Tweet
Further to my comments under "Digital PI" below about the ADuC81x family and the limitations of the 805x core for DSP applications, I should perhaps expand on the subject for non-DSP types...
There is a general issue with doing DSP-type operations efficiently on processors that do not have atomic instructions for key operations. In particular, the so-called bit-reversed indexing operations allow very efficient FFTs (Fast Fourier Transforms) to be carried out, and MAC instructions (Multiply-Accumulate) which are used by FIR and IIR digital filters. Other capabilities, such as doing multiple fetches and preloads are also important.
Of course, any processor can implement FIR/FIR/FFT etc. algorithms, but the above mentioned instructions make a radical difference to the speed that they run at. PICs, AVRs, 805x etc. do not have these, and are consequentially relatively very slow on a cycle for cycle basis. Also, the instruction clock time for these processors tends to be low, e.g. a max of about 20MHz, which is simply not enought to do anything reasonable in DSP terms - a DSP-specific chip will do far far more per instruction than any generic device and anyway, they tend to run at very high clock speeds.
So, lets look at the ADuC81x again. Its like an AD (Analog Devices) version of a PIC - it has ADCs as well as DACs, timers, EEPROM etc., but it has a 20MHz 8051 core rather than a PIC one... Essentially, then, it has the same problems - its not at all well suited to real-time DSP-style processing of data. Even if you drive the ADCs off interrupts in order to get more instruction cycles between samples, you will then have a non-deterministic core processing loop, making accurate timing very difficult for the sample/sample&hold gates.
MicroChip has produced the dsPIC, which I haven't played with yet. These chips are really for low-speed DSP operations, such as motor control. To do significant DSP work (running FIRs etc.) you will require a faster and/or more specialised processor. The problem is that these tend to be power-hungry, and they are not really for amateur use as they tend to be fine-pitch LQFP or BGA packaged, typically requiring a minimum of 4-layer PCBs and 8mil (even 6mil)DRC clearance. Also, DSPs tend to be poor at driving slow-speed peripherals, such as displays - the ADuC81x gets away with this because it too is low speed (relatively) - but a real DSP chip will gag on trying to access a display as a memory or I/O mapped device - it will be hobbled.
A compromise is to use a PIC or similar to drive the user interface, and the DSP to do the grunt work. This is a happy compromise - you can use any number of ways for the sub-systems to communicate, even serial interfaces, I2C, SPI etc.
So, in summary, general purpose processors are good at just that - being general purpose. When it comes to very specialised applications, such as real-time signal processing, they fall down, and DSP-optimised processors are far, far better.
HTH,
Nicko
There is a general issue with doing DSP-type operations efficiently on processors that do not have atomic instructions for key operations. In particular, the so-called bit-reversed indexing operations allow very efficient FFTs (Fast Fourier Transforms) to be carried out, and MAC instructions (Multiply-Accumulate) which are used by FIR and IIR digital filters. Other capabilities, such as doing multiple fetches and preloads are also important.
Of course, any processor can implement FIR/FIR/FFT etc. algorithms, but the above mentioned instructions make a radical difference to the speed that they run at. PICs, AVRs, 805x etc. do not have these, and are consequentially relatively very slow on a cycle for cycle basis. Also, the instruction clock time for these processors tends to be low, e.g. a max of about 20MHz, which is simply not enought to do anything reasonable in DSP terms - a DSP-specific chip will do far far more per instruction than any generic device and anyway, they tend to run at very high clock speeds.
So, lets look at the ADuC81x again. Its like an AD (Analog Devices) version of a PIC - it has ADCs as well as DACs, timers, EEPROM etc., but it has a 20MHz 8051 core rather than a PIC one... Essentially, then, it has the same problems - its not at all well suited to real-time DSP-style processing of data. Even if you drive the ADCs off interrupts in order to get more instruction cycles between samples, you will then have a non-deterministic core processing loop, making accurate timing very difficult for the sample/sample&hold gates.
MicroChip has produced the dsPIC, which I haven't played with yet. These chips are really for low-speed DSP operations, such as motor control. To do significant DSP work (running FIRs etc.) you will require a faster and/or more specialised processor. The problem is that these tend to be power-hungry, and they are not really for amateur use as they tend to be fine-pitch LQFP or BGA packaged, typically requiring a minimum of 4-layer PCBs and 8mil (even 6mil)DRC clearance. Also, DSPs tend to be poor at driving slow-speed peripherals, such as displays - the ADuC81x gets away with this because it too is low speed (relatively) - but a real DSP chip will gag on trying to access a display as a memory or I/O mapped device - it will be hobbled.
A compromise is to use a PIC or similar to drive the user interface, and the DSP to do the grunt work. This is a happy compromise - you can use any number of ways for the sub-systems to communicate, even serial interfaces, I2C, SPI etc.
So, in summary, general purpose processors are good at just that - being general purpose. When it comes to very specialised applications, such as real-time signal processing, they fall down, and DSP-optimised processors are far, far better.
HTH,
Nicko
Comment