02:11.020 --> 00:02:13.080
from the earlier rows.

00:02:13.080 --> 00:02:15.640
Actually there are two types of adder modules.

00:02:15.640 --> 00:02:19.880
The full adder is used when the modules needs
three inputs.

00:02:19.880 --> 00:02:24.970
The simpler half adder is used when only two
inputs are needed.

00:02:24.970 --> 00:02:28.910
The longest path through this circuit takes
a moment to figure out.

00:02:28.910 --> 00:02:33.900
Information is always moving either down a
row or left to the adjacent column.

00:02:33.900 --> 00:02:40.349
Since there are M rows and, in any particular
row, N columns, there are at most N+M modules

00:02:40.349 --> 00:02:43.690
along any path from input to output.

00:02:43.690 --> 00:02:50.810
So the latency is order N, since M and N differ
by just some constant factor.

00:02:50.810 --> 00:02:55.329
Since this is a combinational circuit, the
throughput is just 1/latency.

00:02:55.329 --> 00:02:58.650
And the total amount of hardware is order
N^2.

00:02:58.650 --> 00:03:03.890
In the next section, we'll investigate how
to reduce the hardware costs, or, separately,

00:03:03.890 --> 00:03:06.240
how to increase the throughput.

00:03:06.240 --> 00:03:10.160
But before we do that, let's take a moment
to see how the circuit would change if the

00:03:10.160 --> 00:03:15.730
operands were two's complement integers instead
of unsigned integers.

00:03:15.730 --> 00:03:22.110
With a two's complement multiplier and multiplicand,
the high-order bit of each has negative weight.

00:03:22.110 --> 00:03:26.770
So when adding together the partial products,
we'll need to sign-extend each of the N-bit

00:03:26.770 --> 00:03:31.280
partial products to the full N+M-bit width
of the addition.

00:03:31.280 --> 00:03:35.770
This will ensure that a negative partial product
is properly treated when doing the addition.

00:03:35.770 --> 00:03:40.740
And, of course, since the high-order bit of
the multiplier has a negative weight, we'd

00:03:40.740 --> 00:03:44.989
subtract instead of add the last partial product.

00:03:44.989 --> 00:03:46.590
Now for the clever bit.

00:03:46.590 --> 00:03:51.160
We'll add 1's to various of the columns and
then subtract them later, with the goal of

00:03:51.160 --> 00:03:54.150
eliminating all the extra additions caused
by the sign-extension.

00:03:54.150 --> 00:03:59.500
We'll also rewrite the subtraction of the
last partial product as first complementing

00:03:59.500 --> 00:04:02.900
the partial product and then adding 1.

00:04:02.900 --> 00:04:05.660
This is all a bit mysterious but…

00:04:05.660 --> 00:04:11.019
Here in step 3 we see the effect of all the
step 2 machinations.

00:04:11.019 --> 00:04:16.149
Let's look at the high order bit of the first
partial product X3Y0.

00:04:16.149 --> 00:04:22.599
If that partial product is non-negative, X3Y0
is a 0, so all the sign-extension bits are

00:04:22.599 --> 00:04:24.909
0 and can be removed.

00:04:24.909 --> 00:04:30.689
The effect of adding a 1 in that position
is to simply complement X3Y0.

00:04:30.689 --> 00:04:36.779
On the other hand, if that partial product
is negative, X3Y0 is 1, and all the sign-extension

00:04:36.779 --> 00:04:38.099
bits are 1.

00:04:38.099 --> 00:04:44.990
Now when we add a 1 in that position, we complement
the X3Y0 bit back to 0, but we also get a

00:04:44.990 --> 00:04:46.819
carry-out.

00:04:46.819 --> 00:04:51.740
When that's added to the first sign-extension
bit (which is itself a 1), we get zero with

00:04:51.740 --> 00:04:53.219
another carry-out.

00:04:53.219 --> 00:04:58.999
And so on, with all the sign-extension bits
eventually getting flipped to 0 as the carry

00:04:58.999 --> 00:05:01.580
ripples to the end.

00:05:01.580 --> 00:05:09.240
Again the net effect of adding a 1 in that
position is to simply complement X3Y0.

00:05:09.240 --> 00:05:13.969
We do the same for all the other sign-extended
partial products, leaving us with the results

00:05:13.969 --> 00:05:15.279
shown here.

00:05:15.279 --> 00:05:19.319
In the final step we do a bit of arithmetic
on the remaining constants to end up with

00:05:19.319 --> 00:05:21.729
this table of work to be done.

00:05:21.729 --> 00:05:26.009
Somewhat to our surprise, this isn't much
different than the original table for the

00:05:26.009 --> 00:05:28.169
unsigned multiplication.

00:05:28.169 --> 00:05:32.979
There are a few partial product bits that
need to be complemented, and two 1-bits that

00:05:32.979 --> 00:05:35.499
need to be added to particular columns.

00:05:35.499 --> 00:05:38.159
The resulting circuitry is shown here.

00:05:38.159 --> 00:05:42.860
We've changed some of the AND gates to NAND
gates to perform the necessary complements.

00:05:42.860 --> 00:05:46.819
And we've changed the logic necessary to deal
with the two 1-bits that needed to be added

00:05:46.819 --> 00:05:48.869
in.

00:05:48.869 --> 00:05:54.029
The colored elements show the changes made
from the original unsigned multiplier circuitry.

00:05:54.029 --> 00:05:59.409
Basically, the circuit for multiplying two's
complement operands has the same latency,

00:05:59.409 --> 00:06:02.419
throughput and hardware costs as the original
circuitry.