Re: [PATCH] arm64: dts: qcom: sdm660: Fix CPU capacities

From: Danny Lin
Date: Tue Jan 12 2021 - 23:39:06 EST


On Tue, Jan 12, 2021 at 8:04 pm, AngeloGioacchino Del Regno <angelogioacchino.delregno@xxxxxxxxxxxxxx> wrote:
Il 12/01/21 15:59, Alexey Minnekhanov ha scritto:
Hi!
I always had a feeling something is not right in those cpu
definitions, so cpus with reg 100-103 are little cores, and 0-3 big
ones?
But downstream sdm660.dtsi has a property "efficiency" [1] with values
which are larger for cores 100-103 than for 0-3 cores (1638 > 1024),
I'm confused...


It appears that in downstream, logical CPUs 0-3 are mapped to the little CPUs (physical IDs 0x100-0x103) and logical CPUs 4-7 are mapped to the big CPUs (physical IDs 0x0-0x3), while mainline has it reversed unlike most other Qualcomm platforms. Thanks for catching the discrepancy, I wasn't aware that the CPU reg values actually mattered before this.

All the SDM630, SDM636 and SDM660 smartphones I ever saw are booting off of the BIG cluster (and that's why cpu@100 is CPU0), and at 0x100 you find the first BIG CPU.

It looks like the mainline logical CPU map for SDM660 was originally copied from SDM630 and was not updated to match the conventional layout for SDM660. I'm not familiar with SDM630, but on the SDM660 devices I've seen, the boot CPU is CPU 0 on the little cluster (boot log is from downstream but logical CPU numbers are not relevant here) as indicated by the cpuid:

[ 0.000000] Boot CPU: AArch64 Processor [51af8014]
[ 0.029322] CPU1: Booted secondary processor [51af8014]
[ 0.034276] CPU2: Booted secondary processor [51af8014]
[ 0.039177] CPU3: Booted secondary processor [51af8014]
[ 0.044637] CPU4: Booted secondary processor [51af8002]
[ 0.049645] CPU5: Booted secondary processor [51af8002]
[ 0.054926] CPU6: Booted secondary processor [51af8002]
[ 0.059934] CPU7: Booted secondary processor [51af8002]

See downstream device trees for SDM630 [1] and SDM660 [2] - the cluster order is different and only SDM630 matches the behavior you describe.

In either case, it works as long as the capacities are assigned to the correct logical CPUs, so please disregard this patch if the current CPU map is retained.

[1] https://source.codeaurora.org/quic/la/kernel/msm-4.4/tree/arch/arm/boot/dts/qcom/sdm630.dtsi?h=LA.UM.6.2.c27-03100-sdm660.0#n49
[2] https://source.codeaurora.org/quic/la/kernel/msm-4.4/tree/arch/arm/boot/dts/qcom/sdm660.dtsi?h=LA.UM.6.2.c27-03100-sdm660.0#n49



Property "efficiency" is described in the same tree in [2].

[1] https://source.codeaurora.org/quic/la/kernel/msm-4.4/tree/arch/arm/boot/dts/qcom/sdm660.dtsi?h=LA.UM.7.2.c25#n155
[2] https://source.codeaurora.org/quic/la/kernel/msm-4.4/tree/Documentation/devicetree/bindings/arm/cpus.txt?h=LA.UM.7.2.c25#n216

вт, 12 янв. 2021 г. в 13:51, Danny Lin <danny@xxxxxxxxxxx>:

sdm660 has a big.LITTLE 4+4 CPU setup with CPUs 0-3 being little cores
and CPUs 4-7 being big cores. The big cores have higher IPC, so they
should have the higher capacity-dmips-mhz, not the other way around as
the device tree currently describes it. Fix the incorrect CPU map to
improve EAS scheduling behavior.

While we're at it, let's replace the old DMIPS/MHz values with new
measurements that reflect the exact IPC of the CPUs as reported by
CoreMark.

Performance measurements were made using my freqbench [1]
benchmark coordinator, which isolates, offlines, and disables the timer
tick on test CPUs to maximize accuracy. It uses EEMBC CoreMark [2] as
the workload and measures power usage using the PM660 PMIC's fuel
gauge.

Normalized DMIPS/MHz capacity scale values for each CPU were calculated
from CoreMarks/MHz (CoreMark iterations per second per MHz), which
serves the same purpose. For each CPU, the final capacity-dmips-mhz
value is the C/MHz value of its maximum frequency normalized to
SCHED_CAPACITY_SCALE (1024) for the fastest CPU in the system.

A Xiaomi Redmi Note 7 device running a downstream Qualcomm 4.4 kernel
was used for benchmarking to ensure proper frequency scaling and other
low-level controls.

This is wrong, the downstream kernel may be doing "magic" to switch clusters the other way around, and this is likely... Please, run your benchmark on a upstream kernel: there may be differences.

