Skip to contentSkip to conversation

Add support for Zen 2 CPUs to amd_energy driver

PATCH In review Sandeep Raghuraman 3 participants 9 messages Dec 8, 2020 – Dec 10, 2020 Hardware monitoring

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Add support for Zen 2 CPUs to amd_energy driver

This patch adds support for Zen 2 desktop CPUs to the amd_energy driver.

---
 drivers/hwmon/amd_energy.c | 1 +
 1 file changed, 1 insertion(+)
drivers/hwmon/amd_energy.c +1 -0
index 0dee535e6c85..21341c8e1d6e 100644
@@ -363,6 +363,7 @@ static struct platform_device *amd_energy_platdev;
 
 static const struct x86_cpu_id cpu_ids[] __initconst = {
        X86_MATCH_VENDOR_FAM_MODEL(AMD, 0x17, 0x31, NULL),
+       X86_MATCH_VENDOR_FAM_MODEL(AMD, 0x17, 0x71, NULL),
        {}
 };
 MODULE_DEVICE_TABLE(x86cpu, cpu_ids);

--

maintainer
#1 lore raw
Quoted 3 lines · Sandeep Raghuraman

This patch adds support for Zen 2 desktop CPUs to the amd_energy driver.

This patch will require an Ack from Naveen; support for this model was previously removed due to inconsistent results; see commit f28e360f2903 ("hwmon: (amd_energy) match for supported models"),

Guenter

Quoted 18 lines
---
 drivers/hwmon/amd_energy.c | 1 +
 1 file changed, 1 insertion(+)

--

drivers/hwmon/amd_energy.c +1 -0
index 0dee535e6c85..21341c8e1d6e 100644
@@ -363,6 +363,7 @@ static struct platform_device *amd_energy_platdev;
 
 static const struct x86_cpu_id cpu_ids[] __initconst = {
        X86_MATCH_VENDOR_FAM_MODEL(AMD, 0x17, 0x31, NULL),
+       X86_MATCH_VENDOR_FAM_MODEL(AMD, 0x17, 0x71, NULL),
        {}
 };
 MODULE_DEVICE_TABLE(x86cpu, cpu_ids);
#2 lore raw

[AMD Official Use Only - Approved for External Use]

Hi Guenter,

Quoted 1 line
Quoted 1 line

This patch will require an Ack from Naveen; support for this model was previously removed due to inconsistent results; see commit f28e360f2903

("hwmon: (amd_energy) match for supported models"), I could not get hold of a board with the mentioned processor.

Sandeep,

Earlier, we noticed on some of the family 17h processors the core and socket energy counters are not well calibrated.

While running avx-turbo work load https://github.com/travisdowns/avx-turbo as below, we noticed that for a given time, the sum of the energy consumed by all the cores in a socket is greater than the energy consumed by the socket itself.

Can you run the avx-turbo test, with following options and confirm the results ?

  ./avx-turbo --iters 10000000 --spec avx256_fma_t/8

This will run AVX code on all 8 cores for about 1 second. To run it for longer, increase the --iters argument.

Regards,
Naveenk

-----Original Message-----
From: Guenter Roeck <[email protected]>
Sent: Wednesday, December 9, 2020 2:11 AM
To: Sandeep Raghuraman <[email protected]>
Cc: [email protected]; Chatradhi, Naveen Krishna <[email protected]>
Subject: Re: [PATCH] Add support for Zen 2 CPUs to amd_energy driver

[CAUTION: External Email]

Quoted 3 lines · Sandeep Raghuraman

This patch adds support for Zen 2 desktop CPUs to the amd_energy driver.

