Cool... I'd like to see if Ouping and Mengmeng can work together to get
power & power numbers across a bunch of workloads for these policies.
Ouping & Mengmeng, can you run your usual battery of stuff against each
of these settings and report back the results?
Thanks,
Jesse
On Fri, 26 Jul 2013 23:24:11 +0100
Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> wrote:
> As a complement to the suggestions put forward by Jesse in
>
> drm/i915: "Scotty, I need more power!"
> (1372436290-3297-1-git-send-email-jbarnes@xxxxxxxxxxxxxxxx)
>
> and
>
> drm/i915: boost GPU and CPU freq when leaving idle
> (1372438472-3233-1-git-send-email-jbarnes@xxxxxxxxxxxxxxxx)
>
> we also have the ability to fine tune how we respond to the GPU's
> requests for more power.
>
> This patch introduces the i915.rps_policy module parameter that adjusts
> how we respond to the upclock request:
>
> 1 - Powersaving (current implementation)
> Increase the frequency by one bin on every upclock request.
>
> 2 - Conservative (new default)
> When waking from idle, boost the render frequency to the
> default render frequency (RP1)
>
> 3 - Ondemand
> At every upclock request, increase to halfway between the current
> frequency and maximum. This should give rapid upclocking similar
> to tcp window scaling.
>
> 4 - Performance
> Every time the GPU wants more power, give it everything we have.
>
> Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> ---
> drivers/gpu/drm/i915/i915_drv.c | 8 +++++++
> drivers/gpu/drm/i915/i915_drv.h | 2 ++
> drivers/gpu/drm/i915/i915_irq.c | 46 +++++++++++++++++++++++++++++------------
> drivers/gpu/drm/i915/intel_pm.c | 1 +
> 4 files changed, 44 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> index 1d8cedc..7ad80f4 100644
> --- a/drivers/gpu/drm/i915/i915_drv.c
> +++ b/drivers/gpu/drm/i915/i915_drv.c
> @@ -72,6 +72,14 @@ MODULE_PARM_DESC(i915_enable_rc6,
> "For example, 3 would enable rc6 and deep rc6, and 7 would enable everything. "
> "default: -1 (use per-chip default)");
>
> +int i915_rps_policy __read_mostly = 1;
> +module_param_named(rps_policy, i915_rps_policy, int, 0600);
> +MODULE_PARM_DESC(rps_policy,
> + "Specify a policy to use to respond to requests to change the render clock. "
> + "Different levels of uplocking agressiveness can be selected"
> + "(0 = Powersaving, slowly uplock; 1 = Conservative, when the gpu wakes from idle go straight to RP1/RPe (recommended minimum active GPU frequency); 2 = Ondemand, rapidly increase frequency whilst upclocking; 4 = Performance, always upclock to maximum frequency). "
> + "default: 1 (Conservative)");
> +
> int i915_enable_fbc __read_mostly = -1;
> module_param_named(i915_enable_fbc, i915_enable_fbc, int, 0600);
> MODULE_PARM_DESC(i915_enable_fbc,
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index ec14124..71232bc 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -811,6 +811,7 @@ struct intel_gen6_power_mgmt {
> u8 min_delay;
> u8 max_delay;
> u8 rpe_delay;
> + u8 rp1_delay;
> u8 hw_max;
>
> struct delayed_work delayed_resume_work;
> @@ -1635,6 +1636,7 @@ extern int i915_lvds_channel_mode __read_mostly;
> extern int i915_panel_use_ssc __read_mostly;
> extern int i915_vbt_sdvo_panel_type __read_mostly;
> extern int i915_enable_rc6 __read_mostly;
> +extern int i915_rps_policy __read_mostly;
> extern int i915_enable_fbc __read_mostly;
> extern bool i915_enable_hangcheck __read_mostly;
> extern int i915_enable_ppgtt __read_mostly;
> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> index ee3e49c..edd73c0 100644
> --- a/drivers/gpu/drm/i915/i915_irq.c
> +++ b/drivers/gpu/drm/i915/i915_irq.c
> @@ -728,14 +728,32 @@ static void gen6_pm_rps_work(struct work_struct *work)
> mutex_lock(&dev_priv->rps.hw_lock);
>
> if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) {
> - new_delay = dev_priv->rps.cur_delay + 1;
> + switch (i915_rps_policy) {
> + case 0: /* Powersave, slow increase in frequency */
> + new_delay = dev_priv->rps.cur_delay + 1;
> + break;
> +
> + case 1: /* Conservative, jump directly to RPe */
> + new_delay = dev_priv->rps.rp1_delay;
> + if (dev_priv->rps.cur_delay >= new_delay)
> + new_delay = dev_priv->rps.cur_delay + 1;
> + break;
> +
> + case 2: /* Ondemand, halfway to maximum */
> + new_delay = dev_priv->rps.cur_delay + 1;
> + if (new_delay < dev_priv->rps.rp1_delay)
> + new_delay = dev_priv->rps.rp1_delay;
> + else
> + new_delay = (dev_priv->rps.max_delay + new_delay) >> 1;
> + break;
> +
> + default: /* Performance, straight to maximum */
> + new_delay = dev_priv->rps.max_delay;
> + break;
> + }
>
> - /*
> - * For better performance, jump directly
> - * to RPe if we're below it.
> - */
> if (IS_VALLEYVIEW(dev_priv->dev) &&
> - dev_priv->rps.cur_delay < dev_priv->rps.rpe_delay)
> + new_delay < dev_priv->rps.rpe_delay)
> new_delay = dev_priv->rps.rpe_delay;
> } else
> new_delay = dev_priv->rps.cur_delay - 1;
> @@ -743,13 +761,15 @@ static void gen6_pm_rps_work(struct work_struct *work)
> /* sysfs frequency interfaces may have snuck in while servicing the
> * interrupt
> */
> - if (new_delay >= dev_priv->rps.min_delay &&
> - new_delay <= dev_priv->rps.max_delay) {
> - if (IS_VALLEYVIEW(dev_priv->dev))
> - valleyview_set_rps(dev_priv->dev, new_delay);
> - else
> - gen6_set_rps(dev_priv->dev, new_delay);
> - }
> + if (new_delay < dev_priv->rps.min_delay)
> + new_delay = dev_priv->rps.min_delay;
> + if (new_delay > dev_priv->rps.max_delay)
> + new_delay = dev_priv->rps.max_delay;
> +
> + if (IS_VALLEYVIEW(dev_priv->dev))
> + valleyview_set_rps(dev_priv->dev, new_delay);
> + else
> + gen6_set_rps(dev_priv->dev, new_delay);
>
> if (IS_VALLEYVIEW(dev_priv->dev)) {
> /*
> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> index 4e6d618..e9e467c 100644
> --- a/drivers/gpu/drm/i915/intel_pm.c
> +++ b/drivers/gpu/drm/i915/intel_pm.c
> @@ -3242,6 +3242,7 @@ static void gen6_enable_rps(struct drm_device *dev)
> /* In units of 50MHz */
> dev_priv->rps.hw_max = dev_priv->rps.max_delay = rp_state_cap & 0xff;
> dev_priv->rps.min_delay = (rp_state_cap & 0xff0000) >> 16;
> + dev_priv->rps.rp1_delay = (rp_state_cap & 0xff00) >> 8;
> dev_priv->rps.cur_delay = 0;
>
> /* disable the counters and set deterministic thresholds */
--
Jesse Barnes, Intel Open Source Technology Center
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