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

Frequency domains: cpu1 cpu4
Offline CPUs: cpu1 cpu2 cpu3 cpu4 cpu5 cpu6 cpu7
Baseline power usage: 1130 mW

===== CPU 1 =====
Frequencies: 633 902 1113 1401 1536 1747 1843

633: 2058 3.2 C/MHz 48 mW 5.9 J 42.6 I/mJ 121.5 s
902: 2930 3.2 C/MHz 72 mW 6.2 J 40.6 I/mJ 85.3 s
1113: 3616 3.2 C/MHz 79 mW 5.4 J 46.0 I/mJ 69.1 s
1401: 4551 3.2 C/MHz 125 mW 6.9 J 36.3 I/mJ 54.9 s
1536: 4988 3.2 C/MHz 134 mW 6.7 J 37.1 I/mJ 50.1 s
1747: 5674 3.2 C/MHz 179 mW 7.9 J 31.7 I/mJ 44.1 s
1843: 5986 3.2 C/MHz 228 mW 9.5 J 26.3 I/mJ 41.8 s

===== CPU 4 =====
Frequencies: 1113 1401 1747 1958 2150 2208

1113: 5825 5.2 C/MHz 220 mW 9.4 J 26.5 I/mJ 42.9 s
1401: 7324 5.2 C/MHz 317 mW 10.8 J 23.1 I/mJ 34.1 s
1747: 9135 5.2 C/MHz 474 mW 13.0 J 19.2 I/mJ 27.4 s
1958: 10247 5.2 C/MHz 578 mW 14.1 J 17.7 I/mJ 24.4 s
2150: 11246 5.2 C/MHz 694 mW 15.4 J 16.2 I/mJ 22.2 s
2208: 11551 5.2 C/MHz 736 mW 15.9 J 15.7 I/mJ 21.7 s

[1] https://github.com/kdrag0n/freqbench
[2] https://www.eembc.org/coremark/
[3] https://github.com/kdrag0n/freqbench/tree/master/results/sdm660/main

Signed-off-by: Danny Lin <danny@xxxxxxxxxxx>
---
arch/arm64/boot/dts/qcom/sdm660.dtsi | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/arch/arm64/boot/dts/qcom/sdm660.dtsi b/arch/arm64/boot/dts/qcom/sdm660.dtsi
index 4abbdd03d1e7..ca985c5429db 100644
--- a/arch/arm64/boot/dts/qcom/sdm660.dtsi
+++ b/arch/arm64/boot/dts/qcom/sdm660.dtsi
@@ -40,7 +40,7 @@ CPU0: cpu@100 {
compatible = "qcom,kryo260";
reg = <0x0 0x100>;
enable-method = "psci";
- capacity-dmips-mhz = <1024>;
+ capacity-dmips-mhz = <636>;
next-level-cache = <&L2_1>;
L2_1: l2-cache {
compatible = "cache";
@@ -59,7 +59,7 @@ CPU1: cpu@101 {
compatible = "qcom,kryo260";
reg = <0x0 0x101>;
enable-method = "psci";
- capacity-dmips-mhz = <1024>;
+ capacity-dmips-mhz = <636>;
next-level-cache = <&L2_1>;
L1_I_101: l1-icache {
compatible = "cache";
@@ -74,7 +74,7 @@ CPU2: cpu@102 {
compatible = "qcom,kryo260";
reg = <0x0 0x102>;
enable-method = "psci";
- capacity-dmips-mhz = <1024>;
+ capacity-dmips-mhz = <636>;
next-level-cache = <&L2_1>;
L1_I_102: l1-icache {
compatible = "cache";
@@ -89,7 +89,7 @@ CPU3: cpu@103 {
compatible = "qcom,kryo260";
reg = <0x0 0x103>;
enable-method = "psci";
- capacity-dmips-mhz = <1024>;
+ capacity-dmips-mhz = <636>;
next-level-cache = <&L2_1>;
L1_I_103: l1-icache {
compatible = "cache";
@@ -104,7 +104,7 @@ CPU4: cpu@0 {
compatible = "qcom,kryo260";
reg = <0x0 0x0>;
enable-method = "psci";
- capacity-dmips-mhz = <640>;
+ capacity-dmips-mhz = <1024>;
next-level-cache = <&L2_0>;
L2_0: l2-cache {
compatible = "cache";
@@ -123,7 +123,7 @@ CPU5: cpu@1 {
compatible = "qcom,kryo260";
reg = <0x0 0x1>;
enable-method = "psci";
- capacity-dmips-mhz = <640>;
+ capacity-dmips-mhz = <1024>;
next-level-cache = <&L2_0>;
L1_I_1: l1-icache {
compatible = "cache";
@@ -138,7 +138,7 @@ CPU6: cpu@2 {
compatible = "qcom,kryo260";
reg = <0x0 0x2>;
enable-method = "psci";
- capacity-dmips-mhz = <640>;
+ capacity-dmips-mhz = <1024>;
next-level-cache = <&L2_0>;
L1_I_2: l1-icache {
compatible = "cache";
@@ -153,7 +153,7 @@ CPU7: cpu@3 {
compatible = "qcom,kryo260";
reg = <0x0 0x3>;
enable-method = "psci";
- capacity-dmips-mhz = <640>;
+ capacity-dmips-mhz = <1024>;
next-level-cache = <&L2_0>;
L1_I_3: l1-icache {
compatible = "cache";
--
2.29.2