This patch will require an Ack from Naveen; support for this model was previously removed due to inconsistent results; see commit f28e360f2903 ("hwmon: (amd_energy) match for supported models"),

Guenter

Quoted 18 lines
---
 drivers/hwmon/amd_energy.c | 1 +
 1 file changed, 1 insertion(+)

--

drivers/hwmon/amd_energy.c +1 -0
index 0dee535e6c85..21341c8e1d6e 100644
@@ -363,6 +363,7 @@ static struct platform_device *amd_energy_platdev;
 
 static const struct x86_cpu_id cpu_ids[] __initconst = {
        X86_MATCH_VENDOR_FAM_MODEL(AMD, 0x17, 0x31, NULL),
+       X86_MATCH_VENDOR_FAM_MODEL(AMD, 0x17, 0x71, NULL),
        {}
 };
 MODULE_DEVICE_TABLE(x86cpu, cpu_ids);
author
#3 lore raw
Quoted 16 lines · Naveen Krishna

[AMD Official Use Only - Approved for External Use]

Hi Guenter,

Quoted 1 line
Quoted 1 line

This patch will require an Ack from Naveen; support for this model was previously removed due to inconsistent results; see commit f28e360f2903

("hwmon: (amd_energy) match for supported models"), I could not get hold of a board with the mentioned processor.

Sandeep,

Earlier, we noticed on some of the family 17h processors the core and socket energy counters are not well calibrated.

While running avx-turbo work load https://github.com/travisdowns/avx-turbo as below, we noticed that for a given time, the sum of the energy consumed by all the cores in a socket is greater than the energy consumed by the socket itself.

Quoted 8 lines

Can you run the avx-turbo test, with following options and confirm the results ?

  ./avx-turbo --iters 10000000 --spec avx256_fma_t/8

This will run AVX code on all 8 cores for about 1 second. To run it for longer, increase the --iters argument.

Regards,
Naveenk

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

- Sandeep

author
#4 lore raw
Quoted 25 lines · Sandeep
Quoted 22 lines · Naveen Krishna

[AMD Official Use Only - Approved for External Use]

Hi Guenter,

Quoted 1 line
Quoted 1 line

This patch will require an Ack from Naveen; support for this model was previously removed due to inconsistent results; see commit f28e360f2903

("hwmon: (amd_energy) match for supported models"), I could not get hold of a board with the mentioned processor.

Sandeep,

Earlier, we noticed on some of the family 17h processors the core and socket energy counters are not well calibrated.

While running avx-turbo work load https://github.com/travisdowns/avx-turbo as below, we noticed that for a given time, the sum of the energy consumed by all the cores in a socket is greater than the energy consumed by the socket itself.

Can you run the avx-turbo test, with following options and confirm the results ?

  ./avx-turbo --iters 10000000 --spec avx256_fma_t/8

This will run AVX code on all 8 cores for about 1 second. To run it for longer, increase the --iters argument.

Quoted 8 lines
Quoted 2 lines

Regards,
Naveenk

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

- Sandeep

Ok, I ran the test for both 8 cores and 12 cores - in both cases, I can confirm that the sum of the energy consumed by all the cores in a socket, is NOT greater than the energy consumed by the socket itself.

- Sandeep

#5 lore raw
In reply to Sandeep

[AMD Official Use Only - Approved for External Use]

Hi Sandeep,

Quoted 1 line
Quoted 1 line

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

I'm trying to get hold of a system to test it myself. In the meanwhile, you may run the following command in the back ground and read the counters (cores and socket) at 5 to 10 secs apart and sum of the diff of core counters should be not be > diff of socket counters

sudo ./avx-turbo --iters 1000000000 --spec avx256_fma_t/12

You may refer https://www.spinics.net/lists/linux-hwmon/msg08255.html link for more background of the testing.

Regards,
Naveenk

-----Original Message-----
From: Sandeep <[email protected]>
Sent: Wednesday, December 9, 2020 8:24 PM
To: Chatradhi, Naveen Krishna <[email protected]>
Cc: Guenter Roeck <[email protected]>; [email protected]
Subject: Re: [PATCH] Add support for Zen 2 CPUs to amd_energy driver

[CAUTION: External Email]

Quoted 18 lines · Naveen Krishna

[AMD Official Use Only - Approved for External Use]

Hi Guenter,

Quoted 3 lines
Quoted 3 lines

This patch will require an Ack from Naveen; support for this model was previously removed due to inconsistent results; see commit f28e360f2903

("hwmon: (amd_energy) match for supported models"), I could not get hold of a board with the mentioned processor.

Sandeep,

Earlier, we noticed on some of the family 17h processors the core and socket energy counters are not well calibrated.

