Hi all,
I am doing some motion tuning for a quite big and heavy machine. In general I noticed that having high values of jerk (about 500-600 mm/min) are good enough to travel circular paths smoothly (even tough it depends a lot on the mesh quality), but they are a bit high when printing corners, reducing a lot the effect of the input shaping and introducing lot of ringing.
Since I couldn't find a decent balance of jerks, acceleration and the Input Shaping L parameter, I modified a bit the function DDA::MatchSpeed and the command M566:

Basically my idea is to estimate the angle between two consecutive segments and if it is greater than a certain threshold I use a lower value of jerk.

I added to M566 a parameter J "junctionDeviation" which is the threshold below which an high value of jerk is used (the one set from M566 X-- Y--). For greater values, the minimum jerk multiplied by the parameter S is used.
I consider as threshold the difference between the X components of v1 and v2 in the reference frame of v1, and since v1 and v2 are unit vectors the threshold is equal to (1 - cos(alpha)).

void DDA::MatchSpeeds() noexcept
{
	// v2 needs to be expressed in the reference frame of v1
	// directionVector and next->directionVector are unit vectors

	// The rotation matrix R(-alpha) can be expressed by using the directionVector[0] and directionVector[1] which are the X and Y components of v1 in the absolute reference frame (v1_0).
	const float rotationMatrix[2][2] = {
		{ directionVector[0], directionVector[1] },
		{ -directionVector[1], directionVector[0] }
	};

	// Project v2 in the reference frame of v1 (v2_1)
	const float nextVectorProjected[2] = {
		rotationMatrix[0][0] * next->directionVector[0] + rotationMatrix[0][1] * next->directionVector[1],
		rotationMatrix[1][0] * next->directionVector[0] + rotationMatrix[1][1] * next->directionVector[1]
	};

	// Difference of X and Y components in the reference frame of v1 (v1_1 - v2_1)
	const float relativeFraction[2] = {
		fabsf(1 - nextVectorProjected[0]),
		fabsf(0 - nextVectorProjected[1])
	};

	for (size_t drive = 0; drive < MaxAxesPlusExtruders; ++drive)
	{
		if (directionVector[drive] != 0.0 || next->directionVector[drive] != 0.0)
		{
			const float totalFraction = fabsf(directionVector[drive] - next->directionVector[drive]);
			const float jerk = totalFraction * beforePrepare.targetNextSpeed;
			float allowedJerk = reprap.GetPlatform().GetInstantDv(drive);
			// if dealing with X and Y axes reduce jerk if sharp edges
			if (drive < 2 && relativeFraction[0] > reprap.GetPlatform().JunctionDeviation())
			{
				allowedJerk = ConvertSpeedFromMmPerSec(MinimumJerk * reprap.GetPlatform().MinimumJerkMultiplier());
			}
			if (jerk > allowedJerk)
			{
				beforePrepare.targetNextSpeed = allowedJerk/totalFraction;
			}
		}
	}
}

Of course the code is not optimised, but it was just a first test to see if the idea could work.
I have done a couple of prints with a bit extreme setting (M566 X600 Y600 J0.01 S1) and I noticed a good improvement in print quality. Corners don't induce much ringing, so the input shaping can play a role there, and circular motions are done without too much scattering.

I'd like to have some feedback from this approach, especially if you see weird/dangerous edge cases that I didn't think about, in the code implementation and/or in the machine behaviour.

Thank you in advance to all of you!

Hi @chrishamm, I just noticed in the Git tree that the tag 3.6.0-beta.2 is pointing to a commit in the 3.5-dev branch (72d85117). Maybe the intended commit was 4cc23ddc in the 3.6-dev branch? Both have "Tidied up M307 processing" as message.

@dc42 great news!
The modulation parameter goes from -27 to 39.

@dc42 unfortunately, the IDE doesn't say anything more than the datasheet/application notes.
My understanding here is that the sine wave is encoded as a list of increments/decrements to add to the START_SIN and START_SIN90 registries while increasing the microsteps. Since START_SIN and START_SIN90 represent the current of the two coils at the beginning of the electrical revolution, the value has to be greater or equal than zero.
It can be tuned for some desired behaviours or to make some weird configurations feasible (as in the picture below, where for offset less than 28 the wave is not feasible).
In general, from the examples in the datasheet and application notes, I saw the START_SIN registry always set to zero.

@dc42 here they are. The sine offset should be "the absolute current at entry of the lookup table". It shifts up or down the sine wave, so to have it feasible I had to reduce the microstep amplitude. The maximum offset is not feasible, the higher it gets, the flatter the sinewave must be. Hope it's enough to clarify the behaviour, if you need something else, let me know.

@AndyE3D I followed the driver's datasheet and the application note 26:
https://www.analog.com/en/resources/app-notes/an-026.html
https://www.analog.com/media/en/technical-documentation/data-sheets/TMC5160A_datasheet_rev1.17.pdf
These are settings strongly dependent on the motor itself, and I have only nema24 motors. The configuration I posted is the best I found for this type of motor.

