Since AMDGPU v0.5.0 gridsize represents the number of "workgroups" (or blocks in CUDA) and no longer "workitems workgroups" (or threads blocks in CUDA) as HIP is used for kernel launches instead of HSA.
This means that previous AMDGPU kernel launch param gridsize need to be adapted from gridsize = nx to gridsize = cld(nx, groupsize).
Also, queues and signals are now abstracted into streams.
luraess
commented
1 year ago
Since AMDGPU v0.5.0
gridsizerepresents the number of "workgroups" (or blocks in CUDA) and no longer "workitems workgroups" (or threads blocks in CUDA) as HIP is used for kernel launches instead of HSA.This means that previous AMDGPU kernel launch param
gridsizeneed to be adapted fromgridsize = nxtogridsize = cld(nx, groupsize).Also,
queuesandsignalsare now abstracted intostreams.