Skip to contentSkip to conversation

OMAP: Modifying the frequency comparison logic.

PATCH In review thara gopinath 3 participants 8 messages Aug 14, 2010 – Sep 16, 2010

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

PATCH 1/1 OMAP: Modifying the frequency comparison logic.

From: Thara Gopinath <[email protected]>

In the current opp layer the frequency matching API's tries to match the exact frequency passed in Hz. This leads to problems in cases where dplls are locked to slightly differnet frequencies due to sys clock speed or optimum M,N values. Consider the following scenario
a. dpll_x is supposed to be locked at 266Mhz but is
actually att 266.045 Mhz due to above mentioned reasons.
b. The opp table has the entry as 266000000 hz. c. The user does a clk_get_rate(dpll_x) and passes the
corresponding rate <rate> into the opp API's to get back a opp table entry.
d. In this scenario if the user has called into
opp_find_freq_exact it will return an error. If the user has called into opp_find_freq_ceil it will either return the opp entry corresponding to the next higher frequency in the opp table or return an error if 266 Mhz is the last entry in the table.

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

Now there are two ways of fixing this issue.
a. Put the correct dpll lock frequency in the opp table.
This means opp table will now have entries like 266045000 hz. But this is not scalable as these dpll lock frequencies can change slightly based on sys clk speeds. Like for eg at sys clk speed 38.4 Mhz we will able to get 266.045 Mhz where as at sys clk speed 26 Mhz we will be able to get onyl 266.124 Mhz etc. So depending on the platform we will have to start changing the opp table.

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp layer
API and the rates stored in the opp tables by 1000000 to get the rates in Mhz and do the necessary comparision. In this approach any vague frequency like 266.045Mhz will get mapped to 266 Mhz in the opp table. But if the passed rate is 267 Mhz, the opp framework will still rerturn an error or the next highest opp table entry

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird frequencies due to sys clk being at 38.4 Mhz.

---
 arch/arm/plat-omap/opp.c |   37 ++++++++++++++++++++++++++-----------
 1 files changed, 26 insertions(+), 11 deletions(-)
arch/arm/plat-omap/opp.c +26 -11
index b9b7bda..ba7a554 100644
@@ -212,15 +212,20 @@ struct omap_opp *opp_find_freq_exact(struct device *dev,
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate == freq) {
-			opp = temp_opp;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate == req_freq) {
+				opp = temp_opp;
+				break;
+			}
 		}
 	}
 
@@ -258,16 +263,21 @@ struct omap_opp *opp_find_freq_ceil(struct device *dev, unsigned long *freq)
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = *freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate >= *freq) {
-			opp = temp_opp;
-			*freq = opp->rate;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate >= req_freq) {
+				opp = temp_opp;
+				*freq = opp->rate;
+				break;
+			}
 		}
 	}
 
@@ -305,16 +315,21 @@ struct omap_opp *opp_find_freq_floor(struct device *dev, unsigned long *freq)
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = *freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry_reverse(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate <= *freq) {
-			opp = temp_opp;
-			*freq = opp->rate;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate <= req_freq) {
+				opp = temp_opp;
+				*freq = opp->rate;
+				break;
+			}
 		}
 	}
 
Signature · 1 line
1.7.1.GIT
#1 lore raw
Quoted 42 lines · thara gopinath

From: Thara Gopinath <[email protected]>

In the current opp layer the frequency matching API's tries to match the exact frequency passed in Hz. This leads to problems in cases where dplls are locked to slightly differnet frequencies due to sys clock speed or optimum M,N values. Consider the following scenario
a. dpll_x is supposed to be locked at 266Mhz but is
actually att 266.045 Mhz due to above mentioned reasons.
b. The opp table has the entry as 266000000 hz. c. The user does a clk_get_rate(dpll_x) and passes the
corresponding rate <rate> into the opp API's to get back a opp table entry.
d. In this scenario if the user has called into
opp_find_freq_exact it will return an error. If the user has called into opp_find_freq_ceil it will either return the opp entry corresponding to the next higher frequency in the opp table or return an error if 266 Mhz is the last entry in the table.

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

Now there are two ways of fixing this issue.
a. Put the correct dpll lock frequency in the opp table.
This means opp table will now have entries like 266045000 hz. But this is not scalable as these dpll lock frequencies can change slightly based on sys clk speeds. Like for eg at sys clk speed 38.4 Mhz we will able to get 266.045 Mhz where as at sys clk speed 26 Mhz we will be able to get onyl 266.124 Mhz etc. So depending on the platform we will have to start changing the opp table.

