[Freedreno] [PATCH 5/6] drm/msm/a6xx: Use GMU_ALWAYS_ON_COUNTER for GMU-equipped GPUs in timestamp

[Jonathan Marek]

On 5/17/23 3:08 PM, Konrad Dybcio wrote:
> 
> 
> On 17.05.2023 20:09, Jonathan Marek wrote:
>> AFAIK GMU_ALWAYS_ON_COUNTER does not have the same value as CP_ALWAYS_ON_COUNTER (only the same frequency), so changing this would break userspace expecting to be able to compare the value returned by MSM_PARAM_TIMESTAMP with CP timestamp values.
> FWIW A630 and A730 seem to work fine with this patch. Anything
> in particular I should look out for?

mesa uses it for perfetto tracing to synchronize the GPU/CPU timelines.

> 
> Konrad
>>
>> On 5/17/23 12:50 PM, Konrad Dybcio wrote:
>>> Use the always-on counter provided by the GMU to skip having to
>>> keep the GPU online.
>>>
>>> Signed-off-by: Konrad Dybcio <konrad.dybcio at linaro.org>
>>> ---
>>>    drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 9 +++------
>>>    1 file changed, 3 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>>> index 8707e8b6ac7e..d2a999b90589 100644
>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
>>> @@ -1664,12 +1664,9 @@ static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
>>>          mutex_lock(&a6xx_gpu->gmu.lock);
>>>    -    /* Force the GPU power on so we can read this register */
>>> -    a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
>>> -
>>> -    *value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER);
>>> -
>>> -    a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
>>> +    *value = gmu_read64(&a6xx_gpu->gmu,
>>> +                REG_A6XX_GMU_ALWAYS_ON_COUNTER_L,
>>> +                REG_A6XX_GMU_ALWAYS_ON_COUNTER_H);
>>>          mutex_unlock(&a6xx_gpu->gmu.lock);
>>>