[PATCH RFC 1/3] drm/msm/adreno: Add support for ACD

Akhil P Oommen quic_akhilpo at quicinc.com
Thu Nov 7 14:35:49 UTC 2024


On 11/7/2024 8:01 PM, neil.armstrong at linaro.org wrote:
> On 07/11/2024 13:46, Akhil P Oommen wrote:
>> On 11/7/2024 2:25 PM, neil.armstrong at linaro.org wrote:
>>> On 06/11/2024 02:44, Akhil P Oommen wrote:
>>>> On 11/4/2024 9:14 PM, neil.armstrong at linaro.org wrote:
>>>>> On 11/10/2024 22:29, Akhil P Oommen wrote:
>>>>>> ACD a.k.a Adaptive Clock Distribution is a feature which helps to
>>>>>> reduce
>>>>>> the power consumption. In some chipsets, it is also a requirement to
>>>>>> support higher GPU frequencies. This patch adds support for GPU
>>>>>> ACD by
>>>>>> sending necessary data to GMU and AOSS. The feature support for the
>>>>>> chipset is detected based on devicetree data.
>>>>>>
>>>>>> Signed-off-by: Akhil P Oommen <quic_akhilpo at quicinc.com>
>>>>>> ---
>>>>>>     drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 81 +++++++++++++++++++
>>>>>> ++++++
>>>>>> +++-------
>>>>>>     drivers/gpu/drm/msm/adreno/a6xx_gmu.h |  1 +
>>>>>>     drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 36 ++++++++++++++++
>>>>>>     drivers/gpu/drm/msm/adreno/a6xx_hfi.h | 21 +++++++++
>>>>>>     4 files changed, 124 insertions(+), 15 deletions(-)
>>>>>>
>>>>>
>>>>> <snip>
>>>>>
>>>>>> +
>>>>>> +static int a6xx_hfi_enable_acd(struct a6xx_gmu *gmu)
>>>>>> +{
>>>>>> +    struct a6xx_hfi_acd_table *acd_table = &gmu->acd_table;
>>>>>> +    struct a6xx_hfi_msg_feature_ctrl msg = {
>>>>>> +        .feature = HFI_FEATURE_ACD,
>>>>>> +        .enable = 1,
>>>>>> +        .data = 0,
>>>>>> +    };
>>>>>> +    int ret;
>>>>>> +
>>>>>> +    if (!acd_table->enable_by_level)
>>>>>> +        return 0;
>>>>>> +
>>>>>> +    /* Enable ACD feature at GMU */
>>>>>> +    ret = a6xx_hfi_send_msg(gmu, HFI_H2F_FEATURE_CTRL, &msg,
>>>>>> sizeof(msg), NULL, 0);
>>>>>> +    if (ret) {
>>>>>> +        DRM_DEV_ERROR(gmu->dev, "Unable to enable ACD (%d)\n", ret);
>>>>>> +        return ret;
>>>>>> +    }
>>>>>> +
>>>>>> +    /* Send ACD table to GMU */
>>>>>> +    ret = a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_ACD, &msg, sizeof(msg),
>>>>>> NULL, 0);
>>>>>
>>>>> This looks wrong, in this exact code, you never use the acd_table...
>>>>> perhaps it should be acd_table here
>>>>
>>>> Whoops! Weirdly gmu didn't explode when I tested.
>>>>
>>>> Thanks for your keen eye.
>>>
>>> You're welcome !
>>>
>>> I've been trying to enable this on SM8650, but HFI_H2F_MSG_ACD fails.
>>>
>>> My changes:
>>> ================><================================
>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/
>>> msm/adreno/a6xx_hfi.c
>>> index 7c96d6f8aaa9..bd9d586f245e 100644
>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
>>> @@ -682,7 +682,7 @@ static int a6xx_hfi_enable_acd(struct a6xx_gmu *gmu)
>>>          }
>>>
>>>          /* Send ACD table to GMU */
>>> -       ret = a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_ACD, &acd_table,
>>> sizeof(*acd_table), NULL, 0);
>>> +       ret = a6xx_hfi_send_msg(gmu, HFI_H2F_MSG_ACD, &acd_table,
>>
>> &acd_table -> acd_table here?
> 
> Damn, good catch !
> 
> Ok so it didn't explode anymore, but still fails:
> =====
> [    7.083258] platform 3d6a000.gmu: [drm:a6xx_hfi_send_msg.constprop.0
> [msm]] *ERROR* Message HFI_H2F_MSG_GX_BW_PERF_VOTE id 7 timed out
> waiting for response
> [    7.149709] platform 3d6a000.gmu: [drm:a6xx_hfi_send_msg.constprop.0
> [msm]] *ERROR* Unexpected message id 7 on the response queue
> [    7.149744] platform 3d6a000.gmu: [drm:a6xx_hfi_send_msg.constprop.0
> [msm]] *ERROR* The HFI response queue is unexpectedly empty
> [    7.165163] platform 3d6a000.gmu: [drm:a6xx_hfi_send_msg.constprop.0
> [msm]] *ERROR* Unexpected message id 8 on the response queue
> [    7.165188] platform 3d6a000.gmu: [drm:a6xx_hfi_send_msg.constprop.0
> [msm]] *ERROR* The HFI response queue is unexpectedly empty
> ====
> 
> Seems with ACD enabled, first vote can take up to 100ms, and downstream
> has 1s timeout, with a larger timeout I got it to work !

