Re: [PATCH v3 2/5] drm/amdgpu: add new functions to set GPU power profile

["Lazar, Lijo" <lijo.lazar@amd.com>]

On 9/27/2022 6:23 PM, Sharma, Shashank wrote:
> 
> 
> On 9/27/2022 2:39 PM, Lazar, Lijo wrote:
>>
>>
>> On 9/27/2022 5:53 PM, Sharma, Shashank wrote:
>>>
>>>
>>> On 9/27/2022 2:10 PM, Lazar, Lijo wrote:
>>>>
>>>>
>>>> On 9/27/2022 5:11 PM, Sharma, Shashank wrote:
>>>>>
>>>>>
>>>>> On 9/27/2022 11:58 AM, Lazar, Lijo wrote:
>>>>>>
>>>>>>
>>>>>> On 9/27/2022 3:10 AM, Shashank Sharma wrote:
>>>>>>> This patch adds new functions which will allow a user to
>>>>>>> change the GPU power profile based a GPU workload hint
>>>>>>> flag.
>>>>>>>
>>>>>>> Cc: Alex Deucher <alexander.deucher@amd.com>
>>>>>>> Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
>>>>>>> ---
>>>>>>>   drivers/gpu/drm/amd/amdgpu/Makefile           |  2 +-
>>>>>>>   .../gpu/drm/amd/amdgpu/amdgpu_ctx_workload.c  | 97 
>>>>>>> +++++++++++++++++++
>>>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  1 +
>>>>>>>   .../gpu/drm/amd/include/amdgpu_ctx_workload.h | 54 +++++++++++
>>>>>>>   drivers/gpu/drm/amd/pm/inc/amdgpu_dpm.h       |  5 +
>>>>>>>   5 files changed, 158 insertions(+), 1 deletion(-)
>>>>>>>   create mode 100644 
>>>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ctx_workload.c
>>>>>>>   create mode 100644 
>>>>>>> drivers/gpu/drm/amd/include/amdgpu_ctx_workload.h
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/Makefile 
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/Makefile
>>>>>>> index 5a283d12f8e1..34679c657ecc 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/Makefile
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/Makefile
>>>>>>> @@ -50,7 +50,7 @@ amdgpu-y += amdgpu_device.o amdgpu_kms.o \
>>>>>>>       atombios_dp.o amdgpu_afmt.o amdgpu_trace_points.o \
>>>>>>>       atombios_encoders.o amdgpu_sa.o atombios_i2c.o \
>>>>>>>       amdgpu_dma_buf.o amdgpu_vm.o amdgpu_vm_pt.o amdgpu_ib.o 
>>>>>>> amdgpu_pll.o \
>>>>>>> -    amdgpu_ucode.o amdgpu_bo_list.o amdgpu_ctx.o amdgpu_sync.o \
>>>>>>> +    amdgpu_ucode.o amdgpu_bo_list.o amdgpu_ctx.o 
>>>>>>> amdgpu_ctx_workload.o amdgpu_sync.o \
>>>>>>>       amdgpu_gtt_mgr.o amdgpu_preempt_mgr.o amdgpu_vram_mgr.o 
>>>>>>> amdgpu_virt.o \
>>>>>>>       amdgpu_atomfirmware.o amdgpu_vf_error.o amdgpu_sched.o \
>>>>>>>       amdgpu_debugfs.o amdgpu_ids.o amdgpu_gmc.o \
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx_workload.c 
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx_workload.c
>>>>>>> new file mode 100644
>>>>>>> index 000000000000..a11cf29bc388
>>>>>>> --- /dev/null
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx_workload.c
>>>>>>> @@ -0,0 +1,97 @@
>>>>>>> +/*
>>>>>>> + * Copyright 2022 Advanced Micro Devices, Inc.
>>>>>>> + *
>>>>>>> + * Permission is hereby granted, free of charge, to any person 
>>>>>>> obtaining a
>>>>>>> + * copy of this software and associated documentation files (the 
>>>>>>> "Software"),
>>>>>>> + * to deal in the Software without restriction, including 
>>>>>>> without limitation
>>>>>>> + * the rights to use, copy, modify, merge, publish, distribute, 
>>>>>>> sublicense,
>>>>>>> + * and/or sell copies of the Software, and to permit persons to 
>>>>>>> whom the
>>>>>>> + * Software is furnished to do so, subject to the following 
>>>>>>> conditions:
>>>>>>> + *
>>>>>>> + * The above copyright notice and this permission notice shall 
>>>>>>> be included in
