Skip to contentSkip to conversation

drm/i915: initialize ring frequency scaling table on SNB

PATCH In review Jesse Barnes 3 participants 9 messages Jun 21, 2011 – Jun 23, 2011

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

PATCH 1/1 drm/i915: initialize ring frequency scaling table on SNB

The ring frequency scaling table tells the PCU to treat certain GPU frequencies as if they were a given CPU frequency for purposes of scaling the ring frequency. Normally the PCU will scale the ring frequency based on the CPU P-state, but with the table present, it will also take the GPU frequency into account. The scaling_factor used in this patch may not be ideal, but is enough to increase performance in nexuiz on a 1366x768 panel by about 20%.

The main downside of keeping the ring frequency high while the CPU is at a low frequency (or asleep altogether) is increased power consumption. But then if you're keeping your GPU busy, you probably want the extra performance.

drivers/gpu/drm/i915/i915_reg.h +2 -1
index 2f967af..15ee639 100644
@@ -3433,7 +3433,8 @@
 #define GEN6_PCODE_MAILBOX			0x138124
 #define   GEN6_PCODE_READY			(1<<31)
 #define   GEN6_READ_OC_PARAMS			0xc
-#define   GEN6_PCODE_WRITE_MIN_FREQ_TABLE	0x9
+#define   GEN6_PCODE_WRITE_MIN_FREQ_TABLE	0x8
 #define GEN6_PCODE_DATA				0x138128
+#define   GEN6_PCODE_FREQ_IA_RATIO_SHIFT	8
 
 #endif /* _I915_REG_H_ */
drivers/gpu/drm/i915/i915_suspend.c +3 -1
index 60a94d2..03f0fac 100644
@@ -870,8 +870,10 @@ int i915_restore_state(struct drm_device *dev)
 		intel_init_emon(dev);
 	}
 
-	if (IS_GEN6(dev))
+	if (IS_GEN6(dev)) {
 		gen6_enable_rps(dev_priv);
+		gen6_update_ring_freq(dev_priv);
+	}
 
 	/* Cache mode state */
 	I915_WRITE (CACHE_MODE_0, dev_priv->saveCACHE_MODE_0 | 0xffff0000);
drivers/gpu/drm/i915/intel_display.c +43 -1

57 lines not on this page · raw message

drivers/gpu/drm/i915/intel_drv.h +1 -0
index 9ffa61e..8ac3bd8 100644
@@ -317,6 +317,7 @@ extern void intel_enable_clock_gating(struct drm_device *dev);
 extern void ironlake_enable_drps(struct drm_device *dev);
 extern void ironlake_disable_drps(struct drm_device *dev);
 extern void gen6_enable_rps(struct drm_i915_private *dev_priv);
+extern void gen6_update_ring_freq(struct drm_i915_private *dev_priv);
 extern void gen6_disable_rps(struct drm_device *dev);
 extern void intel_init_emon(struct drm_device *dev);
#1 lore raw
Quoted 7 lines · Jesse Barnes

The ring frequency scaling table tells the PCU to treat certain GPU frequencies as if they were a given CPU frequency for purposes of scaling the ring frequency. Normally the PCU will scale the ring frequency based on the CPU P-state, but with the table present, it will also take the GPU frequency into account. The scaling_factor used in this patch may not be ideal, but is enough to increase performance in nexuiz on a 1366x768 panel by about 20%.

Am I right in thinking that the improvement offered by these new defaults is dependent upon the previous values programmed by the BIOS? On my desktop SNB, I only see a marginal improvement (<~1%) for games and similarly small differences for cairo-gl/xlib.

Plus they also revealed my confusion over dev_priv->min_delay is actually min_freq on SNB. Evil. :-p
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#2 lore raw
Quoted 17 lines · Chris Wilson
Quoted 4 lines · Jesse Barnes

The ring frequency scaling table tells the PCU to treat certain GPU frequencies as if they were a given CPU frequency for purposes of scaling the ring frequency. Normally the PCU will scale the ring frequency based on the CPU P-state, but with the table present, it

will

Quoted 3 lines

also take the GPU frequency into account. The scaling_factor used in this patch may not be ideal, but is enough to increase performance in nexuiz on a 1366x768 panel by about 20%.

Am I right in thinking that the improvement offered by these new defaults
is dependent upon the previous values programmed by the BIOS? On my desktop SNB, I only see a marginal improvement (<~1%) for games and similarly small differences for cairo-gl/xlib.

