Directory Layout

The src/ directory contains the C source for the micro-controller code. The src/avr/ directory contains specific code for Atmel ATmega micro-controllers. The src/sam3x8e/ directory contains code specific to the Arduino Due style ARM micro-controllers. The src/pru/ directory contains code specific to the Beaglebone's on-board PRU micro-controller. The src/simulator/ contains code stubs that allow the micro-controller to be test compiled on other architectures. The src/generic/ directory contains helper code that may be useful across different host architectures. The build arranges for includes of "board/somefile.h" to first look in the current architecture directory (eg, src/avr/somefile.h) and then in the generic directory (eg, src/generic/somefile.h).

The klippy/ directory contains the C and Python source for the host part of the software.

The lib/ directory contains external 3rd-party library code that is necessary to build some targets.

The config/ directory contains example printer configuration files.

The scripts/ directory contains build-time scripts useful for compiling the micro-controller code.

During compilation, the build may create an out/ directory. This contains temporary build time objects. The final micro-controller object that is built is out/klipper.elf.hex on AVR and out/klipper.bin on ARM.

Micro-controller code flow

Execution of the micro-controller code starts in architecture specific code (eg, src/avr/main.c) which ultimately calls sched_main() located in src/sched.c. The sched_main() code starts by running all functions that have been tagged with the DECL_INIT() macro. It then goes on to repeatedly run all functions tagged with the DECL_TASK() macro.

One of the main task functions is command_dispatch() located in src/command.c. This function is called from the board specific input/output code (eg, src/avr/serial.c) and it runs the command functions associated with the commands found in the input stream. Command functions are declared using the DECL_COMMAND() macro (see the protocol document for more information).

Task, init, and command functions always run with interrupts enabled (however, they can temporarily disable interrupts if needed). These functions should never pause, delay, or do any work that lasts more than a few micro-seconds. These functions schedule work at specific times by scheduling timers.

Timer functions are scheduled by calling sched_add_timer() (located in src/sched.c). The scheduler code will arrange for the given function to be called at the requested clock time. Timer interrupts are initially handled in an architecture specific interrupt handler (eg, src/avr/timer.c) which calls sched_timer_dispatch() located in src/sched.c. The timer interrupt leads to execution of schedule timer functions. Timer functions always run with interrupts disabled. The timer functions should always complete within a few micro-seconds. At completion of the timer event, the function may choose to reschedule itself.

In the event an error is detected the code can invoke shutdown() (a macro which calls sched_shutdown() located in src/sched.c). Invoking shutdown() causes all functions tagged with the DECL_SHUTDOWN() macro to be run. Shutdown functions always run with interrupts disabled.

Much of the functionality of the micro-controller involves working with General-Purpose Input/Output pins (GPIO). In order to abstract the low-level architecture specific code from the high-level task code, all GPIO events are implemented in architectures specific wrappers (eg, src/avr/gpio.c). The code is compiled with gcc's "-flto -fwhole-program" optimization which does an excellent job of inlining functions across compilation units, so most of these tiny gpio functions are inlined into their callers, and there is no run-time cost to using them.

Klippy code overview

The host code (Klippy) is intended to run on a low-cost computer (such as a Raspberry Pi) paired with the micro-controller. The code is primarily written in Python, however it does use CFFI to implement some functionality in C code.

Initial execution starts in klippy/klippy.py. This reads the command-line arguments, opens the printer config file, instantiates the main printer objects, and starts the serial connection. The main execution of G-code commands is in the process_commands() method in klippy/gcode.py. This code translates the G-code commands into printer object calls, which frequently translate the actions to commands to be executed on the micro-controller (as declared via the DECL_COMMAND macro in the micro-controller code).

There are four threads in the Klippy host code. The main thread handles incoming gcode commands. A second thread (which resides entirely in the klippy/serialqueue.c C code) handles low-level IO with the serial port. The third thread is used to process response messages from the micro-controller in the Python code (see klippy/serialhdl.py). The fourth thread writes debug messages to the log (see klippy/queuelogger.py) so that the other threads never block on log writes.

Code flow of a move command

A typical printer movement starts when a "G1" command is sent to the Klippy host and it completes when the corresponding step pulses are produced on the micro-controller. This section outlines the code flow of a typical move command. The kinematics document provides further information on the mechanics of moves.

