Skip to contentSkip to conversation

arm64: dts: imx8mq: Restore VPU G2 clock to 600MHz for 4K60fps decoding

PATCH In review [email protected] 6 participants 15 messages Jan 30 – Feb 4

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm64: dts: imx8mq: Restore VPU G2 clock to 600MHz for 4K60fps decoding

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:

  gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
    video/x-matroska ! aiurdemux ! h265parse ! \
    v4l2slh265dec ! video/x-raw,format=NV12 ! \
    queue ! waylandsink
Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")
---
 arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -1
index 607962f807be..731142176625 100644
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {
 									 <&clk IMX8MQ_SYS1_PLL_800M>,
 									 <&clk IMX8MQ_VPU_PLL>;
 						assigned-clock-rates = <600000000>,
-								       <300000000>,
+								       <600000000>,
 								       <800000000>,
 								       <0>;
 					};

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

Signature · 1 line
2.52.0
#1 lore raw

Am Freitag, 30. Januar 2026, 09:41:31 CET schrieb [email protected]:

Quoted 45 lines

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:

  gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
    video/x-matroska ! aiurdemux ! h265parse ! \
    v4l2slh265dec ! video/x-raw,format=NV12 ! \
    queue ! waylandsink
Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")
---
 arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -0
index 607962f807be..731142176625 100644

arch/arm64/boot/dts/freescale/imx8mq.dtsi pgc_vpu: power-domain@6 { · line 963

Quoted 5 lines
arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -1
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {
 									 <&clk IMX8MQ_SYS1_PLL_800M>,
 									 <&clk IMX8MQ_VPU_PLL>;
 						assigned-clock-rates = <600000000>,
-								       <300000000>,
+								       <600000000>,

If I read the Datasheet correctly 600 MHz is only supported by overdrive mode (also depending on the VDD_VPU).
Is this frequency really correct?

Best regards,
Alexander

Quoted 7 lines
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -0
 								       <800000000>,
 								       <0>;
 					};

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

#2 lore raw
Quoted 41 lines · Alexander Stein

Am Freitag, 30. Januar 2026, 09:41:31 CET schrieb [email protected]:

Quoted 40 lines

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:

  gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
    video/x-matroska ! aiurdemux ! h265parse ! \
    v4l2slh265dec ! video/x-raw,format=NV12 ! \
    queue ! waylandsink
Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")
---
 arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines

arch/arm64/boot/dts/freescale/imx8mq.dtsi pgc_vpu: power-domain@6 { · line 963

Quoted 10 lines
arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -1
index 607962f807be..731142176625 100644
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {
 									 <&clk IMX8MQ_SYS1_PLL_800M>,
 									 <&clk IMX8MQ_VPU_PLL>;
 						assigned-clock-rates = <600000000>,
-								       <300000000>,
+								       <600000000>,

If I read the Datasheet correctly 600 MHz is only supported by overdrive mode (also depending on the VDD_VPU).
Is this frequency really correct?

G1 and BUS clk were already set as Overdrive frequency.

This change is to only upgrading G2 from 300M to 600M.

So if your question is should we downgrade all to Nominal mode, I think no. The freq could be override in board dts, or adding a new dts as arch/arm64/boot/dts/freescale/imx8mp-nominal.dtsi

Regards
Peng

Quoted 15 lines

Best regards,
Alexander

Quoted 7 lines
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -0
 								       <800000000>,
 								       <0>;
 					};

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

#3 lore raw

Hi Alexander,

Quoted 59 lines · Peng Fan
Quoted 41 lines · Alexander Stein

Am Freitag, 30. Januar 2026, 09:41:31 CET schrieb [email protected]:

Quoted 40 lines

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:

   gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
     video/x-matroska ! aiurdemux ! h265parse ! \
     v4l2slh265dec ! video/x-raw,format=NV12 ! \
     queue ! waylandsink
Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")
---
  arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 14 lines

arch/arm64/boot/dts/freescale/imx8mq.dtsi pgc_vpu: power-domain@6 { · line 963

Quoted 10 lines
arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -1
index 607962f807be..731142176625 100644
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {
  									 <&clk IMX8MQ_SYS1_PLL_800M>,
  									 <&clk IMX8MQ_VPU_PLL>;
  						assigned-clock-rates = <600000000>,
-								       <300000000>,
+								       <600000000>,

If I read the Datasheet correctly 600 MHz is only supported by overdrive mode (also depending on the VDD_VPU).
Is this frequency really correct?

