It has been two months since I first was able to play with the OsmoSDR hardware prototypes. Back at that time, there was no FPGA code yet, and some hardware bugs still had to be resolved. Nonetheless, the e4k tuner driver could already be implemented and tuning was confirmed by looking at the analog i/q spectrum.
Meanwhile, the hardware has been re-worked by SR-Systems and FPGA VHDL code written by maintech.de. Ever since that, they dropped the ball again with me as I had been careless enough to volunteer for writing the firmware.
And that's what I did or at least tried to do for quite some time during the last two weeks. The main problem was that I didn't have much time. The second problem was that I never was able to get the SSC (synchronous serial controller) receive DMA working.
This was a really odd experience, as I've worked a lot with that very same SSC peripheral before, while writing firmware for the OpenPICC some 6 years ago. However, this was in an at91sam7s, where the SSC is interfaced with the PDC (Peripheral DMA Controller). In the at91sam3u of OsmoSDR, it interfaces with a more modern DMAC/HDMA controller, capable of scatter-gather DMA and other fancy stuff.
Atmel has provided reference code that uses the SSC DMA in transmit mode (for a USB audio device playing back music via the Wolfson codec on the SAM3U-EK board). After thoroughly studying the DMAC/HDMA documentation I set out to write code for DMA-based SSC receiver. And it never worked.
I actually wrote two independent implementations, one from scratch and the other based on Atmel reference code. Neither of them worked. It seemed to be a problem with the hardware hand-shaking between SSC and DMAC. The SSC was successfully receiving data, and that data could be read out from the CPU using a polling or IRQ based driver. But if you're running at something like 32 Mbps and don't have a FIFO, you desperately want to use DMA. When the DMA handshaking was turned off, the DMA code worked, but of course it read the same received word several thousand times before the next data arrived on the SSC.
In the end, I was actually convinced it must be a silicon bug. Until I thought well, maybe they just connected the flow controller to a different ID in Rx and Tx direction. Since there are only 16 such identifiers, it was relatively easy to brute-force all of them and see if it worked. And voila - using the identifier 4, it worked!
So what had happened? The Atmel-provided reference code contained a
#define BOARD_SSC_DMA_HW_SRC_REQ_ID AT91C_HDMA_SRC_PER_3and that was wrong. 3 is valid for SSC TX but not for SSC RX. Unfortunately I never found any of those magic numbers in the SAM3U manual either. They are not documented in the chapter of the SSC, and they are not documented in the chapter about HDMA/DMAC either. And they are not identical with the Peripheral Identifiers that are used all over the chip for the built-in peripherals.
In case anyone else is interested, a patch can be found at my at91lib git repository.
I filed a ticket with Atmel support, and they pointed out in fact there was a table with those identifiers somewhere in the early introductory chapters where you can see a brief summary of the features of each integrated peripheral. Unfortunately they use slightly different naming in that chapter and in the DMAC, so a full-text search also didn't find them. Neither is that table visible in the PDF index.
So about four man-days later it was finally working. Another day was spent on integrating it with the USB DMA for sending high-speed isochronous transfers over the bus into the PC. And ever since I'm happily receiving something like 500,000 or 1,000,000 samples / second from an alsa device, using snd-usb-audio. Luckily, unlike MacOS or Windows, the Linux audio drivers don't make arbitrary restrictions in the sample rate. According to the USB Audio spec, the sample rate can be any 24bit number. So audio devices with 16.7 Ms/s are very much within the spec. I hope some of the other OS driver writers would take that to their heart.
One of the first captures can be found at this link, containing a bzip2ed wave file in S16LE format Stereo (I/Q). It contains a FM audio signal transmitted using a small pocket-sized FM transmitter.
There is no I/Q DC offset calibration yet, but once that is done we're probably able to finally put the design into production.