klipper/docs/Benchmarks.md

16 KiB

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.)}

Benchmarks may be run with the micro-controller code compiled using a "step pulse duration" of zero (the tables below report this as "no delay"). This configuration is believed to be valid in real-world usage when one is solely using Trinamic stepper drivers. The results of these benchmarks are not reported in the Features.md document.

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 invert_step=0
config_stepper oid=1 step_pin=ar27 dir_pin=ar26 invert_step=0
config_stepper oid=2 step_pin=ar23 dir_pin=ar22 invert_step=0
finalize_config crc=0

The test was last run on commit 01d2183f with gcc version avr-gcc (GCC) 5.4.0. 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).

avr ticks
1 stepper 104
2 stepper 296
3 stepper 472

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 invert_step=0
config_stepper oid=1 step_pin=PB26 dir_pin=PC30 invert_step=0
config_stepper oid=2 step_pin=PA21 dir_pin=PC30 invert_step=0
finalize_config crc=0

The test was last run on commit 8d4a5c16 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0.

sam3x8e ticks
1 stepper 388
2 stepper 405
3 stepper 576
1 stepper (no delay) 77
3 stepper (no delay) 299

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 invert_step=0
config_stepper oid=1 step_pin=PC26 dir_pin=PA8 invert_step=0
config_stepper oid=2 step_pin=PC26 dir_pin=PB4 invert_step=0
finalize_config crc=0

The test was last run on commit 8d4a5c16 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0.

sam4s8c ticks
1 stepper 527
2 stepper 535
3 stepper 638
1 stepper (no delay) 70
3 stepper (no delay) 254

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 invert_step=0
config_stepper oid=1 step_pin=PD7 dir_pin=PD12 invert_step=0
config_stepper oid=2 step_pin=PD8 dir_pin=PD13 invert_step=0
config_stepper oid=3 step_pin=PD5 dir_pin=PA1 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].

sam4e8e ticks
1 stepper 519
2 stepper 520
3 stepper 525
4 stepper 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 invert_step=0
config_stepper oid=1 step_pin=P8_15 dir_pin=P8_14 invert_step=0
config_stepper oid=2 step_pin=P8_19 dir_pin=P8_18 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).

pru ticks
1 stepper 861
2 stepper 853
3 stepper 883

STM32F042 step rate benchmark

The following configuration sequence is used on the STM32F042:

allocate_oids count=3
config_stepper oid=0 step_pin=PA1 dir_pin=PA2 invert_step=0
config_stepper oid=1 step_pin=PA3 dir_pin=PA2 invert_step=0
config_stepper oid=2 step_pin=PB8 dir_pin=PA2 invert_step=0
finalize_config crc=0

The test was last run on commit 0b0c47c5 with gcc version arm-none-eabi-gcc (Fedora 9.2.0-1.fc30) 9.2.0.

stm32f042 ticks
1 stepper 247
2 stepper 328
3 stepper 558

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 invert_step=0
config_stepper oid=1 step_pin=PB3 dir_pin=PB6 invert_step=0
config_stepper oid=2 step_pin=PA4 dir_pin=PB7 invert_step=0
finalize_config crc=0

The test was last run on commit 8d4a5c16 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0.

stm32f103 ticks
1 stepper 347
2 stepper 372
3 stepper 600
1 stepper (no delay) 71
3 stepper (no delay) 288

STM32F4 step rate benchmark

The following configuration sequence is used on the STM32F4:

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

The test was last run on commit 8d4a5c16 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0. The STM32F407 results were obtained by running an STM32F407 binary on an STM32F446 (and thus using a 168Mhz clock).

stm32f446 ticks
1 stepper 757
2 stepper 761
3 stepper 757
4 stepper 767
1 stepper (no delay) 51
3 stepper (no delay) 226
stm32f407 ticks
1 stepper 709
2 stepper 714
3 stepper 709
4 stepper 729
1 stepper (no delay) 52
3 stepper (no delay) 226

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 invert_step=0
config_stepper oid=1 step_pin=P1.21 dir_pin=P1.18 invert_step=0
config_stepper oid=2 step_pin=P1.23 dir_pin=P1.18 invert_step=0
finalize_config crc=0

The test was last run on commit 8d4a5c16 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0. The 120Mhz LPC1769 results were obtained by overclocking an LPC1768 to 120Mhz.

lpc1768 ticks
1 stepper 448
2 stepper 450
3 stepper 523
1 stepper (no delay) 56
3 stepper (no delay) 240
lpc1769 ticks
1 stepper 525
2 stepper 526
3 stepper 545
1 stepper (no delay) 56
3 stepper (no delay) 240

