AMD P-State EPP Patches Spun An 8th Time For Helping Out Linux Performance & Efficiency

[HD7950]

@HD7950

Compiled well also with GCC.
Really check if WERROR is enabled or other patches are maybe conflicting to this.

I found the culprit: CONFIG_X86_INTEL_PSTATE was disabled on my personal kernel but it is needed by the AMD EPP driver now.

Thanks!

[HD7950]

https://paste.cachyos.org/p/db1b342.diff

tested on 6.1.0 with clang thinlto.
I will let a GCC kernel also running and report you.

Still can't against 6.1.0 and GCC.

LD  .tmp_vmlinux.kallsyms1
ld: vmlinux.o: in function `show_energy_performance_available_preferences':
amd-pstate.c:(.text+0x7f9d57): undefined reference to `energy_perf_strings'
ld: amd-pstate.c:(.text+0x7f9d87): undefined reference to `energy_perf_strings'
ld: vmlinux.o: in function `store_energy_performance_preference':
amd-pstate.c:(.text+0x7f9f00): undefined reference to `energy_perf_strings'
ld: amd-pstate.c:(.text+0x7f9f23): undefined reference to `epp_values'
ld: vmlinux.o: in function `show_energy_performance_preference':
amd-pstate.c:(.text+0x7fa0e7): undefined reference to `energy_perf_strings'
make[1]: *** [scripts/Makefile.vmlinux:34: vmlinux] Error 1
make: *** [Makefile:1236: vmlinux] Error 2
A failure occurred in build().
Aborting... ​

This is weird as i could compile all previous epp patches. Thanks for your help anyways.

[ptr1337]

@HD7950

Compiled well also with GCC.
Really check if WERROR is enabled or other patches are maybe conflicting to this.

[Mitch]

If your goal is the highest efficiency for the most workloads, it seems like Epp Balance Performance is the best all around choice for efficiency. What's interesting is that Power save is lower PPW, so unless you're working with a bad cooling solution or want minimum fan noise, you should stick with Epp balance performance.

Obviously other benchmarks may prove the p-state on demand will win more often, but just looking at what they posted here, that's my take.

My 6850u seems to rival my M1 air's efficiency sometimes (mostly multicore workloads), so I wonder if these patches could push it over that gap for some scenarios

[ptr1337]

Still can´t compile. Can you share your local patch? Maybe mine is malformed.

https://paste.cachyos.org/p/db1b342.diff

tested on 6.1.0 with clang thinlto.
I will let a GCC kernel also running and report you.

[HD7950]

Yes, you can. It wont harm your kernel. WERROR just means that a warning triggers a failure in the compilation. Has been some time introduced as default enabled. But if you want to test patches from lkml better keep it disabled

Still can´t compile. Can you share your local patch? Maybe mine is malformed.

[ptr1337]

Yes, it's the default option. Should i disable it?

Yes, you can. It wont harm your kernel. WERROR just means that a warning triggers a failure in the compilation. Has been some time introduced as default enabled. But if you want to test patches from lkml better keep it disabled

[HD7950]

Compiled fine at me. No issues. Are you building with WERROR ?

Yes, it's the default option. Should i disable it?

[Anux]

Just at the performance per watt according the benchmarks.

Only the second bech gives a value for throughput and pstate ondemand has the highest throughput while having the same energy consumption, therefore winning in absolute perf and on efficiency. The first bench is useless, as long as we don't know any performance number.

Also those values are probably unique for every single CPU, you can't compare ROME with desktop

I'm kinda amazed how people manage to get their fonts so blurry and impossible to read under Linux.

I don't know what that has got to do with Linux? It's just an image scaled up slightly, that never looks good with pixel art. If you view the original table 1:1 it's as sharp as a dressed man. And your version is also blurry (scaled up).

[ptr1337]

v8 patches are broken. Compile fails on 6.1 and 6.2 kernels.

Compiled fine at me. No issues. Are you building with WERROR ?