Skip to contentSkip to conversation

drm/amd/pm: Fix incorrect avg vcn utilization in gpu_metrics

PATCH In review Boqun Feng 4 participants 8 messages Aug 5 – Aug 14

Subsystems (3)
lore
Share & tools
lore

Boqun Feng

author
#0 lore raw

PATCH 1/1 drm/amd/pm: Fix incorrect avg vcn utilization in gpu_metrics

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -1
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
 	gpu_metrics->temperature_skin = metrics.SkinTemp;
 
 	gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-	gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+	/* metrics.VcnActivity is permyriad, convert it to percentage. */
+	gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
 	memcpy(&gpu_metrics->average_ipu_activity[0],
 		&metrics.IpuBusy[0],
 		sizeof(uint16_t) * 8);
Signature · 1 line
2.50.1 (Apple Git-155)
author
#1 lore raw
Quoted 9 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4

This is incorrect because the value is supposed to be in [0-100] range.

Quoted 8 lines

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

Ping?

Regards,
Boqun

Quoted 21 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -1
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
 	gpu_metrics->temperature_skin = metrics.SkinTemp;
 
 	gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-	gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+	/* metrics.VcnActivity is permyriad, convert it to percentage. */
+	gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
 	memcpy(&gpu_metrics->average_ipu_activity[0],
 		&metrics.IpuBusy[0],
 		sizeof(uint16_t) * 8);
Signature · 1 line
2.50.1 (Apple Git-155)
#2 lore raw
Quoted 9 lines · Boqun Feng
Quoted 7 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4
Quoted 7 lines
Quoted 7 lines

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

I think this is correct as is. smu_v14_0_0_get_smu_metrics_data() is used to fetch data for standardized formats like hwmon while smu_v14_0_0_get_gpu_metrics() is for dumping the raw metrics data.

Alex

Quoted 30 lines
Quoted 3 lines

Ping?

Regards,
Boqun

Quoted 21 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -2
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
      gpu_metrics->temperature_skin = metrics.SkinTemp;
 
      gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-     gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+     /* metrics.VcnActivity is permyriad, convert it to percentage. */
+     gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
      memcpy(&gpu_metrics->average_ipu_activity[0],
              &metrics.IpuBusy[0],
              sizeof(uint16_t) * 8);
--

2.50.1 (Apple Git-155)

author
#3 lore raw
Quoted 18 lines · Alex Deucher
Quoted 9 lines · Boqun Feng
Quoted 7 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4
Quoted 7 lines
Quoted 7 lines

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

Quoted 4 lines

I think this is correct as is. smu_v14_0_0_get_smu_metrics_data() is used to fetch data for standardized formats like hwmon while smu_v14_0_0_get_gpu_metrics() is for dumping the raw metrics data.

Well, then kgd_pp_interface.h needs some help?

struct gpu_metrics_v3_0 {
	...
	/* Utilization */
	/* time filtered GFX busy % [0-100] */
	uint16_t			average_gfx_activity;
	/* time filtered VCN busy % [0-100] */
	uint16_t			average_vcn_activity;
	...
}

Clearly it's not true.

[Cc amdgpu_top]

Regards,
Boqun

Quoted 32 lines

Alex

Quoted 30 lines
Quoted 3 lines

Ping?

Regards,
Boqun

Quoted 21 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -2
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
      gpu_metrics->temperature_skin = metrics.SkinTemp;
 
      gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-     gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+     /* metrics.VcnActivity is permyriad, convert it to percentage. */
+     gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
      memcpy(&gpu_metrics->average_ipu_activity[0],
              &metrics.IpuBusy[0],
              sizeof(uint16_t) * 8);
--

2.50.1 (Apple Git-155)

#4 lore raw
Quoted 37 lines · Boqun Feng
Quoted 18 lines · Alex Deucher
Quoted 9 lines · Boqun Feng
Quoted 7 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4
Quoted 7 lines
Quoted 7 lines

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

Quoted 4 lines

I think this is correct as is. smu_v14_0_0_get_smu_metrics_data() is used to fetch data for standardized formats like hwmon while smu_v14_0_0_get_gpu_metrics() is for dumping the raw metrics data.

Well, then kgd_pp_interface.h needs some help?

struct gpu_metrics_v3_0 {
        ...
        /* Utilization */
        /* time filtered GFX busy % [0-100] */
        uint16_t                        average_gfx_activity;
        /* time filtered VCN busy % [0-100] */
        uint16_t                        average_vcn_activity;
        ...
}
Quoted 1 line

Clearly it's not true.

That was likely copy and pasted from the firmware interface (smu14_driver_if_v14_0_0.h):
uint16_t VcnActivity; //Time filtered VCN busy % [0-100]

Other generations seem to do the same thing (raw data in metrics table and % in the hwmon type interfaces). E.g., smu_v13_0_4_ppt.c:
smu_v13_0_4_get_gpu_metrics() gpu_metrics->average_mm_activity = metrics.UvdActivity;
vs smu_v13_0_4_get_smu_metrics_data():
*value = metrics->UvdActivity / 100;

Alex

Quoted 38 lines

[Cc amdgpu_top]

Regards,
Boqun

Quoted 32 lines

Alex

Quoted 30 lines
Quoted 3 lines

Ping?

