This document describes Klipper benchmarks.

Micro-controller Benchmarks

This section describes the mechanism used to generate the Klipper micro-controller step rate benchmarks.

The primary goal of the benchmarks is to provide a consistent mechanism for measuring the impact of coding changes within the software. A secondary goal is to provide high-level metrics for comparing the performance between chips and between software platforms.

The step rate benchmark is designed to find the maximum stepping rate that the hardware and software can reach. This benchmark stepping rate is not achievable in day-to-day use as Klipper needs to perform other tasks (eg, mcu/host communication, temperature reading, endstop checking) in any real-world usage.

In general, the pins for the benchmark tests are chosen to flash LEDs or other innocuous pins. Always verify that it is safe to drive the configured pins prior to running a benchmark. It is not recommended to drive an actual stepper during a benchmark.

Step rate benchmark test

The test is performed using the console.py tool (described in Debugging.md). The micro-controller is configured for the particular hardware platform (see below) and then the following is cut-and-paste into the console.py terminal window:

SET start_clock {clock+freq}
SET ticks 1000

reset_step_clock oid=0 clock={start_clock}
set_next_step_dir oid=0 dir=0
queue_step oid=0 interval={ticks} count=60000 add=0
set_next_step_dir oid=0 dir=1
queue_step oid=0 interval=3000 count=1 add=0

reset_step_clock oid=1 clock={start_clock}
set_next_step_dir oid=1 dir=0
queue_step oid=1 interval={ticks} count=60000 add=0
set_next_step_dir oid=1 dir=1
queue_step oid=1 interval=3000 count=1 add=0

reset_step_clock oid=2 clock={start_clock}
set_next_step_dir oid=2 dir=0
queue_step oid=2 interval={ticks} count=60000 add=0
set_next_step_dir oid=2 dir=1
queue_step oid=2 interval=3000 count=1 add=0

The above tests three steppers simultaneously stepping. If running the above results in a “Rescheduled timer in the past” or “Stepper too far in past” error then it indicates the ticks parameter is too low (it results in a stepping rate that is too fast). The goal is to find the lowest setting of the ticks parameter that reliably results in a successful completion of the test. It should be possible to bisect the ticks parameter until a stable value is found.

On a failure, one can copy-and-paste the following to clear the error in preparation for the next test:

clear_shutdown

To obtain the single stepper and dual stepper benchmarks, the same configuration sequence is used, but only the first block (for the single stepper case) or first two blocks (for the dual stepper case) of the above test is cut-and-paste into the console.py window.

To produce the benchmarks found in the Features.md document, the total number of steps per second is calculated by multiplying the number of active steppers with the nominal mcu frequency and dividing by the final ticks parameter. The results are rounded to the nearest K. For example, with three active steppers:

ECHO Test result is: {"%.0fK" % (3. * freq / ticks / 1000.)}

AVR step rate benchmark

The following configuration sequence is used on AVR chips:

PINS arduino
allocate_oids count=3
config_stepper oid=0 step_pin=ar29 dir_pin=ar28 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=ar27 dir_pin=ar26 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=ar23 dir_pin=ar22 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit b161a69e with gcc version avr-gcc (GCC) 4.8.1. Both the 16Mhz and 20Mhz tests were run using simulavr configured for an atmega644p (previous tests have confirmed simulavr results match tests on both a 16Mhz at90usb and a 16Mhz atmega2560). On both 16Mhz and 20Mhz the best single stepper result is SET ticks 106, the best dual stepper result is SET ticks 276, and the best three stepper result is SET ticks 481.

Arduino Due step rate benchmark

The following configuration sequence is used on the Due:

allocate_oids count=3
config_stepper oid=0 step_pin=PB27 dir_pin=PA21 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PB26 dir_pin=PC30 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PA21 dir_pin=PC30 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 74c21654 with gcc version arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0. The best single stepper result is SET ticks 388, the best dual stepper result is SET ticks 405, and the best three stepper result is SET ticks 576.

Duet Maestro step rate benchmark

The following configuration sequence is used on the Duet Maestro:

allocate_oids count=3
config_stepper oid=0 step_pin=PC26 dir_pin=PC18 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PC26 dir_pin=PA8 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PC26 dir_pin=PB4 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 74c21654 with gcc version arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0. The best single stepper result is SET ticks 553, the best dual stepper result is SET ticks 563, and the best three stepper result is SET ticks 623.

Duet Wifi step rate benchmark

The following configuration sequence is used on the Duet Wifi:

allocate_oids count=4
config_stepper oid=0 step_pin=PD6 dir_pin=PD11 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PD7 dir_pin=PD12 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PD8 dir_pin=PD13 min_stop_interval=0 invert_step=0
config_stepper oid=3 step_pin=PD5 dir_pin=PA1 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 59a60d68 with gcc version arm-none-eabi-gcc 7.3.1 20180622 (release) [ARM/embedded-7-branch revision 261907]. The best single stepper result is SET ticks 519, the best dual stepper result is SET ticks 520, and the best three stepper result is SET ticks 525, and the best four stepper result is SET ticks 703.

Beaglebone PRU step rate benchmark

The following configuration sequence is used on the PRU:

PINS beaglebone
allocate_oids count=3
config_stepper oid=0 step_pin=P8_13 dir_pin=P8_12 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=P8_15 dir_pin=P8_14 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=P8_19 dir_pin=P8_18 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit b161a69e with gcc version pru-gcc (GCC) 8.0.0 20170530 (experimental). The best single stepper result is SET ticks 861, the best dual stepper result is SET ticks 853, and the best three stepper result is SET ticks 883.