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp layer
API and the rates stored in the opp tables by 1000000 to get the rates in Mhz and do the necessary comparision. In this approach any vague frequency like 266.045Mhz will get mapped to 266 Mhz in the opp table. But if the passed rate is 267 Mhz, the opp framework will still rerturn an error or the next highest opp table entry

Quoted 3 lines

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird frequencies due to sys clk being at 38.4 Mhz.

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find functions should be updated to describe the new matching technique.

Kevin

Quoted 84 lines
---
 arch/arm/plat-omap/opp.c |   37 ++++++++++++++++++++++++++-----------
 1 files changed, 26 insertions(+), 11 deletions(-)
arch/arm/plat-omap/opp.c +26 -11
index b9b7bda..ba7a554 100644
@@ -212,15 +212,20 @@ struct omap_opp *opp_find_freq_exact(struct device *dev,
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate == freq) {
-			opp = temp_opp;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate == req_freq) {
+				opp = temp_opp;
+				break;
+			}
 		}
 	}
 
@@ -258,16 +263,21 @@ struct omap_opp *opp_find_freq_ceil(struct device *dev, unsigned long *freq)
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = *freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate >= *freq) {
-			opp = temp_opp;
-			*freq = opp->rate;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate >= req_freq) {
+				opp = temp_opp;
+				*freq = opp->rate;
+				break;
+			}
 		}
 	}
 
@@ -305,16 +315,21 @@ struct omap_opp *opp_find_freq_floor(struct device *dev, unsigned long *freq)
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = *freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry_reverse(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate <= *freq) {
-			opp = temp_opp;
-			*freq = opp->rate;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate <= req_freq) {
+				opp = temp_opp;
+				*freq = opp->rate;
+				break;
+			}
 		}
 	}
#2 lore raw
Quoted 54 lines

-----Original Message-----
From: Kevin Hilman [mailto:[email protected]]
Sent: Tuesday, August 24, 2010 2:04 PM

Quoted 34 lines · thara gopinath

From: Thara Gopinath <[email protected]>

In the current opp layer the frequency matching API's tries to match the exact frequency passed in Hz. This leads to problems in cases where dplls are locked to slightly differnet frequencies due to sys clock speed or optimum M,N values. Consider the following scenario
a. dpll_x is supposed to be locked at 266Mhz but is
actually att 266.045 Mhz due to above mentioned reasons.
b. The opp table has the entry as 266000000 hz. c. The user does a clk_get_rate(dpll_x) and passes the
corresponding rate <rate> into the opp API's to get back a opp table entry.
d. In this scenario if the user has called into
opp_find_freq_exact it will return an error. If the user has called into opp_find_freq_ceil it will either return the opp entry corresponding to the next higher frequency in the opp table or return an error if 266 Mhz is the last entry in the table.

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

Now there are two ways of fixing this issue.
a. Put the correct dpll lock frequency in the opp table.
This means opp table will now have entries like 266045000 hz. But this is not scalable as these dpll lock frequencies can change slightly based on sys clk speeds. Like for eg at sys clk speed 38.4 Mhz we will able to get 266.045 Mhz where as at sys clk speed 26 Mhz we will be able to get onyl 266.124 Mhz etc. So depending on the platform we will have to start changing the opp table.

Quoted 4 lines

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp layer
API and the rates stored in the opp tables by 1000000 to get the rates in Mhz and do the necessary comparision. In this approach

any

Quoted 6 lines

vague frequency like 266.045Mhz will get mapped to 266 Mhz in the opp table. But if the passed rate is 267 Mhz, the opp framework will still rerturn an error or the next highest opp table entry

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird

Quoted 8 lines

frequencies

Quoted 1 line

due to sys clk being at 38.4 Mhz.

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find functions should be updated to describe the new matching technique.

I agree, I suggest one improvement though - the search accuracy will vary Based on the silicon rev, one size will probably not fit every silicon and Domains we have - I suggest having accuracy as a parameter as part of domain Registration/configurable parameter
e.g.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +1 -0
+			unsigned long rate = temp_opp->rate / 1000000;

Will probably configurable to the "exactness" we expect to handle per domain/silicon family.

