[PATCH] drm/amdgpu: Add GFXOFF auto-tunning algorithm
Sergey Kovalenko
seryoga.engineering at gmail.com
Mon Mar 24 15:14:19 UTC 2025
"Moreover the work only gets scheduled/cancelled
around a ref count of 0" - and this happens more than 600 times
per second under load, as you can see from the table.
Moreover these 600 GFXOFF requests are executed with a very
little interval between them. If we set GFXOFF delay to 0 by
default, it will be enabled and disabled 600 times with pretty
significant performance drop in any test.
Keeping 100ms GFXOFF delay leads to 2 times higher
power consumption in idle, nicely noticeable for laptops.
I don't think this is the behavior the users want to see.
Workarounds introduced a significant overhead, and
to mitigate it there is a simple solution to limit the amount
of those calls.
Regards,
Sergey
On Mon, Mar 24, 2025 at 4:50 PM Alex Deucher <alexdeucher at gmail.com> wrote:
>
> On Mon, Mar 24, 2025 at 10:34 AM Sergey Kovalenko
> <seryoga.engineering at gmail.com> wrote:
> >
> > Hello Alex!
> >
> > "If there are a lot of requests to toggle gfxoff, the worker thread to
> > allow it again gets
> > cancelled and scheduled again, extending the time it's disallowed." -
> > That's true except one thing:
> > cancelling and scheduling also take CPU cycles, and the pause between
> > submissions can exceed
> > 100ms interval, which leads to GFXOFF being enabled even if the GPU is loaded.
> > This simple prediction algorithm eliminates such cases.
> > GFXOFF is enabled immediately when there are only a few submissions
> > ongoing, which
> > means the system is idling. And when the number of compute submissions
> > exceeds the
> > range, the algorithm chooses a longer delay. Finally it simply ignores
> > GFXOFF ON requests
> > to prevent performance drops under high load.
> > This allows keeping idle power consumption low and using stable GPU
> > performance under load.
>
> What performance drops? Your data doesn't show any. Moreover the
> work only gets scheduled/cancelled around a ref count of 0. Plus, your
> change introduces a bunch of additional calculations for this function
> which add overhead which you are concerned about.
>
> Alex
>
> >
> > Regards,
> > Sergey
> >
> > On Mon, Mar 24, 2025 at 4:14 PM Alex Deucher <alexdeucher at gmail.com> wrote:
> > >
> > > On Mon, Mar 24, 2025 at 5:06 AM Sergey Kovalenko
> > > <seryoga.engineering at gmail.com> wrote:
> > > >
> > > > Predict an optimal delay to enable GFXOFF for the next interval
> > > > based on the request count:
> > > > - less than 15 requests per second - zero delay
> > > > - less than 25 requests per second - default delay
> > > > - 25 and more requests per second - don't enable GFXOFF
> > > >
> > > > The algorithm allows maintaining low power consumption in idle,
> > > > as well as using the full GPU power under load by eliminating
> > > > hundreds of extra GFXOFF ON/OFF switches.
> > >
> > > I still don't understand what problem this is solving. This already
> > > happens with the way the code works now. If there are a lot of
> > > requests to toggle gfxoff, the worker thread to allow it again gets
> > > cancelled and scheduled again, extending the time it's disallowed.
