Sunday, March 26, 2017

DCO: SPI and DAC accuracy

A few compromises have to be made when making the DCO. There are however a few things that must be done right.

First of all, we cannot afford to lose any input commands. This will put the SPI input buffer in an unknown state, possibly locking up the DCO.

With this in mind, we can do a few calculations.

First of all - each frequency update is based on a 16bit integer. The SPI buffer is only 8 bits long, so two bytes must be transferred. Each transfer triggers an interrupt, and that interrupt must copy the transferred byte from the SPI buffer to somewhere else before the next byte arrives.

Fortunately, the SPI buffer is, as it says, buffered. This means that it is not written until the last bit of the byte is received. In addition, there is a short delay between bytes.

As an example, when running the SPI bus at 200kHz, a byte is 40uS long. The delay between two bytes is 7.5uS (when output from node.js on a Raspberry PI).

Running the MCU at 20MHz we have an instruction clock of 5MHz, and each instruction takes 0.2uS. This means that with a 200kHz SPI bus and 20MHz clock, we have 237 instruction cycles to handle a received byte.

This is of course much more than we need - as long as the interrupt handler triggers immediately. But for the DCO to be as accurate as possible, we have given priority to the timer that sets the output wave period. The SPI interrupt will be blocked if a timer interrupt has been received and will not be treated until the timer interrupt handler returns. Thus, we have to make sure the time it takes to enter the timer interrupt handler, do whatever it needs to do, return from the handler and then enter the SPI (low priority) interrupt handler, takes less than 237 instructions - if not, we may be inside the timer interrupt handler when the next SPI byte arrives and we'll lose the previous byte.

This sounds all good, right? The time it takes to reset the timer is not that long. Well, we do actually have a problem. If we are to both reset the timer AND change the DAC value (that controls the current that charges the integrator capacitor making the saw wave), we run into trouble. The DAC is written to using SPI, and we have to write three bytes to set it. Unless the DAC SPI runs at several MHz, we do not have time to write to it within the 47uS we have available. In simple terms, we cannot write to the DAC inside the interrupt handler.

So what is the consequence of this? 

The only other place to write to the DAC is inside the main program loop. As we will be at an arbitrary place in the code when a timer interrupt triggers, we will get a delay from when the timer triggers to where the DAC update call is found (+ the time spent writing to the DAC).

To reduce the maximum delay, we may check if a DAC update has been requested and update the DAC at several spots in the main loop.

But how big is the consequence of delaying the DAC?

At 200kHz, updating the DAC takes approximately 135uS = 675 instruction cycles. This delay is unavoidable.

The longest single method within the main loop, recalculating the frequency, takes 723 cycles. If we break this up into 5 intervals, we'll add around 150cycles to the maximum delay, which will then be around 825 cycles or 165uS.

What does this mean?

It means that for 165uS, the DAC will output the wrong voltage, and the capacitor will charge too fast or too slow. If it charges too slow, the capacitor will be at less than 5V when the charging is reset by the period timer. If it charges too fast, it will be at more than 5V. But how much?

Let's look at the two extremes:

Going from 16.5Hz to 20kHz

Here, the DAC will have a very low charge current. The period at 20kHz is 50uS. This means that for a little more than three periods, the charging will be way too low, in fact probably so low that the wave will look completely flat. However, this is only for 3 of 20.000 cycles within a second. Question is if it is audible.

Going from 20kHz to 16.5Hz

The DAC will have a very high charge current. It charges at the same level as when the frequency is 20kHz, which means that after 50uS, the output value is 5V and after 150uS it has reached 15V (but the buffer opamp or other parts of the circuit probably saturates, keeping the maximum at around 14.3V. The cap stays at this level for one full charging cycle, around 61mS. This is not a very ideal situation.

This is of course in the extreme cases. Most often the change in pitch will be small, but a few octaves are still likely.

What can be done to remedy this?

1) Increased DAC SPI speed - pushing the speed to 2MHz will reduce the write delay to 13.5uS, with a total of 43.5uS including the delay before we reach the reset DAC code. This is if the MCU and DAC can handle the increased SPI speed

2) Increase the MCU speed from 20MHz to 32MHz - this will reduce the maximum time before the DAC write is started from 30uS to 18.75uS (as each instruction now takes 0.125uS instead of 0.2uS.

