Who decides what a "thread" or "warp" is in a GPU? I don't think GPU threads are scheduled by the OS, so how does a GPU decide when one instruction stream constitutes a thread that is context-switchable? Is this information provided by a compiler?
This comment was marked helpful 0 times.
kayvonf
@kfc9001: Your question also applies to ISPC's abstractions of a "program instance" (akin to a CUDA thread) and "gang of program instances" (akin to an NVIDIA warp). Once you've had experience with Assignment 1, how do you think it's done?
NOTE: Please see Lecture 6, slide 43 for a description of how CUDA programs run on this processor.
This comment was marked helpful 0 times.
fangyihua
Where is the number 48 warps come from? What is the relationship between the execution contexts and 48 warps?
This comment was marked helpful 0 times.
kayvonf
@fangyihua: In the context of this lecture, a warp corresponds an execution context for an instruction stream. This GPU core can interleave up to 48 unique instruction streams. Each instruction stream operates on 32 pieces of data at a time (think 32-wide SIMD). See the next slide for further details.
Who decides what a "thread" or "warp" is in a GPU? I don't think GPU threads are scheduled by the OS, so how does a GPU decide when one instruction stream constitutes a thread that is context-switchable? Is this information provided by a compiler?
This comment was marked helpful 0 times.
@kfc9001: Your question also applies to ISPC's abstractions of a "program instance" (akin to a CUDA thread) and "gang of program instances" (akin to an NVIDIA warp). Once you've had experience with Assignment 1, how do you think it's done?
NOTE: Please see Lecture 6, slide 43 for a description of how CUDA programs run on this processor.
This comment was marked helpful 0 times.
Where is the number 48 warps come from? What is the relationship between the execution contexts and 48 warps?
This comment was marked helpful 0 times.
@fangyihua: In the context of this lecture, a warp corresponds an execution context for an instruction stream. This GPU core can interleave up to 48 unique instruction streams. Each instruction stream operates on 32 pieces of data at a time (think 32-wide SIMD). See the next slide for further details.
This comment was marked helpful 0 times.