Performance benefit when using DMA for PWM - c

Question

I have a segment of code below as a FreeRTOS task running on an STM32F411RE microcontroller:
static void TaskADCPWM(void *argument)
{
/* Variables used by FreeRTOS to set delays of 50ms periodically */
const TickType_t DelayFrequency = pdMS_TO_TICKS(50);
TickType_t LastActiveTime;
/* Update the variable RawAdcValue through DMA */
HAL_ADC_Start_DMA(&hadc1, (uint32_t*)&RawAdcValue, 1);
#if PWM_DMA_ON
/* Initialize PWM CHANNEL2 with DMA, to automatically change TIMx->CCR by updating a variable */
HAL_TIM_PWM_Start_DMA(&htim3, TIM_CHANNEL_2, (uint32_t*)&RawPWMThresh, 1);
#else
/* If DMA is not used, user must update TIMx->CCRy manually to alter duty cycle */
HAL_TIM_PWM_Start(&htim3, TIM_CHANNEL_2);
#endif
while(1)
{
/* Record last wakeup time and use it to perform blocking delay the next 50ms */
LastActiveTime = xTaskGetTickCount();
vTaskDelayUntil(&LastActiveTime, DelayFrequency);
/* Perform scaling conversion based on ADC input, and feed value into PWM CCR register */
#if PWM_DMA_ON
RawPWMThresh = (uint16_t)((RawAdcValue * MAX_TIM3_PWM_VALUE)/MAX_ADC_12BIT_VALUE);
#else
TIM3->CCR2 = (uint16_t)((RawAdcValue * MAX_TIM3_PWM_VALUE)/MAX_ADC_12BIT_VALUE);
#endif
}
}
The task above uses RawAdcValue value to update a TIM3->CCR2 register either through DMA or manually. The RawAdcValue gets updated periodically through DMA, and the value stored in this variable is 12-bits wide.
I understand how using DMA could benefit reading the ADC samples above as the CPU will not need to poll/wait for the ADC samples, or using the DMA to transfer long streams of data through I2C or SPI. But, is there a significant performance advantage to using DMA to update the TIM3->CCR2 register instead of manually modifying the TIM3->CCR2 register through:
TIM3->CCR2 &= ~0xFFFF;
TIM3->CCR2 |= SomeValue;
What would be the main differences between updating the CCR register through DMA or non-DMA?