G1 and BUS clk were already set as Overdrive frequency.

Quoted 13 lines

This change is to only upgrading G2 from 300M to 600M.

So if your question is should we downgrade all to Nominal mode, I think no. The freq could be override in board dts, or adding a new dts as arch/arm64/boot/dts/freescale/imx8mp-nominal.dtsi

Regards
Peng

Quoted 4 lines

Best regards,
Alexander

Yes, you are right, 600MHz is the Overdriver frequency. However, to achieve the 4K 60fps target, we set the VPU to run in overdrive mode by default, just as Peng said.

Regards,
Ming

Quoted 11 lines
Quoted 11 lines
Quoted 7 lines
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -0
  								       <800000000>,
  								       <0>;
  					};

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

#4 lore raw
In reply to [email protected]

Hi,

Le vendredi 30 janvier 2026 à 16:41 +0800, [email protected] a écrit :

Quoted 9 lines

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Tested on EVK, with the downstream DCSS driver, and this change triggers DCSS underrun (which is related to the DRAM QoS erratas on this SoC). It also sometimes trigger the "not all macroblock decoded" warning I added recently, and we can empty IRQs, but these are handled now.

Quoted 10 lines

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

Quoted 5 lines

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

That probably only true with the upstream DCSS + a small resolution embedded panel ? Can you clarify this setup, because the display drivers mainline are very minimal. Would be nice to show you average DDR read/write bandwidth utilization during this run for comparision.

Another information that bugs me, in the BSP code, the G2 voltage is increased too, which you didn't do here. They also use the thermal 2 zone to kick it down to 300 until it cools down.

Nicolas

Quoted 39 lines

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:
  gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
    video/x-matroska ! aiurdemux ! h265parse ! \
    v4l2slh265dec ! video/x-raw,format=NV12 ! \
    queue ! waylandsink

Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")

---
 arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -0
index 607962f807be..731142176625 100644
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {


<&clk IMX8MQ_SYS1_PLL_800M>,

<&clk IMX8MQ_VPU_PLL>;
  assigned-clock-rates = <600000000>,

arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -1
-								      

<300000000>,

arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -0
+								      

<600000000>,

<800000000>,
         <0>; };

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

#5 lore raw

Hi Nicolas,

Quoted 9 lines · Nicolas Dufresne

Hi,

Quoted 6 lines · Le vendredi 30 janvier 2026 à 16:41 +0800, [email protected] a écrit :

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

Quoted 9 lines
Quoted 3 lines

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Tested on EVK, with the downstream DCSS driver, and this change triggers DCSS underrun (which is related to the DRAM QoS erratas on this SoC). It also sometimes trigger the "not all macroblock decoded" warning I added recently, and we can empty IRQs, but these are handled now.

This doesn't sound like just a VPU issue; it's related to the display or DDR. If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 16 lines
Quoted 10 lines

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

Quoted 5 lines

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

Quoted 4 lines

That probably only true with the upstream DCSS + a small resolution embedded panel ? Can you clarify this setup, because the display drivers mainline are very minimal. Would be nice to show you average DDR read/write bandwidth utilization during this run for comparision.

My display is hdmi, I'll try the DCSS.
And the DDR bandwidth results measured by perf are as follows:

Performance counter stats for 'system wide':

      113303664278      imx8_ddr0/read-cycles/
       82457075530      imx8_ddr0/write-cycles/

     634.892101865 seconds time elapsed
Quoted 5 lines

Another information that bugs me, in the BSP code, the G2 voltage is increased too, which you didn't do here. They also use the thermal 2 zone to kick it down to 300 until it cools down.