>>>>>>> + * all copies or substantial portions of the Software.
>>>>>>> + *
>>>>>>> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY 
>>>>>>> KIND, EXPRESS OR
>>>>>>> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 
>>>>>>> MERCHANTABILITY,
>>>>>>> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO 
>>>>>>> EVENT SHALL
>>>>>>> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, 
>>>>>>> DAMAGES OR
>>>>>>> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 
>>>>>>> OTHERWISE,
>>>>>>> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR 
>>>>>>> THE USE OR
>>>>>>> + * OTHER DEALINGS IN THE SOFTWARE.
>>>>>>> + *
>>>>>>> + */
>>>>>>> +#include <drm/drm.h>
>>>>>>> +#include "kgd_pp_interface.h"
>>>>>>> +#include "amdgpu_ctx_workload.h"
>>>>>>> +
>>>>>>> +static enum PP_SMC_POWER_PROFILE
>>>>>>> +amdgpu_workload_to_power_profile(uint32_t hint)
>>>>>>> +{
>>>>>>> +    switch (hint) {
>>>>>>> +    case AMDGPU_CTX_WORKLOAD_HINT_NONE:
>>>>>>> +    default:
>>>>>>> +        return PP_SMC_POWER_PROFILE_BOOTUP_DEFAULT;
>>>>>>> +
>>>>>>> +    case AMDGPU_CTX_WORKLOAD_HINT_3D:
>>>>>>> +        return PP_SMC_POWER_PROFILE_FULLSCREEN3D;
>>>>>>> +    case AMDGPU_CTX_WORKLOAD_HINT_VIDEO:
>>>>>>> +        return PP_SMC_POWER_PROFILE_VIDEO;
>>>>>>> +    case AMDGPU_CTX_WORKLOAD_HINT_VR:
>>>>>>> +        return PP_SMC_POWER_PROFILE_VR;
>>>>>>> +    case AMDGPU_CTX_WORKLOAD_HINT_COMPUTE:
>>>>>>> +        return PP_SMC_POWER_PROFILE_COMPUTE;
>>>>>>> +    }
>>>>>>> +}
>>>>>>> +
>>>>>>> +int amdgpu_set_workload_profile(struct amdgpu_device *adev,
>>>>>>> +                uint32_t hint)
>>>>>>> +{
>>>>>>> +    int ret = 0;
>>>>>>> +    enum PP_SMC_POWER_PROFILE profile =
>>>>>>> +            amdgpu_workload_to_power_profile(hint);
>>>>>>> +
>>>>>>> +    if (adev->pm.workload_mode == hint)
>>>>>>> +        return 0;
>>>>>>> +
>>>>>>> +    mutex_lock(&adev->pm.smu_workload_lock);
>>>>>>
>>>>>> If it's all about pm subsystem variable accesses, this API should 
>>>>>> rather be inside amd/pm subsystem. No need to expose the variable 
>>>>>> outside pm subsytem. Also currently all amdgpu_dpm* calls are 
>>>>>> protected under one mutex. Then this extra lock won't be needed.
>>>>>>
>>>>>
>>>>> This is tricky, this is not all about PM subsystem. Note that the 
>>>>> job management and scheduling is handled into amdgpu_ctx, so the 
>>>>> workload hint is set in context_management API. The API is consumed 
>>>>> when the job is actually run from amdgpu_run() layer. So its a 
>>>>> joint interface between context and PM.
>>>>>
>>>>
>>>> If you take out amdgpu_workload_to_power_profile() line, everything 
>>>> else looks to touch only pm variables/functions. 
>>>
>>> That's not a line, that function converts a AMGPU_CTX hint to PPM 
>>> profile. And going by that logic, this whole code was kept in the 
>>> amdgpu_ctx.c file as well, coz this code is consuming the PM API. So 
>>> to avoid these conflicts and having a new file is a better idea.
>>>
>>> You could still keep a
>>>> wrapper though. Also dpm_* functions are protected, so the extra 
>>>> mutex can be avoided as well.
>>>>
>>> The lock also protects pm.workload_mode writes.
>>>
>>>>>>> +
>>>>>>> +    if (adev->pm.workload_mode == hint)
>>>>>>> +        goto unlock;
>>>>>>> +
>>>>>>> +    ret = amdgpu_dpm_switch_power_profile(adev, profile, 1);
>>>>>>> +    if (!ret)
>>>>>>> +        adev->pm.workload_mode = hint;
>>>>>>> +    atomic_inc(&adev->pm.workload_switch_ref);
