stm32-data
is a project aiming to produce clean machine-readable data about the STM32 microcontroller
families, including:
- ✔️ Base chip information
- RAM, flash
- Packages
- ✔️ Peripheral addresses and interrupts
- ✔️ Interrupts
- ✔️ GPIO AlternateFunction mappings (for all families except F1)
- ❌ GPIO mappings for F1
- 🚧 Register blocks for all peripherals
- ✔️ DMA stream mappings
- ✔️ Per-package pinouts
- ✔️ Links to applicable reference manuals, datasheets, appnotes PDFs.
✔️ = done, 🚧 = work in progress, ❌ = to do
The generated JSON files are available here in the stm32-data-generated
repo.
If you're looking for the API docs for stm32-metapac
customized to a
particular chip, they are available here: https://docs.embassy.dev/stm32-metapac
The generated PAC is available here in the stm32-data-generated
repo.
These are the data sources currently used.
- STM32Cube database:
- List of all MCUs
- Peripheral versions
- Flash, RAM size
- GPIO AF mappings
- Interrupt mappings (which signals from which peripherals are OR'd into a NVIC interrupt line)
- DMA channel request mappings
- Packages and package pinouts (TODO)
- STM32Cubeprog database
- Flash, RAM size
- STM32 mcufinder:
- Links to documentation PDFs (reference manuals, datasheets...)
- STM32 HAL headers:
- interrupt number, name
- peripheral addresses
- stm32-rs SVDs: register blocks.
For register blocks, YAMLs are intially extracted from SVDs, manually cleaned up and committed. From this point on, they're manually maintained.
We don't maintain "patches" for registers unlike the stm32-rs
project. This has two advantages:
- Fixing mistakes and typos in the SVDs is now much easier (and yes, there are A LOT of those). It doesn't require a complicated patching system, you just edit the YAML to fix the mistake instead of writing a patch that will fix it when applied.
- Ensuring consistency for the same peripherals between chips. (i.e. we check in a single
i2c_v2.yaml
and ensure it is good, vs trying to patch wildly differing SVDs for what's an identical IP block into being consistent). Thestm32-rs
project doesn't have this as a goal, while we do. Writing a single HAL for all stm32 chips (such as theembassy-stm32
HAL) is impossible otherwise.
-
Run
./d download-all
This will download all the required data sources into
sources/
. -
Run
cargo run --release --bin stm32-data-gen
This generates all the intermediate JSON's in
build/data/
.Assignments of registers to peripherals is done in a perimap and fixes to registers can be done in the files located in
data/registers
. -
Run
cargo run --release --bin stm32-metapac-gen
This generates the
stm32-metapac
crate intobuild/stm32-metapac/
.
This will help you add support for a new peripheral to all STM32 families. (Please take the time to add it for all families, even if you personally are only interested in one. It's easier than it looks, and doing all families at once is significantly less work than adding one now then having to revisit everything later when adding more. It also helps massively in catching mistakes and inconsistencies in the source SVDs.)
- First, make sure you can generate the YAMLs following the steps above. You should be able to regenerate the already-committed yamls, and end up with no diff.
- Install chiptool with
./d install-chiptool
- Run
./d extract-all LPUART1
. This'll output a bunch of yamls intmp/LPUART1
. NOTE sometimes peripherals have a number sometimes not (LPUART1
vsLPUART
). You might want to try both and merge the outputted YAMLs into a single directory. - Diff them between themselves, to identify differences. The differences can either be:
- 1: Legitimate differences between families, because there are different LPUART versions. For example, added registers/fields.
- 2: SVD inconsistencies, like different register names for the same register
- 3: SVD mistakes (yes, there are some)
- 4: Missing stuff in SVDs, usually enums or doc descriptions.
- Identify how many actually-different (incompatible) versions of the peripheral exist, as we must not merge them. Name them v1, v2.. (if possible, by order of chip release date, see here.
- For each version, pick the "best" YAML (the one that has less enums/docs missing), place them in
data/registers/lpuart_vX.yaml
- Cleanup the register yamls (see below).
- Minimize the diff between each pair of versions. For example between
lpuart_v1.yaml
andlpuart_v2.yaml
. If one is missing enums or descriptions, copy it from another. - Make sure the block
- Add entries to
perimap
, see below. - Regen, then:
- Check
data/chips/*.yaml
has the rightblock: lpuart_vX/LPUART
fields. - Ensure a successful build of the affected pac. e.g.
cd build/stm32-metapac cargo build --features stm32u5a9nj,memory-x --target thumbv8m.main-none-eabihf
- Check
Please separate manual changes and changes resulting from regen in separate commits. It helps tremendously with review and rebasing/merging.