In our internal code, whenever the frequency of either g1 or g2 reaches 600MHz, the voltage is adjusted to 1.0V. Since g1 is already set to 600 MHz in the upstream DTS, I believe the default version is already 1.0v.

And do you mean vpu-thermal? But it doesn't define the cooling-map, I'm not sure how it works.

		vpu-thermal {
			polling-delay-passive = <250>;
			polling-delay = <2000>;
			thermal-sensors = <&tmu 2>;

			trips {
				vpu-crit {
					temperature = <90000>;
					hysteresis = <2000>;
					type = "critical";
				};
			};
		};

Regards,
Ming

Quoted 41 lines

Nicolas

Quoted 39 lines

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:

   gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
     video/x-matroska ! aiurdemux ! h265parse ! \
     v4l2slh265dec ! video/x-raw,format=NV12 ! \
     queue ! waylandsink
Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")
---
  arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -1
index 607962f807be..731142176625 100644
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {
 
 <&clk IMX8MQ_SYS1_PLL_800M>,
 
 <&clk IMX8MQ_VPU_PLL>;
  						assigned-clock-rates =
 <600000000>,
-								

<300000000>,

arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -0
+								

<600000000>,

<800000000>,
<0>;
};

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

#6 lore raw

Le lundi 02 février 2026 à 15:44 +0800, Ming Qian(OSS) a écrit :

Quoted 22 lines

Hi Nicolas,

Quoted 9 lines · Nicolas Dufresne

Hi,

Quoted 6 lines · Le vendredi 30 janvier 2026 à 16:41 +0800, [email protected] a écrit :

From: Ming Qian <[email protected]>

The VPU G2 clock was reduced from 600MHz to 300MHz in commit b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock") to address pixel errors with high-resolution HEVC postprocessor output.

Quoted 9 lines
Quoted 3 lines

However, testing shows the 300MHz clock rate is insufficient for 4K60fps decoding and the original pixel errors no longer occur at 600MHz with current drivers.

Tested on EVK, with the downstream DCSS driver, and this change triggers DCSS underrun (which is related to the DRAM QoS erratas on this SoC). It also sometimes trigger the "not all macroblock decoded" warning I added recently, and we can empty IRQs, but these are handled now.

Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Didn't you run these tests before sending ? I can try again, but in my internal

Quoted 1 line · notes, I wrote:

Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Quoted 45 lines
Quoted 17 lines
Quoted 11 lines

Test results with 3840x2160@60fps HEVC stream decoded to NV12 (the same scenario that exhibited pixel errors previously):

300MHz performance:
- Severe frame dropping throughout playback
- Only 336 frames rendered in 11:53 (0.471 fps)
- Continuous "A lot of buffers are being dropped" warnings
- Completely unusable for 4K video

Quoted 5 lines

600MHz performance:
- Smooth playback with only 1 frame dropped at startup
- 37981 frames rendered in 10:34 (59.857 fps)
- Achieves target 60fps performance
- No pixel errors or artifacts observed

Quoted 4 lines

That probably only true with the upstream DCSS + a small resolution embedded panel ? Can you clarify this setup, because the display drivers mainline are very minimal. Would be nice to show you average DDR read/write bandwidth utilization during this run for comparision.

My display is hdmi, I'll try the DCSS.
And the DDR bandwidth results measured by perf are as follows:

  Performance counter stats for 'system wide':

       113303664278      imx8_ddr0/read-cycles/
        82457075530      imx8_ddr0/write-cycles/

      634.892101865 seconds time elapsed

Quoted 5 lines

Another information that bugs me, in the BSP code, the G2 voltage is increased too, which you didn't do here. They also use the thermal 2 zone to kick it down to 300 until it cools down.

