Reputation: 157
CUDA: Can a SM concurrently alternate between warps from different blocks?
Let's say a SM has been populated with 8 blocks of 64 threads each.
That gives us 2 warps/block, and 16 warps in total. SMs can alternate between warps in order to hide latencies. Must these warps belong to the same block, or can a warp from block 5 be replaced by a warp from block 8, for example?
Upvotes: 0
Views: 179
Answers (1)
Reputation: 151809
Yes, the SM scheduler can "alternate" or choose warps for scheduling from any that are resident on that SM.
The fact that SMs have a max possible warp load (64, currently, for some GPUs) or thread load (2048, currently, for some GPUs) that exceeds the possible limit of a single block (1024, currently, for all GPUs supported by recent CUDA toolkits) is so that the SM can choose warps from different blocks for scheduling, to improve the possibilities for latency hiding.
Upvotes: 3
Related Questions
- Can a Cuda warp communicate with a different warp without using shared memory?
- CUDA blocks & warps - which can run in parallel on a single SM?
- When a warp is assigned to a portion of a SM, will it stay there until that warp completes?
- CUDA: Selecting one thread per warp
- CUDA shared memory and warp synchronization
- Do the threads in a CUDA warp execute in parallel on a multiprocessor?
- CUDA warp scheduling: another warp from the current block or a warp from another block?
- Can two blocks be executed in the same warp?
- Is changing memory within warp defined?
- Why only one of the warps is executed by a SM in cuda?