Details:
- Added optional flop counting to all level-3 front-ends, which is
enabled via BLIS_ENABLE_FLOP_COUNT. The flop count can be
reset at any time via bli_flop_count_reset() and queried via
bli_flop_count(). Caveats:
- flop counts are approximate for her[2]k, syr[2]k, trmm, and
trsm operations;
- flop counts ignore extra flops due to non-unit alpha;
- flop counts do not account for situations where beta is zero.
Details:
- Implemented a new function, bli_obj_imag_equals(), which compares the
imaginary part of the first argument to the second argument, which may
be a BLIS_CONSTANT or of a regular real datatype.
Details:
- Changed gemm_kc blocksizes to be reduced by two-thirds instead of
half.
- Changed 3m4m/test_gemm.c driver to divide by 3 instead of 2 when
computing the fixed k dimension.
- Fixed runme.sh so that it would use multiple threads for s/dgemm
cases.
Details:
- Added yet another 4m-based implementation for complex domain level-3
operations. This method, which the 3m/4m paper identifies as Algorithm
"4m_1b" fissures the first loop around the micro-kernel so that the
real sub-panel of the current micro-panel of B is multiplied against
(both sub-panels of) all micro-panels of A, before doing the same for
the imaginary sub-panel of the micro-panel of B. For now, only gemm is
supported, and 4m_1b (labeled "4mb" within the framework) is not yet
integrated into the test suite.
Details:
- Fixed a bug in certain structured level-2 operations that manifested
only when the structured matrix was provided to BLIS as matrix stored
with general stride. The bug was introduced in c472993b when the
densify field was removed from the packm control tree node and
associated APIs. Since then, the packed object was unconditionally
marked with an uplo field of BLIS_DENSE. This is fine for level-3
operations where micro-panels are always densified, but in level-2
contexts, the underlying unblocked variant (fused or unfused) of
structured operations (e.g. trmv) still needs to know whether to
execute its "lower" or "upper" branches of code. Since this field
was unconditionally being set to BLIS_DENSE, the unblocked variants
were always executed the "else" branch, which happened to be the
"lower" case code. Thus, running an upper case produced the wrong
answer. This most obviously manifested in the form of failures for
trmm, trmm3, and trsm in the test suite.
The bug was fixed by setting the packed object's uplo field to
BLIS_DENSE only if the schema indicated that micro-panels were to be
packed. Otherwise, we can assume we are packing to regular row or
column storage, as is the case with level-2 packing. Thanks to
Francisco Igual for reporting the testsuite failures and ultimately
leading us to this bug.
Lead thread of the inner thread communicator could exit subproblem, move on the next iteration of the loop and modify a1_pack, b1_pack, or c1_pack while other threads were still using those.
Barriers were inserted to fix this.
Details:
- No longer need to define BLIS_ENABLE_MULTITHREADING manually in
bli_config.h; it now gets defined when BLIS_ENABLE_OPENMP or
BLIS_ENABLE_PTHREADS is defined.
- Added sanity check to prevent both BLIS__ENABLE_OPENMP and
BLIS_ENABLE_PTHREADS from being enabled simultaneously.
- Reorganization of bli_threading*.h header files, which led to
simplification of threading-related part of blis.h.
- added "-fopenmp -lpthread" to LDFLAGS of sandybridge make_defs.mk
file.
Details:
- Since 4674ca8c, the constraint that KC be a multiple of both MR and
NR have been relaxed, and thus it was time to remove the comments
from the top of the bli_kernel.h files of all configurations.
Details:
- Added new micro-kernels for the AMD piledriver architecture (one
for each datatype).
- Updates and tweaks to piledriver configuration.
- Added 3xk packm micro-kernel support.
- Explicitly unrolled some of the smaller packm micro-kernels.
- Added notes to avx/sandybridge and piledriver micro-kernel files
acknowledging the influence of the corresponding kernel code in
OpenBLAS.
Details:
- Changed the recently-added micro-panel alignment macros so that we now
have two sets--one for micro-panels of matrix A and one for micro-
panels of matrix B: BLIS_UPANEL_[AB]_ALIGN_SIZE_?.
- Store each set of alignment values into a separate blksz_t object in
bli_gemm_cntl_init().
- Adjusted packm_init() to use the separate alignment values.
- Added query routines for the new alignment values to bli_info.c.
- Modified test suite output accordingly.
Details:
- Split bli_determine_blocksize_[fb]() into two functions each, the
newer ones ending with the _sub suffix. These new sub-functions are
now called from bli_[gemm|trmm|trsm]_determine_kc_[fb](), which
eliminates redundant code and will allow any future tweaks to the
core sub-functions to automatically be inherited by the operation-
specific versions.
Details:
- These changes were intended for the previous commit.
- Defined bli_gemm_determine_kc_[fb]() and bli_gemm_determine_kc_[fb](),
which determine blocksizes for gemm-based operations, taking special
care to "nudge" the kc dimension up to a multiple of MR or NR for
hemm and symm operations, as needed.
- Changed bli_gemm_blk_var3f.c to call bli_gemm_determine_kc_f().
instead of bli_determine_blocksize_f().
