topic Re:bandwidth of different clImage format and efficient access pattern in Graphics

bandwidth of different clImage format and efficient access pattern

chris_pn — Wed, 07 Apr 2021 12:23:51 GMT

Hi,

Per OpenCL Developer Guide (as shown below), I find that max bandwidth of each sampler is 64 Byte/cyc:

https://software.intel.com/content/www/us/en/develop/documentation/iocl-opg/top/coding-for-the-intel-processor-graphics/memory-hierarchy.html

My questions are:

1. Which clImage format with (efficient access pattern) can achieve max bandwidth?

2. Does different clImage format has a max texel/cyc when read by sampler? If yes, where can I find the data for common-used format?

3. What's efficient access pattern to read clImage by Sampler?

Thanks!

Re:bandwidth of different clImage format and efficient access pattern

VidyalathaB_Intel — Thu, 08 Apr 2021 12:07:37 GMT

Hi Chris,

Thanks for reaching out to us

we are working on your issue. we will get back to you soon

Regards,

Vidya

Re: Re:bandwidth of different clImage format and efficient access pattern

chris_pn — Fri, 09 Apr 2021 02:37:19 GMT

Thanks Vidya. For clImage write operation, I also has the same question. So many thanks If you could also provide answer for clImage write operation. Thanks again!

Re: Re:bandwidth of different clImage format and efficient access pattern

chris_pn — Thu, 15 Apr 2021 12:00:03 GMT

Hi Vidya,

Can you answer my questions now or could you tell me how long should I wait for the answer? Thanks.

Re: Re:bandwidth of different clImage format and efficient access pattern

chris_pn — Wed, 28 Apr 2021 08:36:33 GMT

Hi Vidya,

Do you have any update? I'm looking forward to your answer. Thanks!

Re:bandwidth of different clImage format and efficient access pattern

Dunni_A_Intel — Wed, 14 Jul 2021 23:16:57 GMT

Hi Chris,

Apologies for the delay in providing you with a timely response.

The maximum bandwidth may only be achieved when all accesses hit the cache. In general, accesses to smaller image formats will yield better performance than accesses to larger image formats.
Yes, performance varies with different image formats. However, I am unable to locate any public documents with specific performance numbers.
2D accesses patterns will be efficient based on how images are actually laid out in memory.

Please let me know if this answers your questions and if you have any additional questions.

Re:bandwidth of different clImage format and efficient access pattern

Dunni_A_Intel — Thu, 22 Jul 2021 19:00:44 GMT

Hi Chris,

We will no longer be monitoring this thread. Please feel free to make a new post if you have additional questions.

Re: Re:bandwidth of different clImage format and efficient access pattern

chris_pn — Wed, 28 Jul 2021 03:07:44 GMT

Hi Dunni,

Thanks for your answer. It can be better if there is detailed explaination and samples. Current answer related to clImage provided by Intel official document is too short and rough to understand. Image is also a important memory object of OpenCL for user. Thanks anyway.