Quoted 7 lines

Plus they also revealed my confusion over dev_priv->min_delay is actually
min_freq on SNB. Evil. :-p
-Chris

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

Yes, though on my machines the BIOS programs these values to 0, indicating no minimum ring freq. The perf increase will be proportional to the bandwidth requirements not met by the existing frequency. So for your tests, it's possible that the CPU is being kept busy enough to kee the ring freq up anyway (iow this patch is redundant for those loads) or that your workloads don't have high memory bw requirements to begin with, so the higher ring freq isn't a benefit.

Signature · 1 line
Jesse Barnes, Intel Open Source Technology Center
#3 lore raw
In reply to Jesse Barnes
Quoted 8 lines · Jesse Barnes

The ring frequency scaling table tells the PCU to treat certain GPU frequencies as if they were a given CPU frequency for purposes of scaling the ring frequency. Normally the PCU will scale the ring frequency based on the CPU P-state, but with the table present, it will also take the GPU frequency into account. The scaling_factor used in this patch may not be ideal, but is enough to increase performance in nexuiz on a 1366x768 panel by about 20%.

Quoted 4 lines

The main downside of keeping the ring frequency high while the CPU is at a low frequency (or asleep altogether) is increased power consumption. But then if you're keeping your GPU busy, you probably want the extra performance.

The intent of the patch sounds good, but it doesn't seem to be doing anything here (graphs below). If I run a "while true; do; done" loop to keep a CPU busy during measurement, though, OA gets a 15.9% +/- 0.6% win, and nexuiz around 20%. There are no complaints in dmesg.

x /home/anholt/nexuiz-before
+ /home/anholt/nexuiz-after

+-------------------------------------------------------------------------------+
| x       +                +    x                                x             +|
||_|_______________________M____MA_____A_________________________|_________|    |
+-------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   3     59.217461     59.775295     59.479322     59.490693    0.27909057
+   3     59.287311     59.893383     59.440451     59.540382    0.31515189

No difference proven at 95.0% confidence

x /home/anholt/oa-before
+ /home/anholt/oa-after

+-------------------------------------------------------------------------------+
|x  x                           +  +    +               x      *          +    x|
|    |________________________|_________MA_______A______M__________|________|   |
+-------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   5         184.2         187.2         186.3        185.72     1.3809417
+   5         185.4           187         185.7        186.04    0.71624018

No difference proven at 95.0% confidence

x /home/anholt/taiji-before
+ /home/anholt/taiji-after

+-------------------------------------------------------------------------------+
|x      x                  +     xx       +    x          + ++     +           x|
|    |____________________________A___|______________A______M_|_____|           |
+-------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   6       132.486       139.931       135.608       135.596     2.6966051
+   6       134.974       138.832       138.111     137.41183      1.434609

No difference proven at 95.0% confidence

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#4 lore raw
Quoted 51 lines · On Wed, 22 Jun 2011 11:42:04 -0700 Eric Anholt <[email protected]> wrote:
Quoted 8 lines · Jesse Barnes

The ring frequency scaling table tells the PCU to treat certain GPU frequencies as if they were a given CPU frequency for purposes of scaling the ring frequency. Normally the PCU will scale the ring frequency based on the CPU P-state, but with the table present, it will also take the GPU frequency into account. The scaling_factor used in this patch may not be ideal, but is enough to increase performance in nexuiz on a 1366x768 panel by about 20%.

Quoted 4 lines

The main downside of keeping the ring frequency high while the CPU is at a low frequency (or asleep altogether) is increased power consumption. But then if you're keeping your GPU busy, you probably want the extra performance.

The intent of the patch sounds good, but it doesn't seem to be doing anything here (graphs below). If I run a "while true; do; done" loop to keep a CPU busy during measurement, though, OA gets a 15.9% +/- 0.6% win, and nexuiz around 20%. There are no complaints in dmesg.

x /home/anholt/nexuiz-before
+ /home/anholt/nexuiz-after

+-------------------------------------------------------------------------------+
| x       +                +    x                                x             +|
||_|_______________________M____MA_____A_________________________|_________|    |
+-------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   3     59.217461     59.775295     59.479322     59.490693    0.27909057
+   3     59.287311     59.893383     59.440451     59.540382    0.31515189