> > >
> > > Alex
> > >
> > >
> > > Alex
> > >
> > > >
> > > > Test configuration:
> > > > - Ryzen 5 2500U
> > > > - Ryzen 5 3400G
> > > > - Chromium 134.0.6998.88 Arch Linux
> > > > - Mesa 1:24.3.4-1
> > > > - KDE Plasma 6.3.2
> > > >
> > > > GFXOFF enable requests per second:
> > > > | Test | min | max |
> > > > |-----------------------------------------------------|-----|-----|
> > > > | System idle | 0 | 64 |
> > > > | Web browsing | 0 | 127 |
> > > > | WebGL aquarium | 10 | 236 |
> > > > | Heavy load: glxgears + vkcube + resizing + flipping | 39 | 677 |
> > > >
> > > > Test results, Ryzen 5 2500U:
> > > > | Test | patched-6.13.7.arch1 | 6.13.7.arch1-1 |
> > > > |-----------------------------|----------------------|------------------|
> > > > | System idle (PkgWatt) | ~0.74W | ~1.25W |
> > > > | glxgears (vblank_mode=0) | ~7300 fps, ~7.3W | ~7300 fps, ~7.3W |
> > > > | WebGL aquarium 15.000 fish | 56-60 fps, ~9.8W | 55-60 fps, ~9.8W |
> > > >
> > > > Test results, Ryzen 5 3400G:
> > > > | Test | patched-6.13.7.arch1 | 6.13.7.arch1-1 |
> > > > |-----------------------------|----------------------|------------------|
> > > > | System idle (PkgWatt) | ~3.8W | ~4.3W |
> > > > | glxgears (vblank_mode=0) | ~7200 fps | ~7200 fps |
> > > > | WebGL aquarium 30.000 fish | 37 fps, 47W | 38 fps, 47W |
> > > >
> > > > Signed-off-by: Sergey Kovalenko <seryoga.engineering at gmail.com>
> > > > Tested-by: Liam Fleming <fleming.squared at proton.me>
> > > > ---
> > > > drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 94 +++++++++++++++++--------
> > > > drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h | 3 +
> > > > 2 files changed, 67 insertions(+), 30 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> > > > index c1f35ded684e..5e23b956e0bf 100644
> > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
> > > > @@ -781,55 +781,89 @@ int amdgpu_gfx_enable_kgq(struct amdgpu_device
> > > > *adev, int xcc_id)
> > > > * 3. other client can cancel their request of disable gfx off feature
> > > > * 4. other client should not send request to enable gfx off feature
> > > > before disable gfx off feature.
> > > > */
> > > > -
> > > > void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
> > > > {
> > > > - unsigned long delay = GFX_OFF_DELAY_ENABLE;
> > > > -
> > > > if (!(adev->pm.pp_feature & PP_GFXOFF_MASK))
> > > > return;
> > > >
> > > > mutex_lock(&adev->gfx.gfx_off_mutex);
> > > >
> > > > if (enable) {
> > > > - /* If the count is already 0, it means there's an imbalance bug
> > > > somewhere.
> > > > - * Note that the bug may be in a different caller than the one which
> > > > triggers the
> > > > - * WARN_ON_ONCE.
> > > > - */
> > > > - if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
> > > > + /* This case covers multiple calls from parallel threads */
> > > > + if (!adev->gfx.gfx_off_req_count)
> > > > goto unlock;
> > > >
> > > > - adev->gfx.gfx_off_req_count--;
> > > > + /* Process only if req_count == 0 and GFXOFF is disabled */
> > > > + if (--adev->gfx.gfx_off_req_count || adev->gfx.gfx_off_state)
> > > > + goto unlock;
> > > > +
> > > > + /* If going to s2idle, no need to wait */
> > > > + if (adev->in_s0ix) {
> > > > + if (!amdgpu_dpm_set_powergating_by_smu(
> > > > + adev, AMD_IP_BLOCK_TYPE_GFX, true, 0))
> > > > + adev->gfx.gfx_off_state = true;
> > > > +
> > > > + /* Reset delay flag */
> > > > + adev->gfx.gfx_off_use_delay = 0;
> > > > + goto unlock;
> > > > + }
> > > >
> > > > - if (adev->gfx.gfx_off_req_count == 0 &&
> > > > - !adev->gfx.gfx_off_state) {
> > > > - /* If going to s2idle, no need to wait */
> > > > - if (adev->in_s0ix) {
> > > > - if (!amdgpu_dpm_set_powergating_by_smu(adev,
> > > > - AMD_IP_BLOCK_TYPE_GFX, true, 0))
> > > > - adev->gfx.gfx_off_state = true;
> > > > + ++adev->gfx.gfx_off_counter;
> > > > +
> > > > + uint64_t now = get_jiffies_64();
> > > > + uint64_t delta =
> > > > + jiffies_to_msecs(now - adev->gfx.gfx_off_timestamp);
> > > > +
> > > > + if (delta >= 1000u) {
> > > > + /*
> > > > + * Predict the optimal delay for the next interval
> > > > + * based on the current number of requests:
> > > > + * <15 - idle, no delay
> > > > + * <25 - light/medium load, default delay
> > > > + * 25 and more - high load, GFXOFF disabled
> > > > + */
> > > > + if (adev->gfx.gfx_off_counter < 15u) {
> > > > + adev->gfx.gfx_off_use_delay = 0;
> > > > + } else if (adev->gfx.gfx_off_counter < 25u) {
> > > > + adev->gfx.gfx_off_use_delay = 1;
> > > > } else {
> > > > - schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
> > > > - delay);
> > > > + adev->gfx.gfx_off_use_delay = 2;
> > > > }
> > > > +
> > > > + adev->gfx.gfx_off_counter = 0;
> > > > + adev->gfx.gfx_off_timestamp = now;
> > > > }
> > > > +
> > > > + /* Don't schedule gfxoff under heavy load */
> > > > + if (adev->gfx.gfx_off_use_delay == 2)
> > > > + goto unlock;
> > > > +
> > > > + schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
> > > > + adev->gfx.gfx_off_use_delay ?
