Results 1 to 3 of 3

Thread: Can different kernels run in a single Compute Unit simultaneously?

  1. #1
    Junior Member Newbie
    Join Date
    Nov 2017
    Location
    Berlin, Germany
    Posts
    2

    Can different kernels run in a single Compute Unit simultaneously?

    I'm puzzled about Wavefront after reading some materials.
    I have two questions that confuse me.
    (1) If there are 4-way SIMDs(16-lane) in a single compute unit and more than 64 work-items running within the compute unit, will the wave scheduler launch all the work-items on all of the four SIMDs or just on one SIMD? In this way, do we call the wavefront with the length of 16 or 64?
    (2) If there are 2 kernels running on a GPU, will the threads owned by different kernels(in different command queue) run in the same compute units simultaneously?

    Sincerely for your help!!!

  2. #2
    Senior Member Regular Contributor
    Join Date
    Apr 2015
    Posts
    316
    Quote Originally Posted by kuba15 View Post
    I'm puzzled about Wavefront after reading some materials.
    I have two questions that confuse me.
    (1) If there are 4-way SIMDs(16-lane) in a single compute unit and more than 64 work-items running within the compute unit, will the wave scheduler launch all the work-items on all of the four SIMDs or just on one SIMD? In this way, do we call the wavefront with the length of 16 or 64?
    (2) If there are 2 kernels running on a GPU, will the threads owned by different kernels(in different command queue) run in the same compute units simultaneously?

    Sincerely for your help!!!
    Keep in mind that this is only relevant for AMD GPUs. I believe, a unit of scheduling there is a full wavefront (64 workitems). I cannot answer your second question with certainty though.

  3. #3
    Junior Member Newbie
    Join Date
    Oct 2016
    Posts
    11
    As far as I know, only one wavefront can be scheduled at one time and one wavefront does only contain work items of one kernel.
    Which means the answer to the second question is "No".

Similar Threads

  1. Branching in compute kernels
    By assembler in forum Vulkan
    Replies: 0
    Last Post: 09-19-2016, 08:05 AM
  2. Replies: 0
    Last Post: 10-21-2015, 07:21 AM
  3. Replies: 1
    Last Post: 04-08-2015, 03:44 AM
  4. Replies: 4
    Last Post: 08-17-2011, 11:37 AM
  5. Multiple textures, single texture unit
    By randomnumbers in forum OpenGL: Basic Coding
    Replies: 1
    Last Post: 05-05-2009, 02:22 PM

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •  
Proudly hosted by Digital Ocean