3) Check more often if we need to update the DAC, further reducing the maximum time before the DAC is written to

4) If the DAC SPI speed is sufficiently high, put the write back into the interrupt routine, removing the delay before a write altogether.

If 1) is possible, and the rest of the timer interrupt handler takes less than 47.5uS - 13.5uS = 34uS, 4) is certainly the best option.


Input SPI speed

We chose 200kHz as the SPI speed at the start of this post. At this speed, 16 bytes will take 47.5uS to transfer (inc delay between bytes), giving an effective transfer rate of 21kB/s.  The Xonik M8 voice controller runs at less than 4k updates/second, so this is sufficient for updating tree DCOs at that rate. The PIC32 has a multi-byte SPI buffer, so we may actually write all bytes to the SPI buffer and forget about them, leaving the voice card MCU to do other duties while the buffer is written (though we have to write to the separate DCOs one at the time as the chip select lines must be changed between the DCO writes.

To further improve the performance of the main MCU (and reduce the delay between when a frequency update is requested and a new frequency is visible on the output), we should calculate the maximum time spent in the period timer interrupt handler and optimize the SPI speed such that the time spent transferring a byte is just barely longer than this.


Maximum delay

The maximum delay from a byte request has been received at the main voice controller to it is visible at the DCO input may be calculated as follows:

2 x matrix calculation time
+ 2 x byte transfer time
+ 1 x transfer delay
+ 1 x time spent calculating new frequency params
+ 1 x delay writing to timer registers
+ 1 x time that must be left for an update
+ 1 x longest cycle time
+ 1 x time within period interrupt handler
+ 1 x maximum delay before DAC write starts
+ 1 x DAC write time.

2 x matrix calc time because the command may arrive just after a calc has started and be written at the end of a calc update


Update: SPI and DAC speeds

The MAX5216 SPI can run at 50Mhz. The PIC SPI runs at maximum Fosc/4, which equals 5MHz with a 20MHz crystal, 8MHz with a 32Mhz. In any case, transferring 24 bits takes approx 27 instruction cycles, so setting the DAC in the interrupt routine is well within what is feasible. Thus, the delay from period start to the DAC is set is 27 * 0.125uS = 3.4uS + any cycles between the timer reset and the call to set the DAC - 7 cycles + the time it takes to call the SPI function - an estimate may be 40 cycles = 5uS in total (at 32MHz).

This means that in the case of a too-high charging current, the cap will be charged at max for about 1/10th of the charge time at 20kHz, contributing about 0.5V too much to the end result (5.5V instead of 5V). This is probably tolerable.

Similarly, the whole interrupt function will take 33 cycles (timer) + 40 cycles (DAC). Add 10 cycles to enter and store the SPI byte and the total is 83 cycles or 10.4uS. Transmission speed will then have to be more than 1.3uS per bit, which gives a max SPI speed of 769kHz (at 32MHz)

Update 2:

The only times we should have to update the charging current will be right after we have calculated and set a new frequency. For most frequencies, this update will happen right after an SPI write (for low frequencies we will get a delay, though). If we keep the DAC update outside of the interrupt handler, we will still normally not be inside the calculation routine for a new frequency when the interrupt returns, so the delay between timer update and DAC update will be minimal. We may still encounter a large delay as it may take up to one period from frequency calculation to the frequency has been set. If SPI bytes arrive faster than this (as they most likely will if the main controller runs at 2-4kHz), we may still be inside the calculation routine.

By moving the DAC update we'll shorten the delay to 43 cycles = 5.38uS, giving us a bit time of 0.673uS and a max SPI speed of 1.49MHz (at 32MHz clock speed)

Ultimately it all comes down to what is most important - a high SPI speed to minimize the time the master spends writing or a short delay before the DAC is set to assure a high max voltage accuracy.

As mentioned previously, we may use the multi-byte SPI buffer on the PIC32 based voice controller, so the exact write speed is not all that important (as we do writes as fire-and-forget) but faster is still better - we will have to write some special code that writes to the DCOs at regular intervals through one matrix recalculation cycle as we cannot afford wasting calculation time waiting for a write to complete before filling the SPI buffer with the next DCO values (Though perhaps it would be possible to come up with a chaining scheme where six bytes are clocked through the three DCOs before they are "commited" and the values read - perhaps by a chip select line going low or similar. The PIC32 is capable of generating an interrupt once the SPI buffer is empty, so it's easy to add a single line change when transmission is complete).

No comments:

Post a Comment