Also enabled weighted partitioning for herk, trmm
Fixed bug where multiple threads would try to modify the same state in the internal level 3 functions
Correctly computed a_next and b_next for gemm, herk macrokernels
a_next and b_next point to the current micropanels in trmm
Details:
- Defined new INSERT_GENTFUNC macros so that the macro always takes
exactly the number of arguments needed for the particular operation or
variant being defined. Many operations were using INSERT_GENTFUNC
macros that expected one auxiliary argument even though none were
needed. Those instances have now been updated. Most of these instances
were in the level-0 and -1v operations, as well as some operations
defined in frame/util.
Details:
- Simplified some code that would have allowed the diagonal of a trmm
or trsm triangular matrix to intersect the short end of a micro-panel.
This is disallowed via higher-level constraints on cache blocksizes, so
this code was never needed and only served to obfuscate.
- Updated some comments in trmm, trsm macro-kernels.
Details:
- Completely reoganized norm operations:
- Renames:
- fnormsc, fnormv, fnormm -> normfsc, normfv, normfm (2-norm)
- absumv -> norm1v (vector 1-norm)
- New operations:
- norm1m (matrix 1-norm)
- normiv, normim (infinity-norm)
- amaxv (BLAS-like absolute maximum value index)
- asumv (BLAS-like absolute sum)
- Deprecated absumm, as it did not correspond to any actual norm.
(However, an inlined version now exists in the testsuite module for
randm.)
The environment variables all follow the format BLIS_X_NT,
where X is the index of the loop as described in our paper
Anatomy of High Performance Many-Threaded Matrix Multiplication.
These indices are IR, JR, IC, KC, and JC.
Also enabled parallelism for hemm and symm, but these are currently untested.
This change makes each operation have its own thread info type,
allowing more fine control of threading in operations that have different types of suboperations
Details:
- Minor update to bli_sumsqv_unb_var1() to bring it up-to-date with
LAPACK 3.5.0's zlassq.f, which, starting with 3.4.2, returns NaN when
the vector (or matrix) contains a NaN.
- Minor change to bli_abmaxv_unb_var1() to more closely mimic the
behavior of netlib BLAS's izamax(). There, a "less than or equal to"
operator is used in the search instead of "less than", which would
change the element index returned if there were multiple maximum values.
- Added macro function definitions for bli_isinf() and bli_isnan(), which
are currently implemented in terms of isinf() and isnan() from math.h.
Details:
- Fixed various bugs in packm_*_cxk(), the 4m/3m micro-kernels, and
elsewhere in the framework that were not yet set up to work properly
when BLIS_ENABLE_C99_COMPLEX is defined in bli_config.h
- Extensive changes to f2c-derived files in frame/compat/f2c to allow
C99 complex storage. Most of these changes center around accessing
real and imaginary components via bli_?real()/bli_?imag() accessor
macros, and setting of values via bli_?sets() assignment macros.
(Thanks to Vladimir Sukarev for pointing out that _ENABLE_C99_COMPLEX
was broken.)
In packm variant 1, the variable p_begin was incremented each iteration, causing a dependency.
This dependeny was removed, allowing each iteration to be executed in parallel.
Somewhere in bli_threading.c, I was allocating an array of pointers instead of an array of structs.
Added a multithreading infrastructure that should be independent of multithreading implementation in the future.
Currently, gemm blocked variants 1f and 2f, and packm variant blocked variant 1 is parallelized.
Details:
- Removed BLIS_CONTIG_STRIDE_ALIGN_SIZE from bli_config.h of all
configurations. It was already going unused in packm_init() since the
recent 4m/3m commit. This setting was rarely, if ever, useful, and its
existence only posed a potential risk for 4m/3m-based implementations.
- Removed BLIS_CONTIG_STRIDE_ALIGN_SIZE usage from mem_pool_macro_defs.h.
- Updated comments regarding CONTIG_STRIDE_ALIGN_SIZE in template
micro-kernels.
Details:
- Standard names for reference kernels (levels-1v, -1f and 3) are now
macro constants. Examples:
BLIS_SAXPYV_KERNEL_REF
BLIS_DDOTXF_KERNEL_REF
BLIS_ZGEMM_UKERNEL_REF
- Developers no longer have to name all datatype instances of a kernel
with a common base name; [sdcz] datatype flavors of each kernel or
micro-kernel (level-1v, -1f, or 3) may now be named independently.
This means you can now, if you wish, encode the datatype-specific
register blocksizes in the name of the micro-kernel functions.
- Any datatype instances of any kernel (1v, 1f, or 3) that is left
undefined in bli_kernel.h will default to the corresponding reference
implementation. For example, if BLIS_DGEMM_UKERNEL is left undefined,
it will be defined to be BLIS_DGEMM_UKERNEL_REF.
