Repository navigation
hisi: name the GK7206 family from the eFuse model code (V11A confirmed, rest guessed) - #289
Conversation
Both report chip ID 0x72060110, so a GK7206V11A was shown as 7206V11. What differs is the model code in the low byte of eFuse word 0x10090030, whose high half is "GK" (0x474B): the vendor modules check it (xm_common accepts only 0x474B00B7..0x474B00DD; xm_vi and xm_venc single out 0xCA among other codes). Nothing in the SDK or in XM's firmware names the codes -- XM's own image hard-codes its product name -- so only the code read on a known part is listed: 0xCA on an IVG-K8 (XM IPC_GK7206V11A). Any other code stays 7206V11 until one is confirmed. The eFuse reads 0x2 while its clock (CRG 0x1BC bit 1) is off; it is turned on for the read and back off if it was off. On the camera: model 7206V11A, CRG 0x1BC 0x284 before and after. Co-authored-by: OpenIPC AI team <ai@openipc.org>
PR Summary by QodoIdentify GK7206V11A using its eFuse model code
AI Description
Merge Risk
|
Code Review by Qodo
1.
|
Extends the V11A check to every GK7206 chip ID, with a table that is
guessed except for 0xCA.
xm_vi and xm_venc sort the valid codes into capability classes (their
limits are stored as (x ^ 0xA64BD) - 11): {B7, BC, C7, CC} 3840x2160 with
a 3840 line, {CA, D8..DB, DD} 3840x2160 with a 4032 line, everything else
3840x3840. Nothing names them. Taking the high nibble as the DDR tier the
SDK's load script gives each part (B 64 MB: V10, V10B; C 128 MB: V11,
V11A, V11AT; D 256 MB: V12A) and the low nibble as the part in the tier
(7 base, C the other, A the A part) fits both classes and the one known
code:
B7 V10, BC V10B, C7 V11, CA V11A (confirmed), CC V11AT, D8..DB/DD V12A
The 3840x3840 class matches no SDK part and keeps the chip-ID name. The
source comment and each case say which are guessed.
Co-authored-by: OpenIPC AI team <ai@openipc.org>
… word Re-read CRG 0x1BC before turning the eFuse clock back off, and leave it if it no longer holds the word written for the read: whoever changed it in between owns it now, as sensor_crg_restore() does. (Qodo) Co-authored-by: OpenIPC AI team <ai@openipc.org>
|
Qodo's "Concurrent clock changes get overwritten": fixed in f3bce33. The probe re-reads CRG |
GK7206V11 and V11A both report chip ID
0x72060110, so a V11A was shown as7206V11. The variant is in the low byte of eFuse word0x10090030(high half0x474B, "GK").What the vendor modules show.
xm_commonaccepts0x474B00B7..BD,C7..CD,D7..DD.xm_viandxm_vencsort the codes into capability classes (their limits are stored obfuscated as(x ^ 0xA64BD) - 11):B7BCC7CCCAD8–DBDDNothing in the SDK (modules, libraries, U-Boot, kernel, docs) or in XM's stock firmware names the codes.
The table — only
0xCAis confirmed, the rest is guessed (and says so in the source): the high nibble taken as the DDR tier the SDK's load script gives each part (B = 64 MB: V10, V10B; C = 128 MB: V11, V11A, V11AT; D = 256 MB: V12A), the low nibble as the part within the tier (7 base, C the tier's other part, A the A part).0xB70xBC0xC70xCAIPC_GK7206V11A)0xCC0xD8–0xDB,0xDDThe 3840×3840 class matches no SDK part and keeps the chip-ID name, as does any read without the "GK" marker. Applied to all three GK7206 chip IDs. The eFuse reads
0x2while its clock (CRG0x120101BCbit 1) is off; it is turned on for the read and back off if it was off.Tested on the camera:
model: 7206V11A, CRG0x1BC0x284before and after.hal_hisi_testcovers the confirmed code, each guessed tier, the fallback class, a gated read, and a clock already on.