Quoted 75 lines

Kevin

arch/arm/plat-omap/opp.c struct omap_opp *opp_find_freq_exact(struct device

Quoted 10 lines
---
 arch/arm/plat-omap/opp.c |   37 ++++++++++++++++++++++++++-----------
 1 files changed, 26 insertions(+), 11 deletions(-)
arch/arm/plat-omap/opp.c +0 -0
index b9b7bda..ba7a554 100644
@@ -212,15 +212,20 @@ struct omap_opp *opp_find_freq_exact(struct device

*dev,

arch/arm/plat-omap/opp.c struct omap_opp *opp_find_freq_ceil(struct device

Quoted 24 lines
arch/arm/plat-omap/opp.c +8 -3
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate == freq) {
-			opp = temp_opp;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate == req_freq) {
+				opp = temp_opp;
+				break;
+			}
 		}
 	}
 
@@ -258,16 +263,21 @@ struct omap_opp *opp_find_freq_ceil(struct device

*dev, unsigned long *freq)

arch/arm/plat-omap/opp.c struct omap_opp *opp_find_freq_floor(struct device

Quoted 26 lines
arch/arm/plat-omap/opp.c +9 -4
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = *freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
 	list_for_each_entry(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate >= *freq) {
-			opp = temp_opp;
-			*freq = opp->rate;
-			break;
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate >= req_freq) {
+				opp = temp_opp;
+				*freq = opp->rate;
+				break;
+			}
 		}
 	}
 
@@ -305,16 +315,21 @@ struct omap_opp *opp_find_freq_floor(struct device

*dev, unsigned long *freq)

Quoted 9 lines
arch/arm/plat-omap/opp.c +1 -0
 {
 	struct device_opp *dev_opp;
 	struct omap_opp *temp_opp, *opp = ERR_PTR(-ENODEV);
+	unsigned long req_freq = *freq / 1000000;
 
 	dev_opp = find_device_opp(dev);
 	if (IS_ERR(dev_opp))
 		return opp;
 
Quoted 15 lines
Quoted 5 lines
arch/arm/plat-omap/opp.c +0 -4
 	list_for_each_entry_reverse(temp_opp, &dev_opp->opp_list, node) {
-		if (temp_opp->enabled && temp_opp->rate <= *freq) {
-			opp = temp_opp;
-			*freq = opp->rate;
-			break;

arch/arm/plat-omap/opp.c struct omap_opp *opp_find_freq_floor(struct device *dev, unsigned long *freq) · line 334

Quoted 10 lines
arch/arm/plat-omap/opp.c +8 -0
+		if (temp_opp->enabled) {
+			unsigned long rate = temp_opp->rate / 1000000;
+
+			if (rate <= req_freq) {
+				opp = temp_opp;
+				*freq = opp->rate;
+				break;
+			}
 		}
 	}

Regards,
Nishanth Menon

#3 lore raw
Quoted 57 lines · Nishanth
Quoted 48 lines
Quoted 34 lines · thara gopinath

From: Thara Gopinath <[email protected]>

In the current opp layer the frequency matching API's tries to match the exact frequency passed in Hz. This leads to problems in cases where dplls are locked to slightly differnet frequencies due to sys clock speed or optimum M,N values. Consider the following scenario
a. dpll_x is supposed to be locked at 266Mhz but is
actually att 266.045 Mhz due to above mentioned reasons.
b. The opp table has the entry as 266000000 hz. c. The user does a clk_get_rate(dpll_x) and passes the
corresponding rate <rate> into the opp API's to get back a opp table entry.
d. In this scenario if the user has called into
opp_find_freq_exact it will return an error. If the user has called into opp_find_freq_ceil it will either return the opp entry corresponding to the next higher frequency in the opp table or return an error if 266 Mhz is the last entry in the table.

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

Now there are two ways of fixing this issue.
a. Put the correct dpll lock frequency in the opp table.
This means opp table will now have entries like 266045000 hz. But this is not scalable as these dpll lock frequencies can change slightly based on sys clk speeds. Like for eg at sys clk speed 38.4 Mhz we will able to get 266.045 Mhz where as at sys clk speed 26 Mhz we will be able to get onyl 266.124 Mhz etc. So depending on the platform we will have to start changing the opp table.

Quoted 4 lines

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp layer
API and the rates stored in the opp tables by 1000000 to get the rates in Mhz and do the necessary comparision. In this approach

any

Quoted 6 lines

vague frequency like 266.045Mhz will get mapped to 266 Mhz in the opp table. But if the passed rate is 267 Mhz, the opp framework will still rerturn an error or the next highest opp table entry

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird

Quoted 8 lines

frequencies

Quoted 1 line

due to sys clk being at 38.4 Mhz.

