Stm32h750 Weact

WeAct STM32H750VBT6 core board

Board-specific firmware knowledge. The two reference files hold the detail — read the one you need rather than guessing, because most of the failure modes below are silent (the board runs, just wrongly).

• reference/board-hardware.md — complete hardware reference: schematic-level pin map, every connector, power tree, memory map, SDK/examples inventory, plus a full development guide (Part II: §12 PlatformIO, §13 clocks, §14 peripheral cookbook, §15 M7 gotchas, §16 flashing, §17 pitfall table).
• reference/recipes.md — copy-paste-ready code: platformio.ini, clock config, LCD driver with the fast blit path, backlight PWM, ADC temperature sensor, USB CDC, DWT frame pacing.
• template/ — a complete working project that builds and flashes, plus a scaffolding script. It is where every recipe was extracted from; see template/README.md.

Orientation

Rules that prevent the expensive mistakes

Check these before writing code — each one produces a failure that looks like something else.

1. -DHSE_VALUE=25000000 in build_flags. The stock framework-stm32cubeh7 CMSIS template already defaults HSE_VALUE to 25 MHz — matching this board's crystal by coincidence, the same way the F4 template does for the Black Pill — so this currently works without the flag. It stops working the moment a CubeMX-generated stm32h7xx_hal_conf.h from a different H7 board (most Nucleo/Eval H7 boards run their HSE from an 8 MHz ST-Link MCO) gets dropped in: the PLL still locks, but every derived timing is wrong by ~3×. Set it explicitly so the project doesn't depend on which hal_conf.h happens to be in play.
2. PWR_LDO_SUPPLY — this board has no SMPS. Configuring one leaves the core under-supplied and unresponsive to SWD until a BOOT0 power cycle. The setting latches on first write after reset.
3. Fix the board definition's swapped sizes in platformio.ini: board_upload.maximum_size = 131072 (flash), board_upload.maximum_ram_size = 524288 (RAM). Otherwise nothing warns you when an image exceeds the real 128 KB.
4. board_build.stm32cube.disable_embedded_libs = yes if you vendor an ST7735 driver into lib/ — Cube's embedded BSP ships its own and they collide at link time.
5. FLASH_LATENCY_1 at HCLK 120 MHz. Too few wait states → hard fault on the first flash read.
6. Cache enable before HAL_Init(), clock config before any peripheral init.
7. D-cache vs DMA: buffers __attribute__((aligned(32))), size a multiple of 32, clean before TX / invalidate after RX. And DMA1/DMA2 cannot reach DTCM or ITCM — a buffer there silently never updates. Default linker puts everything in RAM_D1 (AXI SRAM), which is fine.
8. Backlight is TIM1_CH2N, a complementary output: HAL_TIMEx_PWMN_Start(), not HAL_TIM_PWM_Start(), and HAL_TIMEx_ConfigBreakDeadTime() + HAL_TIM_MspPostInit() are mandatory or the pin never drives.
9. USB needs HAL_PWREx_EnableUSBVoltageDetector() plus HSI48State = RCC_HSI48_ON. Missing it means the host sees nothing at all. 9a. The AF macro is GPIO_AF10_OTG2_FS, not GPIO_AF10_OTG1_FS. This part has two USB_OTG controllers, so stm32h7xx_hal_gpio_ex.h defines USB2_OTG_FS and compiles out the OTG1_FS macro name entirely — code copied from an H7A3/H72x example (which has only one controller) fails to build with an undefined-identifier error that doesn't say why.
10. LCD and camera reset lines are tied to the board reset net — no GPIO control, ever. A wedged panel needs a full MCU reset.

When the task is LCD performance

The stock ST7735 driver is ~200× slower than the bus allows: ST7735_FillRGBRect() calls SetCursor() per pixel row, and SetCursor() sends each address byte as its own CS-framed SPI transaction. A 13-character line costs ~1660 SPI transactions.

Fix by composing a full-width band in RAM (pixels pre-byte-swapped — the panel is big-endian RGB565) and pushing it with one SetCursor + one burst: 8 transactions, ~1.4 ms for 160×16 at 30 MHz. That is what makes a 120 Hz UI loop fit. reference/recipes.md has the working implementation.

Bus speed is secondary but real: SPI_BAUDRATEPRESCALER_4 = 30 MHz works on this board although the ST7735S spec implies 15 MHz; halve it the moment pixels come out garbled.

Panel refresh itself is programmable via FRMCTR1 (f = f_osc / ((RTNA*2+40) * (LINE+FPA+BPA)), f_osc ≈ 850 kHz, LINE = 160): stock ≈ 80 Hz, (0, 2, 2) ≈ 130 Hz. Worth setting an expectation though — pixel response on this panel is ~100 ms, so an fps readout is more convincing than the animation.

Starting a new project

Do not hand-assemble one. template/ is a verified-working project — scaffold from it:

~/.claude/skills/stm32h750-weact/template/variants/new-project.sh <target-dir> [--full|--minimal]
cd <target-dir> && pio run

• --full (default) — ST7735 UI, USB CDC console, ADC temperature sensor. 40.9 KB of 128 KB flash.
• --minimal — LED + K1 only, same clock tree and startup order. 5.7 KB. Use it to prove the toolchain and the flashing route before adding anything.

Both variants build as-is (verified). Nothing is generated and no paths are embedded, so copying the tree by hand works equally well. template/README.md lists which files belong to which subsystem, so a --full scaffold can be stripped back cleanly.

When the user already has a project, prefer bringing it in line with the template's platformio.ini and board.h over rewriting their code.

Flashing

DFU over the Type-C cable, no probe: hold BOOT0, tap NRST, release after ~0.5 s → the board enumerates as 0483:DF11 → pio run -t upload. The board never resets itself into DFU; the sequence is manual every time. SWD via header P3 when you need breakpoints.

When 128 KB of flash runs out, the board's answer is QSPI XIP at 0x90000000 with a small loader in internal flash — a decision to make early, since the linker script and flashing procedure both change. See reference/board-hardware.md §6, §7.3, §15.3.

Reporting

State honestly what was verified on hardware versus derived from the datasheet. Several settings here (30 MHz SPI, raised FRMCTR1) are deliberately outside published specs and work on this specific board — say so rather than presenting them as safe defaults.