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 invert_step=0
config_stepper oid=1 step_pin=PB3 dir_pin=PA21 invert_step=0
config_stepper oid=2 step_pin=PA17 dir_pin=PA21 invert_step=0
finalize_config crc=0

The test was last run on commit 8d4a5c16 with gcc version arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0 on a SAMD21G18 micro-controller.

samd21 ticks
1 stepper 277
2 stepper 410
3 stepper 664
1 stepper (no delay) 83
3 stepper (no delay) 321

SAMD51 step rate benchmark

The following configuration sequence is used on the SAMD51:

allocate_oids count=5
config_stepper oid=0 step_pin=PA22 dir_pin=PA20 invert_step=0
config_stepper oid=1 step_pin=PA22 dir_pin=PA21 invert_step=0
config_stepper oid=2 step_pin=PA22 dir_pin=PA19 invert_step=0
config_stepper oid=3 step_pin=PA22 dir_pin=PA18 invert_step=0
config_stepper oid=4 step_pin=PA23 dir_pin=PA17 invert_step=0
finalize_config crc=0

The test was last run on commit 524ebbc7 with gcc version arm-none-eabi-gcc (Fedora 9.2.0-1.fc30) 9.2.0 on a SAMD51J19A micro-controller.

samd51 ticks
1 stepper 516
2 stepper 520
3 stepper 520
4 stepper 631
1 stepper (200Mhz) 839
2 stepper (200Mhz) 838
3 stepper (200Mhz) 838
4 stepper (200Mhz) 838
5 stepper (200Mhz) 891
1 stepper (no delay) 42
3 stepper (no delay) 194

RP2040 step rate benchmark

The following configuration sequence is used on the RP2040:

allocate_oids count=4
config_stepper oid=0 step_pin=gpio25 dir_pin=gpio3 invert_step=0
config_stepper oid=1 step_pin=gpio26 dir_pin=gpio4 invert_step=0
config_stepper oid=2 step_pin=gpio27 dir_pin=gpio5 invert_step=0
config_stepper oid=3 step_pin=gpio28 dir_pin=gpio6 invert_step=0
finalize_config crc=0

The test was last run on commit c5667193 with gcc version arm-none-eabi-gcc (Fedora 10.2.0-4.fc34) 10.2.0 on a Raspberry Pi Pico board.

rp2040 ticks
1 stepper 52
2 stepper 52
3 stepper 52
4 stepper 66
1 stepper (no delay) 5
3 stepper (no delay) 22

Linux MCU step rate benchmark

The following configuration sequence is used on a Raspberry Pi:

allocate_oids count=3
config_stepper oid=0 step_pin=gpio2 dir_pin=gpio3 invert_step=0
config_stepper oid=1 step_pin=gpio4 dir_pin=gpio5 invert_step=0
config_stepper oid=2 step_pin=gpio6 dir_pin=gpio7 invert_step=0
finalize_config crc=0

The test was last run on commit db0fb5d5 with gcc version gcc (Raspbian 6.3.0-18+rpi1+deb9u1) 6.3.0 20170516 on a Raspberry Pi 3 (revision a22082).

Linux (RPi3) ticks
1 stepper 349
2 stepper 350
3 stepper 400

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 debug_nop
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. If running on a Raspberry Pi, Beaglebone, or similar host computer then increase the delay (eg, DELAY {clock + 20*freq} get_uptime). Where applicable, the benchmarks below are with console.py running on a desktop class machine with the device connected via a high-speed hub.

MCU Rate Build Build compiler
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 01d2183f avr-gcc (GCC) 5.4.0
samd21 (USB) 223K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
pru (shared memory) 260K c5968a08 pru-gcc (GCC) 8.0.0 20170530 (experimental)
stm32f103 (USB) 355K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
sam3x8e (USB) 418K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
lpc1768 (USB) 534K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
lpc1769 (USB) 628K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
sam4s8c (USB) 650K 8d4a5c16 arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
samd51 (USB) 864K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
stm32f446 (USB) 870K 01d2183f arm-none-eabi-gcc (Fedora 7.4.0-1.fc30) 7.4.0
rp2040 (USB) 873K c5667193 arm-none-eabi-gcc (Fedora 10.2.0-4.fc34) 10.2.0

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-cartesian.cfg -i something_complex.gcode -o /dev/null -d out/klipper.dict