In our internal code, whenever the frequency of either g1 or g2 reaches 600MHz, the voltage is adjusted to 1.0V. Since g1 is already set to 600 MHz in the upstream DTS, I believe the default version is already 1.0v.

And do you mean vpu-thermal? But it doesn't define the cooling-map, I'm not sure how it works.

Quoted 13 lines
		vpu-thermal {
			polling-delay-passive = <250>;
			polling-delay = <2000>;
			thermal-sensors = <&tmu 2>;

			trips {
				vpu-crit {
					temperature = <90000>;
					hysteresis = <2000>;
					type = "critical";
				};
			};
		};

Its not:

$> cat /sys/kernel/debug/regulator/regulator_summary | grep SW1C
SW1C 1 1 0 unknown 900mV 0mA 825mV 1100

Before I gave up on 60Hz on this SoC, I did test raising it to 1v with this patch (hopefully there is a way to do that in DT, would be more elegant):

drivers/pmdomain/imx/gpcv2.c +1 -0
index 4b828d74a606..2f2b85ca6fd2 100644
@@ -639,6 +639,7 @@ static const struct imx_pgc_domain imx8m_pgc_domains[] = {
 		},
 		.pgc   = BIT(IMX8M_PGC_VPU),
 		.keep_clocks = true,
+		.voltage   = 1000000,
 	},
 
 	[IMX8M_POWER_DOMAIN_DISP] = {

I would also like to remind you your own erratas, in the errata document you state that DRAM QoS is broken, and all transactions are treated with the same priority. If you overload the bandwidth, it becomes fatal for the display controller. We tried to workaround with changing the NoC configuration, but it did not work. It feels like that NoC granularity is not enough to prevent underrun of the display controller (where the QoS would work, since its done at transaction level, not by measuring bandwidth).

Nicolas

Quoted 45 lines

Regards,
Ming

Quoted 41 lines

Nicolas

Quoted 39 lines

Restore the clock to 600MHz to enable proper 4K60fps decoding capability while maintaining stability.

Test pipeline:
   gst-launch-1.0 filesrc location=<4K60_HEVC.mkv> ! \
     video/x-matroska ! aiurdemux ! h265parse ! \
     v4l2slh265dec ! video/x-raw,format=NV12 ! \
     queue ! waylandsink

Fixes b27bfc5103c7 ("arm64: dts: freescale: Fix VPU G2 clock")

---
  arch/arm64/boot/dts/freescale/imx8mq.dtsi | 2 +-

  1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -0
index 607962f807be..731142176625 100644
@@ -960,7 +960,7 @@ pgc_vpu: power-domain@6 {


<&clk IMX8MQ_SYS1_PLL_800M>,

<&clk IMX8MQ_VPU_PLL>;
   assigned-clock-rates = <600000000>,

arch/arm64/boot/dts/freescale/imx8mq.dtsi +0 -1
-								

<300000000>,

arch/arm64/boot/dts/freescale/imx8mq.dtsi +1 -0
+								

<600000000>,

<800000000>,
          <0>; };

base-commit: c824345288d11e269ce41b36c105715bc2286050
prerequisite-patch-id: 0000000000000000000000000000000000000000

#7 lore raw

Hi,

Le lundi 02 février 2026 à 13:44 -0500, Nicolas Dufresne a écrit :

Quoted 5 lines
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal
notes, I wrote:

  > Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Nicolas

#8 lore raw

Hi Nicolas,

Quoted 37 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal

Quoted 1 line · notes, I wrote:

Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Quoted 5 lines

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz. And results remained unchanged after multiple tests.

I'm not sure what caused the differences between us.

