Details:
- Removed explicit reference to The University of Texas at Austin in the
third clause of the license comment blocks of all relevant files and
replaced it with a more all-encompassing "copyright holder(s)".
- Removed duplicate words ("derived") from a few kernels' license
comment blocks.
- Homogenized license comment block in kernels/zen/3/bli_gemm_small.c
with format of all other comment blocks.
Details:
- Consolidated the *sl.c and *rr.c level-3 macrokernels into a single
file per sl/rr pair, with those files named as they were before
c92762e. The consolidation does not take away the *option* of using
slab or round-robin assignment of micropanels to threads; it merely
*hides* the choice within the definitions of functions such as
bli_thread_range_jrir(), bli_packm_my_iter(), and bli_is_last_iter()
rather than expose that choice explicitly in the code. The choice of
slab or rr is not always hidden, however; there are some cases
involving herk and trmm, for example, that require some part of the
computation to use rr unconditionally. (The --thread-part-jrir option
controls the partitioning in all other cases.)
- Note: Originally, the sl and rr macrokernels were separated out for
clarity. However, aside from the additional binary code bloat, I later
deemed that clarity not worth the price of maintaining the additional
(mostly similar) codes.
Details:
- Implemented support for gemm where A, B, and C may have different
storage datatypes, as well as a computational precision (and implied
computation domain) that may be different from the storage precision
of either A or B. This results in 128 different combinations, all
which are implemented within this commit. (For now, the mixed-datatype
functionality is only supported via the object API.) If desired, the
mixed-datatype support may be disabled at configure-time.
- Added a memory-intensive optimization to certain mixed-datatype cases
that requires a single m-by-n matrix be allocated (temporarily) per
call to gemm. This optimization aims to avoid the overhead involved in
repeatedly updating C with general stride, or updating C after a
typecast from the computation precision. This memory optimization may
be disabled at configure-time (provided that the mixed-datatype
support is enabled in the first place).
- Added support for testing mixed-datatype combinations to testsuite.
The user may test gemm with mixed domains, precisions, both, or
neither.
- Added a standalone test driver directory for building and running
mixed-datatype performance experiments.
- Defined a new variation of castm, castnzm, which operates like castm
except that imaginary values are not touched when casting a real
operand to a complex operand. (By contrast, in these situations castm
sets the imaginary components of the destination matrix to zero.)
- Defined bli_obj_imag_is_zero() and substituted calls in lieu of all
usages of bli_obj_imag_equals() that tested against BLIS_ZERO, and
also simplified the implementation of bli_obj_imag_equals().
- Fixed bad behavior from bli_obj_is_real() and bli_obj_is_complex()
when given BLIS_CONSTANT objects.
- Disabled dt_on_output field in auxinfo_t structure as well as all
accessor functions. Also commented out all usage of accessor
functions within macrokernels. (Typecasting in the microkernel is
still feasible, though probably unrealistic for now given the
additional complexity required.)
- Use void function pointer type (instead of void*) for storing function
pointers in bli_l0_fpa.c.
- Added documentation for using gemm with mixed datatypes in
docs/MixedDatatypes.md and example code in examples/oapi/11gemm_md.c.
- Defined level-1d operation xpbyd and level-1m operation xpbym.
- Added xpbym test module to testsuite.
- Updated frame/include/bli_x86_asm_macros.h with additional macros
(courtsey of Devin Matthews).
Details:
- Updated existing macrokernel function names and definitions to
explicitly use slab assignment of micropanels to threads, then created
duplicate versions of macrokernels that explicitly use round-robin
assignment instead of slab. NOTE: As in ac18949, trsm_r macrokernels
were not substantially updated in this commit because they are
currently disabled in bli_trsm_front.c.
- Updated existing packing function (in blk_packm_blk_var1.c) to
explicitly use slab partitioning, and then duplicated for round-robin.
