arm64: dts: qcom: sc7180: Base dynamic CPU power coefficients in reality
The sc7180's dynamic-power-coefficient violates the device tree bindings. The bindings (arm/cpus.yaml) say that the units for the dynamic-power-coefficient are supposed to be "uW/MHz/V^2". The ones for sc7180 aren't this. Qualcomm arbitrarily picked 100 for the "little" CPUs and then picked a number for the big CPU based on this. At the time, there was a giant dicussion about this. Apparently Qualcomm Engineers were instructed not to share the actual numbers here. As part of the discussion, I pointed out [1] that these numbers shouldn't really be secret since once a device is shipping anyone can just run a script and produce them. This patch is the result of running the script I posted in that discussion on sc7180-trogdor-coachz, which is currently available for purchase by consumers. [1] https://lore.kernel.org/r/CAD=FV=U1FP0e3_AVHpauUUZtD-5X3XCwh5aT9fH_8S_FFML2Uw@mail.gmail.com/ I ran the script four times, measuring little, big, little, big. I used the 64-bit version of dhrystone 2.2 in my test. I got these results: 576 kHz, 596 mV, 20 mW, 88 Cx 768 kHz, 596 mV, 32 mW, 122 Cx 1017 kHz, 660 mV, 45 mW, 97 Cx 1248 kHz, 720 mV, 87 mW, 139 Cx 1324 kHz, 756 mV, 109 mW, 148 Cx 1516 kHz, 828 mV, 150 mW, 148 Cx 1612 kHz, 884 mV, 182 mW, 147 Cx 1708 kHz, 884 mV, 192 mW, 146 Cx 1804 kHz, 884 mV, 207 mW, 149 Cx Your dynamic-power-coefficient for cpu 0: 132 825 kHz, 596 mV, 142 mW, 401 Cx 979 kHz, 628 mV, 183 mW, 427 Cx 1113 kHz, 656 mV, 224 mW, 433 Cx 1267 kHz, 688 mV, 282 mW, 449 Cx 1555 kHz, 812 mV, 475 mW, 450 Cx 1708 kHz, 828 mV, 566 mW, 478 Cx 1843 kHz, 884 mV, 692 mW, 476 Cx 1900 kHz, 884 mV, 722 mW, 482 Cx 1996 kHz, 916 mV, 814 mW, 482 Cx 2112 kHz, 916 mV, 862 mW, 483 Cx 2208 kHz, 916 mV, 962 mW, 521 Cx 2323 kHz, 940 mV, 1060 mW, 517 Cx 2400 kHz, 956 mV, 1133 mW, 518 Cx Your dynamic-power-coefficient for cpu 6: 471 576 kHz, 596 mV, 26 mW, 103 Cx 768 kHz, 596 mV, 40 mW, 147 Cx 1017 kHz, 660 mV, 54 mW, 114 Cx 1248 kHz, 720 mV, 97 mW, 151 Cx 1324 kHz, 756 mV, 113 mW, 150 Cx 1516 kHz, 828 mV, 154 mW, 148 Cx 1612 kHz, 884 mV, 194 mW, 155 Cx 1708 kHz, 884 mV, 203 mW, 152 Cx 1804 kHz, 884 mV, 219 mW, 155 Cx Your dynamic-power-coefficient for cpu 0: 142 825 kHz, 596 mV, 148 mW, 530 Cx 979 kHz, 628 mV, 189 mW, 475 Cx 1113 kHz, 656 mV, 230 mW, 461 Cx 1267 kHz, 688 mV, 287 mW, 466 Cx 1555 kHz, 812 mV, 469 mW, 445 Cx 1708 kHz, 828 mV, 567 mW, 480 Cx 1843 kHz, 884 mV, 699 mW, 482 Cx 1900 kHz, 884 mV, 719 mW, 480 Cx 1996 kHz, 916 mV, 814 mW, 484 Cx 2112 kHz, 916 mV, 861 mW, 483 Cx 2208 kHz, 916 mV, 963 mW, 522 Cx 2323 kHz, 940 mV, 1063 mW, 520 Cx 2400 kHz, 956 mV, 1135 mW, 519 Cx Your dynamic-power-coefficient for cpu 6: 489 As you can see, the calculations aren't perfectly consistent but roughly you could say about 480 for big and 137 for little. The ratio between these numbers isn't quite the same as the ratio between the two numbers that Qualcomm used. Perhaps this is because Qualcomm measured something slightly different than the 64-bit version of dhrystone 2.2 or perhaps it's because they fudged these numbers a bit (and fudged the capacity-dmips-mhz). As per discussion [2], let's use the numbers I came up with and also un-fudge capacity-dmips-mhz. While unfudging capacity-dmips-mhz, let's scale it so that bigs are 1024 which seems to be the common practice. In general these numbers don't need to be perfectly exact. In fact, they can't be since the CPU power depends a lot on what's being run on the CPU and the big/little CPUs are each more or less efficient in different operations. Historically running the 32-bit vs. 64-bit versions of dhrystone produced notably different numbers, though I didn't test this time. We also need to scale all of the sustainable-power numbers by the same amount. I scale ones related to the big CPUs by the adjustment I made to the big dynamic-power-coefficient and the ones related to the little CPUs by the adjustment I made to the little dynamic-power-coefficient. [2] https://lore.kernel.org/r/0a865b6e-be34-6371-f9f2-9913ee1c5608@codeaurora.org/ Fixes: 71f87316 ("arm64: dts: qcom: sc7180: Add dynamic CPU power coefficients") Signed-off-by:Douglas Anderson <dianders@chromium.org> Signed-off-by:
Bjorn Andersson <bjorn.andersson@linaro.org> Link: https://lore.kernel.org/r/20210902145127.v2.1.I049b30065f3c715234b6303f55d72c059c8625eb@changeid
Showing
Please register or sign in to comment