Regards,
Boqun

Quoted 21 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -2
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
      gpu_metrics->temperature_skin = metrics.SkinTemp;
 
      gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-     gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+     /* metrics.VcnActivity is permyriad, convert it to percentage. */
+     gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
      memcpy(&gpu_metrics->average_ipu_activity[0],
              &metrics.IpuBusy[0],
              sizeof(uint16_t) * 8);
--

2.50.1 (Apple Git-155)

author
#5 lore raw
Quoted 44 lines · Alex Deucher
Quoted 37 lines · Boqun Feng
Quoted 18 lines · Alex Deucher
Quoted 9 lines · Boqun Feng
Quoted 7 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4
Quoted 7 lines
Quoted 7 lines

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

Quoted 4 lines

I think this is correct as is. smu_v14_0_0_get_smu_metrics_data() is used to fetch data for standardized formats like hwmon while smu_v14_0_0_get_gpu_metrics() is for dumping the raw metrics data.

Well, then kgd_pp_interface.h needs some help?

struct gpu_metrics_v3_0 {
        ...
        /* Utilization */
        /* time filtered GFX busy % [0-100] */
        uint16_t                        average_gfx_activity;
        /* time filtered VCN busy % [0-100] */
        uint16_t                        average_vcn_activity;
        ...
}
Quoted 1 line

Clearly it's not true.

That was likely copy and pasted from the firmware interface (smu14_driver_if_v14_0_0.h):
uint16_t VcnActivity; //Time filtered VCN busy % [0-100]

Quoted 7 lines

Other generations seem to do the same thing (raw data in metrics table and % in the hwmon type interfaces). E.g., smu_v13_0_4_ppt.c:
smu_v13_0_4_get_gpu_metrics() gpu_metrics->average_mm_activity = metrics.UvdActivity;
vs smu_v13_0_4_get_smu_metrics_data():
*value = metrics->UvdActivity / 100;

Could you help fix the document while you're at it? I cannot confirm by myself the valid range of that raw data. Thanks!

Regards,
Boqun

Quoted 40 lines

Alex

Quoted 38 lines

[Cc amdgpu_top]

Regards,
Boqun

Quoted 32 lines

Alex

Quoted 30 lines
Quoted 3 lines

Ping?

Regards,
Boqun

Quoted 21 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -2
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
      gpu_metrics->temperature_skin = metrics.SkinTemp;
 
      gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-     gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+     /* metrics.VcnActivity is permyriad, convert it to percentage. */
+     gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
      memcpy(&gpu_metrics->average_ipu_activity[0],
              &metrics.IpuBusy[0],
              sizeof(uint16_t) * 8);
--

2.50.1 (Apple Git-155)

#6 lore raw
In reply to Alex Deucher

AMD General

Quoted 62 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Friday, August 14, 2026 8:51 AM
To: Boqun Feng <[email protected]>
Cc: [email protected]; Mark Janes <[email protected]>; Feng, Kenneth <[email protected]>; Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; David Airlie <[email protected]>; Simona Vetter <[email protected]>; Lazar, Lijo <[email protected]>; Kamal, Asad <[email protected]>; Wang, Yang(Kevin) <[email protected]>; Hosur, Priya <[email protected]>; M, Mythila <[email protected]>; Limonciello, Mario <[email protected]>; dri- [email protected]; [email protected]; Umio Yasuno <[email protected]>
Subject: Re: [PATCH] drm/amd/pm: Fix incorrect avg vcn utilization in gpu_metrics

Quoted 38 lines · Boqun Feng
Quoted 19 lines · Alex Deucher
Quoted 9 lines · Boqun Feng
Quoted 7 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4
Quoted 8 lines
Quoted 8 lines

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by
smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

Quoted 4 lines

I think this is correct as is. smu_v14_0_0_get_smu_metrics_data() is used to fetch data for standardized formats like hwmon while smu_v14_0_0_get_gpu_metrics() is for dumping the raw metrics data.

Well, then kgd_pp_interface.h needs some help?

struct gpu_metrics_v3_0 {
        ...
        /* Utilization */
        /* time filtered GFX busy % [0-100] */
        uint16_t                        average_gfx_activity;
        /* time filtered VCN busy % [0-100] */
        uint16_t                        average_vcn_activity;
        ...
}
Quoted 1 line

Clearly it's not true.

