The DSP
(digital signal processor) are made to the treatment of numerical signs.
The DSP is
a calculator only for numerical signals, and on the sample the signal is
however converted before being calculated.
The
processor DSP can today be used for mixture, equalisation, limitation of dynamics of audio signals and all at once,
for the coding of numerical audio data.
• In the
field of telecommunications: modems, multiplexers, numerical recievers DTMF, fax machines, GSM word coder, and radio
modems
• In the field of vocal interfaces: vocal converters for answering machines,
automatic word recognition, vocal synthesis.
• In the field of the multimedia and general public: compression of the audio
signals (CD) compression of pictures, multimedia cards for PC, musical
synthesis
• In the field of instrumentation: spectrum analyzers, function generation and
seismic signs interpretation.
. The
peculiarities of DSP
As a classical
microprocessor DSP is implemented by linking him memory (RAM, ROM) and
peripherals. Typical DSP is rather used in systems of autonomous treatments. It
usually comes as a microcontroller inserting according to brands and
constructors gammas of the memory of timers, quick synchronous ports, DMA
controllers, ports of various E/S.
2.1 The heart of a system of numerical
treatment of the signç
DSP is donc conceived to optimize this time of counting. Guide
with effect, they have optimized functions allowing to calculate much more
fast.
2.2 Access to memory
Another characteristic of DSPs is their capacities to accomplish several memory
access in a single cycle. Nevertheless
on certain basic DSP this simultaneous type of operation is in general limited
to special instructions. Those instructions use a restricted addressing mode
concerning only the random access memory inserted into the DSP
A DSP may posses many logical units of address
generation working in parallel with the logic of the heart of the DSP. A
logical unit of address generation is only once configured using the
appropriate registers.
This allows
not only to accomplish access to simultaneous memories in a single cycle as it
was said before but also to increment generated address automatically. This
mode of addressing particular usually called indirect addressing
by register with post (or pre) increment is very used to perform repetitive
counting on various datas saved sequentially in the memory. Let’s take the TMS320C64XX as an example . If we consider the operations required for the
execution of an instruction may be described as follows:
To save time when running series of instructions it is necessary to optimize
the different steps in the parallelizing or in the pipeline. The standard retained is the one
used in the production plants which consists decreasing the job in essential objectives:
While reading of this picture we can see
that starting at the 5th instruction every new cycle produces a new instruction is realised instead of
an instruction every 5 cycles produced. A program is therefore executed about 5
times as quickly!
. Architecture
of the processor
The
architecture of a microprocessor and therefore DSP is an important element that
conditions the performances of a processor directly. There are two fundamental
types of said structures "Von Neumann" and "Harvard" such
as introduced by the following image:
3.1 Structure Of VON NEUMAN
A standard microprocessor on a structure Von Neumann stock programs and
information in the same memory zone. An instruction contains surgical code and
address of the operand. This type of microprocessor incorporates mainly two
standard logical units:
•
Arithmetical and logical Unit (or ALU) responsible for accomplishing central
operations (as multiplications, addition, subtractions, rotation, etc)
• Unit in charge of entries / exits which orders the flux of data between the
heart of the microprocessor and memories or ports.
3.2 Structure of Harvard
Due to the fact that memory programs and data are separated this allows to
transfer an instruction and data at the same time wich ameliorates performance.
Harvard is
rather used in microprocessors specialized for real time applications as DSP.
(Examples: The Motorola families 68XXX, the Intel family 80X86)
Some DSP use “Modified Structure Harvard ". DSP offers only a bus of data
and a bus of address. The data transfer between the external and internal buses is performed by temporal
multiplexing. It is the case of TMS320C6711.
. Formats of
data use in DSP.
• DSP with fixed comma: fractional numbers with fixed comma which uses the
method of « CCPL » who allows easily the binary addition of well negative numbers.
• DSP with floating comma: data is represented by using a mantissa and an
exhibitor according to expression: n = 2 displaying mantissa x 2exposant
4.1 DSP with floating comma
DSP with floating comma is more flexible and easier to program than a DSP with
fixed comma. The big dynamic offered by DSP has floating comma allow virtually
not to care about limits of results calculated during the conception of a
program.
. Precision is searched on whole important dynamic range
4.2 DSP with fixed comma
DSP has fixed comma is a little more complicated to program than a DSP with
floating comma.
On a DSP
calculations are performed with 32-bit accumulators. When the results must be
stored in memory the 16 least significant bits are lost. It is always possible
to separately store in memory the least significant 16-bit then the most
significant 16-bit if there are more registers
open at a calculation step.
This feature is not always available on all DSPs. In this situation,
calculations that require absolutely a double precision are made in pairs of
special instructions that handle 16-bit single precision at the expense of
execution time.
The accuracy of the calculations is a critical point of fixed comma DSP, as the
program developer must remain vigilant at every stage of a calculation. He must
seek the greatest possible dynamic to maintain good accuracy of calculations
while avoiding as much as these little overflows of accumulators
The fixed comma DSP are most used because they are less expensive than floating
comma DSP. They are found in all mass market products where cost is an
important factor..
. The
performances of DSP
Most DSPs are particularly intended
for "real time" applications and specialized that means applications where the processing
time is of course important but where the
variety of events to be processed is not significantly important.
However, there are applications where
the DSP performs both functions of digital signal processing and general functions of
a microprocessor at the heart of a conventional computer system.
5.1 Measure of pure speed of counting
classical
method to test the performance of DSP is to be based on its speed of execution.
This basic
method consists therefore in counting the number of instructions performed a
second. An obstacle appears then because an instruction does not mean forcing
the same thing from a family of DSP to other one.
5.2 Measure of the time of execution ("
Benchmark ")
The pure speed
of counting of DSP is not an universal indication, methods and results
different from DSP to other one
Certain DSP
offer in fact modes of more competitive addressing than others
Finally the
time of access to memory is another inevitable parameter. Some DSP have
integrated quick random access memory blocks. This memory is put in the addressing space of DSP in the same capacity as
external random access memory. This allows to line up data and programs without
having to perform permanent transfers towards the outside. The possible cycles
of wait that can be necessary to address a slow external memory are so avoided.
The measure
of performances by benchmark consists in measuring the time that puts the DSP
to run standard treatment of the sign programs
Families latest Texas
Instruments of DSP are:
· TRMS320C54x, DSP fixed format;
· TRMS320C20x, DSP fixed format;
· TRMS320C24x, DSP fixed format;
· TRMS320C62x, fixed format architecture DSP VLIW (Very Long InstructionWord);
· TRMS320C67x, formatting floating DSP VLIW architecture;
These new
families are grouped
into three classes called platforms. These threeclasses are called:
· TMS320C6000, C62x formed and C67x families;
· TMS320C5000, C54x family formed and C54xx;
· TMS320C2000, formed C20x and C24x families;
Figure 9 summarizes the main characteristics of
these three classes.
