Availability of IWOCL poster: "Accelerating SGEMM with Subgroups"

[allanmac1]

Will this poster be made available online:

Accelerating SGEMM with Subgroups

The concept of a subgroup was introduced in the OpenCL 2.0 spec and is an optional Khronos OpenCL extension. This poster will describe work done at Intel to accelerate the SGEMM matrix multiplication algorithm on Intel GPUs using subgroups. Using subgroups, we were able to achieve SGEMM performance results that were comparable to our best hand-written assembler results.

?

[Robert_I_Intel]

Hi Allen,

See the document below. Let me know if you have further questions.

461878

[allanmac1]

Thanks, this is a nice result!

[Robert_I_Intel]

Hi Allan,

Check out this article (the sample code is at the end of it). https://software.intel.com/en-us/articles/sgemm-for-intel-processor-graphics

Let me know what you think!

[allanmac1]

Nice article!

A possible followup article could demonstrate FP16 half float multiply and FP32 accumulate on Gen8 (if 'half' support is added).

There is a lot of interest in FP16 since there are performance, power and bandwidth benefits while still being precise enough for certain applications.

[Robert_I_Intel]

Allan,

FYI: we are working on FP16 example - coming soon :)

[allanmac1]

Robert -- how about some HD 580 Skull Canyon SGEMM benchmarks?

[Robert_I_Intel · Accepted solution]

Hi Allan,

We just got out first Skull Canyon today :) - I can try SGEMM there: probably next week, since I am working from home this week, and will let you know how it goes.