| ↑back Search ←Prev date Next date→ Show only urls | (Click on time to select a line by its url) |
| Who | What | When |
|---|---|---|
| *** | Whoopie has quit IRC (Ping timeout: 252 seconds) | [05:40] |
| ......... (idle for 43mn) | ||
| tensa has quit IRC (Ping timeout: 248 seconds) | [06:23] | |
| ...... (idle for 27mn) | ||
| jernej | paulk-leonov: I would be surprised if it wouldn't be any faster :)
paulk-leonov: I found HW quirk and reason why test.mkv is not properly decoded from what I can tell, it seems necessary to call at least one stream parsing command but fortunately, skip bits works well, so instead of setting offset directly, I'm using skip bits trigger to set it like in this patch: http://sprunge.us/kTtNGV I guess there is some internal VPU state which doesn't get initialized properly if at least one stream parsing command is executed currently I didn't found any pattern which would suggest when this problem would occur I have only two samples which need this fix but workaround works with all videos | [06:50] |
| ...... (idle for 25mn) | ||
| paulk-leonov | jernej, oh interesting | [07:21] |
| jernej | another possible explanation is that only 8-bit aligned offsets work (both, CedarX and libvdpau-sunxi, use 8-bit aligned offsets only) | [07:25] |
| ...... (idle for 25mn) | ||
| paulk-leonov: I also find out that VPU on H3 and others SoC, presumably everywhere where CMA node for VPU is not needed, doesn't need DMA offset
so I think we should add CEDRUS_QUIRK_NO_DMA_OFFSET flag to them but that needs more testing | [07:50] | |
| ................................... (idle for 2h50mn) | ||
| *** | harrow has quit IRC (Quit: Leaving) | [10:41] |
| .......... (idle for 46mn) | ||
| paulk-leonov | jernej, yeah there was still some uncertainty about DMA offset
there's also the fact that DRAM can mapped continuously, so under some (fairly common conditions), it could look like there's no need for a DMA offset IIRC the thing to try was 2 GiB setups where planets aren't aligned as well as 512 MiB or 1 GiB | [11:27] |
| ndufresne | ezequielg, so, example 1 works, example 2 is not zero copy, we get a EPERM on the dmabuf import, example 3 works | [11:31] |
| .... (idle for 15mn) | ||
| hverkuil | paulk-leonov: note my review comments for the hevc patches. I'll postpone merging the hevc series until that's fixed. I plan to make a PR for the h264 patches today. | [11:46] |
| ........................ (idle for 1h55mn) | ||
| ezequielg | ndufresne: I am not entirely following. The three worked for me. | [13:41] |
| .............................................. (idle for 3h45mn) | ||
| ndufresne | ezequielg, setup GST_DEBUG, you'll see some dmabuf import error on the second
ezequielg, then trace it to the drm driver (if you want to figure-out), that bit I haven't done yet but I think EPERM in this context is the wrong E number, or there is a secret real meaning for that | [17:26] |
| ...................................................... (idle for 4h27mn) | ||
| ezequielg | ndufresne: so the pipeline fails and then v4l2src internally fallsback to a working io-mode? | [21:54] |
| ndufresne | ezequielg, no, it's kmssink that will bounce the buffers
ezequielg, but should not matter, we should investigate why we get this eperm in the first place, what it means ezequielg, I know the dmabuf is backed by vmalloc allocator in v4l2, could it be the virtio cannot handle having scattered memory ? | [22:07] |
| ezequielg | GST_DEBUG=kmssink:8 ?
let me find the EPERM first :-) hm, i see this: kmssink gstkmssink.c:406:get_drm_caps:<kmssink0> prime import (✗) / prime export (✓) / async page flip (✗) no prime import, so no dmabuf export from vivid | [22:14] |
| hm, I had an old kernel. | [22:27] | |
| .... (idle for 16mn) | ||
| ok.. I found the EPERM. | [22:43] | |
| ndufresne: drmPrimeFDToHandle will return -1 and set errno on failure, that's why we see EPERM.
According to DRM debug, the error is -19 ENODEV? | [22:53] | |
| ndufresne: ok, nailed it.
virtgpu_gem_prime_import_sg_table is implemented as a stub that returns ENODEV. shouldn't be too hard to fix, i'll add this to my queue. | [23:04] | |
| ......... (idle for 44mn) | ||
| *** | pfallenop has quit IRC (Ping timeout: 245 seconds) | [23:50] |
| ↑back Search ←Prev date Next date→ Show only urls | (Click on time to select a line by its url) |