Soft-Core Processor Design - CiteSeer

Another argument in favour of simple applications is the Nios instruction set, which supports
only simple arithmetic and logic operations. Therefore, the benchmark set should be based on a
set of simple programs. We base our benchmark set on MiBench [56], a freely available
benchmark suite representing applications for commercial embedded systems. The MiBench is
divided into six categories according to the application field, which include automotive and
industrial control, network, security, consumer devices, office automation, and
telecommunications. Many of the benchmarks in the MiBench suite perform computationally
intensive operations such as the Fast Fourier Transform (FFT), image
compression/decompression, or GSM encoding/decoding. The UT Nios benchmark set does not
include any computationally intensive applications. It also does not include benchmarks that use
floating point numbers, because we believe that most applications running on Nios-based systems
will not use floating point numbers. Some of the programs in the MiBench suite that originally
used the floating point numbers were adapted to use only the integer types.
According to the established criteria, the following MiBench benchmarks were selected:
• Bitcount: assesses the bit manipulation capabilities of a processor. The program
counts the number of bits in an array of integers with an equal number of 1s and 0s.
Five different algorithms on two dataset sizes are used. Algorithms include 1-bit per
loop counter, recursive bit count by nibbles, non-recursive bit count by nibbles using
a table look-up, non-recursive bit count by bytes using a table look-up, and shift and
count bits. A small dataset consists of 75,000 integers, while large dataset contains
1,125,000 integers.
• CRC32: calculates a 32-bit Cyclic Redundancy Check (CRC) on an input text (305
Kbytes).
• Dijkstra: calculates the shortest path between the pairs of nodes in a graph given in the
form of an adjacency matrix. Two versions of the benchmark are used, differing in the
number of the pairs of nodes for which the shortest paths are calculated. Small version
calculates 20, while large version calculates 100 shortest paths. The adjacency matrix
has size 100 X 100 for both benchmark versions.
• Patricia: searches for a node with a given key in the Patricia trie data structure, and
inserts a new node if the one does not already exist. Patricia trie data structure is often
used to represent the routing tables in network applications. The dataset is a list of
11,000 entries representing the IP traffic of a web server [56].
48