SVDs have some widespread annoyances that should be fixed when adding register YAMLs to this repo. Check out chiptool
transforms, they can help in speeding up the cleanups.
- Remove "useless prefixes". For example if all regs in the
RNG
peripheral are namedRNG_FOO
,RNG_BAR
, theRNG_
peripheral is not conveying any useful information at all, and must go. - Remove "useless enums". Useless enums is one of the biggest cause of slow compilation times in STM32 PACs.
- 0=disabled, 1=enabled. Common in
xxEN
andxxIE
fields. If a field says "enable foo" and is one bit, it's obvious "true" means enabled and "false" means disabled. - "Write 0/1 to clear" enums, common in
xxIF
fields. - Check out the
DeleteEnums
chiptool transforms.
- 0=disabled, 1=enabled. Common in
- Convert repeated registers or fields (like
FOO0 FOO1, FOO2, FOO3
) to arraysFOO[n]
.- Check out the
MakeRegisterArray
,MakeFieldArray
chiptool transforms.
- Check out the
- Use
chiptool fmt
on each of the register yamls.
NOTE: At the time of writing all families are supported, so this is only useful in particular situations, for example if you want to regenerate an RCC register yaml from scratch.
Adding support for a new familiy is mostly a matter of adding support for RCC.
Now extract the RCC peripheral registers: ./d extract-all RCC --transform ./transform-RCC.yaml
Note that we have used a transform to mechanically clean up some of the RCC
definitions. This will produce a YAML file for each chip model in ./tmp/RCC
.
Sometimes the peripheral name will not match the name defined in the SVD files, check the SVD file for the correct peripheral name.
RCC registers are usually identical within a family, except they have different subsets of the enable/reset bits because different chips have different amounts of peripherals. (i.e. one particular bit position in one particular register is either present or not, but will never have different meanings in different chips of the same family.)
To verify they're indeed subsets, choose the model with the largest peripheral set possible (e.g. the STM32G081) and compare its YAML against each of the other models' to verify that they are all mutually consistent.
Finally, we can merge
./merge_regs.py tmp/RCC/g0*.yaml
This will produce regs_merged.yaml
, which we can copy into its final resting
place:
mv regs_merged.yaml data/registers/rcc_g0.yaml
To assign these newly generated registers to peripherals, utilize the mapping done in parse.py
.
An example mapping can be seen in the following snippet
('STM32G0.*:RCC:.*', 'rcc_g0/RCC'),
such mapping assignes the rcc_g0/RCC
register block to the RCC
peripheral in every device from the STM32G0
family.
The stm32-data-gen
binary has a map to match peripherals to the right version in all chips, the perimap.
When parsing a chip, for each peripheral a "key" string is constructed using this format: CHIP:PERIPHERAL_NAME:IP_NAME:IP_VERSION
, where:
CHIP
: full chip name, for exampleSTM32L443CC
PERIPHERAL_NAME
: peripheral name, for exampleSPI3
. Corresponds toIP.InstanceName
in the MCU XML.IP_NAME
: IP name, for exampleSPI
. Corresponds toIP.Name
in the MCU XML.IP_VERSION
: IP version, for examplespi2s1_v3_3_Cube
, Corresponds toIP.Version
in the MCU XML.
perimap
entries are regexes matching on the above "key" string. First regex that matches wins.
The IP version is particularly useful. It is an ST-internal "IP version" that's incremented every time changes are made to the peripheral, so it correlates very well to changes in the peripheral's register interface.
You should prefer matching peripherals by IP version whenever possible. For example:
('.*:SPI:spi2s1_v2_2', 'spi_v1/SPI'),
('.*:SPI:spi2s1_v3_2', 'spi_v2/SPI'),
Sometimes it's not possible to map by IP version, and we have to map by chip name. For example:
('STM32H7.*:FLASH:.*', 'flash_h7/FLASH'),
('STM32F0.*:FLASH:.*', 'flash_f0/FLASH'),
('STM32F1.*:FLASH:.*', 'flash_f1/FLASH'),
('STM32F3.*:FLASH:.*', 'flash_f3/FLASH'),
('STM32F4.*:FLASH:.*', 'flash_f4/FLASH'),
...etc
Sometimes even the same IP name+version in the same chip family has different registers (different instances of the IP are configured differently), so we have to map by chip name AND peripheral instance name. This should be the last resort. For example:
('STM32F7.*:TIM1:.*', 'timer_v1/TIM_ADV'),
('STM32F7.*:TIM8:.*', 'timer_v1/TIM_ADV'),
('.*TIM\d.*:gptimer.*', 'timer_v1/TIM_GP16'),
The versions of peripherals can be found in the table here.