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find functions should be updated to describe the new matching technique.

Quoted 8 lines

I agree, I suggest one improvement though - the search accuracy will vary Based on the silicon rev, one size will probably not fit every silicon and Domains we have - I suggest having accuracy as a parameter as part of domain Registration/configurable parameter
e.g.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +1 -0
+			unsigned long rate = temp_opp->rate / 1000000;

Will probably configurable to the "exactness" we expect to handle per domain/silicon family.

The more I think about, I think we should leave the 'exact' find the way it is, especially as we move to device OPPs we will probably want to have more precise matching.

What about adding another function that does a "find closest"?

Kevin

#4 lore raw
Quoted 12 lines

From: Kevin Hilman [mailto:[email protected]]
Sent: Tuesday, August 24, 2010 4:14 PM

Quoted 5 lines · Nishanth
Quoted 5 lines
Quoted 2 lines · thara gopinath

From: Thara Gopinath <[email protected]>

[...]

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

[...]

Quoted 31 lines
Quoted 15 lines
Quoted 15 lines
Quoted 5 lines

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp layer
API and the rates stored in the opp tables by 1000000 to get the rates in Mhz and do the necessary comparision. In this approach

any

Quoted 6 lines

vague frequency like 266.045Mhz will get mapped to 266 Mhz in the opp table. But if the passed rate is 267 Mhz, the opp framework will still rerturn an error or the next highest opp table entry

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird

frequencies

Quoted 1 line

due to sys clk being at 38.4 Mhz.

Quoted 7 lines
Quoted 5 lines

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find functions should be updated to describe the new matching technique.

I agree, I suggest one improvement though - the search accuracy will

vary

Quoted 1 line

Based on the silicon rev, one size will probably not fit every silicon

and

Quoted 1 line

Domains we have - I suggest having accuracy as a parameter as part of

domain

Quoted 4 lines

Registration/configurable parameter
e.g.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +1 -0
+			unsigned long rate = temp_opp->rate / 1000000;

Will probably configurable to the "exactness" we expect to handle per

Quoted 8 lines

domain/silicon family.

Quoted 1 line

The more I think about, I think we should leave the 'exact' find the way it is, especially as we move to device OPPs we will probably want to have more precise matching.

What about adding another function that does a "find closest"?

Just my 2cents:
With accuracy as a param? Then we fall back to the question - who would be
the users of the "_ceil, _floor" functions? Probably exact might should mean exact, and ceil and floor should be using as they were originally intended..

But, the real users of ceil and floor are guys from cpufreq and frequency searchers - they don't really want to know the minor deltas b/w 19.2MHz Vs 26MHz Vs 38.4MHz sysclk variations caused to precise clock definitions.. They like to do ciel(1Ghz) and get opp corresponding to 1Ghz if 19.2Mhz sysclk causes this to be 999,999Mhz it will fail in the current logic..

Also, we would like to users of these apis to remain consistent and not Know about the variation of freq deltas. E.g.:
Omapx 12MHz to 38.4Mhz causes fluctuations around 100Mhz OMAPy 19.2MHz to 38.4Mhz causes fluctuations around 10Mhz The caller should not do:
If cpu_is_omapx()

	Find_closest_ciel(freq,100)
Else if cpu_is_omapy()
	Find_closest_ciel(freq,10)

That'd be nightmare..

Instead as part of cpu domain registration, we mention what is the accuracy And the callers do find_ciel(freq) and that will "automagically" translate To accuracy needed for that silicon on that domain..

Regards,
Nishanth Menon

#5 lore raw
Quoted 54 lines · Nishanth
Quoted 12 lines

From: Kevin Hilman [mailto:[email protected]]
Sent: Tuesday, August 24, 2010 4:14 PM

Quoted 5 lines · Nishanth
Quoted 5 lines
Quoted 2 lines · thara gopinath

From: Thara Gopinath <[email protected]>

[...]

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

[...]

Quoted 34 lines
Quoted 15 lines
Quoted 15 lines
Quoted 5 lines

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp layer
API and the rates stored in the opp tables by 1000000 to get the rates in Mhz and do the necessary comparision. In this approach

any

Quoted 6 lines

vague frequency like 266.045Mhz will get mapped to 266 Mhz in the opp table. But if the passed rate is 267 Mhz, the opp framework will still rerturn an error or the next highest opp table entry

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird

frequencies

Quoted 1 line

due to sys clk being at 38.4 Mhz.