- Updated control tree initialization to use the appropriate macrokernel
and packm function pointers depending on which method (slab or rr) was
enabled at configure-time.
- Updated configure script to accept new --thread-part-jrir=[slab|rr]
option (-m [slab|rr] for short), which allows the user to explicitly
request either slab or round-robin assignment (partitioning) of
micropanels to threads.
- Updated sandbox/ref99 according to above changes.
- Minor updates to build/add-copyright.py.
Details:
- Removed four trailing spaces after "BLIS" that occurs in most files'
commented-out license headers.
- Added UT copyright lines to some files. (These files previously had
only AMD copyright lines but were contributed to by both UT and AMD.)
- In some files' copyright lines, expanded 'The University of Texas' to
'The University of Texas at Austin'.
- Fixed various typos/misspellings in some license headers.
Details:
- Defined a new struct datatype, rntm_t (runtime), to house the thrloop
field of the cntx_t (context). The thrloop array holds the number of
ways of parallelism (thread "splits") to extract per level-3
algorithmic loop until those values can be used to create a
corresponding node in the thread control tree (thrinfo_t structure),
which (for any given level-3 invocation) usually happens by the time
the macrokernel is called for the first time.
- Relocating the thrloop from the cntx_t remedies a thread-safety issue
when invoking level-3 operations from two or more application threads.
The race condition existed because the cntx_t, a pointer to which is
usually queried from the global kernel structure (gks), is supposed to
be a read-only. However, the previous code would write to the cntx_t's
thrloop field *after* it had been queried, thus violating its read-only
status. In practice, this would not cause a problem when a sequential
application made a multithreaded call to BLIS, nor when two or more
application threads used the same parallelization scheme when calling
BLIS, because in either case all application theads would be using
the same ways of parallelism for each loop. The true effects of the
race condition were limited to situations where two or more application
theads used *different* parallelization schemes for any given level-3
call.
- In remedying the above race condition, the application or calling
library can now specify the parallelization scheme on a per-call basis.
All that is required is that the thread encode its request for
parallelism into the rntm_t struct prior to passing the address of the
rntm_t to one of the expert interfaces of either the typed or object
APIs. This allows, for example, one application thread to extract 4-way
parallelism from a call to gemm while another application thread
requests 2-way parallelism. Or, two threads could each request 4-way
parallelism, but from different loops.
- A rntm_t* parameter has been added to the function signatures of most
of the level-3 implementation stack (with the most notable exception
being packm) as well as all level-1v, -1d, -1f, -1m, and -2 expert
APIs. (A few internal functions gained the rntm_t* parameter even
though they currently have no use for it, such as bli_l3_packm().)
This required some internal calls to some of those functions to
be updated since BLIS was already using those operations internally
via the expert interfaces. For situations where a rntm_t object is
not available, such as within packm/unpackm implementations, NULL is
passed in to the relevant expert interfaces. This is acceptable for
now since parallelism is not obtained for non-level-3 operations.
- Revamped how global parallelism is encoded. First, the conventional
environment variables such as BLIS_NUM_THREADS and BLIS_*_NT are only
read once, at library initialization. (Thanks to Nathaniel Smith for
suggesting this to avoid repeated calls getenv(), which can be slow.)
Those values are recorded to a global rntm_t object. Public APIs, in
bli_thread.c, are still available to get/set these values from the
global rntm_t, though now the "set" functions have additional logic
to ensure that the values are set in a synchronous manner via a mutex.
If/when NULL is passed into an expert API (meaning the user opted to
not provide a custom rntm_t), the values from the global rntm_t are
copied to a local rntm_t, which is then passed down the function stack.
Calling a basic API is equivalent to calling the expert APIs with NULL
for the cntx and rntm parameters, which means the semantic behavior of
these basic APIs (vis-a-vis multithreading) is unchanged from before.
