DMA and the Blackfin DSP

Every day there's more to like about the Analog Devices Blackfin DSP. Today the big wins are DMA transfers and built in libraries.

DMA

So what is DMA? If you aren't familiar with DMA, it's the ability of the processor to transfer memory to or from other memory locations or peripherals without the attention of the CPU. This means that there is a separate DMA engine sitting inside the chip acting like a gremlim, moving things about while no one is looking. And gremlin is an appropriate image since DMA in the past had a bit of a reputation for being complex, error prone, and unreliable at least among the embedded developers I know. The upside of DMA is that it can free the processor from huge amounts of work it would normally have to do itself.

For example, many chips have UARTS for serial communication that can be accessed via DMA. If you had a long string of data to read in from the port there are only three choices: poll, interrupts, or DMA. In polling you periodically check on the UARTS buffer and copy out any values that have arrived since the last time you checked. This requires you to check back often and to be careful lest you not check back quickly enough and the buffer overflows and you lose data. With interrupts you write an interrupt handler that gets called every time N new values have arrived in the buffer. Interrupts guarantee that you get there in time but they still take processing time away from your main process, and let's not even delve in what happens when higher priority interrupts are occuring rapidly at the same time. With DMA you set the DMA engine off and running and at the end you can probably setup an interrupt to be triggered once the whole buffer has arrived, meanwhile you have 100% of the processor time to allocate to your main process. Now compound this with the fact that many DMAs can run multiple different kinds of transfers at the same time and you can see how powerful they become.

On a PC the only time you would encounter DMA is if you were doing device driver development. In the embedded world you encounter them much more often because your programming much closer to the metal and more often than not working in a constrained environment where you need to squeeze the most performance from your chip.

With anything as powerful as DMA there are issues, the first being that they are complex to setup. This is something the designers of the TI DSPs seemed to revel in. The setup of a DMA transfer requires setting up two 32-bit control registers where each bit affects the transfer in some way or another. On top of that there are a dozen or so other registers that may be used, depending, of course, on the setup of those 64 bits in the control registers. But even here, they just couldn't stop themselves because they must have decided their DMA engine was getting too large, so what did they do? Reduce the number of options? No. Simplify? No. What they did is decide that having all those other registers was the problem, so different DMA transfers would share those registers. Share them. And of course since they're sharing you have to use more control bits to the control registers to select which of the shared registers you are using.

Compare this to the Analog Devices Blackfin DMA controller. There is a single 16-bit DMA control register and each transfer has a unique set of other registers. No sharing. How did they do this? By concentrating on making the easy things easy and the hard things possible. You see, the basic DMA covers most of my needs and is configured easily. If a more complex DMA transfer is needed they actually used a clever technique where you setup structures in memory with control information and chain them together, a linked list of structures, then you feed the first structure to the DMA engine and off it goes following the instructions at each structure then automatically jumping to the next one. It's a great setup and makes it even more flexible than the TI DSP's DMA with its 64-bits of configuration.

The usual caveat should be applied here that I am only talking about a single family of the TI DSP and I don't have experience with other families which may have different schemes.

Libraries

Now if DMAs aren't your thing how about FFTs? Well if you're using a DSP you'd be quite likely to want to do such things. TI handled this by having a large download section with plug in modules that dropped into the development environment to cover basic signal processing algorithms. They have constructed a whole mechanism where a provider can write a library and have it work across a range of their DSP families.

This is how I expected Analog Devices works with their DSPs so I go on line and try to download the required modules from their site, but for the life of me I can't find any place to download the modules. I searched and searched and in a last act of desperation I typed 'FFT' into the developement tools help system. Now, that last statement could be read as, "in a last act of desperation I read the manual", but we won't go there. As it turns out I didn't need to go to their site, nor did I need to use any special module loading, the Blackfin developement tools ships with DSP library that includes real FFTs, complex FFTs, fir filters, auto-correlation, etc., etc. Built in. To the default library. For a DSP. A Digital Signal Processor. In retrospect it just seems so obvious.

At this point I'm still learning about this chip and it's capabilities, I may yet run across something I don't like, but so far I'm impressed both with the robustness of the product and the obvious thought that went into it's design.