Below are my test results:

600Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 
-j2 -t 90
	****************************************************************************************************
	Running test suite JCT-VC-HEVC_V1 with decoder 

GStreamer-H.265-V4L2SL-Gst1.0
Using 2 parallel job(s)
****************************************************************************************************

Ran 139/147 tests successfully in 505.434 secs
Ran 139/147 tests successfully in 505.350 secs
Ran 139/147 tests successfully in 507.540 secs

600Mhz, 1.0v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 1000mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 

-j2 -t 90
Ran 139/147 tests successfully in 506.901 secs

300Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	300000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 

-j2 -t 90
Ran 139/147 tests successfully in 506.063 secs

Downstream v4l2 driver
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 2 0 unknown 1000mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2-Gst1.0 -j2 

-t 90
Ran 136/147 tests successfully in 460.435 secs

Regards,
Ming

#9 lore raw

Hi,

sorry for jumping in.

Quoted 44 lines · Ming Qian(OSS)

Hi Nicolas,

Quoted 41 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal

Quoted 1 line · notes, I wrote:

Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Quoted 8 lines
Quoted 1 line

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz.
And results remained unchanged after multiple tests.

I'm not sure what caused the differences between us.

Once it comes to system stability, you need to ensure that your bootstack is aligned e.g. same TF-A version and sometimes same bootloader since there might be workarounds/erratum applied by the boot firmware.

Regards,
Marco

Quoted 59 lines

Below are my test results:

600Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t
90
	****************************************************************************************************
	Running test suite JCT-VC-HEVC_V1 with decoder

GStreamer-H.265-V4L2SL-Gst1.0
Using 2 parallel job(s)
****************************************************************************************************

Ran 139/147 tests successfully in 505.434 secs
Ran 139/147 tests successfully in 505.350 secs
Ran 139/147 tests successfully in 507.540 secs

600Mhz, 1.0v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 1000mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
Ran 139/147 tests successfully in 506.901 secs

300Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	300000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
Ran 139/147 tests successfully in 506.063 secs

Downstream v4l2 driver
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 2 0 unknown 1000mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

	./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2-Gst1.0 -j2 -t

90
Ran 136/147 tests successfully in 460.435 secs

Regards,
Ming

Signature · 7 lines
#gernperDu 
#CallMeByMyFirstName

Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | https://www.pengutronix.de/ |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-9    |
#10 lore raw

Hi Marco,

Quoted 60 lines · Marco Felsch

[You don't often get email from [email protected]. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]

Hi,

sorry for jumping in.

Quoted 44 lines · Ming Qian(OSS)

Hi Nicolas,

Quoted 41 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal

Quoted 1 line · notes, I wrote:

Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Quoted 8 lines
Quoted 1 line

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz.
And results remained unchanged after multiple tests.

I'm not sure what caused the differences between us.

Quoted 8 lines

Once it comes to system stability, you need to ensure that your bootstack is aligned e.g. same TF-A version and sometimes same bootloader since there might be workarounds/erratum applied by the boot firmware.

Regards,
Marco

Thanks for the reminder, and I agree.
I think we need to align our board environment first.

Regards,
Ming

Quoted 68 lines
Quoted 52 lines

Below are my test results:

600Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       600000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
****************************************************************************************************
Running test suite JCT-VC-HEVC_V1 with decoder
GStreamer-H.265-V4L2SL-Gst1.0
Using 2 parallel job(s)
****************************************************************************************************

Ran 139/147 tests successfully in 505.434 secs
Ran 139/147 tests successfully in 505.350 secs
Ran 139/147 tests successfully in 507.540 secs

600Mhz, 1.0v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 1000mV 0mA
825mV 1100mV

       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       600000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
Ran 139/147 tests successfully in 506.901 secs

300Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       300000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
Ran 139/147 tests successfully in 506.063 secs

Downstream v4l2 driver
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 2 0 unknown 1000mV 0mA
825mV 1100mV

       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       600000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2-Gst1.0 -j2 -t
Quoted 7 lines

90
Ran 136/147 tests successfully in 460.435 secs

Regards,
Ming

--
#gernperDu
#CallMeByMyFirstName

Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | https://www.pengutronix.de/ |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-9    |
#11 lore raw

Le mardi 03 février 2026 à 16:53 +0800, Ming Qian(OSS) a écrit :

Quoted 60 lines

Hi Marco,

Quoted 57 lines · Marco Felsch

[You don't often get email from [email protected]. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]

Hi,

sorry for jumping in.

