Summary
This RFC obsoletes RFC [0].
This RFC depends on RFC [1].
This pre-RFC aims to discuss the details and collect inputs about how TVMC (TVM CLI tool) can be extended to support microTVM targets, considering the variety of platforms supported by microTVM, like Zephyr and Arduino, and also considering future platforms, taking into account the use of custom/adhoc platforms provided by the users at their convenience.
The interface here proposed relies on the new Project API available on microTVM.
PR 9229 [2] implements the interface in question.
[0] [RFC] TVMC: Add support for ”TVM â [RFC] TVMC: Add support for ”TVM
[1] [RFC][Project API] Extend metadata in ProjectOption â https://github.com/apache/tvm-rfcs/pull/33
[2] [TVMC] Add new micro context â https://github.com/apache/tvm/pull/9229
Motivation
Currently if a microTVM user wants to compile, build, flash, and run a model for a micro target available on a supported microTVM platform â like Zephyr and Arduino â itâs necessary to drive the process end-to-end via Python scripts. This is ok when creating tests and developing or experimenting with new microTVM code, for example, but for some users that just want to try a new model of interest on a micro target that way can be slow.
TVMC is a command-line driver for TVM (tvmc
) that allows users to control the compilation and execution of models on a plethora of targets, using only the CLI, however it currently does not support microTVM targets, so it can not drive a similar workflow (compile, run, etc) on microcontrollers.
This RFC proposes to extend TVMC to support the microTVM targets and platforms so one can easily use tvmc
to compile, build (a device image with the compiled model of interest), flash, and run a model on a micro target using only a command-line interface.
Guide-level explanation
Micro targets have unique requirements which justifies the creation of a new context in tvmc
called micro
to accommodate them.
For instance, a given micro target can exist in more than one supported platform, like the STM32F746 MCU that can be supported by Zephyr
and by FreeRTOS
. Hence a platform must be selected in addition to the MCU and/or the board.
Moreover, the build and flash steps, although might not be totally exclusive on micro targets, are quite specific (like using the toolchain and the flash tool supported by the platform), so much it justifies them to exist inside the new micro context, i.e. under tvmc micro
.
In that sense, the following subcommands are available under tvmc micro
:
1. create-project
2. build
3. flash
create-project
purpose is for creating a project directory based on (a) an already compiled model kept in a Model Library Format (MLF) archive and (b) a template associated to a platform, like Zephyr, allowing one to specify several options specific to the selected platform, like the board and extra files to be included in the project directory to will be created based on the template directory. For convenience, an alias (create
) will exist for this subcommand.
build
can be used for building a firmware image ready to be flashed to the board specified. For example, if Zephyr platform is selected build a zephyr.{elf,hex} image. Pertinent options for âbuildâ given the platform selected as an option under âbuildâ will become available.
flash
can be used for effectively flashing, to the selected board, the firmware built. Again, pertinent options for âflashâ context considering the platform selected will be made available for the user, like for example options regarding the serial port number to be used when flashing the image to the board.
build
and flash
subcommads are intended to help with debugging inference flows built on top of Project API infrastructure. Hence theyâre not intended to replace the userâs workflow (e.g. users can still build and flash independently of these subcommands).
For all these subcommads (1, 2, and 3) the platform template can be selected among the default platforms (zephyr
and arduino
) or can be a custom one, in that case the platform template must be selected instead. Upon selecting the template platform option -d TEMPLATE_DIR
will become available and one can so specify the adhoc template dir (which must contain an adhoc Project API server as per the Project API specification).
A key-value pair is used to specify the options specific to the platforms. If the option is of type âboolâ the values available are âtrueâ and âfalseâ. Details on project option types are described in RFC-0020.
An example for that interface (using Zephyr) follows:
-
create-project
subcommand:
$ tvmc micro
usage: tvmc micro [-h] {create-project,build,flash} ...
tvmc micro: error: the following arguments are required: subcommand
$ tvmc micro -h
usage: tvmc micro [-h] {create-project,build,flash} ...
optional arguments:
-h, --help show this help message and exit
subcommands:
{create-project,build,flash}
create-project create a project template of a given type or given a template dir.
build build a project dir, generally creating an image to be flashed, e.g. zephyr.elf.
flash flash the built image on a given micro target.
