Hardware reference
Clock and reset
The block at 0x0aa8_0000 holds the eight memory remap windows, three clock mask words and a one-time control write that changes the packet-RAM mapping.
Registers
| Offset | Address | Register | Type | Description |
|---|---|---|---|---|
+0x04 |
0x0aa8_0004 |
CONTROL_04 |
u32 WO | Written 0x200 once, during DMA and clock set-up. The stack must be in DTCM; see the transition on this page. |
+0x08 |
0x0aa8_0008 |
WINDOW_SELECTOR |
u32 RW | Selects the remap window read through WINDOW_VALUE. |
+0x18 |
0x0aa8_0018 |
FEATURE_CONTROL |
u32 RW | The memory probe ORs in 0x0018_0200 and restores the saved value afterwards. |
+0x2c |
0x0aa8_002c |
WINDOW_VALUE |
u32 RO | The selected remap window. All eight are sampled once at startup. |
+0x40 |
0x0aa8_0040 |
MASK_40 |
u32 RW | DMA and clock set-up ORs in 0x00f0_0000. |
+0x44 |
0x0aa8_0044 |
MASK_44 |
u32 RW | Interrupt set-up ANDs with 0x7fff_f777, clearing bits 31, 11, 7 and 3. |
+0x48 |
0x0aa8_0048 |
MASK_48 |
u32 RW | Interrupt set-up ANDs with 0x777f_ffff; DMA and clock set-up later ORs in 0xf0. |
Source: ClockReset · 76 bytes
WINDOW_SELECTOR
| Bits | Field | Description |
|---|---|---|
| 4:2 | WINDOW |
Remap window (0..7) read back through WINDOW_VALUE. The firmware preserves the other bits. |
Source: WindowSelector · 4 bytes
WINDOW_VALUE
One remap window descriptor, as the relocation arithmetic decodes it. Window 5’s bits 11:0 and window 4’s HIGH form a third page number that is relocated the same way. Window 0 doubles as the remap control word: when its bit 27 is clear the memory probe runs.
| Bits | Field | Description |
|---|---|---|
| 31:28 | HIGH |
With window 5’s bits 11:0, a third page number (window 4 only). |
| 27:12 | PAGE |
Page number. Windows 3 and 4 are relocated by adding 0x166. |
| 11:0 | LOW |
Preserved by relocation in windows 3 and 4; in window 5 it is part of the third page number. |
Source: RemapWindow · 4 bytes
Remap windows
At start-up the firmware selects each window in turn through WINDOW_SELECTOR and samples it from WINDOW_VALUE. If any window is nonzero, windows 3, 4 and 5 are stored relocated. Three page numbers move up by 0x166: window 3’s bits 27:12, window 4’s bits 27:12, and a third formed from window 5’s bits 11:0 above window 4’s bits 31:28. The firmware reproduces the vendor arithmetic exactly:
t3 = ((w3 & 0x0fff_ffff) >> 12) + 0x166
t4 = ((w4 & 0x0fff_ffff) >> 12) + 0x166
th = ((w4 >> 28) | (w5 << 4)) + 0x166
w3 = (w3 & 0xf000_0fff) | (t3 << 12)
w4 = (w4 & 0x0000_0fff) | (t4 << 12) | (th << 28)
w5 = (w5 & 0xffff_f000) | ((th & 0xffff) >> 4)
Window 0 then serves as the remap control word: if its bit 27 is clear, startup mode becomes 3 and the memory probe runs. Whether any window is nonzero also decides a later PlatformControl.CONFIGURATION update.
The CONTROL_04 transition
CONTROL_04 is written 0x200 exactly once, during DMA and clock set-up. After this write the packet-RAM word at 0x0900_ff98, in the download control block, may no longer reflect what was written there. That is a mapping change, not a CPU stop.
Keep the stack in DTCM across this write. With the stack in high SRAM the firmware fails at this write, in a way that looks like a packet-memory fault.