- Comment updates to bli_trmm_blocksize.c, bli_trsm_blocksize.c.
Details:
- Relaxed a long-held requirement in register blocksizes that required
the kernel programmer to choose a KC that was divisible by both MR
and NR. This was very constraining on some architectures that did not
use register blocksizes that were powers of two. The constraint is
now enforced only for trmm and trsm, where it is needed, and it is
now handled by "nudging" kc upward at runtime, if necessary, to be a
multiple of MR or NR, as needed.
- Defined bli_trmm_determine_kc_[fb]() and bli_trsm_determine_kc_[fb](),
which determine blocksizes for trmm and trsm, taking special care to
"nudge" the kc dimension up to a multiple of MR or NR, as needed.
- Changed bli_trmm_blk_var3[fb].c to call bli_trmm_determine_kc_[fb]()
instead of bli_determine_blocksize_[fb]().
- Added safeguard to bli_align_dim_to_mult() that returns the dimension
unmodified if the dimension multiple is zero (to avoid division by
zero).
- Removed cpp guard/check for KC % MR == 0 and KC % NR == 0 from
bli_kernel_macro_defs.h.
- Whitespace, variable name changes to bli_blocksize.c.
- Removed old commented code from bli_gemm_cntl.c.
Details:
- This commit re-implements a feature that was removed in commit
c2b2ab62. It was removed because, at the time, I wasn't sure how the
micro-panel alignment feature would interact with the 4m method (when
applied at the micro-kernrel level), and so it seemed safer to disable
the feature entirely rather than allow possible breakage. This commit
revisits the issue and safely re-implements the feature in a way that
is compatible with 4m, 3m, 4mh, and 3mh (and native execution).
- Modified the static memory pool to account for micro-panel alignment
space.
- Modified packm_init and blocked variants to align whole micro-panels
by a datatype-specific alignment value that may be set by the
configuration. (If it is not set by the configuration, it will default
to BLIS_SIZEOF_?.)
- Modified macro-kernels so that:
- storage stride is handled properly given the new micro-panel
alignment behavior;
- indexing through 3m/4m/rih-type sub-panels, as is done by trmm and
trsm, is more robust (e.g. will work if the applicable packing
register blocksize is odd);
- imaginary strides are computed and stored within auxinfo_t structs,
which allows the virtual micro-kernels to more easily determine how
to index into the micro-panel operands.
- Modified virtual 3m and 4m micro-kernels to use the imaginary strides
within the auxinfo_t structs instead of panel strides.
- Deprecated the panel stride fields from the auxinfo_t structs.
- Updated test suite to print out the micro-panel alignment values.
Details:
- Added a modified test driver for [cz]gemm that will test all 3m/4m
as well as assembly-based and OpenBLAS implementations of gemm
in single and multithreaded modes.
Details:
- Redefined bli_is_last_iter() to take thread_id and num_thread
arguments, which allows the macro to correctly compute whether a
given iteration is the last that the thread will compute in that
particular loop. The new definition, however, remains disabled
(commented out) until someone can look at this more closely, as
the new definition seems to actually hurt performance slightly.
- Whitespace and related updates to level-3 macro-kernels.
- Updated test suite so that performance results in the hundreds of
gigaflops does not disrupt the column alignment of the output.
Details:
- Fixed a bug in the test suite for the trsm_ukr and gemmtrsm_ukr test
modules whereby the uplo bits of some packed matrix objects were not
being set properly, resulting in false FAILURE results for those
tests. Thanks to Tyler Smith for bringing this issue to my attention.
- Fixed a bug in bli_obj_alloc_buffer() that caused an unnecessary
"not yet implemented" abort() when creating a 1x1 object with non-unit
strides.
Details:
- Expanded the BLIS_PACK_SCHEMA_BITS value in bli_type_defs.h to
include all six bits presently used in the pack schema bitfield of
the info field of obj_t structs. Prior to this commit, the macro
constant only included the lowest five bits, which excluded the
"is or is not packed" bit. This manifested as a strange bug in
probably many level-2 codes that invoked packing, though we only
observed it in ger before fixing. Thanks to Devin Matthews for
finding and reporting this bug.
Details:
- Updated bli_kernel_*_macro_defs.h headers to include default
definitions for 30xk packm kernels.
- Extended function pointer arrays in bli_packm_cxk_*() out to 31 and
included 30xk kernels.
- Addex 30xk kernels to frame/1m/packm/ukernels/bli_packm_ref_cxk_*.c.
Details:
- Applied a patch that allows the top-level Makefile to work on certain
systems. The patch simply separates out the source-to-object code
generation rules for .c and .S files into two separate rules. Thanks
to Devin Matthews for submitting this patch.
Details:
- We actually need to check alignment of lda*sizeof(double) and NOT
a+lda because in the latter case, alignment could cancel out and
still allow the optimized code to run when it shouldn't. Thanks
to Devin for pointing this out.
Details:
- The bugfix in a68b316c was inadvertantly checkin alignment of the
leading dimension itself, rather than the byte size of the leading
dimension. Now, we simply check alignment of a+lda.
