Re: [RFC PATCH v6 7/9] accel/rocket: add RK3576 NPU (RKNN) support

Hello Jiaxing, 

On 8/8/2026 5:16 AM, Jiaxing Hu wrote:
> On 07/08/2026 1:56 pm, Robin Murphy wrote:
>> If the interrupt never fires at all then possibly the signal depends on
>> some additional clock or power domain in order to propagate, or it's
>> just described incorrectly; or if other interrupt sources within the
>> NPU/IOMMU do still work then maybe there's some additional masking
>> control that's been overlooked, or perhaps it it just terminally broken.
> 
> Robin, Diederik, I owe you both a correction before you spend any more
> time on this. The premise is wrong. The interrupt is fine, and the
> polling in patch 7 should not exist.
> 
> The completion interrupt does reach the GIC on RK3576. With the fix
> below and the hrtimer disabled, so that only a real interrupt can retire
> a job, a single int8 convolution runs correctly three times out of three
> on three different inputs with zero timeouts and /proc/interrupts
> counting up. No extra clock, no extra power domain, no extra mask.
> 
> What was actually wrong is one register write. PC_TASK_CON packs the
> task number, and rocket_registers.h is derived from RK3588, where that
> field is 12 bits wide with TASK_PP_EN, TASK_COUNT_CLEAR and RESERVED_0
> above it at bits 12, 13 and 14. RK3576 uses a 16 bit task number, so
> those three controls sit at bits 16, 17 and 18 instead:
>

Sorry, I haven't had time to check this series. Here is some
information for your reference:

For RK3588:

- BIT[11:0] task_number
- BIT[12] task_pp_en
- BIT[13] task_count_clear

For RK3576:

- BIT[15:0] task_number
- BIT[16] task_pp_en
- BIT[17] task_count_clear
- BIT[18] task_last_layer_clear

>   rocket, v1 through v6:  TASK_CON = 0x00007001
>   vendor driver, RK3576:  TASK_CON = 0x00070001
> 
> So the PC read our word as task_number = 0x7001, that is 28673 tasks,
> with the count clear landing on nothing. It never signalled completion
> because by its own count it was never finished, and only a full reset
> ever cleared the counter. That also explains the other symptom in the
> cover letter, that only the first job after a reset computed anything.
> 
>   rocket_pc_writel(core, TASK_CON, (0x7u << 16) | task_count);
> 
> I found it by taking an ordered trace of every register write our driver
> makes during one submit and diffing it against the same trace from the
> vendor driver on the same board. Exactly one value differed. I should
> have done that before writing a workaround, and before describing a
> hardware limitation I had not established.
> 
> Two guesses I made the same night, a per job IOMMU teardown and the
> vendor's post completion sequence, were both wrong, which is the other
> half of the lesson.
> 
> So for v7: patch 7 loses the polling and the "the interrupt does not
> arrive" text. I would rather not carry a bounded poll at all. Jobs that
> compute incorrectly do still time out, but that is a driver bug on my
> side rather than something the hardware needs help with, and the
> scheduler timeout already covers it. Diederik, this is what you warned
> me about off list, that a poll reads as a workaround for an
> undetermined problem. You were right and the problem is now determined.
> 
> The rest of v7 follows what you both asked for: the job_lock fix becomes
> its own patch with a Fixes tag and leaves the RFC series, the
> rk3588_soc_data change is separated from adding rk3576_soc_data,
> refactoring comes before the new support rather than inside it, and I
> will stop editing a comment in the patch after the one that added it.
> 
> On the GIC question, for the record, so the thread has it in one place:
> RK3576 is GICv2, gic-400 in the upstream DT, not GICv3. It made no
> difference here, as you said it would not.
> 
> Thanks for digging into this, and sorry for pointing you at a fault that
> was mine.
> 
> Jiaxing
> 
> 

-- 
Best, 
Chaoyi

_______________________________________________
Linux-rockchip mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-rockchip