That was likely copy and pasted from the firmware interface (smu14_driver_if_v14_0_0.h):
uint16_t VcnActivity; //Time filtered VCN busy % [0-100]

Quoted 6 lines

Other generations seem to do the same thing (raw data in metrics table and % in the hwmon type interfaces). E.g., smu_v13_0_4_ppt.c:
smu_v13_0_4_get_gpu_metrics() gpu_metrics->average_mm_activity = metrics.UvdActivity; vs smu_v13_0_4_get_smu_metrics_data():
*value = metrics->UvdActivity / 100;

For SMU 14.0.0, the VCN activity value returned from PMFW ranges from 0 to 10000, and the driver shall divide this value by 100 before reporting it. This scaling does **not** apply to GFX activity.

Best Regards,
Kevin

Quoted 45 lines

Alex

Quoted 21 lines

[Cc amdgpu_top]

Regards,
Boqun

Quoted 15 lines

Alex

Quoted 13 lines
Quoted 3 lines

Ping?

Regards,
Boqun

Quoted 4 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c static ssize_t

Quoted 7 lines

diff --git a/drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c b/drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c index 568665265a13..1058338344b0 100644

drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +0 -0
@@ -566,7 +566,8 @@ static ssize_t

smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +3 -2
      gpu_metrics->temperature_skin = metrics.SkinTemp;
 
      gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-     gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+     /* metrics.VcnActivity is permyriad, convert it to percentage. */
+     gpu_metrics->average_vcn_activity = metrics.VcnActivity /
+ 100;
      memcpy(&gpu_metrics->average_ipu_activity[0],
              &metrics.IpuBusy[0],
              sizeof(uint16_t) * 8);
--

2.50.1 (Apple Git-155)

#7 lore raw
In reply to Boqun Feng

Apologies, I forgot to set the CC. Below is my opinion that I wrote outside the thread.

---
amdgpu_top already divides average_vcn_activity in gpu_metrics_v3_0 by 100, so this patch would unfortunately end up affecting user-space applications.

In my view, if users want to obtain media engine usage that is properly corrected based on the firmware (or firmware version) in the AMDGPU driver, the best option is to use the vcn_busy_percent sysfs interface. However, that adds extra reading cost.

Alternatively, since gpu_metrics is designed as an interface that provides raw data[^1] (but undocumented it[^2]), it would be better to have the correction applied on the user-space.

[^1]: https://gitlab.freedesktop.org/drm/amd/-/work_items/3251#note_2904104
[^2]: https://www.kernel.org/doc/html/latest/gpu/amdgpu/thermal.html#gpu-metrics
---

Quoted 76 lines · Boqun Feng
Quoted 18 lines · Alex Deucher
Quoted 9 lines · Boqun Feng
Quoted 7 lines · Boqun Feng

In a Strix Halo system, a >100 average_vcn_activity in gpu_metrics can be observed by running the following command:

    ffmpeg -hwaccel vappi -vappi_device /dev/dri/renderD128 -i \
    ~/sample-10s.mp4 -vf hwupload,scale_vappi=format=nv12 -c:v \
    h264_vappi /tmp/output.mp4
Quoted 7 lines
Quoted 7 lines

This is incorrect because the value is supposed to be in [0-100] range.

The cause of the issue is that the metrics.VcnActivity reported by smu_cmn_get_metrics_table() is a permyriad value instead of a percentage value (e.g. smu_v14_0_0_get_smu_metrics_data() has a division by 100 to report the vcn_busy_percent). Therefore fix this by applying the same division in smu_v14_0_0_get_gpu_metrics().

Quoted 4 lines

I think this is correct as is. smu_v14_0_0_get_smu_metrics_data() is used to fetch data for standardized formats like hwmon while smu_v14_0_0_get_gpu_metrics() is for dumping the raw metrics data.

Well, then kgd_pp_interface.h needs some help?

struct gpu_metrics_v3_0 {
	...
	/* Utilization */
	/* time filtered GFX busy % [0-100] */
	uint16_t			average_gfx_activity;
	/* time filtered VCN busy % [0-100] */
	uint16_t			average_vcn_activity;
	...
}

Clearly it's not true.

[Cc amdgpu_top]

Regards,
Boqun

Quoted 32 lines

Alex

Quoted 30 lines
Quoted 3 lines

Ping?

Regards,
Boqun

Quoted 21 lines
---
 drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/pm/swsmu/smu14/smu_v14_0_0_ppt.c +2 -2
index 568665265a13..1058338344b0 100644
@@ -566,7 +566,8 @@ static ssize_t smu_v14_0_0_get_gpu_metrics(struct smu_context *smu,
      gpu_metrics->temperature_skin = metrics.SkinTemp;
 
      gpu_metrics->average_gfx_activity = metrics.GfxActivity;
-     gpu_metrics->average_vcn_activity = metrics.VcnActivity;
+     /* metrics.VcnActivity is permyriad, convert it to percentage. */
+     gpu_metrics->average_vcn_activity = metrics.VcnActivity / 100;
      memcpy(&gpu_metrics->average_ipu_activity[0],
              &metrics.IpuBusy[0],
              sizeof(uint16_t) * 8);
--

2.50.1 (Apple Git-155)