STM32F042 step rate benchmark

The following configuration sequence is used on the STM32F042:

allocate_oids count=3
config_stepper oid=0 step_pin=PA0 dir_pin=PA1 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PA2 dir_pin=PA3 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PA6 dir_pin=PA7 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit c105adc8 with gcc version arm-none-eabi-gcc (GNU Tools 7-2018-q3-update) 7.3.1. The best single stepper result is SET ticks 308, the best dual stepper result is SET ticks 638, and the best three stepper result is SET ticks 1021.

STM32F103 step rate benchmark

The following configuration sequence is used on the STM32F103:

allocate_oids count=3
config_stepper oid=0 step_pin=PC13 dir_pin=PB5 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PB3 dir_pin=PB6 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PA4 dir_pin=PB7 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 9f3517fd with gcc version arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0. The best single stepper result is SET ticks 345, the best dual stepper result is SET ticks 365, and the best three stepper result is SET ticks 606.

STM32F446 step rate benchmark

The following configuration sequence is used on the STM32F446:

allocate_oids count=4
config_stepper oid=0 step_pin=PA5 dir_pin=PB5 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PB2 dir_pin=PB6 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PB3 dir_pin=PB7 min_stop_interval=0 invert_step=0
config_stepper oid=3 step_pin=PB3 dir_pin=PB8 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 1487f8a2 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0. The best single stepper result is SET ticks 757, the best dual stepper result is SET ticks 761, the best three stepper result is SET ticks 757, and the best four stepper result is SET ticks 767.

LPC176x step rate benchmark

The following configuration sequence is used on the LPC176x:

allocate_oids count=3
config_stepper oid=0 step_pin=P1.20 dir_pin=P1.18 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=P1.21 dir_pin=P1.18 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=P1.23 dir_pin=P1.18 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 9f3517fd with gcc version arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0. For the 100Mhz LPC1768, the best single stepper result is SET ticks 448, the best dual stepper result is SET ticks 450, and the best three stepper result is SET ticks 523. The 120Mhz LPC1769 results were obtained by overclocking an LPC1768 to 120Mhz - the best single stepper result is SET ticks 525, the best dual stepper result is SET ticks 526, and the best three stepper result is SET ticks 545.

SAMD21 step rate benchmark

The following configuration sequence is used on the SAMD21:

allocate_oids count=3
config_stepper oid=0 step_pin=PA27 dir_pin=PA20 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PB3 dir_pin=PA21 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PA17 dir_pin=PA21 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 9f3517fd with gcc version arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0. The best single stepper result is SET ticks 277, the best dual stepper result is SET ticks 410, and the best three stepper result is SET ticks 664.

SAMD51 step rate benchmark

The following configuration sequence is used on the SAMD51:

allocate_oids count=4
config_stepper oid=0 step_pin=PA22 dir_pin=PA20 min_stop_interval=0 invert_step=0
config_stepper oid=1 step_pin=PA22 dir_pin=PA21 min_stop_interval=0 invert_step=0
config_stepper oid=2 step_pin=PA22 dir_pin=PA19 min_stop_interval=0 invert_step=0
config_stepper oid=3 step_pin=PA22 dir_pin=PA18 min_stop_interval=0 invert_step=0
finalize_config crc=0

The test was last run on commit 9f3517fd with gcc version arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0 on a SAMD51G19A micro-controller. The best single stepper result is SET ticks 516, the best dual stepper result is SET ticks 520, the best three stepper result is SET ticks 519, and the best four stepper result is SET ticks 655.

Command dispatch benchmark

The command dispatch benchmark tests how many “dummy” commands the micro-controller can process. It is primarily a test of the hardware communication mechanism. The test is run using the console.py tool (described in Debugging.md). The following is cut-and-paste into the console.py terminal window:

DELAY {clock + 2*freq} get_uptime
FLOOD 100000 0.0 end_group
get_uptime

When the test completes, determine the difference between the clocks reported in the two “uptime” response messages. The total number of commands per second is then 100000 * mcu_frequency / clock_diff.

Note that this test may saturate the USB/CPU capacity of a Raspberry Pi. The benchmarks below are with console.py running on a desktop class machine.

MCU Rate Build Build compiler
pru (shared memory) 5K b161a69e pru-gcc (GCC) 8.0.0 20170530 (experimental)
stm32f042 (CAN) 18K c105adc8 arm-none-eabi-gcc (GNU Tools 7-2018-q3-update) 7.3.1
atmega2560 (serial) 23K b161a69e avr-gcc (GCC) 4.8.1
sam3x8e (serial) 23K b161a69e arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
at90usb1286 (USB) 75K b161a69e avr-gcc (GCC) 4.8.1
samd21 (USB) 238K b161a69e arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
stm32f103 (USB) 335K b161a69e arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
sam3x8e (USB) 450K a5aede52 arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
lpc1768 (USB) 546K b161a69e arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
sam4s8c (USB) 619K a5aede52 arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
lpc1769 (USB) 619K b161a69e arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
samd51 (USB) 620K 8cd83b4c arm-none-eabi-gcc (Fedora 7.1.0-5.fc27) 7.1.0
stm32f446 (USB) 621K aac51bdb arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0

Host Benchmarks

It is possible to run timing tests on the host software using the “batch mode” processing mechanism (described in Debugging.md). This is typically done by choosing a large and complex G-Code file and timing how long it takes for the host software to process it. For example:

time ~/klippy-env/bin/python ./klippy/klippy.py config/example.cfg -i something_complex.gcode -o /dev/null -d out/klipper.dict