@AndyE3D I am using the TMC5160 evaluation kit (https://www.mouser.de/ProductDetail/ADI-Trinamic/TMC5160-EVAL-KIT?qs=TiOZkKH1s2TJebvRqsmGsQ%3D%3D).
Once you connect the eval-kit to the pc you will find something called "sinus tool" and you can play with it. This is what in the user manual is called "microstepping look-up table"... I'm not sure if with "phase modulation" you're referring to the same thing.

@dc42 Thinking in abstract terms, to completely describe and have full control of such a transition, the total amount of parameters might be: max speed change, min speed change (square corner speed), angle threshold (to determine at which angle the curve has to drop), and the slope of the curve.
Focusing on the parameters and not on the type of curve now, I agree that for simplicity two parameters are preferable, but if we use max speed change and min speed change, the user doesn't have control over the transition between the two speeds. I think the slope of the curve can be hard coded (with a parameter or by the type of the curve itself), but if I had to choose a second parameter, between a min speed change or an angle threshold, I would rather prefer to have control on the angle threshold and have the min speed change decided by the FW (it might also be the minJerk).

@dc42 thanks for the feedback!

@gloomyandy here it is. It's not much more complicate than the original DDA::MatchSpeed function. It can also be simplified by removing the rotation matrix definition and just compute the X coordinate of v2 in the reference frame of v1.
I also decided to implement it only for xy moves and keeping the possibility to disable it by setting the JerkDecay parameter to zero.

void DDA::MatchSpeeds() noexcept
{
	// v2 needs to be expressed in the reference frame of v1
	// directionVector and next->directionVector are unit vectors

	// The rotation matrix R(-angle) can be expressed by using the directionVector[0] and directionVector[1] which are the X and Y components of v1 in the absolute reference frame (v1_0).
	const float rotationMatrix[2][2] = {
		{ directionVector[0], directionVector[1] },
		{ -directionVector[1], directionVector[0] }
	};

	// Project v2 in the reference frame of v1 (v2_1)
	const float nextVectorProjected[2] = {
		rotationMatrix[0][0] * next->directionVector[0] + rotationMatrix[0][1] * next->directionVector[1],
		rotationMatrix[1][0] * next->directionVector[0] + rotationMatrix[1][1] * next->directionVector[1]
	};

	debugPrintf("JUNCTION: START targetNextSpeed [mm/s]: %f \n", (double)InverseConvertSpeedToMmPerSec(beforePrepare.targetNextSpeed));
	DebugPrintVector("JUNCTION: directionVector \n", directionVector, 2);
	DebugPrintVector("JUNCTION: next->directionVector ", next->directionVector, 2);
	DebugPrintVector("JUNCTION: nextVectorProjected ", nextVectorProjected, 2);

	for (size_t drive = 0; drive < MaxAxesPlusExtruders; ++drive)
	{
		if (directionVector[drive] != 0.0 || next->directionVector[drive] != 0.0)
		{
			const float totalFraction = fabsf(directionVector[drive] - next->directionVector[drive]);
			const float jerk = totalFraction * beforePrepare.targetNextSpeed;
			float allowedJerk = reprap.GetPlatform().GetInstantDv(drive);

			if (flags.xyMoving && flags.isPrintingMove && reprap.GetPlatform().JerkDecay() != 0)
			{
				const float A = (reprap.GetPlatform().GetInstantDv(drive) - ConvertSpeedFromMmPerSec(MinimumJerk));
				const float B = reprap.GetPlatform().JerkDecay();
				const float C = ConvertSpeedFromMmPerSec(MinimumJerk);

				allowedJerk = A / (1 + (float)exp(B * (-nextVectorProjected[0] + reprap.GetPlatform().JunctionDeviation()))) + C;
				debugPrintf("JUNCTION: allowedJerk for drive %i: %f \n", drive, (double)InverseConvertSpeedToMmPerSec(allowedJerk));
			}

			if (jerk > allowedJerk)
			{
				beforePrepare.targetNextSpeed = allowedJerk/totalFraction;
			}
		}
	}
	debugPrintf("JUNCTION: junctionSpeed [mm/s]: %f \n", (double)InverseConvertSpeedToMmPerSec(beforePrepare.targetNextSpeed));

}

@gloomyandy your interpretation is correct. However, I have a couple of parameters that allow to change the behaviour of the curve.
The function is:

allowedJerk = A / (1 + e^(B*(cosAngle + cosAngleThreshold))))+ C

where:

• A = maxJerk (defined by M566) - minJerk
• B = tunable parameter to define the slope of the decay from maxJerk to minJerk
• C = minJerk

Below the plot of the function with maxJerk = 600 mm/min, minJerk = 6 mm/min cosAngleThreshold = -0.707 (following the first picture, alpha=45deg) and increasing B.

In that configuration, vectors at 45deg will have an allowableJerk of 300mm/min, while for 90deg 6mm/min. I think having such a low jerk for sharp corners is not bad, also it allows the input shaping to work better since you have a bigger acceleration profile.

This approach gave me more flexibility, and with a proper setting of parameters you can have pretty much the same behaviour of the junction deviation. I can post the code of course if anyone is interested.

The weak point I see in both approaches (or working only inside DDA::MatchSpeeds) is when the round corners have a very fine mesh and the angle between two consecutive vectors is small. In that case, you may have the "illusion" of traveling a linear path while actually doing a circle and then end up going too fast.