https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98215
Bug ID: 98215
Summary: Coalescing memory in target region creates slower code
Product: gcc
Version: 10.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: rene.jacobsen at deic dot dk
CC: jakub at gcc dot gnu.org
Target Milestone: ---
Created attachment 49714
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49714&action=edit
Code that produces the bug
Exact compiler version: g++ (Ubuntu 10.2.0-13ubuntu1) 10.2.0
System: Ubuntu 20.10
Command line: g++ -fopenmp -fcf-protection=none -foffload=nvptx-none
-fno-stack-protector -foffload=-misa=sm_35 -foffload=-lm gcc_coalescing_bug.cpp
The attached code shows two possible ways of running code on the GPU. The
coalesced function should be faster, due to coalesced memory access, but is ~4x
slower.
When running it on our system we get the following output:
non_coalesced:
Elapsed time: 0.13381
coalesced:
Elapsed time: 0.48244
non_coalesced:
Elapsed time: 0.133868
coalesced:
Elapsed time: 0.481802
non_coalesced:
Elapsed time: 0.133794
coalesced:
Elapsed time: 0.481685
non_coalesced:
Elapsed time: 0.133875
coalesced:
Elapsed time: 0.481841
