Details:
- Fixed bugs in trmv_l and trsv_u due to backwards iteration resulting in
unaligned subpartitions. We were already going out of our way a bit to
handle edge cases in the first iteration for blocked variants, and this
was simply the unblocked-fused extension of that idea.
- Fixed control tree handling in her/her2/syr/syr2 that was not taking
into account how the choice of variant needed to be altered for
upper-stored matrices (given that only lower-stored algorithms are
explicitly implemented).
- Added bli_determine_blocksize_dim_f(), bli_determine_blocksize_dim_b()
macros to provide inlined versions of bli_determine_blocksize_[fb]() for
use by unblocked-fused variants.
- Integrated new blocksize_dim macros into gemv/hemv unf variants for
consistency with that of the bugfix for trmv/trsv (both of which now
use the same macros).
- Modified bli_obj_vector_inc() so that 1 is returned if the object is a
vector of length 1 (ie: 1 x 1). This fixes a bug whereby under certain
conditions (e.g. dotv_opt_var1), an invalid increment was returned, which
was invalid only because the code was expecting 1 (for purposes of
performing contiguous vector loads) but got a value greater than 1 because
the column stride of the object (e.g. rho) was inflated for alignment
purposes (albeit unnecessarily since there is only one element in the
object).
- Replaced some old invocations of set0 with set0s.
- Added alpha parameter to gemmtrsm ukernels for x86_64 and use accordingly.
- Fixed increment bug in cleanup loop of gemm ukernel for x86_64.
- Added safeguard to test modules so that testing a problem with a zero
dimension does not result in a failure.
- Tweaked handling of zero dimensions in level-2 and level-3 operations'
internal back-ends to correctly handle cases where output operand still
needs to be scaled (e.g. by beta, in the case of gemm with k = 0).
Details:
- Renamed the following macros:
bli_obj_set_trans() -> bli_obj_set_onlytrans()
bli_obj_trans_status() -> bli_obj_onlytrans_status()
to remove ambiguity as to which bits are read/updated.
Details:
- Changed bli_mem_acquire_m() in bli_mem.c so that we still check if the
memory pool is exhausted before checking out and returning a block, even
if BLIS error checking has been disabled. These errors are useful because
they likely indicate that BLIS was improperly configured for the code
being run.
Details:
- Revamped some parts of commit b6ef84fad1 by adding blocksize extension
fields to the blksz_t object rather than have them as separate structs.
- Updated all packm interfaces/invocations according to above change.
- Generalized bli_determine_blocksize_?() so that edge case optimization
happens if and only if cache blocksizes are created with non-zero
extensions.
- Updated comments in bli_kernel.h files to indicate that the edge case
blocksize extension mechanism is now available for use.
Details:
- Updated _cntl.c files for each level-3 operation to induce blocked
algorithms that first paritition in the n dimension with a blocksize
of NC. Typically this is not an issue since only very large problems
exceed that of NC. But developers often run very large problems, and
so this extra blocking should be the default.
- Removed some recently introduced but now unused macros from
bli_param_macro_defs.h.
Details:
- Replaced multi-type invocations of copys_mxn, xpbys_mxn, etc. (PASTEMAC2
and PASTEMAC3) with those that only use a single type (PASTEMAC).
- Added extra macros to bli_adds_mxn_uplo.h and bli_xpbys_mxn_uplo.h to
accommodate above change.
- Fixed comment typo in bli_config.h files.
- Added .nfs* pattern to .gitignore.
Details:
- Restructured herk_l and herk_u macro-kernels in the imagine of trmm
and trsm, in that the edge cases are captured by the main loop, rather
than trying to have "cleanup" sections that result in four distinct
parts (interior, bottom edge, right edge, bottom-right edge) of the
code.
- Fixed the way b_next was being computed in the non-gemm level-3
macro-kernels (herk, trmm, trsm). The way they are computed now matches
that of gemm.
Details:
- Fixed a potential bug in herk_l at the m_left edge case. If MR was
chosen to be much larger than NR, then one could encounter edge cases
in the the MC dimension that fall entirely below the diagonal, which
the previous implementation of the herk_l macro-kernel was not allowing
for.
Details:
- Added two more arguments to the gemm and gemmtrsm microkernels: the
addresses of the next micro-panels of A and B. By passing these
pointers into the micro-kernel, we allow the micro-kernel author to
prefetch micro-panels of A and B as necessary (though this is
completely optional; these addresses may also be safely ignored).
- Updated all seven macro-kernels so that they compute and pass in
a_next and b_next. Note that ONLY the gemm macro-kernel computes
a_next and b_next with the precise semantics we want. I will go back
and fix the other macro-kernels in the near future.