Processing for a move command starts in gcode.py. The goal of gcode.py is to translate G-code into internal calls. Changes in origin (eg, G92), changes in relative vs absolute positions (eg, G90), and unit changes (eg, F6000=100mm/s) are handled here. The code path for a move is: process_data() -> process_commands() -> cmd_G1(). Ultimately the ToolHead class is invoked to execute the actual request: cmd_G1() -> ToolHead.move()
The ToolHead class (in toolhead.py) handles "look-ahead" and tracks the timing of printing actions. The codepath for a move is: ToolHead.move() -> MoveQueue.add_move() -> MoveQueue.flush() -> Move.set_junction() -> Move.move().
- ToolHead.move() creates a Move() object with the parameters of the move (in cartesian space and in units of seconds and millimeters).
- MoveQueue.add_move() places the move object on the "look-ahead" queue.
- MoveQueue.flush() determines the start and end velocities of each move.
- Move.set_junction() implements the "trapezoid generator" on a move. The "trapezoid generator" breaks every move into three parts: a constant acceleration phase, followed by a constant velocity phase, followed by a constant deceleration phase. Every move contains these three phases in this order, but some phases may be of zero duration.
- When Move.move() is called, everything about the move is known - its start location, its end location, its acceleration, its start/crusing/end velocity, and distance traveled during acceleration/cruising/deceleration. All the information is stored in the Move() class and is in cartesian space in units of millimeters and seconds.
The move is then handed off to the kinematics classes: Move.move() -> kin.move()
The goal of the kinematics classes is to translate the movement in cartesian space to movement on each stepper. The kinematics classes are in cartesian.py, corexy.py, delta.py, and extruder.py. The kinematic class is given a chance to audit the move (ToolHead.move() -> kin.check_move()) before it goes on the look-ahead queue, but once the move arrives in kin.move() the kinematic class is required to handle the move as specified. The kinematic classes translate the three parts of each move (acceleration, constant "cruising" velocity, and deceleration) to the associated movement on each stepper. Note that the extruder is handled in its own kinematic class. Since the Move() class specifies the exact movement time and since step pulses are sent to the micro-controller with specific timing, stepper movements produced by the extruder class will be in sync with head movement even though the code is kept separate.
For efficiency reasons, the stepper pulse times are generated in C code. The code flow is: kin.move() -> MCU_Stepper.step_const() -> stepcompress_push_const(), or for delta kinematics: DeltaKinematics.move() -> MCU_Stepper.step_delta() -> stepcompress_push_delta(). The MCU_Stepper code just performs unit and axis transformation (millimeters to step distances), and calls the C code. The C code calculates the stepper step times for each movement and fills an array (struct stepcompress.queue) with the corresponding micro-controller clock counter times for every step. Here the "micro-controller clock counter" value directly corresponds to the micro-controller's hardware counter - it is relative to when the micro-controller was last powered up.
The next major step is to compress the steps: stepcompress_flush() -> compress_bisect_add() (in stepcompress.c). This code generates and encodes a series of micro-controller "queue_step" commands that correspond to the list of stepper step times built in the previous stage. These "queue_step" commands are then queued, prioritized, and sent to the micro-controller (via stepcompress.c:steppersync and serialqueue.c:serialqueue).
Processing of the queue_step commands on the micro-controller starts in command.c which parses the command and calls command_queue_step(). The command_queue_step() code (in stepper.c) just appends the parameters of each queue_step command to a per stepper queue. Under normal operation the queue_step command is parsed and queued at least 100ms before the time of its first step. Finally, the generation of stepper events is done in stepper_event(). It's called from the hardware timer interrupt at the scheduled time of the first step. The stepper_event() code generates a step pulse and then reschedules itself to run at the time of the next step pulse for the given queue_step parameters. The parameters for each queue_step command are "interval", "count", and "add". At a high-level, stepper_event() runs the following, 'count' times: do_step(); next_wake_time = last_wake_time + interval; interval += add;

The above may seem like a lot of complexity to execute a movement. However, the only really interesting parts are in the ToolHead and kinematic classes. It's this part of the code which specifies the movements and their timings. The remaining parts of the processing is mostly just communication and plumbing.

Adding new kinematics

This section provides some tips on adding support to Klipper for additional types of printer kinematics. This type of activity requires excellent understanding of the math formulas for the target kinematics. It also requires software development skills - though one should only need to update the host software (which is written in Python).

Useful steps:

