Hardware Reference
In-Depth Information
32-Bit register
A B C D
E F G H
I J K L
M N O P
A E I M
B F J N
C G K O
D H L P
R2
R3
R4
R5
A B C D
E F G H
I J K L
M N O P
R2
R3
R4
R5
A E I M
B F J N
C G K O
D H L P
(a)
(b)
(c)
(d)
Figure 8-6.
(a) An array of 8-bit elements. (b) The transposed array. (c) The
original array fetched into four registers. (d) The transposed array in four regis-
ters.
One way to do the transposition is to use 12 operations that each load a byte
into a different register, followed by 12 more operations that each store a byte in its
correct location. (
Note
: the four bytes on the diagonal do not move in the transposi-
tion.) The problem with this approach is that it requires 24 (long and slow) opera-
tions that reference memory.
An alternative approach is to start with four operations that each load a word
into four different registers,
R2
through
R5
, as shown in Fig. 8-6(c). Then the four
output words are assembled by masking and shifting operations to achieve the de-
sired output, as shown in Fig. 8-6(d). Finally, these words are stored in memory.
Although this way of doing it reduces the number of memory references from 24 to
8, the masking and shifting is expensive due to the many operations required to
extract and insert each byte in the correct position.
The TriMedia provides a better solution than either of these. It begins by
fetching the four words into registers. However, instead of building the output
using masking and shifting, special operations that extract and insert bytes within
registers are used to build the output. The result is that with eight memory refer-
ences and eight of these special multimedia operations, the transposition can be
accomplished. The code first contains an operation with two load operations in
slots 4 and 5, respectively, to load words into
R2
and
R3
, followed by another such
operation to load
R4
and
R5
.
The instructions holding these operations can use slots 1, 2, and 3 for other
purposes. When all the words have been loaded, the eight special multimedia op-
erations can be packed into two instructions to build the output, followed by two
operations to store them. Only six instructions are needed, and 14 of the 30 slots
in these instructions are available for other operations. In effect, the entire job can
be done with the effective equivalent of about three or so instructions. Other multi-
media operations are also efficient. Between these powerful operations and the
five issue slots per instruction, the TriMedia is highly efficient at doing the kinds of
calculations needed in multimedia processing.