>>>>>>
>>>>>> Why is this reference kept? The swtiching happens inside a lock 
>>>>>> and there is already a check not to switch if the hint matches 
>>>>>> with current workload.
>>>>>>
>>>>>
>>>>> This reference is kept so that we would not reset the PM mode to 
>>>>> DEFAULT when some other context has switched the PP mode. If you 
>>>>> see the 4th patch, the PM mode will be changed when the job in that 
>>>>> context is run, and a pm_reset function will be scheduled when the 
>>>>> job is done. But in between if another job from another context has 
>>>>> changed the PM mode, the refrence count will prevent us from 
>>>>> resetting the PM mode.
>>>>>
>>>>
>>>> This helps only if multiple jobs request the same mode. If they 
>>>> request different modes, then this is not helping much.
>>>
>>> No that's certainly not the case. It's a counter, whose aim is to 
>>> allow a PP reset only when the counter is 0. Do note that the reset() 
>>> happens only in the job_free_cb(), which gets schedule later. If this 
>>> counter is not zero, which means another work has changed the profile 
>>> in between, and we should not reset it.
>>>
>>>>
>>>> It could be useful to profile some apps assuming it has exclusive 
>>>> access.
>>>>
>>>> However, in general, the API is not reliable from a user point as 
>>>> the mode requested can be overridden by some other job. Then a 
>>>> better thing to do is to document that and avoid the extra stuff 
>>>> around it.
>>>>
>>> As I mentioned before, like any PM feature, the benefits can be seen 
>>> only while running consistant workloads for long time. I an still add 
>>> a doc note in the UAPI page.
>>>
>>
>>
>> a) What is the goal of the API? Is it guaranteeing the job to run 
>> under a workprofile mode or something else?
> 
> No, it does not guarentee anything. If you see the cover letter, it just 
> provides an interface to an app to submit workload under a power profile 
> which can be more suitable for its workload type. As I mentioned, it 
> could be very useful for many scenarios like fullscreen 3D / fullscreen 
> MM scenarios. It could also allow a system-gfx-manager to shift load 
> balance towards one type of workload. There are many applications, once 
> the UAPI is in place.
> 
>>
>> b) If it's to guarantee work profile mode, does it really guarantee 
>> that - the answer is NO when some other job is running. It may or may 
>> not work is the answer.
>>
>> c) What is the difference between one job resetting the profile mode 
>> to NONE vs another job change the mode to say VIDEO when the original 
>> request is for COMPUTE? While that is the case, what is the use of any 
>> sort of 'pseudo-protection' other than running some code to do extra 
>> lock/unlock stuff.
>>
> 
> Your understanding of protection is wrong here. There is intentionally 
> no protection for a job changing another job's set workload profile, coz 
> in that was we will end up seriazling/bottlenecking workload submission 
> until PM profile is ready to be changed, which takes away benefit of 
> having multiple queues of parallel submission.
> 
> The protection provided by the ref counter is to avoid the clearing of 
> the profile (to NONE), while another workload is in execution. The 
> difference between NONE and VIDEO is still that NONE is the default 
> profile without any fine tuning, and VIDEO is still fine tuned for VIDEO 
> type of workloads.
> 