- Developers no longer need to name level-1v/-1f kernels with multiple
datatype chars to match the number of types the kernel WOULD take in
a mixed type environment, as in bli_dddaxpyv_opt(). Now, one char is
sufficient, as in bli_daxpyv_opt().
- There is no longer a need to define an obj_t wrapper to go along with
your level-1v/-1f kernels. The framework now prvides a _kernel()
function which serves as the obj_t wrapper for whatever kernels are
specified (or defaulted to) via bli_kernel.h
- Developers no longer need to prototype their kernels, and thus no
longer need to include any prototyping headers from within
bli_kernel.h. The framework now generates kernel prototypes, with the
proper type signature, based on the kernel names defined (or defaulted
to) via bli_kernel.h.
- If the complex datatype x (of [cz]) implementation of the gemm micro-
kernel is left undefined by bli_kernel.h, but its same-precision real
domain equivalent IS defined, BLIS will use a 4m-based implementation
for the datatype x implementations of all level-3 operations, using
only the real gemm micro-kernel.
Details:
- Added the ability to induce complex domain level-3 operations via new
virtual complex micro-kernels which are implemented via only real
domain micro-kernels. Two new implementations are provided: 4m and 3m.
4m implements complex matrix multiplication in terms of four real
matrix multiplications, where as 3m uses only three and thus is
capable of even higher (than peak) performance. However, the 3m method
has somewhat weaker numerical properties, making it less desirable
in general.
- Further refined packing routines, which were recently revamped, and
added packing functionality for 4m and 3m.
- Some modifications to trmm and trsm macro-kernels to facilitate indexing
into micro-panels which were packed for 4m/3m virtual kernels.
- Added 4m and 3m interfaces for each level-3 operation.
- Various other minor changes to facilitate 4m/3m methods.
Details:
- Removed use of bli_min() and bli_max() that were only being used to
try to support situations where the diagonal would intersect the
short end of some micro-panels, which is situation that is disallowed
at a higher level by various constraints on the register and cache
blocksize. This only affected trsm_ll and trsm_lu.
- Use panel stride as passed into the macro-kernel rather than compute
it via k and PACKMR/PACKNR. This affects all macro-kernels of trsm.
Details:
- Fixed an obscure bug in left-hand trmm that would only manifest when
non-zero register blocksize extensions (PACKMR > MR or PACKNR > NR)
are used.
- Removed use of bli_min() and bli_max() that were only being used to
try to support situations where the diagonal would intersect the
short end of some micro-panels, which is situation that is disallowed
at a higher level by various constraints on the register and cache
blocksize. This only affected trmm_ll and trmm_lu.
- Use panel stride as passed into the macro-kernel rather than compute
it via k and PACKMR/PACKNR. This affects all macro-kernels of trmm.
Details:
- Fixed a bug in packm_cxk() whereby the packm ukernel was being chosen
from ldp, which is always equal to PACKMR or PACKNR. The problem with
this is that the pack ukernels were implicitly assuming that the
panel dimension of the panel being packed was equal to ldp, which
is not the case when the register blocksizes extensions are non-zero
(ie: when PACKMR > MR or PACKNR > NR, whichever is applicable). This
problem has been fixed by passing ldp into the pack ukernels, which
now walk through the packed micro-panel region by incrementing by this
value, rather than incrementing by the inherent panel dimension value
assumed by each packm ukernel (e.g. 4 in the case of packm_ref_4xk).
- Also fixed a very minor edge case inefficiency whereby pack ukernels
smaller than the default were not being used in edge cases, and instead
those situations were being handled by scal2m. This is related to the
issue above, because the pack ukernel itself was being chosen based on
ldp instead of the panel dimension.
Details:
- Consolidated the functionality previously supported by packm_blk_var2()
and packm_blk_var3() into a new variant, packm_blk_var1().
- Updates to packm_gen_cxk(), packm_herm_cxk.c(), and packm_tri_cxk()
to accommodate above changes.
- Removed packm_blk_var3() and retired packm_blk_var2() to
frame/1m/packm/old.
- Updated all level-3 _cntl_init() functions so that the new, more
versatile packm_blk_var1 is used for all level-3 matrix packing.
Details:
- Revised packm_blk_var2() and _var3() by encapsulating the general,
hermitian/symmetric, and triangular panel-packing subproblems into
separate functions: packm_gen_cxk(), packm_herm_cxk(), and
packm_tri_cxk(), respectively. Also, homogenized the packm code as
well as the new specialized packm_*_cxk() code to further improve
readability.
