Skip to contentSkip to conversation

drm/i915: Print ring min freq scaling

PATCH In review Ben Widawsky 3 participants 5 messages Sep 25, 2013 – Sep 26, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 drm/i915: Print ring min freq scaling

This allows us to keep track of the values being set if we want to tweak this code.

---
 drivers/gpu/drm/i915/intel_pm.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/gpu/drm/i915/intel_pm.c +6 -0
index d27eda6..31cf188 100644
@@ -3671,6 +3671,9 @@ void gen6_update_ring_freq(struct drm_device *dev)
 			ring_freq = (gpu_freq * 5 + 3) / 4;
 			ring_freq = max(min_ring_freq, ring_freq);
 			/* leave ia_freq as the default, chosen by cpufreq */
+
+			DRM_DEBUG_DRIVER("Setup min ring frequency %dMHz for GT freq %dMHz\n",
+					 ring_freq * 100, gpu_freq * 50);
 		} else {
 			/* On older processors, there is no separate ring
 			 * clock domain, so in order to boost the bandwidth
@@ -3684,6 +3687,9 @@ void gen6_update_ring_freq(struct drm_device *dev)
 			else
 				ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
 			ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
+
+			DRM_DEBUG_DRIVER("Setup min ring frequency %dMHz for GT freq %dMHz\n",
+					 ia_freq * 100, gpu_freq * 50);
 		}
 
 		sandybridge_pcode_write(dev_priv,
Signature · 1 line
1.8.4
#1 lore raw

PATCH 2/2 drm/i915: Use the real cpu max frequency for ring scaling

The policy's max frequency is not equal to the CPU's max frequency. The ring frequency is derived from the CPU frequency, and not the policy frequency.

One example of how this may differ through sysfs. If the sysfs max frequency is modified, that will be used for the max ring frequency calculation. (/sys/devices/system/cpu/cpu0/cpufreq/scaling_max_freq). As far as I know, no current governor uses anything but max as the default, but in theory, they could. Similarly distributions might set policy as part of their init process.

It's ideal to use the real frequency because when we're currently scaled up on the GPU. In this case we likely want to race to idle, and using a less than max ring frequency is non-optimal for this situation.

AFAIK, this patch should have no impact on a majority of people.

This behavior hasn't been changed since it was first introduced:
commit 23b2f8bb92feb83127679c53633def32d3108e70

Author: Jesse Barnes <[email protected]>
Date:   Tue Jun 28 13:04:16 2011 -0700

drm/i915: load a ring frequency scaling table v3

---
 drivers/gpu/drm/i915/intel_pm.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/intel_pm.c +11 -6
index 31cf188..3212d3b 100644
@@ -3639,16 +3639,21 @@ void gen6_update_ring_freq(struct drm_device *dev)
 	unsigned int gpu_freq;
 	unsigned int max_ia_freq, min_ring_freq;
 	int scaling_factor = 180;
+	struct cpufreq_policy *policy;
 
 	WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
 
-	max_ia_freq = cpufreq_quick_get_max(0);
-	/*
-	 * Default to measured freq if none found, PCU will ensure we don't go
-	 * over
-	 */
-	if (!max_ia_freq)
+	policy = cpufreq_cpu_get(0);
+	if (policy) {
+		max_ia_freq = policy->cpuinfo.max_freq;
+		cpufreq_cpu_put(policy);
+	} else {
+		/*
+		 * Default to measured freq if none found, PCU will ensure we
+		 * don't go over
+		 */
 		max_ia_freq = tsc_khz;
+	}
 
 	/* Convert from kHz to MHz */
 	max_ia_freq /= 1000;
Signature · 1 line
1.8.4
Reviewed-by on v1 2/2
#2 lore raw

Re: [PATCH 2/2] drm/i915: Use the real cpu max frequency for ring scaling

Quoted 11 lines · Ben Widawsky

The policy's max frequency is not equal to the CPU's max frequency. The ring frequency is derived from the CPU frequency, and not the policy frequency.

One example of how this may differ through sysfs. If the sysfs max frequency is modified, that will be used for the max ring frequency calculation. (/sys/devices/system/cpu/cpu0/cpufreq/scaling_max_freq). As far as I know, no current governor uses anything but max as the default, but in theory, they could. Similarly distributions might set policy as part of their init process.

Actually this seems to differ based on policy as well. The pstate may disable turbo mode, in which case policy->max <= policy->cpuinfo.max. But excluding user intervention, it seems that max is what we want, and it is possibly arguable that we do not want to push the core to turbo for a GPU bound workload - though I'm not sure if the ring frequency continues to scale with turbo cpu frequencies.