Quoted 50 lines · Ming Qian(OSS)

Hi Nicolas,

Quoted 37 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal
notes, I wrote:

    > Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out. [  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Quoted 5 lines

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz.
And results remained unchanged after multiple tests.

After more testing, the fluster test is stable for NV12/NV15 tiled output for me too. I'm running the tests with linear NV12/P010, which imply an extra set of buffer. I will check if I can give you a easy way to test the linear formats. I also have couple of streams that systematically breaks at specific spot (high complexity scenes) with the provided patch. As most licensed content, this is not sharable as-is. I will try and see if I can find a way to share something.

Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

I'm not sure what caused the differences between us.

Once it comes to system stability, you need to ensure that your bootstack is aligned e.g. same TF-A version and sometimes same bootloader since there might be workarounds/erratum applied by the boot firmware.

Quoted 6 lines
Quoted 3 lines

Regards,
   Marco

Thanks for the reminder, and I agree.
I think we need to align our board environment first.

I do likely have slightly different bootchain, and of course all the HDMI component are downstream, but I can't really isolate the dramatic issue of this overclock without a display component of some sort. Its a huge differentiator in the bandwidth consumption which is the main challenge on this SoC so far. 10bit videos makes things a lot worse fwiw.

We did review latest IMX vendor firmware package and can confirm we are running the latest memory training blob and HDMI firmware.

Nicolas

Quoted 71 lines

Regards,
Ming

Quoted 68 lines
Quoted 51 lines

Below are my test results:

600Mhz, 0.9v
       cat /sys/kernel/debug/regulator/regulator_summary  |grep SW1C SW1C                             0    1      0 unknown   900mV     0mA 825mV  1100mV
       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       600000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t 90
       ****************************************************************************************************
       Running test suite JCT-VC-HEVC_V1 with decoder GStreamer-H.265-V4L2SL-Gst1.0
       Using 2 parallel job(s)
       ****************************************************************************************************

       Ran 139/147 tests successfully               in 505.434 secs Ran 139/147 tests successfully               in 505.350 secs Ran 139/147 tests successfully               in 507.540 secs

600Mhz, 1.0v
       cat /sys/kernel/debug/regulator/regulator_summary  |grep SW1C SW1C                             0    1      0 unknown  1000mV     0mA 825mV  1100mV
       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       600000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t 90
       Ran 139/147 tests successfully               in 506.901 secs

300Mhz, 0.9v
       cat /sys/kernel/debug/regulator/regulator_summary  |grep SW1C SW1C                             0    1      0 unknown   900mV     0mA 825mV  1100mV
       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       300000000

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t 90
       Ran 139/147 tests successfully               in 506.063 secs

Downstream v4l2 driver
       cat /sys/kernel/debug/regulator/regulator_summary  |grep SW1C SW1C                             0    2      0 unknown  1000mV     0mA 825mV  1100mV
       cat /sys/kernel/debug/clk/vpu_g2/clk_rate
       600000000

Quoted 8 lines

       ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2-Gst1.0 -j2 -t 90
       Ran 136/147 tests successfully               in 460.435 secs

Regards,
Ming

--
#gernperDu
#CallMeByMyFirstName

Pengutronix e.K.                           |                             | Steuerwalder Str. 21                       | https://www.pengutronix.de/ | 31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    | Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-9    |

#12 lore raw

Hi Nicolas,

Quoted 62 lines · Nicolas Dufresne
Quoted 60 lines · Ming Qian(OSS)

Hi Marco,

Quoted 57 lines · Marco Felsch

[You don't often get email from [email protected]. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]

Hi,

sorry for jumping in.

Quoted 50 lines · Ming Qian(OSS)

Hi Nicolas,

Quoted 37 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal
notes, I wrote:

> Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Quoted 5 lines

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz.
And results remained unchanged after multiple tests.