While running avx-turbo work load
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Ftravisdowns%2Favx-turbo&amp;data=04%7C01%7CNaveenKrishna.Chatradhi%40amd.com%7C8cd70f66176547ea9d8208d89c525af5%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637431224816944682%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=mJmMnL3Y4Wwi%2BLCqZmn9AjxGeIKdxgoLrjz%2FqASnl68%3D&amp;reserved=0 as below, we noticed that for a given time, the sum of the energy consumed by all the cores in a socket is greater than the energy consumed by the socket itself.

Quoted 8 lines

Can you run the avx-turbo test, with following options and confirm the results ?

  ./avx-turbo --iters 10000000 --spec avx256_fma_t/8

This will run AVX code on all 8 cores for about 1 second. To run it for longer, increase the --iters argument.

Regards,
Naveenk

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

- Sandeep

author
#6 lore raw
Quoted 11 lines · Naveen Krishna

[AMD Official Use Only - Approved for External Use]

Hi Sandeep,

Quoted 1 line
Quoted 1 line

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

I'm trying to get hold of a system to test it myself. In the meanwhile, you may run the following command in the back ground and read the counters (cores and socket) at 5 to 10 secs apart and sum of the diff of core counters should be not be > diff of socket counters

Quoted 7 lines

sudo ./avx-turbo --iters 1000000000 --spec avx256_fma_t/12

You may refer https://www.spinics.net/lists/linux-hwmon/msg08255.html link for more background of the testing.

Regards,
Naveenk

Do you already have a script or something that I can run to automatically gather this information (i.e check values every 5 seconds, calculate differences and check if the results line up)? Doing this manually, sounds like a lot of work that I really don't want to get into.

- Sandeep

author
#7 lore raw
Quoted 22 lines · Sandeep
Quoted 11 lines · Naveen Krishna

[AMD Official Use Only - Approved for External Use]

Hi Sandeep,

Quoted 1 line
Quoted 1 line

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

I'm trying to get hold of a system to test it myself. In the meanwhile, you may run the following command in the back ground and read the counters (cores and socket) at 5 to 10 secs apart and sum of the diff of core counters should be not be > diff of socket counters

Quoted 7 lines

sudo ./avx-turbo --iters 1000000000 --spec avx256_fma_t/12

You may refer https://www.spinics.net/lists/linux-hwmon/msg08255.html link for more background of the testing.

Regards,
Naveenk

Quoted 7 lines

Do you already have a script or something that I can run to automatically gather this information (i.e check values every 5 seconds, calculate differences and check if the results line up)? Doing this manually, sounds like a lot of work that I really don't want to get into.

- Sandeep

Ok, I was able to write a small program to automate the test, and it looks like the following condition "sum of the diff of core counters

Quoted 1 line

diff of socket counters" is indeed true. Looks like the values are

indeed incorrect sometimes.
- Sandeep

author
#8 lore raw
Quoted 30 lines · Sandeep
Quoted 28 lines · Sandeep
Quoted 11 lines · Naveen Krishna

[AMD Official Use Only - Approved for External Use]

Hi Sandeep,

Quoted 1 line
Quoted 1 line

Sure, I'll run that test. I have a 3900X though, which is 12 core, so do I have to adjust anything for that CPU, or just run with the parameters you've specified?

I'm trying to get hold of a system to test it myself. In the meanwhile, you may run the following command in the back ground and read the counters (cores and socket) at 5 to 10 secs apart and sum of the diff of core counters should be not be > diff of socket counters

Quoted 7 lines

sudo ./avx-turbo --iters 1000000000 --spec avx256_fma_t/12

You may refer https://www.spinics.net/lists/linux-hwmon/msg08255.html link for more background of the testing.

Regards,
Naveenk

Do you already have a script or something that I can run to automatically gather this information (i.e check values every 5 seconds, calculate differences and check if the results line up)? Doing this manually, sounds like a lot of work that I really don't want to get into.

Quoted 7 lines
Quoted 1 line

- Sandeep

Ok, I was able to write a small program to automate the test, and it looks like the following condition "sum of the diff of core counters

Quoted 1 line

diff of socket counters" is indeed true. Looks like the values are

indeed incorrect sometimes.
- Sandeep

I've uploaded the test program I wrote here -
https://github.com/sandy-8925/amd-energy-tester - in case it's useful for anyone else. Let me know if you spot any bugs in that code.

- Sandeep