Thanh Le / bench notes
EMBEDDED · BOARD BPI-M3 · DONE

DRAM init over FEL, and the probe that froze the boot ROM

xfel ddr doesn't know the A83T, so I ported U-Boot's LPDDR3 init into a 4 KiB SRAM program

Result2 GiB LPDDR3, 294 896 words tested, 0 errors, 3.44 s
Last change2026-10-04: 2 GiB up and tested, blink runs from 0x40000000
Boardall stories on this board →

DRAM init over FEL, and the probe that froze the boot ROM

Back to the Banana Pi M3 hub

  1. Situation
  2. Task
  3. Action
  4. Result
  5. What I’d do next

Situation

Same night as the blink. It ran from SRAM, but SRAM A2 is tiny. On the D1, xfel ddr brings up DRAM for you. On the A83T:

$ xfel ddr
Failed to initial ddr controller
$ xfel ddr lpddr3
Failed to initial ddr controller

No luck, xfel has no DRAM init for this chip.

ddr_init in SRAM brings up DRAM; reading DRAM before that hangs the ROM The plan, and the hole I fell into on the way: a 4196-byte init in SRAM, DRAM only after it says DDR!. My early hexdump 0x40000000 cost a power cycle.

Task

Bring up the 2 GB of LPDDR3 myself, from a program in SRAM, prove it works, then run the unchanged blink from DRAM. Everything over FEL, no SD card.

Action

U-Boot already knows how to do it: dram_sun8i_a83t.c, 471 lines. I copied it and changed only how registers are addressed (U-Boot uses struct pointers, I use plain addresses), plus a small compat.h. The first compile warned about a missing max() and a pointer/integer mix; fixed, 2560 bytes of code.

Around it, my own main does what SPL would normally do first:

  1. DCDC5 (the DRAM supply) to 1.2 V over RSB: 32 + (1200 - 1140) / 20 = 0x23, U-Boot’s formula.
  2. PLL6, MBUS, PLL5, gates.
  3. U-Boot’s init: timing, ODT, ZQ, PHY training, size detection.
  4. A memory test, then a magic word DDR! in the result table if it all passed.

ddr_init.bin came out at 4196 bytes, linked at 0x44000 like the blink.

Then I hurt myself. Before running the init I wanted a “before” picture, so I did xfel hexdump 0x40000000 16. That read never returned, timeout killed it, and from then on the ROM answered nothing:

$ xfel hexdump 0x40000000 16
Terminated
usb bulk send error
rc=124

Reading DRAM while the DRAM controller has no clock leaves the CPU’s bus transaction waiting forever. Power cycle, back in FEL.

Two smaller bugs on the way: a field in my result table read “MfCe”, because there’s no C runtime clearing .bss (so don’t trust .bss). And my shell guard comparing the magic failed, because xfel read32 prints 0x21524444\r\n. A tr -d '\r' fixed it.

A small run.sh does the whole thing with a guard: init, check the magic is DDR!, and only then write to DRAM.

$ sh run.sh 4 250 2

Result

DRAM came up with all 2 GiB, controller config 0x004719E5 (2 ranks, 8 banks, 16-bit bus, LPDDR3), and both byte lanes trained clean (DXnGSR0 = 0x0046A300 on lanes 0 and 1). The memory test checked 294 896 words, dense over the first 1 MiB plus one word every 64 KiB, with 0 errors. Init plus test took 3.44 s, and then the CPU was back in FEL.

The unchanged blink, relinked at 0x40000000, ran from DRAM and wrote DONE at 0x40000F00, an address only DRAM has. 8 writes, so 4 blue blinks, in 2.0 s.

Coverage of the 294 896-word memory test over 2 GiB What my “0 errors” actually covers: every word of the first MiB, then one word per 64 KiB. It’s a smoke test, not a full memtest.

Not measured: DRAM at full speed or with marginal timing. The CPU was at the ROM’s clock with caches off.

What I’d do next

The obvious one: put something bigger than a blink in all that DRAM. That became the bare-metal CLI.