Quoted 7 lines

After more testing, the fluster test is stable for NV12/NV15 tiled output for me too. I'm running the tests with linear NV12/P010, which imply an extra set of buffer. I will check if I can give you a easy way to test the linear formats. I also have couple of streams that systematically breaks at specific spot (high complexity scenes) with the provided patch. As most licensed content, this is not sharable as-is. I will try and see if I can find a way to share something.

That would be very helpful. Thank you very much.

Quoted 21 lines
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

I'm not sure what caused the differences between us.

Once it comes to system stability, you need to ensure that your bootstack is aligned e.g. same TF-A version and sometimes same bootloader since there might be workarounds/erratum applied by the boot firmware.

Quoted 6 lines
Quoted 3 lines

Regards,
Marco

Thanks for the reminder, and I agree.
I think we need to align our board environment first.

I do likely have slightly different bootchain, and of course all the HDMI component are downstream, but I can't really isolate the dramatic issue of this overclock without a display component of some sort. Its a huge differentiator in the bandwidth consumption which is the main challenge on this SoC so far. 10bit videos makes things a lot worse fwiw.

Quoted 4 lines

We did review latest IMX vendor firmware package and can confirm we are running the latest memory training blob and HDMI firmware.

Nicolas

Since the datasheet clearly states that the VPU G2 requires a voltage increase to 1.0V to run at 600MHz, (I thought it was already like this, but it actually wasn't.)
I think we can align our test conditions to this.

I will increase the G2 voltage in V2.

Regards,
Ming

Quoted 72 lines
Quoted 72 lines

Regards,
Ming

Quoted 68 lines
Quoted 52 lines

Below are my test results:

600Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

        cat /sys/kernel/debug/clk/vpu_g2/clk_rate
        600000000

        ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
****************************************************************************************************
Running test suite JCT-VC-HEVC_V1 with decoder
GStreamer-H.265-V4L2SL-Gst1.0
Using 2 parallel job(s)
****************************************************************************************************

Ran 139/147 tests successfully in 505.434 secs
Ran 139/147 tests successfully in 505.350 secs
Ran 139/147 tests successfully in 507.540 secs

600Mhz, 1.0v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 1000mV 0mA
825mV 1100mV

        cat /sys/kernel/debug/clk/vpu_g2/clk_rate
        600000000

        ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
Ran 139/147 tests successfully in 506.901 secs

300Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

        cat /sys/kernel/debug/clk/vpu_g2/clk_rate
        300000000

        ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2SL-Gst1.0 -j2 -t

90
Ran 139/147 tests successfully in 506.063 secs

Downstream v4l2 driver
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 2 0 unknown 1000mV 0mA
825mV 1100mV

        cat /sys/kernel/debug/clk/vpu_g2/clk_rate
        600000000

        ./fluster.py run -ts JCT-VC-HEVC_V1 -d GStreamer-H.265-V4L2-Gst1.0 -j2 -t
Quoted 7 lines

90
Ran 136/147 tests successfully in 460.435 secs

Regards,
Ming

--
#gernperDu
#CallMeByMyFirstName

Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | https://www.pengutronix.de/ |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-9    |
#13 lore raw
In reply to Ming Qian(OSS)

Hi,

Am Dienstag, dem 03.02.2026 um 15:13 +0800 schrieb Ming Qian(OSS):

Quoted 53 lines

Hi Nicolas,

Quoted 37 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal

Quoted 1 line · notes, I wrote:

Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Quoted 5 lines

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz.
And results remained unchanged after multiple tests.

I'm not sure what caused the differences between us.

Quoted 8 lines

Below are my test results:

600Mhz, 0.9v
cat /sys/kernel/debug/regulator/regulator_summary |grep SW1C
SW1C 0 1 0 unknown 900mV 0mA
825mV 1100mV

	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

You are driving the SoC out of spec. The datasheet clearly states that you need a 1000mV typical voltage for 600MHz VPU clock.

