[Intel-gfx] [RFC 0/4] GPU/CPU timestamps correlation for relating OA samples with system events

[Sagar Arun Kamble]

On 12/7/2017 1:32 AM, Lionel Landwerlin wrote:
> I've put together some trival IGT tests : 
> https://github.com/djdeath/intel-gpu-tools/commits/wip/djdeath/cpu-timestamps
> With a few changes which I pointed in the review : 
> https://github.com/djdeath/linux/commit/d0e4cf4d3f464491b4ffe97d112284d1ce73656d
>
> Put together it seems to work relatively well.
> There is still a small drift happening between the 2 timestamps. I've 
> noticed over a 160ms of OA reports, there is a accumulated difference 
> of ~35us between the GPU timestamp and cpu timestamps.
> I may be doing something wrong with the scaling in the tests, or maybe 
> there is an issue in the kernel, or both.
Went through the testcase. scaled_gpu_delta calculation is same as what 
timecounter does in kernel for calculating system_time corresponding
to gpu timestamp hence we don't see much of delta between these two 
times/time deltas.
Ideally we should be testing the system time delta by sampling system 
time and gpu timestamp atomically (which isn't feasible unless we do 
some precise adjustments)
I have attempted to check these two times/delta by sampling them in 
debugfs and reading over variable periods manually and checking the delta.
https://github.com/sakamble/i915-timestamp-support/commit/03be3056752d7b05a02cd01f5c20b3fcfcf18395

It is showing that delta is less than 10us in most cases for 30min and 
occasionally around 20-30us. Again this delta might be because of 
initial time setting as well.
timecounter initializing system time and gpu clocks as pair should be 
highly accurate for which I have currently taken 350us start offset.

I think gpu timestamp clock is highly stable as seen in my testing on 
SKL. I think this clock is used for calculations in GuC too so will
be good to get confirmation about the stability of this clock from them 
and HW team.

    systemd-udevd-169   [001] ....     3.035812: i915_driver_load: sys 
start time: 1512308011156099790

    systemd-udevd-169   [001] d...     3.036012: i915_cyclecounter_read: 
52025098974

              cat-1654  [001] ....    52.407957: i915_cyclecounter_read: 
52617562292

              cat-1654  [001] ....    52.407958: i915_timestamp_info: 
sys time: 1512308060527894638

              cat-1654  [001] ....    52.407958: i915_timestamp_info: 
ts: 52617562292 device time: 1512308060528043050

              cat-1684  [001] ....   177.239733: i915_cyclecounter_read: 
54115543581

              cat-1684  [001] ....   177.239736: i915_timestamp_info: 
sys time: 1512308185359666602

              cat-1684  [001] ....   177.239737: i915_timestamp_info: 
ts: 54115543581 device time: 1512308185359817372

              cat-1693  [001] ....   329.820374: i915_cyclecounter_read: 
55946511277

              cat-1693  [001] ....   329.820377: i915_timestamp_info: 
sys time: 1512308337940301732

              cat-1693  [001] ....   329.820378: i915_timestamp_info: 
ts: 55946511277 device time: 1512308337940458996


<delta between system time delta and gpu time delta for above two 
samples (177, 329) = 6494ns>


              cat-1702  [001] ....   506.980313: i915_cyclecounter_read: 
58072430542

              cat-1702  [001] ....   506.980315: i915_timestamp_info: 
sys time: 1512308515100233102

              cat-1702  [001] ....   506.980317: i915_timestamp_info: 
ts: 58072430542 device time: 1512308515100398084

<delta between system time delta and gpu time delta for above two 
samples (329, 506) = 6494ns>

>
> I'll build the GPUTop parts and see if the results make sense.
>
> Thanks!,
>
> -
> Lionel
>
> On 15/11/17 12:13, Sagar Arun Kamble wrote:
>> We can compute system time corresponding to GPU timestamp by taking a
>> reference point (CPU monotonic time, GPU timestamp) and then adding
>> delta time computed using timecounter/cyclecounter support in kernel.
>> We have to configure cyclecounter with the GPU timestamp frequency.
>> Earlier approach that was based on cross-timestamp is not needed. It
>> was being used to approximate the frequency based on invalid assumptions
>> (possibly drift was being seen in the time due to precision issue).
>> The precision of time from GPU clocks is already in ns and timecounter
>> takes care of it as verified over variable durations.
>>
>> This series adds base timecounter/cyclecounter changes and changes to
>> get GPU and CPU timestamps in OA samples.
>>
>> Sagar Arun Kamble (1):
>>    drm/i915/perf: Add support to correlate GPU timestamp with system 
>> time
>>
>> Sourab Gupta (3):
>>    drm/i915/perf: Add support for collecting 64 bit timestamps with OA
>>      reports
>>    drm/i915/perf: Extract raw GPU timestamps from OA reports
>>    drm/i915/perf: Send system clock monotonic time in perf samples
>>
>>   drivers/gpu/drm/i915/i915_drv.h  |  11 ++++
>>   drivers/gpu/drm/i915/i915_perf.c | 124 
>> ++++++++++++++++++++++++++++++++++++++-
>>   drivers/gpu/drm/i915/i915_reg.h  |   6 ++
>>   include/uapi/drm/i915_drm.h      |  14 +++++
>>   4 files changed, 154 insertions(+), 1 deletion(-)
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-gfx/attachments/20171222/e3a982cf/attachment-0001.html>