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

Sharma, Shashank shashank.sharma at amd.com
Tue Sep 27 12:23:41 UTC 2022



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 at amd.com>
>>>> Signed-off-by: Shashank Sharma <shashank.sharma at 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.

- 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;
>>>>


More information about the amd-gfx mailing list