- Added 'restrict' to various micro-kernels from which it was missing.
Disabled:
- Edited bli_determine_blocksize_b() to include experimental (and
currently disabled) code that computes extended blocks.
- Updated commnts relate to above changes.
- Enabled use of x86 gemmtrsm ukernel in config/flame/bli_kernel.h.
Details:
- Updated frame/util/dupl/bli_dupl_unb_var1.c to utilize PACKNR and NR
explicitly so navigate b1 so that situations where PACKNR > NR are
supported.
- Moved the 4x2 and 4x4 reference micro-kernels in frame/3/gemm/ukernels and
frame/3/trsm/ukernels to kernels/c99/.
- Updated clarksville and flame configurations.
Details:
- Temporarily disabled checks that ensure that enough memory will be allocated
by the contiguous memory allocator for all types, given that the values for
double precision real are the ones used to allocate the space. These checks
can easily go awry in certain situations, especially if you are developing for
only one datatype. So for now, they are probably more trouble than they are
worth.
Details:
- Made substantial changes throughout the framework to decouple the leading
dimension (row or column stride) used within each packed micro-panel from
the corresponding register blocksize. It appears advantageous on some
systems to use, for example, packed micro-panels of A where the column
stride is greater than MR (whereas previously it was always equal to MR).
- Changes include:
- Added BLIS_EXTEND_[MNK]R_? macros, which specify how much extra padding
to use when packing micro-panels of A and B.
- Adjusted all packing routines and macro-kernels to use PACKMR and PACKNR
where appropriate, instead of MR and NR.
- Added pd field (panel dimension) to obj_t.
- New interface to bli_packm_cntl_obj_create().
- Renamed bli_obj_packed_length()/_width() macros to
bli_obj_padded_length()/_width().
- Removed local #defines for cache/register blocksizes in level-3 *_cntl.c.
- Print out new cache and register blocksize extensions in test suite.
- Also added new BLIS_EXTEND_[MNK]C_? macros for future use in using a larger
blocksize for edge cases, which can improve performance at the margins.
Details:
- Fixed a bug in the BLAS compatibility layer, specifically in bla_her2k.c
and bla_syr2k.c, that caused incorrect computation to occur when the BLAS
interface caller requests the [conjugate-]transpose case. Thanks to Bryan
Marker for reporting the behavior that led to this bug.
Details:
- Changed bli_packm_init() to react differently to a situation where a pack
obj_t has an already-allocated mem_t entry that has a buffer that is smaller
than what will be needed to hold the block/panel that now needs to be
packed. Previously, this situation was treated with an abort() since I
assumed something was horribly wrong. I have changed the code so that it now
reacts by releasing the previous mem_t and re-acquires a new mem_t with the
new information. (This change was done at the request of Bryan Marker to
facilitate code generation via DxT.)
Details:
- Fixed a bug in bli_packm_blk_var2() that affected the packing functionality
of hemm and symm. The bug occurs whenever attempting to pack a Hermitian or
symmetric matrix where the block of A being packed intersects the diagonal,
but some of its micro-panels do not intersect the diagonal and lie completely
in the unstored region. Thanks to Francisco Igual for reporting this bug.
- Comment updates to both _blk_var2.c and _blk_var3.c.
Details:
- Removed the overly-paranoid bli_abort() from the end of
bli_packm_acquire_mpart_t2b(), to allow others to experiment with
partitioning through packed blocks of A. Also, and more importantly,
changed an earlier check that was causing an erroneous (but
coincidentally redundant) abort(). Also, updated some of the comments
in bli_packm_part.c.
Details:
- Modified bli_obj_alloc_buffer() to allow allocating an empty buffer, and
modified bli_adjust_strides() to explicitly handle m = n = 0.
- Updated bli_check_matrix_strides() to allow cases where m = n = 0.
Details:
- Fixed an obscure bug in the front-ends for herk, her2k, syrk, and syr2k,
that manifested as the incorrect triangle being updated. It occurred when
the user would pass in a matrix object that was correctly marked as
symmetric/Hermitian and lower-stored, but whose root object was never marked
as lower (or upper). We now alias and re-assign root status for matrix C
within the front-ends. Note that trmm and trsm were already doing this,
albeit for a slightly different reason (to allow the internal back-end to
choose which algorithm to run--lower or upper--based on the uplo of the root
object for both left and right side cases). Thanks to Bryan Marker for
leading me to this bug.
Details:
- Relaxed type checking in getsc so that the input object could be a constant
and not just a proper floating-point type. (If it is a constant, default to
extracting the dcomplex values.) Thanks to Bryan Marker for reporting this
bug.
