[Mesa-dev] [PATCH v2 3/5] i965: perf: read slice/unslice frequencies from OA reports

[Lionel Landwerlin]

On 17/04/18 23:03, Kenneth Graunke wrote:
> On Tuesday, April 17, 2018 4:45:13 PM PDT Lionel Landwerlin wrote:
>> On 16/04/18 23:16, Kenneth Graunke wrote:
>>> On Tuesday, April 3, 2018 7:48:11 AM PDT Lionel Landwerlin wrote:
>>>> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
>>>> ---
>>>>    src/mesa/drivers/dri/i965/brw_performance_query.c | 37 +++++++++++++++++++++++
>>>>    src/mesa/drivers/dri/i965/brw_performance_query.h | 12 ++++++++
>>>>    2 files changed, 49 insertions(+)
>>>>
>>>> diff --git a/src/mesa/drivers/dri/i965/brw_performance_query.c b/src/mesa/drivers/dri/i965/brw_performance_query.c
>>>> index 32cf96a333d..d20bc58ee33 100644
>>>> --- a/src/mesa/drivers/dri/i965/brw_performance_query.c
>>>> +++ b/src/mesa/drivers/dri/i965/brw_performance_query.c
>>>> @@ -1341,6 +1341,42 @@ brw_is_perf_query_ready(struct gl_context *ctx,
>>>>       return false;
>>>>    }
>>>>    
>>>> +static void
>>>> +gen8_read_report_clock_ratios(const uint32_t *report,
>>>> +                              uint64_t *slice_freq_hz,
>>>> +                              uint64_t *unslice_freq_hz)
>>>> +{
>>>> +   uint32_t unslice_freq = report[0] & 0x1ff;
>>>> +   uint32_t slice_freq_low = (report[0] >> 25) & 0x7f;
>>>> +   uint32_t slice_freq_high = (report[0] >> 9) & 0x3;
>>>> +   uint32_t slice_freq = slice_freq_low | (slice_freq_high << 7);
>>>> +
>>>> +   *slice_freq_hz = slice_freq * 16666000ULL;
>>>> +   *unslice_freq_hz = unslice_freq * 16666000ULL;
>>> What is this reading?  Could we add a comment to the function?
>>> Otherwise, I'm not sure where to even begin to look for this in docs.
>> Apologies, I've just added a fairly descriptive comment. I should have
>> put that in the first version.
>>
>>>> +}
>>>> +
>>>> +static void
>>>> +read_slice_unslice_frequencies(struct brw_context *brw,
>>>> +                               struct brw_perf_query_object *obj)
>>>> +{
>>>> +   const struct gen_device_info *devinfo = &brw->screen->devinfo;
>>>> +   uint32_t *begin_report, *end_report;
>>>> +
>>>> +   /* No data available before gen8. */
>>>> +   if (devinfo->gen < 8)
>>>> +      return;
>>>> +
>>>> +   begin_report = obj->oa.map;
>>>> +   end_report = obj->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
>>> I'm probably missing something obvious here, but...this looks fishy.
>>>
>>> What actually writes these values to the OA buffer?  And here, you're
>>> reading the starting value from DWord 0 of the buffer...is that really
>>> where those values live?
>>>
>>> I guess I would have expected to see a capture_uncore_frequencies()
>>> and MI_SLICE_FREQ_START_OFFSET_BYTES or something, like patch 2/5 does.
>> We read these values from the OA reports. It's bits mashed into the
>> RPT_ID field.
>> RPT_ID is the first dword indeed.
> Oh, wow, so it is.  I hadn't realized until talking to you today and
> reading this that RPT_ID has all kinds of crazy information jammed into
> it.  I thought it was just an integer identifier.
>
> This is what I get for not reading those docs in a few generations :)
>
> Looking at the docs for RPT_ID, I'm still not fully tracking, so I
> guess I'll look forward to reading your comment...
>
> Thanks for helping me understand all this.
No worries, if it's not obvious from the patch, it probably needs 
explaining.
I sent a v3 (in reply to the v2 patch), that should have some 
explanation which is somewhat clearer than what is in the documentation.

-
Lionel