No difference proven at 95.0% confidence

x /home/anholt/oa-before
+ /home/anholt/oa-after

+-------------------------------------------------------------------------------+
|x  x                           +  +    +               x      *          +    x|
|    |________________________|_________MA_______A______M__________|________|   |
+-------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   5         184.2         187.2         186.3        185.72     1.3809417
+   5         185.4           187         185.7        186.04    0.71624018

No difference proven at 95.0% confidence

x /home/anholt/taiji-before
+ /home/anholt/taiji-after

+-------------------------------------------------------------------------------+
|x      x                  +     xx       +    x          + ++     +           x|
|    |____________________________A___|______________A______M_|_____|           |
+-------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   6       132.486       139.931       135.608       135.596     2.6966051
+   6       134.974       138.832       138.111     137.41183      1.434609

No difference proven at 95.0% confidence

#5 lore raw
In reply to Eric Anholt
Quoted 14 lines · On Wed, 22 Jun 2011 11:42:04 -0700 Eric Anholt <[email protected]> wrote:
Quoted 8 lines · Jesse Barnes

The ring frequency scaling table tells the PCU to treat certain GPU frequencies as if they were a given CPU frequency for purposes of scaling the ring frequency. Normally the PCU will scale the ring frequency based on the CPU P-state, but with the table present, it will also take the GPU frequency into account. The scaling_factor used in this patch may not be ideal, but is enough to increase performance in nexuiz on a 1366x768 panel by about 20%.

Quoted 4 lines

The main downside of keeping the ring frequency high while the CPU is at a low frequency (or asleep altogether) is increased power consumption. But then if you're keeping your GPU busy, you probably want the extra performance.

Quoted 4 lines

The intent of the patch sounds good, but it doesn't seem to be doing anything here (graphs below). If I run a "while true; do; done" loop to keep a CPU busy during measurement, though, OA gets a 15.9% +/- 0.6% win, and nexuiz around 20%. There are no complaints in dmesg.

Ok I'll checck it out; I admit I tested it then made a few changes then sent it out, so I must have broken something...

Jesse

#6 lore raw

I tested again and things were working as expected. This patch adds a debugfs file for dumping the ring freq table, can you guys test it out (try disabling the call to update the table, dump it, make sure it's 0, then re-enable and make sure it's programmed correctly).

Signature · 1 line
Jesse Barnes, Intel Open Source Technology Center
drivers/gpu/drm/i915/i915_debugfs.c +43 -0
index 4d46441..79394cd 100644
@@ -1123,6 +1123,48 @@ static int i915_emon_status(struct seq_file *m, void *unused)
 	return 0;
 }
 
+static int i915_ring_freq_table(struct seq_file *m, void *unused)
+{
+	struct drm_info_node *node = (struct drm_info_node *) m->private;
+	struct drm_device *dev = node->minor->dev;
+	drm_i915_private_t *dev_priv = dev->dev_private;
+	int ret;
+	int gpu_freq, ia_freq;
+
+	if (!IS_GEN6(dev)) {
+		seq_printf(m, "unsupported on this chipset\n");
+		return 0;
+	}
+
+	ret = mutex_lock_interruptible(&dev->struct_mutex);
+	if (ret)
+		return ret;
+
+	gen6_gt_force_wake_get(dev_priv);
+
+	seq_printf(m, "GPU freq\tEffective CPU freq\n");
+
+	for (gpu_freq = dev_priv->min_delay; gpu_freq <= dev_priv->max_delay;
+	     gpu_freq++) {
+		I915_WRITE(GEN6_PCODE_DATA, gpu_freq);
+		I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY |
+			   GEN6_PCODE_READ_MIN_FREQ_TABLE);
+		if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) &
+			      GEN6_PCODE_READY) == 0, 10)) {
+			DRM_ERROR("pcode write of freq table timed out\n");
+			continue;
+		}
+		ia_freq = I915_READ(GEN6_PCODE_DATA);
+		seq_printf(m, "%d\t\t%d\n", gpu_freq * 50, ia_freq * 100);
+	}
+
+	gen6_gt_force_wake_put(dev_priv);
+
+	mutex_unlock(&dev->struct_mutex);
+
+	return 0;
+}
+
 static int i915_gfxec(struct seq_file *m, void *unused)
 {
 	struct drm_info_node *node = (struct drm_info_node *) m->private;
@@ -1426,6 +1468,7 @@ static struct drm_info_list i915_debugfs_list[] = {
 	{"i915_inttoext_table", i915_inttoext_table, 0},
 	{"i915_drpc_info", i915_drpc_info, 0},
 	{"i915_emon_status", i915_emon_status, 0},
+	{"i915_ring_freq_table", i915_ring_freq_table, 0},
 	{"i915_gfxec", i915_gfxec, 0},
 	{"i915_fbc_status", i915_fbc_status, 0},
 	{"i915_sr_status", i915_sr_status, 0},
drivers/gpu/drm/i915/i915_reg.h +3 -1
index 2f967af..757e024 100644
@@ -3433,7 +3433,9 @@
 #define GEN6_PCODE_MAILBOX			0x138124
 #define   GEN6_PCODE_READY			(1<<31)
 #define   GEN6_READ_OC_PARAMS			0xc
-#define   GEN6_PCODE_WRITE_MIN_FREQ_TABLE	0x9
+#define   GEN6_PCODE_WRITE_MIN_FREQ_TABLE	0x8
+#define   GEN6_PCODE_READ_MIN_FREQ_TABLE	0x9
 #define GEN6_PCODE_DATA				0x138128
+#define   GEN6_PCODE_FREQ_IA_RATIO_SHIFT	8
 