- Added definition for bli_is_constant() in bli_param_macro_defs.h
- Comment updates to various level-0 scalar routines.
Details:
- This macro is used to determine whether the partitioning routines should
call a corresponding packm_part routine instead. However, it was
unintentionally catching matrices that were marked as "packed" by virtue
of them simply being marked as BLIS_PACKED_UNSPEC in, say, bli_gemv().
The macro has now been renamed to bli_obj_is_panel_packed(), and now only
checks for row or column panel packing. (Note that I first attempted to
fix this bug in a571af816d72.) Thanks to Bryan Marker for reporting the
erroneous behavior that led me to this bug.
Details:
- Removed locally defined gemm microkernel blocksize macros from _mxn
reference microkernel definition and header. Meant to include this in
a recent/previous commit (0020ef7c82).
Details:
- Added missing statement to set structure field of local objects in
top-level BLIS (BLAS-like) API wrappers. Thanks to Bryan Marker for
reporting this bug.
Details:
- Fixed a bug in reference implementation of dupl (bli_dupl_unb_var1.c),
which resulted in incorrect duplication.
- Updated old test drivers according to recently updated packm control tree
creation interface.
- Added 'restrict' to x86 gemm microkernel interface.
Details:
- Delayed #include of bli_kernel.h in blis.h to prevent a situation where
_kernel.h includes an optimized microkernel header, which uses BLIS types
such as dim_t and inc_t, which would precede the definition of those types
in bli_type_defs.h.
- Moved the #include of bli_kernel_macro_defs.h in bli_macro_defs.h to blis.h
(immediately after that of bli_kernel.h).
Details:
- Modified gemmtrsm micro-kernel wrappers to use new aliased blocksize macros
instead of operation-specific ones.
- Removed local, gemmtrsm-specific blocksize macro definitions found in
micro-kernel header files.
(Meant to include above changes in 31b100e7bf4a.)
- Added comments to reference gemmtrsm micro-kernel wrapper implementation.
Details:
- Added new memory alignment constants:
BLIS_HEAP_STRIDE_ALIGN_SIZE (previously assumed to be same as SYSTEM_MEM)
BLIS_CONTIG_ADDR_ALIGN_SIZE (previously assumed to be same as PAGE_SIZE)
BLIS_STACK_BUF_ALIGN_SIZE (previously not enforced)
and renamed existing ones
BLIS_SYSTEM_MEM_ALIGN_SIZE -> BLIS_HEAP_ADDR_ALIGN_SIZE
BLIS_CONTIG_MEM_ALIGN_SIZE -> BLIS_CONTIG_STRIDE_ALIGN_SIZE
to better convey what the alignment factor is used for (and what it is
not used for).
- Removed BLIS_ENABLE_SYSTEM_MEM_ALIGN. Dynamic memory alignment is now
disabled by setting BLIS_HEAP_STRIDE_ALIGN_SIZE to 1.
- Inserted instances of __attribute__((aligned(BLIS_STACK_BUF_ALIGN_SIZE)))
into macro-kernels to specify stack alignment of temporary buffers.
- Modified test suite driver to output new constants.
- Removed bli_align_dim_to_sys() and bli_align_dim_to_cmem(). Instead, we now
use bli_align_dim_to_size(), which takes a third argument (the desired
alignment).
Details:
- Fixed bug whereby axpyv and axpym were incorrectly simplifying to a copy,
rather than an add, when alpha = 1. Thanks to Bryan Marker for identifying
this bug.
Details:
- Renamed abs, min, max, dmin, and dmax macros in bli_f2c.h so that they
would not conflict with anything defined by the user (or the language).
Thanks to Devin Matthews for suggesting this fix.
- Updated all instances of the above macros accordingly.
Details:
- Added new macros that alias level-3 cache and register blocksize macros
to names that can be constructed via the PASTEMAC macro. These aliased
macro definitions live inside bli_kernel_macro_defs.h, which is now
#included after bli_kernel.h.
- Modified macro-kernels to use new aliased blocksize macros instead of
operation-specific ones.
- Removed local, operation-specific kernel blocksize macro definitions
(found in macro-kernel header files).
Details:
- Replaced scalar constant macro definitions in bli_const_defs.h with a single,
simplier macro in bli_obj_macro_defs.h.
- Updated invocations of old macros accordingly.
- Removed bli_const_defs.h.
