PTX does not differentiate between read and write fences. Hence, these a
lowered to a mem_fence call. The mem_fence function compiles to the
“member.cta” instruction, which commits all outstanding reads and writes
of a thread such that these become visible to all other threads in the same
CTA (i.e., work-group). The instruction does not differentiate between
global and local memory. Hence, the flags parameter is ignored, except
for deciding whether a “member.cta” instruction should be issued at all.
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 315235
This generates a "bar.sync 0” instruction, which not only causes the
threads to wait, but does acts as a memory fence, as required by
OpenCL. The fence does not differentiate between local and global
memory. Unfortunately, there is no similar instruction which does
not include a memory fence. Hence, we cannot optimize the case
where neither CLK_LOCAL_MEM_FENCE nor CLK_GLOBAL_MEM_FENCE is
passed.
llvm-svn: 315228
Also fix get_global_id to consider offset
No idea how to add this for ptx, so they are stuck with the old get_global_id
implementation.
v2: split to a separate patch
v3: Switch R600 to use implictarg.ptr
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 276443
Jeroen Ketema