Protection 1 is - mutex_lock(&adev->pm.smu_workload_lock);

The line that follows is amdgpu_dpm_switch_power_profile() - this one 
will allow only single client use- two jobs won't be able to switch at 
the same time. All *dpm* APIs are protected like that.

Protection 2 is - ref counter.

It helps only in this kind of scenario when two jobs requested the same 
mode successively -
	Job 1 requested compute
	Job 2 requested compute
	Job 1 ends (doesnt't reset)

Scenario - 2
	Job 1 requested compute
	Job 2 requested compute
	Job 3 requested 3D
	Job 1 ends (doesnt't reset, it continues in 3D)

In this mixed scenario case, I would say NONE is much more optimized as 
it's under FW control. Actually, it does much more fine tuning because 
of its background data collection.

> In the end, *again* the actual benefit comes when consistant workload is 
> submitted for a long time, like fullscreen 3D game playback, fullscreen 
> Video movie playback, and so on.
> 

"only under consistent", doesn't justify any software protection logic. 
Again, if the workload is consistent most likely PMFW could be managing 
it better.

Thanks,
Lijo

> - Shashank
> 
>> Thanks,
>> Lijo
>>
>>> - Shashank
>>>
>>>> Thanks,
>>>> Lijo
>>>>
>>>>> - Shashank
>>>>>
>>>>>> Thanks,
>>>>>> Lijo
>>>>>>
>>>>>>> +
>>>>>>> +unlock:
>>>>>>> +    mutex_unlock(&adev->pm.smu_workload_lock);
>>>>>>> +    return ret;
>>>>>>> +}
>>>>>>> +
>>>>>>> +int amdgpu_clear_workload_profile(struct amdgpu_device *adev,
>>>>>>> +                  uint32_t hint)
>>>>>>> +{
>>>>>>> +    int ret = 0;
>>>>>>> +    enum PP_SMC_POWER_PROFILE profile =
>>>>>>> +            amdgpu_workload_to_power_profile(hint);
>>>>>>> +
>>>>>>> +    if (hint == AMDGPU_CTX_WORKLOAD_HINT_NONE)
>>>>>>> +        return 0;
>>>>>>> +
>>>>>>> +    /* Do not reset GPU power profile if another reset is coming */
>>>>>>> +    if (atomic_dec_return(&adev->pm.workload_switch_ref) > 0)
>>>>>>> +        return 0;
>>>>>>> +
>>>>>>> +    mutex_lock(&adev->pm.smu_workload_lock);
>>>>>>> +
>>>>>>> +    if (adev->pm.workload_mode != hint)
>>>>>>> +        goto unlock;
>>>>>>> +
>>>>>>> +    ret = amdgpu_dpm_switch_power_profile(adev, profile, 0);
>>>>>>> +    if (!ret)
>>>>>>> +        adev->pm.workload_mode = AMDGPU_CTX_WORKLOAD_HINT_NONE;
>>>>>>> +
>>>>>>> +unlock:
>>>>>>> +    mutex_unlock(&adev->pm.smu_workload_lock);
>>>>>>> +    return ret;
>>>>>>> +}
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>>> index be7aff2d4a57..1f0f64662c04 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>>> @@ -3554,6 +3554,7 @@ int amdgpu_device_init(struct amdgpu_device 
>>>>>>> *adev,
>>>>>>>       mutex_init(&adev->psp.mutex);
>>>>>>>       mutex_init(&adev->notifier_lock);
>>>>>>>       mutex_init(&adev->pm.stable_pstate_ctx_lock);
>>>>>>> +    mutex_init(&adev->pm.smu_workload_lock);
>>>>>>>       mutex_init(&adev->benchmark_mutex);
>>>>>>>       amdgpu_device_init_apu_flags(adev);
>>>>>>> diff --git a/drivers/gpu/drm/amd/include/amdgpu_ctx_workload.h 
>>>>>>> b/drivers/gpu/drm/amd/include/amdgpu_ctx_workload.h
>>>>>>> new file mode 100644
>>>>>>> index 000000000000..6060fc53c3b0
>>>>>>> --- /dev/null
>>>>>>> +++ b/drivers/gpu/drm/amd/include/amdgpu_ctx_workload.h
>>>>>>> @@ -0,0 +1,54 @@