Yes, there is an additional overhead during first perf vote. Thanks for
the heads up. I am yet to test with fixes.

-Akhil.

> 
> Thanks,
> Neil
>>
>> -Akhil
>>
>>> sizeof(struct a6xx_hfi_acd_table), NULL, 0);
>>>          if (ret) {
>>>                  DRM_DEV_ERROR(gmu->dev, "Unable to send ACD table
>>> (%d)\n", ret);
>>>                  return ret;
>>> ================><================================
>>>
>>> with the appropriate qcom,opp-acd-level in DT taken from downstream,
>>> I get:
>>> [    6.946184] platform 3d6a000.gmu: [drm:a6xx_hfi_send_msg.constprop.0
>>> [msm]] *ERROR* Message (null) id 4 timed out waiting for response
>>> [    6.958697] platform 3d6a000.gmu: [drm:a6xx_hfi_start [msm]] *ERROR*
>>> Unable to send ACD table (-110)
>>>
>>> is there something missing ?
>>>
>>> Neil
>>>
>>>>
>>>> -Akhil.
>>>>
>>>>>
>>>>>> +    if (ret) {
>>>>>> +        DRM_DEV_ERROR(gmu->dev, "Unable to ACD table (%d)\n", ret);
>>>>>> +        return ret;
>>>>>> +    }
>>>>>> +
>>>>>> +    return 0;
>>>>>> +}
>>>>>> +
>>>>>>     static int a6xx_hfi_send_test(struct a6xx_gmu *gmu)
>>>>>>     {
>>>>>>         struct a6xx_hfi_msg_test msg = { 0 };
>>>>>> @@ -756,6 +788,10 @@ int a6xx_hfi_start(struct a6xx_gmu *gmu, int
>>>>>> boot_state)
>>>>>>         if (ret)
>>>>>>             return ret;
>>>>>>     +    ret = a6xx_hfi_enable_acd(gmu);
>>>>>> +    if (ret)
>>>>>> +        return ret;
>>>>>> +
>>>>>>         ret = a6xx_hfi_send_core_fw_start(gmu);
>>>>>>         if (ret)
>>>>>>             return ret;
>>>>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.h b/drivers/gpu/drm/
>>>>>> msm/adreno/a6xx_hfi.h
>>>>>> index 528110169398..51864c8ad0e6 100644
>>>>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.h
>>>>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.h
>>>>>> @@ -151,12 +151,33 @@ struct a6xx_hfi_msg_test {
>>>>>>         u32 header;
>>>>>>     };
>>>>>>     +#define HFI_H2F_MSG_ACD 7
>>>>>> +#define MAX_ACD_STRIDE 2
>>>>>> +
>>>>>> +struct a6xx_hfi_acd_table {
>>>>>> +    u32 header;
>>>>>> +    u32 version;
>>>>>> +    u32 enable_by_level;
>>>>>> +    u32 stride;
>>>>>> +    u32 num_levels;
>>>>>> +    u32 data[16 * MAX_ACD_STRIDE];
>>>>>> +};
>>>>>> +
>>>>>>     #define HFI_H2F_MSG_START 10
>>>>>>       struct a6xx_hfi_msg_start {
>>>>>>         u32 header;
>>>>>>     };
>>>>>>     +#define HFI_H2F_FEATURE_CTRL 11
>>>>>> +
>>>>>> +struct a6xx_hfi_msg_feature_ctrl {
>>>>>> +    u32 header;
>>>>>> +    u32 feature;
>>>>>> +    u32 enable;
>>>>>> +    u32 data;
>>>>>> +};
>>>>>> +
>>>>>>     #define HFI_H2F_MSG_CORE_FW_START 14
>>>>>>       struct a6xx_hfi_msg_core_fw_start {
>>>>>>
>>>>>
>>>>> Thanks,
>>>>> Neil
>>>>
>>>
>>
> 



More information about the Freedreno mailing list