Hardware Reference
In-Depth Information
1
1
2
2
3
3
4
4
¥
D
DROTL
ROTL
D
+
~MSB
~MSB
¥
D
D
D
-
-
y
e
r
y
o
r
x
e
r
0
0
D
D
-
-
D
DROTL
ROTL
1
x
o
r
1
¥
D
DROTL
ROTL
D
+
y
e
i
y
o
i
x
e
i
x
o
i
MSB
2
2
3
3
MSB
¥
D
D
D
+
+
a
r
b
r
4
D
DROTL
ROTL
D
D
-
-
4
5
5
¥
D
DROTL
ROTL
D
+
~MSB
6
a
i
6
7
b
i
7
¥
~MSB
D
D
D
-
-
D
D
-
-
D
DROTL
ROTL
8
w
r
8
9
9
w
i
¥
D
DROTL
ROTL
D
+
MSB
Local memory
w/ LS cells
Local memory
w/ LS cells
MSB
¥
D
D
D
+
+
D
DROTL
ROTL
D
D
-
-
Input (LDINC)
Output (STINC)
Operation cell array
Fig. 3.59
Mapping of parallelized eight-point FFT on FE-GA
numbers. In other words, the input data are divided into two groups with even num-
bers and odd numbers, and they are stored in different banks (bank 0 and bank 2 for
the even numbers, bank 1 and bank 3 for the odd numbers) of the local memory
(Fig.
3.59
). Also, the two different input data to the butterfly,
x
and
y
, are respec-
tively stored on the first half and the latter half of the same bank of the local mem-
ory. Since each bank is a dual-port memory, these two data items can be read
simultaneously, and they are provided to two operation cells at the same time by the
crossbar's multicast operation. Operation results are stored on different banks (bank
4-7) of the local memory.
Since the FFT algorithm is modified to obtain an identical mapping of the
butterfly calculation, the total number of threads depends on the cell configurations
related to data input and output. Figure
3.60
describes the defined threads and their
sequence for 1,024-point FFT. The 1,024-point FFT has 10 stages of the butterfly
calculations. The configuration of the cell array that includes the ALU and MLT
cells is common among all the stages. Input data and output data are divided in order
to be stored in ten banks of the local memory. One stage places its output data in a
bank of the local memory, and the next stage uses the output data in the bank as
input data. In other words, two types of configurations for the LS cells (L1 and L2
in the figure) are defined and alternatively used. The twiddle factors are placed in
Search WWH ::
Custom Search