>>>>>>> +/*
>>>>>>> + * Copyright 2022 Advanced Micro Devices, Inc.
>>>>>>> + *
>>>>>>> + * Permission is hereby granted, free of charge, to any person 
>>>>>>> obtaining a
>>>>>>> + * copy of this software and associated documentation files (the 
>>>>>>> "Software"),
>>>>>>> + * to deal in the Software without restriction, including 
>>>>>>> without limitation
>>>>>>> + * the rights to use, copy, modify, merge, publish, distribute, 
>>>>>>> sublicense,
>>>>>>> + * and/or sell copies of the Software, and to permit persons to 
>>>>>>> whom the
>>>>>>> + * Software is furnished to do so, subject to the following 
>>>>>>> conditions:
>>>>>>> + *
>>>>>>> + * The above copyright notice and this permission notice shall 
>>>>>>> be included in
>>>>>>> + * all copies or substantial portions of the Software.
>>>>>>> + *
>>>>>>> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY 
>>>>>>> KIND, EXPRESS OR
>>>>>>> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 
>>>>>>> MERCHANTABILITY,
>>>>>>> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO 
>>>>>>> EVENT SHALL
>>>>>>> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, 
>>>>>>> DAMAGES OR
>>>>>>> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 
>>>>>>> OTHERWISE,
>>>>>>> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR 
>>>>>>> THE USE OR
>>>>>>> + * OTHER DEALINGS IN THE SOFTWARE.
>>>>>>> + *
>>>>>>> + */
>>>>>>> +#ifndef _AMDGPU_CTX_WL_H_
>>>>>>> +#define _AMDGPU_CTX_WL_H_
>>>>>>> +#include <drm/amdgpu_drm.h>
>>>>>>> +#include "amdgpu.h"
>>>>>>> +
>>>>>>> +/* Workload mode names */
>>>>>>> +static const char * const amdgpu_workload_mode_name[] = {
>>>>>>> +    "None",
>>>>>>> +    "3D",
>>>>>>> +    "Video",
>>>>>>> +    "VR",
>>>>>>> +    "Compute",
>>>>>>> +    "Unknown",
>>>>>>> +};
>>>>>>> +
>>>>>>> +static inline const
>>>>>>> +char *amdgpu_workload_profile_name(uint32_t profile)
>>>>>>> +{
>>>>>>> +    if (profile >= AMDGPU_CTX_WORKLOAD_HINT_NONE &&
>>>>>>> +        profile < AMDGPU_CTX_WORKLOAD_HINT_MAX)
>>>>>>> +        return 
>>>>>>> amdgpu_workload_mode_name[AMDGPU_CTX_WORKLOAD_INDEX(profile)];
>>>>>>> +
>>>>>>> +    return amdgpu_workload_mode_name[AMDGPU_CTX_WORKLOAD_HINT_MAX];
>>>>>>> +}
>>>>>>> +
>>>>>>> +int amdgpu_clear_workload_profile(struct amdgpu_device *adev,
>>>>>>> +                uint32_t hint);
>>>>>>> +
>>>>>>> +int amdgpu_set_workload_profile(struct amdgpu_device *adev,
>>>>>>> +                uint32_t hint);
>>>>>>> +
>>>>>>> +#endif
>>>>>>> diff --git a/drivers/gpu/drm/amd/pm/inc/amdgpu_dpm.h 
>>>>>>> b/drivers/gpu/drm/amd/pm/inc/amdgpu_dpm.h
>>>>>>> index 65624d091ed2..565131f789d0 100644
>>>>>>> --- a/drivers/gpu/drm/amd/pm/inc/amdgpu_dpm.h
>>>>>>> +++ b/drivers/gpu/drm/amd/pm/inc/amdgpu_dpm.h
>>>>>>> @@ -361,6 +361,11 @@ struct amdgpu_pm {
>>>>>>>       struct mutex            stable_pstate_ctx_lock;
>>>>>>>       struct amdgpu_ctx       *stable_pstate_ctx;
>>>>>>> +    /* SMU workload mode */
>>>>>>> +    struct mutex smu_workload_lock;
>>>>>>> +    uint32_t workload_mode;
>>>>>>> +    atomic_t workload_switch_ref;
>>>>>>> +
>>>>>>>       struct config_table_setting config_table;
>>>>>>>       /* runtime mode */
>>>>>>>       enum amdgpu_runpm_mode rpm_mode;
>>>>>>>