Answer 1

Let's start by assuming you need to achieve "N samples per second". E.g. for audio this might be 44100 samples per second.
For PWM, you need to change the state of the output multiple times per sample. For example; for audio this might mean writing to the CCR around four times per sample, or "4*44100 = 176400" times per second.
Now look at what vTaskDelayUntil() does - most likely it sets up a timer and does a task switch, then (when the timer expires) you get an IRQ followed by a second task switch. It might add up to a total overhead of 500 CPU cycles each time you change the CCR. You can convert this into a percentage. E.g. (continuing the audio example), "176400 CCR updates per second * 500 cycles per update = about 88.2 million cycles per second of overhead", then, for 100 MHz CPU, you can do "88.2 million / 100 million = 88.2% of all CPU time wasted because you didn't use DMA".
The next step is to figure out where the CPU time comes from. There's 2 possibilities:
a) If your task is the highest priority task in the system (including being higher priority than all IRQs, etc); then every other task will become victims of your time consumption. In this case you've single-handedly ruined any point of bothering with a real time OS (probably better to just use a faster/more efficient non-real-time OS that optimizes "average case" instead of optimizing "worst case", and using DMA, and using a less powerful/cheaper CPU, to get a much better end result at a reduced "cost in $").
b) If your task isn't the highest priority task in the system, then the code shown above is broken. Specifically, an IRQ (and possibly a task switch/preemption) can occur immediately after the vTaskDelayUntil(&LastActiveTime, DelayFrequency);, causing theTIM3->CCR2 = (uint16_t)((RawAdcValue * MAX_TIM3_PWM_VALUE)/MAX_ADC_12BIT_VALUE); to occur at the wrong time (much later than intended). In pathological cases (e.g. where some other event like disk or network just happens to occur at a similar related frequency - e.g. at half your "CCR update frequency") this can easily become completely unusable (e.g. because turning the output on is often delayed more than intended and turning the output off is not).
However...
All of this depends on how many samples per second (or better, how many CCR updates per second) you actually need. For some purposes (e.g. controlling an electric motor's speed in a system that changes the angle of a solar panel to track the position of the sun throughout the day); maybe you only need 1 sample per minute and all the problems caused by using CPU disappear. For other purposes (e.g. AM radio transmissions) DMA probably won't be good enough either.
WARNING
Unfortunately, I can't/didn't find any documentation for HAL_ADC_Start_DMA(), HAL_TIM_PWM_Start() or HAL_TIM_PWM_Start_DMA() online, and don't know what the parameters are or how the DMA is actually being used. When I first wrote this answer I simply relied on a "likely assumption" that may have been a false assumption.
Typically, for DMA you have a block of many pieces of data (e.g. for audio, maybe you have a block 176400 values - enough for a whole second of sound at "4 values per sample, 44100 samples per second"); and while that transfer is happening the CPU is free to do other work (and not wasted). For continuous operation, the CPU might prepare the next block of data while the DMA transfer is happening, and when the DMA transfer completes the hardware would generate an IRQ and the IRQ handler will start the next DMA transfer for the next block of values (alternatively, the DMA channel could be configured for "auto-repeat" and the block of data might be a circular buffer). In that way, the "88.2% of all CPU time wasted because you didn't use DMA" would be "almost zero CPU time used because DMA controller is doing almost everything"; and the whole thing would be immune to most timing problems (an IRQ or higher priority task preempting can not influence the DMA controller's timing).
This is what I assumed the code is doing when it uses DMA. Specifically, I assumed that the every "N nanoseconds" the DMA would take the next raw value from a large block of raw values and use that next raw value (representing the width of the pulse) to set a timer's threshold to a value from 0 to N nanoseconds.
In hindsight; it's possibly more likely that the code sets up the DMA transfer for "1 value per transfer, with continual auto-repeat". In that case the DMA controller would be continually pumping whatever value happens to be in RawPWMThresh to the timer at a (possibly high) frequency, and then the code in the while(1) loop would be changing the value in RawPWMThresh at a (possibly much lower) frequency. For example (continuing the audio example); it could be like doing "16 values per sample (via. the DMA controller), with 44100 samples per second (via. the while(1) loop)". In that case; if something (an unrelated IRQ, etc) causes an unexpected extra delay after the vTaskDelayUntil(); then it's not a huge catastrophe (the DMA controller simply repeats the existing value for a little longer).
If that is the case; then the real difference could be "X values per sample with 20 samples per second" (with DMA) vs. "1 value per sample with 20 samples per second" (without DMA); where the overhead is the same regardless, but the quality of the output is much better with DMA.
However; without knowing what the code actually does (e.g. without knowing the frequency of the DMA channel and how things like the timer's prescaler are configured) it's also technically possible that when using DMA the "X values per sample with 20 samples per second" is actually "1 value per sample with 20 samples per second" (with X == 1). In that case, using DMA would be almost pointless (none of the performance benefits I originally assumed; and almost none of the "output quality" benefits I'm tempted to assume in hindsight, except for the "repeat old value if there's unexpected extra delay after the vTaskDelayUntil()").

Answer 2

First, remember that premature optimization is the cause of uncountably many problems. The question you need to ask is "what ELSE does the processor need to do?". If the processor has nothing better to do, then just poll and save yourself some programming effort.
If the processor does have something better to do (or you are running from batteries and want to save power) then you need to time how long the processor spends waiting between each thing that it needs to do.
In your case, you are using an operating system context switch in place of "waiting". You can time the cost of the switch-write-to-pwm-switch-back cycle by measuring the performance of some other thread.
Set up a system with two threads. Perform some task that you know the performance of in one thread, eg, some fixed computation or processor benchmark. Now set up the other thread to do your timer business above. Measure the performance of the first thread.
Next set up a similar system with only the first thread plus DMA doing the PWM. Measure the performance change, you have you answer.
Obviously this all depends very much on your exact system. There is no general answer that can be given. The closer your test is to your real system the more accurate the answer you will get.
PS: Your PWM will glitch using the above code. Replace the two writes with a single one:
TIM3->CCR2 &= ~0xFFFF;
TIM3->CCR2 |= SomeValue;
should be:
TIM3->CCR2 = ((TIM3->CCR2 & ~0xFFFF) | SomeValue);