If you drive the SoC outside of those ratings it squarely depends on the individual SoC if it will tolerate the too low voltage without errors. Some SoCs land on the better side of PVT curve and will run at the higher speed without issues, but some will not and will exhibit random issues outside of the datasheet provided specs.

There isn't much to discuss here. The upstream DT for the i.MX8MQ runs all the clocks at a rate to meet the nominal drive voltage specs. If some peripheral clock does violate this, this is a bug not a feature to replicate in new patches.

Regards,
Lucas

#14 lore raw

Hi Lucas,

Quoted 74 lines · Lucas Stach

Hi,

Am Dienstag, dem 03.02.2026 um 15:13 +0800 schrieb Ming Qian(OSS):

Quoted 53 lines

Hi Nicolas,

Quoted 37 lines · Nicolas Dufresne

Hi,

Quoted 5 lines · Nicolas Dufresne
Quoted 4 lines

This doesn't sound like just a VPU issue; it's related to the display or DDR.
If not displayed, do the fluster test cases yield different results at 600MHz and 300MHz?

Quoted 8 lines

Didn't you run these tests before sending ? I can try again, but in my internal

Quoted 1 line · notes, I wrote:

Tested that, and everything becomes unstable

That was before I figure-out the IRQ handler didn't handle exception bits that didn't stop the decoder (or dry IRQ, which strangely is common from the G2).

Ran some fluster tests now. With this patch the results is not consistent anymore. Then I ran it with weston being started, and in the middle of the test the display turned black. Matches my past observation. We did reproduce this on BSP kernel too. When the display goes black, the recent hantro drivers reports:

[  827.581586] hantro-vpu 38310000.video-codec: frame decode timed out.
[  827.720201] hantro-vpu 38310000.video-codec: not all macroblocks were

decoded.

I have local patches to reduce the cascade of errors, so it likely survived longer then last time. I will send these patches soon. The "not all macroblocks were decoded." is triggered by a bit in the status register that is not documented in NXP TRM. I found that bit in some VC8000D documentation (the sucessor of G2). I concluded it was the same meaning after looking at the failed buffer visually, it is indeed missing couple of macroblocks near th end. Each time we see this error, the DCSS gives up and turn either black, or sometimes other color. The second case has been tracked to a DCSS Scaler underrun, the first we don't know.

Quoted 5 lines

Fluster command ran (two threads, never completes):

./fluster.py run -d GStreamer-H.265-V4L2SL-Gst1.0 -ts JCT-VC-HEVC_V1 -j2 -t90

Nicolas

My test results for fluster differ from yours.
On my end, the results for JCT-VC-HEVC_V1 are consistent at both 300MHz and 600MHz.
And results remained unchanged after multiple tests.

I'm not sure what caused the differences between us.

Quoted 8 lines

Below are my test results:

600Mhz, 0.9v

	cat /sys/kernel/debug/regulator/regulator_summary  |grep SW1C
	 SW1C                             0    1      0 unknown   900mV     0mA
    825mV  1100mV
	cat /sys/kernel/debug/clk/vpu_g2/clk_rate
	600000000

You are driving the SoC out of spec. The datasheet clearly states that you need a 1000mV typical voltage for 600MHz VPU clock.

If you drive the SoC outside of those ratings it squarely depends on the individual SoC if it will tolerate the too low voltage without errors. Some SoCs land on the better side of PVT curve and will run at the higher speed without issues, but some will not and will exhibit random issues outside of the datasheet provided specs.

Quoted 7 lines

There isn't much to discuss here. The upstream DT for the i.MX8MQ runs all the clocks at a rate to meet the nominal drive voltage specs. If some peripheral clock does violate this, this is a bug not a feature to replicate in new patches.

Regards,
Lucas

I agree with you, it's meaningless that test vpu with overdriver clock frequency and nominal drive voltage.
We should focus on the overdrive mode at a frequency of 600 MHz and a voltage of 1.0 V.

It is my mistake that not to adjust the voltage in this patch.

Regards,
Ming