Details:
- Fixed a bug in the reference implementations of the gemmtrsm wrappers
(bli_gemmtrsm_l_ref_mxn.c and bli_gemmtrsm_u_ref_mxn.c) whereby the
reference gemm microkernel was hard-coded, and thus always called, even
when GEMM_UKERNEL was defined to point to an optimzied microkernel. This
manifested as artificially low trsm performance for all problem sizes, but
especially for small problem sizes as it only affected blocks of A that
intersected the diagonal. Thanks to Mike Kistler of IBM for helping me
find this bug.
Details:
- Changed the definition of bli_is_packed_object() so that it keys off of the
value of the pack schema bits in the info field of obj_t, rather than
comparing the obj_t buffer with that of the mem_t entry. This was the cause
of a very low probability bug whereby uninitialized memory caused the macro
to evaluate to TRUE even though the object in question was not packed.
Thanks to Vernon Austel of IBM for helping discover this bug.
- Changed an abort() in bli_packm_part() to a not-yet-implemented.
Details:
- Fixed a bug present in bli_herk_l_ker_var2() and bli_herk_u_ker_var2() that
only manifests when BLIS is configured such that MR != NR. The bug involves
incorrectly detecting edge cases, which resulted in some parts of matrix C
potentially being skipped and not updated, depending on the problem size.
- Updated the default values of MR and NR in config/reference/bli_kernel.h to
8 and 4, respectively, so that I can better stress the framework on a
day-to-day basis. (The fact that they were both equal to 4 for so long is
why I did not stumble upon this bug much sooner.)
Details:
- Added a new set of reference gemm, gemmtrsm, and trsm micro-kernels that
contain explicit loops over MR and NR, thus allowing them to be used
unmodified by developers who want to build a reference library with
custom register blocksizes.
- Changed config/reference/bli_kernel.h to use above ukernels by default.
- Changed interfaces of new and existing gemm, gemmtrsm, and trsm micro-kernels
to use 'restrict' keyword.
- Added -funroll-loops option to config/reference/make_defs.mk.
- Updated comments in bli_kernel.h describing constraints on register and
cache blocksizes.
- Updated _adds_mxn.h, _copys_mxn.h, and _xpbys_mxn.h macros files so that
single-char macros are also defined.
Details:
- Implemented amax operation in BLIS.
- Activated BLAS2BLIS routine mapping for new amax BLIS implementation.
- Added integer support to [f]printv, [f]printm.
- Added integer support to level-0 copys macros.
- Updated printing of configuration information in test suite driver.
- Comment changes to _config.h files.
- Added comments to bla_dot.c to reminder reader what sdsdot()/dsdot() are
used for.
Details:
- Renamed BLIS_MEMORY_ALIGNMENT_SIZE to BLIS_CONTIG_MEM_ALIGN_SIZE.
- Renamed BLIS_ENABLE_MEMORY_ALIGNMENT to BLIS_ENABLE_SYSTEM_MEM_ALIGN.
- Added BLIS_SYSTEM_MEM_ALIGN_SIZE, which controls only the alignment
passed into posix_memalign() or equivalent.
- Defined new function, bli_align_dim_to_cmem(), which applies the
contiguous memory alignment (rather than the system/malloc alignment).
Details:
- Pass panel strides through bli_align_dim_to_sys() to ensure that each
subsequent packed panel of A and B begins at an aligned address. (The
first panel is presumably aligned to system alignment because it is
aligned to a page boundary, which is typically much larger.)
- Rearranged code in packm_init_pack() to prevent additional conditional
blocks as a result of the aforementioned change.
- Adjusted contiguous memory allocator so that the system memory alignment
is used to allocate enough space for each block no matter what kind of
register blocking is used (even if register blocksize is unit and every
row/column needs maximal padding).
- Adjusted default blocksizes in reference configuration so that MC*KC
and KC*NC result in identical footprints for all datatypes.
Details:
- Changed all filename and function prefixes from 'bl2' to 'bli'.
- Changed the "blis2.h" header filename to "blis.h" and changed all
corresponding #include statements accordingly.
- Fixed incorrect association for Fran in CREDITS file.
Details:
- Removed #include of "blis2.h" from various lower-level, operation-specific
header files throughout the framework. Given that these low-level headers
are included within #blis2.h in a very specific order, #include'ing blis2.h
within them directly is unnecessary.
Details:
- Added cpp guards around the definitions of dim_t, scomplex, and dcomplex.
This is a temporary hack to allow interoperability with libflame. (Similarly
temporary changes are being made to libflame's type definitions file.)
Details:
- Renamed BLIS_MAX_PREFETCH_BYTE_OFFSET to
BLIS_MAX_PRELOAD_BYTE_OFFSET since "prefetch" is kind of a loaded word
(e.g. "prefetch" instructions, which are different than the particular
kind of prefetching/preloading referred to by this constant).