$ tvmc micro create-project
usage: tvmc micro create-project [-h] [-f] PROJECT_DIR MLF {zephyr,arduino,template} ...
tvmc micro create-project: error: the following arguments are required: PROJECT_DIR, MLF, platform
$ tvmc micro create-project /tmp/x100 ./sine_model.tar zephyr
usage: tvmc micro create-project PROJECT_DIR MLF zephyr [--list-options] -o OPTION=VALUE [OPTION=VALUE ...]
tvmc micro create-project PROJECT_DIR MLF zephyr: error: the following arguments are required: -o
$ tvmc micro create-project /tmp/x100 ./sine_model.tar zephyr --list-options
usage: tvmc micro create-project PROJECT_DIR MLF zephyr [--list-options] -o OPTION=VALUE [OPTION=VALUE ...]
optional arguments:
--list-options show all options/values for selected platforms/template.
-o OPTION=VALUE [OPTION=VALUE ...]
extra_files_tar=EXTRA_FILES_TAR
if given, during generate_project, uncompress the tarball at this path into the project dir.
project_type={aot_demo, host_driven}
type of project to generate. (required)
west_cmd=WEST_CMD
path to the west tool. If given, supersedes both the zephyr_base option and ZEPHYR_BASE environment variable.
zephyr_board={mimxrt1050_evk, mps2_an521, nrf5340dk_nrf5340_cpuapp,
nucleo_f746zg, nucleo_l4r5zi, qemu_cortex_r5, qemu_riscv32,
qemu_riscv64, qemu_x86, stm32f746g_disco}
name of the Zephyr board to build for. (required)
config_main_stack_size=CONFIG_MAIN_STACK_SIZE
sets CONFIG_MAIN_STACK_SIZE for Zephyr board.
$ tvmc micro create-project /tmp/x200 ./sine_model.tar zephyr -o zephyr_board=stm32f746g_disco project_type=host_driven
$
-
build
subcommand:
$ tvmc micro build
usage: tvmc micro build [-h] [-f] PROJECT_DIR {zephyr,arduino,template} ...
tvmc micro build: error: the following arguments are required: PROJECT_DIR, platform
$ tvmc micro build -h
usage: tvmc micro build [-h] [-f] PROJECT_DIR {zephyr,arduino,template} ...
positional arguments:
PROJECT_DIR Project dir to build.
optional arguments:
-h, --help show this help message and exit
-f, --force Force rebuild.
platforms:
{zephyr,arduino,template}
you must selected a platform from the list. You can pass '-h' for a selected platform to list its options.
zephyr select Zephyr platform.
arduino select Arduino platform.
template select an adhoc template.
$ tvmc micro build /tmp/x200 zephyr
usage: tvmc micro build PROJECT_DIR zephyr [--list-options] -o OPTION=VALUE [OPTION=VALUE ...]
tvmc micro build PROJECT_DIR zephyr: error: the following arguments are required: -o
$ tvmc micro build /tmp/x200 zephyr --list-options
usage: tvmc micro build PROJECT_DIR zephyr [--list-options] -o OPTION=VALUE [OPTION=VALUE ...]
optional arguments:
--list-options show all options/values for selected platforms/template.
-o OPTION=VALUE [OPTION=VALUE ...]
verbose={true, false}
run build with verbose output.
zephyr_base=ZEPHYR_BASE
path to the zephyr base directory.
zephyr_board={mimxrt1050_evk, mps2_an521, nrf5340dk_nrf5340_cpuapp,
nucleo_f746zg, nucleo_l4r5zi, qemu_cortex_r5, qemu_riscv32,
qemu_riscv64, qemu_x86, stm32f746g_disco}
name of the Zephyr board to build for. (required)
$ tvmc micro build /tmp/x200 zephyr -o zephyr_board=stm32f746g_disco
$
-
flash
subcommand:
$ tvmc micro flash
usage: tvmc micro flash [-h] PROJECT_DIR {zephyr,arduino,template} ...
tvmc micro flash: error: the following arguments are required: PROJECT_DIR, platform
$ tvmc micro flash -h
usage: tvmc micro flash [-h] PROJECT_DIR {zephyr,arduino,template} ...
positional arguments:
PROJECT_DIR Project dir with a image built.
optional arguments:
-h, --help show this help message and exit
platforms:
{zephyr,arduino,template}
you must selected a platform from the list. You can pass '-h' for a selected platform to list its options.
zephyr select Zephyr platform.
arduino select Arduino platform.
template select an adhoc template.
$ tvmc micro flash /tmp/x200 zephyr
usage: tvmc micro flash PROJECT_DIR zephyr [--list-options] -o OPTION=VALUE [OPTION=VALUE ...]
tvmc micro flash PROJECT_DIR zephyr: error: the following arguments are required: -o
$ tvmc micro flash /tmp/x200 zephyr --list-options
usage: tvmc micro flash PROJECT_DIR zephyr [--list-options] -o OPTION=VALUE [OPTION=VALUE ...]
optional arguments:
--list-options show all options/values for selected platforms/template.
