On 2025-01-15 at 23:46 +09, Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
> On Wed, Jan 15, 2025 at 10:30 PM Mikhail Rudenko <mike.rudenko@xxxxxxxxx> wrote:
>>
>> Hi Tomasz,
>>
>> On 2025-01-15 at 17:31 +09, Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
>>
>> > Hi Mikhail and Laurent,
>> >
>> > On Wed, Jan 15, 2025 at 2:07 AM Mikhail Rudenko <mike.rudenko@xxxxxxxxx> wrote:
>> >>
>> >>
>> >> Hi Laurent,
>> >>
>> >> On 2025-01-03 at 17:23 +02, Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx> wrote:
>> >>
>> >> > On Thu, Jan 02, 2025 at 06:35:00PM +0300, Mikhail Rudenko wrote:
>> >> >> Currently, the rkisp1 driver always uses coherent DMA allocations for
>> >> >> video capture buffers. However, on some platforms, using non-coherent
>> >> >> buffers can improve performance, especially when CPU processing of
>> >> >> MMAP'ed video buffers is required.
>> >> >>
>> >> >> For example, on the Rockchip RK3399 running at maximum CPU frequency,
>> >> >> the time to memcpy a frame from a 1280x720 XRGB32 MMAP'ed buffer to a
>> >> >> malloc'ed userspace buffer decreases from 7.7 ms to 1.1 ms when using
>> >> >> non-coherent DMA allocation. CPU usage also decreases accordingly.
>> >> >
>> >> > What's the time taken by the cache management operations ?
>> >>
>> >> Sorry for the late reply, your question turned out a little more
>> >> interesting than I expected initially. :)
>> >>
>> >> When capturing using Yavta with MMAP buffers under the conditions mentioned
>> >> in the commit message, ftrace gives 437.6 +- 1.1 us for
>> >> dma_sync_sgtable_for_cpu and 409 +- 14 us for
>> >> dma_sync_sgtable_for_device. Thus, it looks like using non-coherent
>> >> buffers in this case is more CPU-efficient even when considering cache
>> >> management overhead.
>> >>
>> >> When trying to do the same measurements with libcamera, I failed. In a
>> >> typical libcamera use case when MMAP buffers are allocated from a
>> >> device, exported as dmabufs and then used for capture on the same device
>> >> with DMABUF memory type, cache management in kernel is skipped [1]
>> >> [2]. Also, vb2_dc_dmabuf_ops_{begin,end}_cpu_access are no-ops [3], so
>> >> DMA_BUF_IOCTL_SYNC from userspace does not work either.
>> >
>> > Oops, so I believe this is a bug. When an MMAP buffer is allocated in
>> > the non-coherent mode, those ops should perform proper cache
>> > maintenance.
>>
>> Thanks for pointing this out!
>>
>> > Let me send a patch to fix this in a couple of days unless someone
>> > does it earlier.
>>
>> Now that we know that this is a bug, not an API misuse from my side, I
>> can fix this myself and send a v2. Would this be okay for you?
>
> I'd be more than happy :)
Done, see [1]. A review would be appreciated. :)
[1] https://lore.kernel.org/all/20250115-b4-rkisp-noncoherent-v2-0-0853e1a24012@xxxxxxxxx/
--
Best regards,
Mikhail Rudenko
