[PATCH] ARM: dts: exynos: Add energy model for Exynos 5250

This energy model enables the use of energy- and capacity-aware CPU
frequency scaling.

Power and performance measurements were made using my freqbench [1]
benchmark coordinator, which uses EEMBC CoreMark [2] as the workload
and measures power usage using the integrated PMIC's fuel gauge (DS2784
in this case).

The energy model dynamic-power-coefficient values were calculated with
    DPC = µW / MHz / V^2
for each OPP, and averaged across all OPPs within each cluster for the
final coefficient.

A Google Nexus 10 device running a downstream 3.4 kernel was used for
benchmarking to ensure proper frequency scaling and other low-level
controls.

Raw benchmark results can be found in the freqbench repository [3].
Below is a human-readable summary:

===== CPU 1 =====
Frequencies: 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700

 200:   909     4.5 C/MHz    132 mW   36.2 J    6.9 I/mJ   275.0 s
 300:  1366     4.6 C/MHz    212 mW   38.7 J    6.5 I/mJ   183.0 s
 400:  1821     4.6 C/MHz    286 mW   39.3 J    6.4 I/mJ   137.3 s
 500:  2253     4.5 C/MHz    375 mW   41.7 J    6.0 I/mJ   111.0 s
 600:  2740     4.6 C/MHz    446 mW   40.7 J    6.1 I/mJ    91.2 s
 700:  3199     4.6 C/MHz    513 mW   40.1 J    6.2 I/mJ    78.2 s
 800:  3673     4.6 C/MHz    678 mW   46.1 J    5.4 I/mJ    68.1 s
 900:  4090     4.5 C/MHz    764 mW   46.7 J    5.4 I/mJ    61.1 s
1000:  4586     4.6 C/MHz    878 mW   47.9 J    5.2 I/mJ    54.5 s
1100:  5060     4.6 C/MHz   1084 mW   53.6 J    4.7 I/mJ    49.4 s
1200:  5515     4.6 C/MHz   1225 mW   55.5 J    4.5 I/mJ    45.3 s
1300:  5933     4.6 C/MHz   1396 mW   58.9 J    4.2 I/mJ    42.1 s
1400:  6395     4.6 C/MHz   1662 mW   65.0 J    3.8 I/mJ    39.1 s
1500:  6897     4.6 C/MHz   1895 mW   68.7 J    3.6 I/mJ    36.3 s
1600:  7332     4.6 C/MHz   2198 mW   75.0 J    3.3 I/mJ    34.1 s
1700:  7826     4.6 C/MHz   2497 mW   79.8 J    3.1 I/mJ    31.9 s

[1] https://github.com/kdrag0n/freqbench
[2] https://www.eembc.org/coremark/
[3] https://github.com/kdrag0n/freqbench/tree/master/results/exynos5250/main

Signed-off-by: Danny Lin <redacted>
---
 arch/arm/boot/dts/exynos5250.dtsi | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/arch/arm/boot/dts/exynos5250.dtsi b/arch/arm/boot/dts/exynos5250.dtsi
index 2ea2caaca4e2..cc2fe0afcfc7 100644
--- a/arch/arm/boot/dts/exynos5250.dtsi
+++ b/arch/arm/boot/dts/exynos5250.dtsi

@@ -58,6 +58,8 @@ cpu0: cpu@0 {
 			clock-names = "cpu";
 			operating-points-v2 = <&cpu0_opp_table>;
 			#cooling-cells = <2>; /* min followed by max */
+			capacity-dmips-mhz = <1024>;
+			dynamic-power-coefficient = <800>;
 		};
 		cpu1: cpu@1 {
 			device_type = "cpu";

@@ -67,6 +69,20 @@ cpu1: cpu@1 {
 			clock-names = "cpu";
 			operating-points-v2 = <&cpu0_opp_table>;
 			#cooling-cells = <2>; /* min followed by max */
+			capacity-dmips-mhz = <1024>;
+			dynamic-power-coefficient = <800>;
+		};
+
+		cpu-map {
+			cluster0 {
+				core0 {
+					cpu = <&cpu0>;
+				};
+
+				core1 {
+					cpu = <&cpu1>;
+				};
+			};
 		};
 	};
 

-- 
2.29.2

On Mon, Jan 11, 2021 at 03:57:16PM -0800, Danny Lin wrote:

This energy model enables the use of energy- and capacity-aware CPU
frequency scaling.

Power and performance measurements were made using my freqbench [1]
benchmark coordinator, which uses EEMBC CoreMark [2] as the workload
and measures power usage using the integrated PMIC's fuel gauge (DS2784
in this case).

Thanks for the patch.

The tested setup is not representative here. The Nexus 10 is not
supported by mainline and it might use specific revision of Exynos5250,
different than present on other mainlined boards. One could assume that
ratio of dynamic power coefficient of the cores should be similar... but
there is no ratio here, as this is not a big.LITTLE platform.

Another problem is the v3.4 vendor kernel with a lot of out-of-tree
code. This means it might use some different scheduler, different
drivers, different voltages and a lot more of unknown differences.
Vendor kernel should not matter that much in measurement of DPC but it
makes the results not possible to reproduce.

You were also measuring the power delivered to entire system, no to CPUs,
so you included static power in the data. Static power of CPUs and of
the entire system.

The energy model dynamic-power-coefficient values were calculated with
    DPC = µW / MHz / V^2
for each OPP, and averaged across all OPPs within each cluster for the
final coefficient.

A Google Nexus 10 device running a downstream 3.4 kernel was used for
benchmarking to ensure proper frequency scaling and other low-level
controls.

Raw benchmark results can be found in the freqbench repository [3].
Below is a human-readable summary:

===== CPU 1 =====
Frequencies: 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
 200:   909     4.5 C/MHz    132 mW   36.2 J    6.9 I/mJ   275.0 s

What are the columns here? I would expect that fuel gauge gives you the
current, but it's not here.

The origin of this 1024 value should be explained.

That's a second patch.

Best regards,
Krzysztof

+			cluster0 {
+				core0 {
+					cpu = <&cpu0>;
+				};
+
+				core1 {
+					cpu = <&cpu1>;
+				};
+			};
 		};
 	};
 
-- 
2.29.2