-o OPTION=VALUE [OPTION=VALUE ...]
zephyr_board={mimxrt1050_evk, mps2_an521, nrf5340dk_nrf5340_cpuapp,
nucleo_f746zg, nucleo_l4r5zi, qemu_cortex_r5, qemu_riscv32,
qemu_riscv64, qemu_x86, stm32f746g_disco}
name of the Zephyr board to build for. (required)
$ tvmc micro flash /tmp/x200 zephyr -o zephyr_board=stm32f746g_disco
$
Running a model
As showed above by using tvmc micro
subcommands create-project
, build
, and
flash
itâs possible to get a model flashed and ready to run on a micro target.
Running a model is naturally common to all targets (not exclusive on MCUs),
so commands and options for running a model on microcontrollers are accommodated
under the existing tvmc run
command. To select a micro target the device
âmicroâ needs to be specified with --device micro
. When it happens PATH
argument can be used not to select a compiled model module file (like
currently it happens on non-micro targets) but to specify the project directory
used for flashing the model of interest:
$ tvmc run --device micro -h
usage: tvmc run [-h] [--device {cpu,cuda,cl,metal,vulkan,rocm,micro}] [--fill-mode {zeros,ones,random}] [-i INPUTS] [-o OUTPUTS] [--print-time] [--print-top N] [--profile] [--repeat N] [--number N]
[--rpc-key RPC_KEY] [--rpc-tracker RPC_TRACKER]
PATH
positional arguments:
PATH path to the compiled module file or to the project directory (micro devices only).
optional arguments:
-h, --help show this help message and exit
--device {cpu,cuda,cl,metal,vulkan,rocm,micro}
target device to run the compiled module. Defaults to 'cpu'
[...]
Once both micro
device and project directory are specified --list-options
can be used to list all the options available (optional and required ones) for
running the model on the micro target:
$ tvmc run --device micro /tmp/x200 --list-options
usage: tvmc run [-h] [--device {cpu,cuda,cl,metal,vulkan,rocm,micro}] [--fill-mode {zeros,ones,random}] [-i INPUTS] [-o OUTPUTS] [--print-time] [--print-top N] [--profile] [--repeat N] [--number N]
[--rpc-key RPC_KEY] [--rpc-tracker RPC_TRACKER] [--list-options] --options OPTION=VALUE [OPTION=VALUE ...]
PATH
positional arguments:
PATH path to the compiled module file or to the project directory (micro devices only).
optional arguments:
-h, --help show this help message and exit
--device {cpu,cuda,cl,metal,vulkan,rocm,micro}
target device to run the compiled module. Defaults to 'cpu'
--fill-mode {zeros,ones,random}
fill all input tensors with values. In case --inputs/-i is provided, they will take precedence over --fill-mode. Any remaining inputs will be filled using the chosen fill mode. Defaults to 'random'
-i INPUTS, --inputs INPUTS
path to the .npz input file
-o OUTPUTS, --outputs OUTPUTS
path to the .npz output file
--print-time record and print the execution time(s)
--print-top N print the top n values and indices of the output tensor
--profile generate profiling data from the runtime execution. Using --profile requires the Graph Executor Debug enabled on TVM. Profiling may also have an impact on inference time, making it take longer to be generated.
--repeat N run the model n times. Defaults to '1'
--number N repeat the run n times. Defaults to '1'
--rpc-key RPC_KEY the RPC tracker key of the target device
--rpc-tracker RPC_TRACKER
hostname (required) and port (optional, defaults to 9090) of the RPC tracker, e.g. '192.168.0.100:9999'
--list-options show all options/values for selected platforms/template.
--options OPTION=VALUE [OPTION=VALUE ...]
gdbserver_port=GDBSERVER_PORT
if given, port number to use when running the local gdbserver.
nrfjprog_snr=NRFJPROG_SNR
when used with nRF targets, serial # of the attached board to use, from nrfjprog.
openocd_serial=OPENOCD_SERIAL
when used with OpenOCD targets, serial # of the attached board to use.
zephyr_base=ZEPHYR_BASE
path to the zephyr base directory.
zephyr_board={mimxrt1050_evk, mps2_an521, nrf5340dk_nrf5340_cpuapp,
nucleo_f746zg, nucleo_l4r5zi, qemu_cortex_r5, qemu_riscv32,
qemu_riscv64, qemu_x86, stm32f746g_disco}
name of the Zephyr board to build for. (required)
Already existing options that are quite useful when running or testing a model
become automatically available when running a model on microcontrollers, like, for
instance, --print-top
and --fill-mode
options.