Quoted 7 lines
Quoted 5 lines

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find functions should be updated to describe the new matching technique.

I agree, I suggest one improvement though - the search accuracy will

vary

Quoted 1 line

Based on the silicon rev, one size will probably not fit every silicon

and

Quoted 1 line

Domains we have - I suggest having accuracy as a parameter as part of

domain

Quoted 4 lines

Registration/configurable parameter
e.g.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +1 -0
+			unsigned long rate = temp_opp->rate / 1000000;

Will probably configurable to the "exactness" we expect to handle per

domain/silicon family.

Quoted 1 line
Quoted 7 lines
Quoted 5 lines

The more I think about, I think we should leave the 'exact' find the way it is, especially as we move to device OPPs we will probably want to have more precise matching.

What about adding another function that does a "find closest"?

Just my 2cents: With accuracy as a param?

Why would you need accuracy for "find closest"?

Quoted 3 lines

Then we fall back to the question - who would be the users of the "_ceil, _floor" functions? Probably exact might should mean exact, and ceil and floor should be using as they were originally intended..

Yeah, I'm thinking so too.

Quoted 5 lines

But, the real users of ceil and floor are guys from cpufreq and frequency searchers - they don't really want to know the minor deltas b/w 19.2MHz Vs 26MHz Vs 38.4MHz sysclk variations caused to precise clock definitions.. They like to do ciel(1Ghz) and get opp corresponding to 1Ghz if 19.2Mhz sysclk causes this to be 999,999Mhz it will fail in the current logic..

Yes, that's why I proposed "closest".

Quoted 11 lines

Also, we would like to users of these apis to remain consistent and not Know about the variation of freq deltas. E.g.:
Omapx 12MHz to 38.4Mhz causes fluctuations around 100Mhz OMAPy 19.2MHz to 38.4Mhz causes fluctuations around 10Mhz The caller should not do:
If cpu_is_omapx()

	Find_closest_ciel(freq,100)
Else if cpu_is_omapy()
	Find_closest_ciel(freq,10)

That'd be nightmare..

Agreed.

Kevin

Quoted 3 lines

Instead as part of cpu domain registration, we mention what is the accuracy And the callers do find_ciel(freq) and that will "automagically" translate To accuracy needed for that silicon on that domain..

#6 lore raw
Quoted 74 lines

-----Original Message-----
From: Kevin Hilman [mailto:[email protected]]
Sent: Wednesday, August 25, 2010 5:41 PM
To: Menon, Nishanth
Cc: Gopinath, Thara; [email protected]
Subject: Re: [PM-OPP][PATCH] OMAP: Modifying the frequency comparison logic.

Quoted 15 lines · Nishanth
Quoted 12 lines

From: Kevin Hilman [mailto:[email protected]]
Sent: Tuesday, August 24, 2010 4:14 PM

Quoted 5 lines · Nishanth
Quoted 5 lines
Quoted 2 lines · thara gopinath

From: Thara Gopinath <[email protected]>

[...]

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

[...]

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp

layer

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

API and the rates stored in the opp tables by 1000000 to get

the

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

rates in Mhz and do the necessary comparision. In this

approach

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

any

Quoted 1 line

vague frequency like 266.045Mhz will get mapped to 266 Mhz

in the

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

opp table. But if the passed rate is 267 Mhz, the opp

framework

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

will still rerturn an error or the next highest opp table

entry

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 3 lines

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird

frequencies

Quoted 1 line

due to sys clk being at 38.4 Mhz.

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find

functions

Quoted 5 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

should be updated to describe the new matching technique.

I agree, I suggest one improvement though - the search accuracy will

vary

Quoted 1 line

Based on the silicon rev, one size will probably not fit every

silicon

Quoted 11 lines
Quoted 11 lines

and

Quoted 1 line

Domains we have - I suggest having accuracy as a parameter as part of

domain

Quoted 4 lines

Registration/configurable parameter
e.g.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +1 -0
+			unsigned long rate = temp_opp->rate / 1000000;

Will probably configurable to the "exactness" we expect to handle per

domain/silicon family.

Quoted 1 line

The more I think about, I think we should leave the 'exact' find the

way

Quoted 9 lines
Quoted 6 lines
Quoted 4 lines

it is, especially as we move to device OPPs we will probably want to have more precise matching.

What about adding another function that does a "find closest"?

