Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Details:
- This is an "omnibus" commit, consisting of multiple medium-sized
commits that affect non-trivial aspects of BLIS. The major highlights:
- Relocated the pba, sba pool (from the rntm_t), and mem_t (from the
cntl_t) to the thrinfo_t object. This allows the rntm_t to be
effectively const (although it is sometimes copied internally and
modified to reflect different ways of parallelism). Moving the mem_t
sets the stage for sharing a global control tree amongst all
threads.
- De-templatized the macrokernels for gemmt, trmm, and trsm to match
the macrokernel for gemm, which has been de-templatized since
54fa28b.
- Reimplemented bli_l3_determine_kc() by separating out the logic for
adjusting KC based on MR/NR for triangular A and/or B into a new
function, bli_l3_adjust_kc(). For now, this function is still called
from bli_l3_determine_kc(), but in the future we plan to have it
called once when constructing the control tree.
- Refactored the level-3 thread decorator into two parts:
- One part deals only with launching threads, each one calling a
generic thread entry function. This code resides in frame/thread
and constitutes the definition of bli_thread_launch(). Note that
it is specific to the threading implementation (OpenMP, pthreads,
single, etc.)
- The other part deals with passing the matrix operands and related
information into bli_thread_launch(). This is the "l3 decorator"
and now resides in frame/3. It is agnostic to the threading
implementation.
- Modified the "level" of the thread control tree passed in at each
operation. Previously, each operation (e.g. bli_gemm_blk_var1()) was
passed in a communicator representing the active thread teams which
would share the available work. Now, the *parent* thread comm is
passed in. The operation then grabs the child comm and uses it to
partition the work. The difference is in bli_trsm_blk_var1(), where
there are now two children nodes for this single operation (i.e. the
thread control tree is split one level above where the control tree
is). The sub-prenode is used for the trsm subproblem while the
normal sub-node is used for the gemm part. Importantly, the parent
comm is used for the barrier between them.
- Removed cntl_t* arguments from bli_*_front() functions. These will be
added back in the future when the control tree's creation is moved so
that it happens much sooner (provided that bli_*_front() have not been
absorbed into their respective bli_*_ex() functions).
- Renamed various bli_thread_*() query functions to bli_thrinfo_*(),
for consistency. This includes _num_threads(), _thread_id(), _n_way(),
_work_id(), _sba_pool(), _pba(), _mem(), _barrier(), _broadcast(), and
_am_chief().
- Removed extraneous barrier from _blk_var3() of gemm and trsm.
- Fixed a typo in bli_type_defs.h where BLIS_BLAS_INT_TYPE_SIZE was
misspelled.
- (cherry picked from commit aeb5f0c)
Fixed performance bug caused by redundant packing. (#680)
Details:
- Fixed a performance bug whereby multiple threads were redundantly
packing the same (rather than separate) micropanels. This bug was
caused by different parts of the code using the num_threads/thread_id
field of the thrinfo_t vs. the n_way/work_id fields. The fix was to
standardize on the latter and provide a "fake" thrinfo_t sub-prenode
in the thrinfo tree which consists of single-member thread teams. The
single team with multiple threads node is still required since it and
only it can be used to perform barriers and broadcasts (e.g. of the
packed buffer pointer).
- (cherry picked from commit 29f79f0)- Loading branch information