> > > > + GFX_OFF_DELAY_ENABLE :
> > > > + GFX_OFF_NO_DELAY);
> > > > } else {
> > > > - if (adev->gfx.gfx_off_req_count == 0) {
> > > > - cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
> > > > + /* GFXOFF was enabled when req_count == 0 */
> > > > + if (++adev->gfx.gfx_off_req_count != 1)
> > > > + goto unlock;
> > > >
> > > > - if (adev->gfx.gfx_off_state &&
> > > > - !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX,
> > > > false, 0)) {
> > > > - adev->gfx.gfx_off_state = false;
> > > > + /* Nothing to do if the work wasn't scheduled */
> > > > + if (adev->gfx.gfx_off_use_delay == 2)
> > > > + goto unlock;
> > > >
> > > > - if (adev->gfx.funcs->init_spm_golden) {
> > > > - dev_dbg(adev->dev,
> > > > - "GFXOFF is disabled, re-init SPM golden settings\n");
> > > > - amdgpu_gfx_init_spm_golden(adev);
> > > > - }
> > > > + cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
> > > > +
> > > > + if (adev->gfx.gfx_off_state &&
> > > > + !amdgpu_dpm_set_powergating_by_smu(
> > > > + adev, AMD_IP_BLOCK_TYPE_GFX, false, 0)) {
> > > > + adev->gfx.gfx_off_state = false;
> > > > +
> > > > + if (adev->gfx.funcs->init_spm_golden) {
> > > > + dev_dbg(adev->dev,
> > > > + "GFXOFF is disabled, re-init SPM golden settings\n");
> > > > + amdgpu_gfx_init_spm_golden(adev);
> > > > }
> > > > }
> > > > -
> > > > - adev->gfx.gfx_off_req_count++;
> > > > }
> > > >
> > > > unlock:
> > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> > > > index 8b5bd63b5773..38fd445a353b 100644
> > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> > > > @@ -430,7 +430,10 @@ struct amdgpu_gfx {
> > > > /* gfx off */
> > > > bool gfx_off_state; /* true: enabled,
> > > > false: disabled */
> > > > struct mutex gfx_off_mutex; /* mutex to
> > > > change gfxoff state */
> > > > + uint64_t gfx_off_timestamp; /* gfxoff enable call timestamp */
> > > > + uint32_t gfx_off_use_delay; /* flag to choose the delay range */
> > > > uint32_t gfx_off_req_count; /* default 1,
> > > > enable gfx off: dec 1, disable gfx off: add 1 */
> > > > + uint32_t gfx_off_counter; /* count of gfxoff enable calls */
> > > > struct delayed_work gfx_off_delay_work; /* async work to
> > > > set gfx block off */
> > > > uint32_t gfx_off_residency; /* last logged
> > > > residency */
> > > > uint64_t gfx_off_entrycount; /* count of times
> > > > GPU has get into GFXOFF state */
> > > > --
> > > > 2.49.0
More information about the amd-gfx
mailing list