Quoted 3 lines

It's ideal to use the real frequency because when we're currently scaled up on the GPU. In this case we likely want to race to idle, and using a less than max ring frequency is non-optimal for this situation.

Also note that scaling the cpu frequency cuts into our thermal headroom, so for ULT devices it may not be clearly beneficial.

Quoted 1 line

AFAIK, this patch should have no impact on a majority of people.

And also obsolete since HSW.

Quoted 9 lines

This behavior hasn't been changed since it was first introduced:
commit 23b2f8bb92feb83127679c53633def32d3108e70

Author: Jesse Barnes <[email protected]>
Date:   Tue Jun 28 13:04:16 2011 -0700

drm/i915: load a ring frequency scaling table v3

FWIW,

I think using the real range of the CPU frequency rather than the adjusted policy maximum is sensible. Just the hw implementation is just a bit silly and it is not clear exactly what the best policy should be. -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#3 lore raw
In reply to Ben Widawsky
Quoted 23 lines · Ben Widawsky

This allows us to keep track of the values being set if we want to tweak this code.

---
 drivers/gpu/drm/i915/intel_pm.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/gpu/drm/i915/intel_pm.c +3 -0
index d27eda6..31cf188 100644
@@ -3671,6 +3671,9 @@ void gen6_update_ring_freq(struct drm_device *dev)
 			ring_freq = (gpu_freq * 5 + 3) / 4;
 			ring_freq = max(min_ring_freq, ring_freq);
 			/* leave ia_freq as the default, chosen by cpufreq */
+
+			DRM_DEBUG_DRIVER("Setup min ring frequency %dMHz for GT freq %dMHz\n",
+					 ring_freq * 100, gpu_freq * 50);
 		} else {
 			/* On older processors, there is no separate ring
 			 * clock domain, so in order to boost the bandwidth

drivers/gpu/drm/i915/intel_pm.c void gen6_update_ring_freq(struct drm_device *dev) · line 3692

Quoted 6 lines
drivers/gpu/drm/i915/intel_pm.c +3 -0
@@ -3684,6 +3687,9 @@ void gen6_update_ring_freq(struct drm_device *dev)
 			else
 				ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
 			ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
+
+			DRM_DEBUG_DRIVER("Setup min ring frequency %dMHz for GT freq %dMHz\n",
+					 ia_freq * 100, gpu_freq * 50);

This is not a ring freq, but a cpu core freq.
Also would be wise to point out that this information is in /sys/kernel/debug/dri/0/i915_ring_freq_table -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#4 lore raw
Quoted 27 lines · Chris Wilson
Quoted 26 lines · Ben Widawsky

This allows us to keep track of the values being set if we want to tweak this code.

---
 drivers/gpu/drm/i915/intel_pm.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/gpu/drm/i915/intel_pm.c +3 -0
index d27eda6..31cf188 100644
@@ -3671,6 +3671,9 @@ void gen6_update_ring_freq(struct drm_device *dev)
 			ring_freq = (gpu_freq * 5 + 3) / 4;
 			ring_freq = max(min_ring_freq, ring_freq);
 			/* leave ia_freq as the default, chosen by cpufreq */
+
+			DRM_DEBUG_DRIVER("Setup min ring frequency %dMHz for GT freq %dMHz\n",
+					 ring_freq * 100, gpu_freq * 50);
 		} else {
 			/* On older processors, there is no separate ring
 			 * clock domain, so in order to boost the bandwidth
@@ -3684,6 +3687,9 @@ void gen6_update_ring_freq(struct drm_device *dev)
 			else
 				ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
 			ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
Quoted 11 lines

drivers/gpu/drm/i915/intel_pm.c void gen6_update_ring_freq(struct drm_device *dev) · line 3692

Quoted 3 lines
drivers/gpu/drm/i915/intel_pm.c +3 -0
+
+			DRM_DEBUG_DRIVER("Setup min ring frequency %dMHz for GT freq %dMHz\n",
+					 ia_freq * 100, gpu_freq * 50);

This is not a ring freq, but a cpu core freq.
Also would be wise to point out that this information is in /sys/kernel/debug/dri/0/i915_ring_freq_table -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre

Ooops. I forgot about that interface. Ignore this patch.

You are incorrect in saying it's _just_ the CPU core frequency. This is the ring and IA frequency on SNB + IVB. The frequency is tied to the CPU core on SNB + IVB. It is not tied to the cpu core on Haswell.

To make the print more generic, and correct across the board it makes sense to say we're setting the ring frequency (to me). Even the programming guides typically call this, "IA/ring" frequency.

Above is only for posterity. Again, forget the patch.

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center