Start by studying the above section and the Kinematics document.
Review the existing kinematic classes in cartesian.py, corexy.py, and delta.py. The kinematic classes are tasked with converting a move in cartesian coordinates to the movement on each stepper. One should be able to copy one of these files as a starting point.
Implement the get_postion() method in the new kinematics class. This method converts the current stepper position of each stepper axis (stored in millimeters) to a position in cartesian space (also in millimeters).
Implement the set_postion() method. This is the inverse of get_position() - it sets each axis position (in millimeters) given a position in cartesian coordinates.
Implement the move() method. The goal of the move() method is to convert a move defined in cartesian space to a series of stepper step times that implement the requested movement.
- The move() method is passed a "print_time" parameter (which stores a time in seconds) and a "move" class instance that fully defines the movement. The goal is to repeatedly invoke the stepper.step() method with the time (relative to print_time) that each stepper should step at to obtain the desired motion.
- One "trick" to help with the movement calculations is to imagine there is a physical rail between move.start_pos and move.end_pos that confines the print head so that it can only move along this straight line of motion. Then, if the head is confined to that imaginary rail, the head is at move.start_pos, only one stepper is enabled (all other steppers can move freely), and the given stepper is stepped a single step, then one can imagine that the head will move along the line of movement some distance. Determine the formula converting this step distance to distance along the line of movement. Once one has the distance along the line of movement, one can figure out the time that the head should be at that position (using the standard formulas for velocity and acceleration). This time is the ideal step time for the given stepper and it can be passed to the stepper.step() method.
- The stepper.step() method must always be called with an increasing time for a given stepper (steps must be scheduled in the order they are to be executed). A common error during kinematic development is to receive an "Internal error in stepcompress" failure - this is generally due to the step() method being invoked with a time earlier than the last scheduled step. For example, if the last step in move1 is scheduled at a time greater than the first step in move2 it will generally result in the above error.
- Fractional steps. Be aware that a move request is given in cartesian space and it is not confined to discreet locations. Thus a move's start and end locations may translate to a location on a stepper axis that is between two steps (a fractional step). The code must handle this. The preferred approach is to schedule the next step at the time a move would position the stepper axis at least half way towards the next possible step location. Incorrect handling of fractional steps is a common cause of "Internal error in stepcompress" failures.
Other methods. The home(), check_move(), and other methods should also be implemented. However, at the start of development one can use empty code here.
Implement test cases. Create a g-code file with a series of moves that can test important cases for the given kinematics. Follow the debugging documentation to convert this g-code file to micro-controller commands. This is useful to exercise corner cases and to check for regressions.
Optimize if needed. One may notice that the existing kinematic classes do not call stepper.step(). This is purely an optimization - the inner loop of the kinematic calculations were moved to C to reduce load on the host cpu. All of the existing kinematic classes started development using stepper.step() and then were later optimized. The g-code to mcu command translation (described in the previous step) is a useful tool during optimization - if a code change is purely an optimization then it should not impact the resulting text representation of the mcu commands (though minor changes in output due to floating point rounding are possible). So, one can use this system to detect regressions.

Time

Fundamental to the operation of Klipper is the handling of clocks, times, and timestamps. Klipper executes actions on the printer by scheduling events to occur in the near future. For example, to turn on a fan, the code might schedule a change to a GPIO pin in a 100ms. It is rare for the code to attempt to take an instantaneous action. Thus, the handling of time within Klipper is critical to correct operation.

There are three types of times tracked internally in the Klipper host software:

System time. The system time uses the system's monotonic clock - it is a floating point number stored as seconds and it is (generally) relative to when the host computer was last started. System times have limited use in the software - they are primarily used when interacting with the operating system. Within the host code, system times are frequently stored in variables named eventtime or curtime.
Print time. The print time is synchronized to the main micro-controller clock (the micro-controller defined in the "[mcu]" config section). It is a floating point number stored as seconds and is relative to when the main mcu was last restarted. It is possible to convert from a "print time" to the main micro-controller's hardware clock by multiplying the print time by the mcu's statically configured frequency rate. The high-level host code uses print times to calculates almost all physical actions (eg, head movement, heater changes, etc.). Within the host code, print times are generally stored in variables named print_time or move_time.
MCU clock. This is the hardware clock counter on each micro-controller. It is stored as an integer and its update rate is relative to the frequency of the given micro-controller. The host software translates its internal times to clocks before transmission to the mcu. The mcu code only ever tracks time in clock ticks. Within the host code, clock values are tracked as 64bit integers, while the mcu code uses 32bit integers. Within the host code, clocks are generally stored in variables with names containing clock or ticks.

Conversion between the different time formats is primarily implemented in the klippy/clocksync.py code.

Some things to be aware of when reviewing the code:

32bit and 64bit clocks: To reduce bandwidth and to improve micro-controller efficiency, clocks on the micro-controller are tracked as 32bit integers. When comparing two clocks in the mcu code, the timer_is_before() function must always be used to ensure integer rollovers are handled properly. The host software converts 32bit clocks to 64bit clocks by appending the high-order bits from the last mcu timestamp it has received - no message from the mcu is ever more than 2^31 clock ticks in the future or past so this conversion is never ambiguous. The host converts from 64bit clocks to 32bit clocks by simply truncating the high-order bits. To ensure there is no ambiguity in this conversion, the klippy/serialqueue.c code will buffer messages until they are within 2^31 clock ticks of their target time.
Multiple micro-controllers: The host software supports using multiple micro-controllers on a single printer. In this case, the "MCU clock" of each micro-controller is tracked separately. The clocksync.py code handles clock drift between micro-controllers by modifying the way it converts from "print time" to "MCU clock". On secondary mcus, the mcu frequency that is used in this conversion is regularly updated to account for measured drift.

19 KiB Raw Blame History