 #endif /* _I915_REG_H_ */
drivers/gpu/drm/i915/i915_suspend.c +3 -1
index 60a94d2..03f0fac 100644
@@ -870,8 +870,10 @@ int i915_restore_state(struct drm_device *dev)
 		intel_init_emon(dev);
 	}
 
-	if (IS_GEN6(dev))
+	if (IS_GEN6(dev)) {
 		gen6_enable_rps(dev_priv);
+		gen6_update_ring_freq(dev_priv);
+	}
 
 	/* Cache mode state */
 	I915_WRITE (CACHE_MODE_0, dev_priv->saveCACHE_MODE_0 | 0xffff0000);
drivers/gpu/drm/i915/intel_display.c +43 -1

57 lines not on this page · raw message

drivers/gpu/drm/i915/intel_drv.h +1 -0
index 9ffa61e..8ac3bd8 100644
@@ -317,6 +317,7 @@ extern void intel_enable_clock_gating(struct drm_device *dev);
 extern void ironlake_enable_drps(struct drm_device *dev);
 extern void ironlake_disable_drps(struct drm_device *dev);
 extern void gen6_enable_rps(struct drm_i915_private *dev_priv);
+extern void gen6_update_ring_freq(struct drm_i915_private *dev_priv);
 extern void gen6_disable_rps(struct drm_device *dev);
 extern void intel_init_emon(struct drm_device *dev);
#7 lore raw
Quoted 4 lines · Jesse Barnes

I tested again and things were working as expected. This patch adds a debugfs file for dumping the ring freq table, can you guys test it out (try disabling the call to update the table, dump it, make sure it's 0, then re-enable and make sure it's programmed correctly).

According to the debugfs, your defaults should be effective:

sandybridge:~$ cat /sys/kernel/debug/dri/1/i915_ring_freq_table 
GPU freq	Effective CPU freq
850		0 -> 1600
900		0 -> 1700
950		0 -> 1800
1000		0 -> 1800
1050		0 -> 1900
1100		0 -> 2000

I'll give the tyres another kick and see if I can spot any differences. -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#8 lore raw
Quoted 17 lines · On Thu, 23 Jun 2011 21:08:02 +0100 Chris Wilson <[email protected]>…
Quoted 4 lines · Jesse Barnes

I tested again and things were working as expected. This patch adds a debugfs file for dumping the ring freq table, can you guys test it out (try disabling the call to update the table, dump it, make sure it's 0, then re-enable and make sure it's programmed correctly).

According to the debugfs, your defaults should be effective:

sandybridge:~$ cat /sys/kernel/debug/dri/1/i915_ring_freq_table 
GPU freq	Effective CPU freq
850		0 -> 1600
900		0 -> 1700
950		0 -> 1800
1000		0 -> 1800
1050		0 -> 1900
1100		0 -> 2000
Quoted 1 line

I'll give the tyres another kick and see if I can spot any differences.

Eric was seeing an increase, but not as large as expected or hoped for. He saw a greater increase in performance than this patch provides by simply keeping the CPU busy during his benchmark. That could be due to additional ring frequency when the CPU is in a turbo mode. You may be able to approach that by making the effective CPU frequency much higher (maybe going up to 4GHz).

Signature · 1 line
Jesse Barnes, Intel Open Source Technology Center