Just my 2cents: With accuracy as a param?

Why would you need accuracy for "find closest"?

How close should "closest" be? There has to be a parameter for this - something like as close as 10Mhz, but beyond the match fails. Currently this patch introduces a hardcoded accuracy which could be Carried forward.

But, On different silicon the deltas due to sysclk can vary - there is no silver bullet algo we can write which will be accurate for all silicon and domain combination..

Further, is'nt ceil and floor functions we have a more specific Implementation of "closest"? in the sense, it says closest to which Direction - up or down..

[...]
Rest of the discussion seems aligned.

Regards,
Nishanth Menon

#7 lore raw
Quoted 102 lines
Quoted 102 lines

-----Original Message-----
From: Menon, Nishanth
Sent: Thursday, August 26, 2010 6:00 AM
To: Kevin Hilman
Cc: Gopinath, Thara; [email protected]
Subject: RE: [PM-OPP][PATCH] OMAP: Modifying the frequency comparison logic.

Quoted 74 lines

-----Original Message-----
From: Kevin Hilman [mailto:[email protected]]
Sent: Wednesday, August 25, 2010 5:41 PM
To: Menon, Nishanth
Cc: Gopinath, Thara; [email protected]
Subject: Re: [PM-OPP][PATCH] OMAP: Modifying the frequency comparison logic.

Quoted 15 lines · Nishanth
Quoted 12 lines

From: Kevin Hilman [mailto:[email protected]]
Sent: Tuesday, August 24, 2010 4:14 PM

Quoted 5 lines · Nishanth
Quoted 5 lines
Quoted 2 lines · thara gopinath

From: Thara Gopinath <[email protected]>

[...]

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

The above is not correct as we expect the framework to return back the opp table entry corresponding to 266 Mhz.

[...]

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

b. Do the comparison in Mhz in the opp layer rather than in Hz.
This would mean we will divide the rate passed into the opp

layer

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

API and the rates stored in the opp tables by 1000000 to get

the

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

rates in Mhz and do the necessary comparision. In this

approach

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

any

Quoted 1 line

vague frequency like 266.045Mhz will get mapped to 266 Mhz

in the

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

opp table. But if the passed rate is 267 Mhz, the opp

framework

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

will still rerturn an error or the next highest opp table

entry

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 3 lines

This patch implements solution b. The scenario mentioned above is esp true for OMAP4 dpll_iva where we do end up with such weird

frequencies

Quoted 1 line

due to sys clk being at 38.4 Mhz.

I agree that solution b is better, although it makes the '_exact' function a bit less exact. :/

solution b is fine with me, but the kerneldoc for these find

functions

Quoted 5 lines
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

should be updated to describe the new matching technique.

I agree, I suggest one improvement though - the search accuracy will

vary

Quoted 1 line

Based on the silicon rev, one size will probably not fit every

silicon

Quoted 11 lines
Quoted 11 lines

and

Quoted 1 line

Domains we have - I suggest having accuracy as a parameter as part of

domain

Quoted 4 lines

Registration/configurable parameter
e.g.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +1 -0
+			unsigned long rate = temp_opp->rate / 1000000;

Will probably configurable to the "exactness" we expect to handle per

domain/silicon family.

Quoted 1 line

The more I think about, I think we should leave the 'exact' find the

way

Quoted 9 lines
Quoted 6 lines
Quoted 4 lines

it is, especially as we move to device OPPs we will probably want to have more precise matching.

What about adding another function that does a "find closest"?

Just my 2cents: With accuracy as a param?

Why would you need accuracy for "find closest"?

How close should "closest" be? There has to be a parameter for this - something like as close as 10Mhz, but beyond the match fails. Currently this patch introduces a hardcoded accuracy which could be Carried forward.

But, On different silicon the deltas due to sysclk can vary - there is no silver bullet algo we can write which will be accurate for all silicon and domain combination..

Quoted 6 lines
Quoted 6 lines

Further, is'nt ceil and floor functions we have a more specific Implementation of "closest"? in the sense, it says closest to which Direction - up or down..

[...]
Rest of the discussion seems aligned.

Hello Kevin,

I am not sure about the conclusion of this discussion.
For me we should avoid putting possible weird frequencies like 266789501.. Hence the round off to /10^6. If on some platforms due to sys clk we have frequency 267 instead of 266, I would prefer to edit the opp table through the opp framework APIs. This is my opinion. I may not have fully understood the concerns discussed in this forum.

Regards
Thara