[PATCH v2 5/8] RFC drm/xe/uapi: Add configs for Engine busyness

[Riana Tauro]

Hi Umesh

On 12/21/2023 7:59 AM, Umesh Nerlige Ramappa wrote:
> On Thu, Dec 07, 2023 at 06:27:59PM +0530, Riana Tauro wrote:
>> GuC provides engine busyness ticks as a 64 bit counter which count
>> as clock ticks.
>>
>> Add configs to the uapi to expose Engine busyness via PMU.
>>
>> v2: add "__" prefix for internal helpers
>>    add a simple helper for application usage (Aravind)
>>
>> Cc: Aravind Iddamsetty <aravind.iddamsetty at linux.intel.com>
>> Signed-off-by: Riana Tauro <riana.tauro at intel.com>
>> ---
>> include/uapi/drm/xe_drm.h | 23 +++++++++++++++++++++++
>> 1 file changed, 23 insertions(+)
>>
>> diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
>> index b5e7a4f673fa..074d63cb79df 100644
>> --- a/include/uapi/drm/xe_drm.h
>> +++ b/include/uapi/drm/xe_drm.h
>> @@ -1107,6 +1107,10 @@ struct drm_xe_wait_user_fence {
>>  *    fd = syscall(__NR_perf_event_open, &attr, -1, cpu, -1, 0);
>>  */
>>
>> +enum drm_xe_pmu_engine_sample {
>> +    DRM_XE_PMU_SAMPLE_BUSY_TICKS = 0,
>> +};
>> +
>> /*
>>  * Top bits of every counter are GT id.
>>  */
>> @@ -1115,6 +1119,25 @@ struct drm_xe_wait_user_fence {
>> #define ___DRM_XE_PMU_OTHER(gt, x) \
>>     (((__u64)(x)) | ((__u64)(gt) << __DRM_XE_PMU_GT_SHIFT))
>>
>> +#define __DRM_XE_PMU_SAMPLE_BITS (4)
>> +#define __DRM_XE_PMU_SAMPLE_INSTANCE_BITS (8)
>> +#define __DRM_XE_PMU_CLASS_SHIFT \
>> +    (__DRM_XE_PMU_SAMPLE_BITS + __DRM_XE_PMU_SAMPLE_INSTANCE_BITS)
>> +
>> +/*
>> + * Engine configs offset - 0x1000
>> + */
>> +#define __DRM_XE_PMU_ENGINE_OFFSET(gt) \
>> +    (___DRM_XE_PMU_OTHER(gt, 0xfff) + 1)
> 
> Thinking this should be (___DRM_XE_PMU_OTHER(gt, 0xfffff) + 1) because 
> class is also 8 bits. This part is not any different from i915, so not 
> sure if you intended to change it for XE.
> 
>> +
>> +#define __DRM_XE_PMU_ENGINE(gt, class, instance, sample) \
>> +    (((class) << __DRM_XE_PMU_CLASS_SHIFT | \
>> +    (instance) << __DRM_XE_PMU_SAMPLE_BITS | \
>> +    (sample)) + __DRM_XE_PMU_ENGINE_OFFSET(gt))
> 
> What's the idea here? Engines are limited to a specific GT and the user 
> is also passing in the GT in the config?
Yeah, the engine numbering is same for all GT's. So configs will be same 
if we don't consider the GT. Using GT to create different config ID.

Rev3 has changed the config id's without engine offset as group busyness 
was removed

https://patchwork.freedesktop.org/patch/572085/?series=126919&rev=3
> 
> I think the gt should just be shifted to __DRM_XE_PMU_GT_SHIFT and ORed 
> with the engine counter, more like:
> 
> #define ___DRM_XE_PMU_GT_EVENT(gt, x) \
>      (((__u64)(x)) | ((__u64)(gt) << __DRM_XE_PMU_GT_SHIFT))
> 
> 
> #define __DRM_XE_PMU_ENGINE_EVENT(class, instance, sample) \
>      (((class) << __DRM_XE_PMU_CLASS_SHIFT | \
>      (instance) << __DRM_XE_PMU_SAMPLE_BITS | \
>      (sample)))
> 
> #define DRM_XE_PMU_ENGINE_BUSY_TICKS(gt, class, instance) \
>      ___DRM_XE_PMU_GT_EVENT(gt, __DRM_XE_PMU_ENGINE_EVENT(class, 
> instance, DRM_XE_PMU_SAMPLE_BUSY_TICKS))
> 
> Just an example, naming is up to you.
Will modify this similar to the example.
> 
> For group counters, the logic you have is fine as long as you set class 
> to 8 bits:

Group busyness was added by Aravind but has been removed now from pmu. 
In the next rev will add similar to this for total active ticks

Thanks
Riana

> 
> #define DRM_XE_PMU_RENDER_GROUP_BUSY(gt) \
>      ___DRM_XE_PMU_GT_EVENT(gt, ___DRM_XE_PMU_GT_EVENT(gt, 0xfffff) + 1 
> + 0)
> 
> #define DRM_XE_PMU_COPY_GROUP_BUSY(gt) \
>      ___DRM_XE_PMU_GT_EVENT(gt, ___DRM_XE_PMU_GT_EVENT(gt, 0xfffff) + 1 
> + 1)
> 
> ___DRM_XE_PMU_GT_EVENT(gt, 0xfffff) is same as 
> __DRM_XE_PMU_ENGINE_EVENT(0xff, 0xff, 0xf).
> 
> Thanks,
> Umesh
> 
> 
>> +
>> +#define DRM_XE_PMU_ENGINE_BUSY_TICKS(gt, class, instance) \
>> +    __DRM_XE_PMU_ENGINE(gt, class, instance, 
>> DRM_XE_PMU_SAMPLE_BUSY_TICKS)
>> +
>> #define DRM_XE_PMU_RENDER_GROUP_BUSY(gt)    ___DRM_XE_PMU_OTHER(gt, 0)
>> #define DRM_XE_PMU_COPY_GROUP_BUSY(gt)        ___DRM_XE_PMU_OTHER(gt, 1)
>> #define DRM_XE_PMU_MEDIA_GROUP_BUSY(gt)        ___DRM_XE_PMU_OTHER(gt, 2)
>> -- 
>> 2.40.0
>>