G6200 － OpenCL memory copy speed is extremely slow

PowerVR Insider
, , ,
21

Thanks, this may be useful for our investigations as we continue to look into this.

22

[quote quote=49504]Thanks, this may be useful for our investigations as we continue to look into this.

[/quote]

If the issue on OpenCL drivers had been fixed.
Can i get a patched version of OpenCL to test ?
Or updated OpenCL SDK will be released on https://pvrsupport.imgtec.com/downloads/24-opencl-sdk ?

23

Unfortunately, we cannot directly supply drivers for our customers’ platforms. We deliver the source code of our reference driver to customers. It’s then their responsibility to integrate the driver with their platforms and distribute binaries to their customers.

24

[quote quote=49518]Unfortunately, we cannot directly supply drivers for our customers’ platforms. We deliver the source code of our reference driver to customers. It’s then their responsibility to integrate the driver with their platforms and distribute binaries to their customers.

[/quote]

Ok . I am the customer of AllWinner which had use G6200.
Can you please notice me at the same time when you deliver the updated source code of driver to AllWinner?
So i can contact AllWinner to get the patched OpenCL driver as soon as possible.

25

If you report the issue to AllWinner they can then contact our driver team through the Partner Portal to discuss the performance issue and determine if a patch or newer driver can be shared with you.

This will also ensure the correct priority is assigned by AllWinner to the investigation.

26

[quote quote=49541]If you report the issue to AllWinner they can then contact our driver team through the Partner Portal to discuss the performance issue and determine if a patch or newer driver can be shared with you.

This will also ensure the correct priority is assigned by AllWinner to the investigation.

[/quote]
Does RDI5386 ticket won‘t be handled if no customer involved？

27

The ticket is in our system, so it will be resolved eventually regardless.

28

[quote quote=49663]The ticket is in our system, so it will be resolved eventually regardless.

[/quote]
So the situation is：
There are no arrangement to assign staff to handle it,because no customer sends a request to increase the priority.

Is that right ?

29

We will resolve the issue in our reference drivers as soon as we can, however it is then up to our customers to deliver the updated drivers for their platforms. The priority that our customers may assign to the issue is out of our control.

30

[quote quote=49675]We will resolve the issue in our reference drivers as soon as we can, however it is then up to our customers to deliver the updated drivers for their platforms. The priority that our customers may assign to the issue is out of our control.

[/quote]
What we said driver is the libPVROCL.so ?
Is it possible to release libPVROCL.so to OpenCL SDK : https://pvrsupport.imgtec.com/downloads/24-opencl-sdk
So we can APK can use the updated libPVROCL.so.

Many thanks .

31

Drivers are specific to each chipset, we don’t release any ourselves.

32

AllWinner had improved the priority of RDI5386 in the ticketing system. Hope the issue can be fixed asap.

33

Any update for the driver fix progress ?

34

From what I can see, the investigation is still ongoing from our driver team regarding our reference drivers. There was some communication with Allwinner, you would have to query them for the progress on their drivers.

35

Any update for the driver fix progress ？

36

Hi,

I’ve contacted our OpenCL team but haven’t had an answer yet. Once we receive a status update, I’ll share it with you.

Thanks,
Joe

37

Hi,

There is an IMG extension that was added to our 1.4 DDK (1.4@3238774) that may improve the performance of this test in your benchmark (cl_img_cached_allocations). The extension adds CL_MEM_USE_UNCACHED_CPU_MEMORY_IMG and CL_MEM_USE_CACHED_CPU_MEMORY_IMG flags that can be passed to clCreateBuffer and clCreateImage. It’s recommended to apply the flag on case-by-case basis to see if it improves performance.

The extension hasn’t been added to the Khronos registry yet, but our OpenCL team have told me it will be uploaded there soon.

Please note that the device you originally reported the performance issue on is running a 1.3 driver that does not support this feature.