For example, one can run the flashed example model shown above this way:
$ tvmc run --device micro /tmp/x200 --options zephyr_board=stm32f746g_disco --print-top 1 --fill-mode random
[[ 0. ]
[-0.48289177]]
$
tvmc run
will take care of opening a transport channel (usually via serial)
with the target device where the model is, then it will write the necessary
input data to the run the model and collect back the output, showing it on the
command-line interface if that is requested.
Reference-level explanation
The proposed interface relies on Project API server_info_query
method to
obtain all the options available for a specified platform.
argparse
is used to build the command-line parser, but a dynamic approach is
used to build a sort of dynamic parser: first one parser is built to parse the
platform type; then, once the platform is parsed from the given command line
a query is made using the Project API server_info_query
method to obtain all the options
available for the specified platform and the parser is incremented to parse the options
based on the selected platform and subcommand.
By using metadata in ProjectOption associated to every project option returned by
server_info_query
itâs possible to build the second parser to parse the
options by platform and the selected subcommand (like create-project
, build
, and run
).
The first (main) reason for using that âdynamic parserâ mechanism is because itâs possible for the user to supply their own platform (when selecting the template
adhoc platform and passing a --template-dir
directory), so itâs not possible to enumerate all platformsâ options beforehand.
The second reason is due to the fact that querying the options for a platform can take about 0.4 seconds. So with the 2 default platforms currently available (Zephyr and Arduino) it takes about 1 second to query the options and build a parser accordingly. Thus with 3 or more default platforms the user would start to feel some slowness
when using tvmc
.
The current implementation avoids querying the options for all the default platforms available and so will always spend ~0.4s to query and build the proper parsers to parse the platform options no matter how many platforms might exist.
Default project templates
Default project templates must be detected and made available as platform
options automatically.
The mechanism for automatically detecting the default template directories must consider basically two different environments: when TVM is built from sources and when TVM is installed via a Python package (like TVM tlcpack).
The mechanism here suggested is one similar to the one used by
tvm.micro.get_standalone_crt_dir()
, that relies on
_ffi.libinfo.find_lib_path()
and so can correctly find the libraries taking
into account if TVM being used is from sources or from a installed package.
Hence for the specific case of the template directories a new function named
tvm.micro.get_template_dirs()
which returns all available default platform
template directories will be created. The templates themselves will be added to
the wheel package accordingly, just like the standalone_crt/
is added.
Drawbacks
argparse
unfortunately only supports two mutually exclusive groups through
add_mutually_exclusive_group()
. This complicates a smooth integration of new
micro target specific options into the already existing tvmc run
command.
For example, --rpc-key
and --rpc-tracker
do not apply to micro targets, but
canât be automatically excluded by argparse
if micro target specific option
--list-options
is used. This maybe can be solved by enforcing the mutual
exclusive groups in the code instead of in the parser as itâs done in the
current implementation and mark these options with something like (not available for micro targets)
, just as required options are marked with
(required)
.
Unresolved questions
-
Currently required options that are passed when creating a project, i.e. passed to
create-project
, need to be repeated forbuild
,flash
, andrun
. This is ok if one, for whatever reason, wants to change it from one subcommand to another, but might be a burden to the user that will need to type the key-value again for each subsequent subcommand. Maybe the required values passed whencreate-project
is used could be recorded in the new project directory (in a JSON file â.tvmc.json
) and a flag called--auto
can be added to the subcommands (build
,flash, and
run`) and if the required options where already passed and are recorded in the JSON file then TVMC would pick up and use them automatically. -
Similarly itâs possible to remove the platform selection for other subcommands after the platform is specified once in
create-project
. In that case the project options would be discovered automatically based on the project dir, not based on the template dir, as itâs done in the current implementation.
Future possibilities
-
The current implementation and RFC does not address
tvmc tune
for micro targets. It will be addressed later, once the interface here proposed is discussed and crystallizes. The âtvmc tuneâ will be extended probably along the lines as thetvmc run
command is extended. -
For creating a project one departs from a compiled model kept in a Model Library Format archive. This is fine and gives to the user total control over which settings would be used when compiling a model (using
tvmc compile
). On the other hand the CPU model for the target board, for instance, needs to be specified explicitly, and that maybe annoying for some users. Thus it would be good in the future ifcreate-project
could depart (pass as input instead of a MLF archive) from a model not compiled (i.e. supply a model in one of the frameworksâ format like TFLite, ONNX, etc) and TVMC would deduce a default working set of options to compile the model based on the specified board. In that case TVMC would need to keep a table to convert from specified board to a suitable default compiling options to be used for compiling the model for the specified board. No change would be necessary fortvmc compile
andtvmc create-project
would use it under the hood.