- Renamed bli_cntx_set_thrloop_from_env() to bli_rntm_set_ways_for_op()
and reimplemented, with the function now being able to treat the
incoming rntm_t in a manner agnostic to its origin--whether it came
from the application or is an internal copy of the global rntm_t.
- Removed various global runtime APIs for setting the number of ways of
parallelism for individual loops (e.g. bli_thread_set_*_nt()) as well
as the corresponding "get" functions. The new model simplifies these
interfaces so that one must either set the total number of threads, OR
set all of the ways of parallelism for each loop simultaneously (in a
single function call).
- Updated sandbox/ref99 according to above changes.
- Rewrote/augmented docs/Multithreading.md to document the three methods
(and two specific ways within each method) of requesting parallelism
in BLIS.
- Removed old, disabled code from bli_l3_thrinfo.c.
- Whitespace changes to code (e.g. bli_obj.c) and docs/BuildSystem.md.
Details:
- Changed the way virtual microkernels are handled in the context.
Previously, there were query routines such as bli_cntx_get_l3_ukr_dt()
which returned the native ukernel for a datatype if the method was
equal to BLIS_NAT, or the virtual ukernel for that datatype if the
method was some other value. Going forward, the context native and
virtual ukernel slots will both be initialized to native ukernel
function pointers for native execution, and for non-native execution
the virtual ukernel pointer will be something else. This allows us
to always query the virtual ukernel slot (from within, say, the
macrokernel) without needing any logic in the query routine to decide
which function pointer (native or virtual) to return. (Essentially,
the logic has been shifted to init-time instead of compute-time.)
This scheme will also allow generalized virtual ukernels as a way
to insert extra logic in between the macrokernel and the native
microkernel.
- Initialize native contexts (in bli_cntx_ref.c) with native ukernel
function addresses stored to the virtual ukernel slots pursuant to
the above policy change.
- Renamed all static functions that were native/virtual-ambiguous, such
as bli_cntx_get_l3_ukr_dt() or bli_cntx_l3_ukr_prefers_cols_dt()
pursuant to the above polilcy change. Those routines now use the
substring "get_l3_vir_ukr" in their name instead of "get_l3_ukr". All
of these functions were static functions defined in bli_cntx.h, and
most uses were in level-3 front-ends and macrokernels.
- Deprecated anti_pref bool_t in context, along with related functions
such as bli_cntx_l3_ukr_eff_dislikes_storage_of(), now that 1m's
panel-block execution is disabled.
Details:
- Renamed several macros defined in bli_l3_thrinfo.h designed to compute
the values of a_next and b_next to insert into an auxinfo_t struct in
level-3 macrokernels. (Previously, the macros did not use a bli_
prefix.)
- Updated instances of above macro usage within various macrokernels.
Details:
- Converted most C preprocessor macros in bli_param_macro_defs.h and
bli_obj_macro_defs.h to static functions.
- Reshuffled some functions/macros to bli_misc_macro_defs.h and also
between bli_param_macro_defs.h and bli_obj_macro_defs.h.
- Changed obj_t-initializing macros in bli_type_defs.h to static
functions.
- Removed some old references to BLIS_TWO and BLIS_MINUS_TWO from
bli_constants.h.
- Whitespace changes in select files (four spaces to single tab).
Details:
- Reimplemented several sets of get/set-style preprocessor macros with
static functions, including those in the following frame/base headers:
auxinfo, cntl, mbool, mem, membrk, opid, and pool. A few headers in
frame/thread were touched as well: mutex_*, thrcomm, and thrinfo.
Details:
- Updated the changes introduced in 618f433 so that the strides of the
temporary microtile ct used in the macrokernels is determined based
on the storage preference of the microkernel (via the new functions
below), rather than the strides of c. In almost all cases, presently,
this change results in no net effect, as a high-level optimization
in the _front() functions aligns the storage of c to that of the
microkernel's preference. However, I encountered some cases where
this is not always the case in some development code that has yet
to be committed, and therefore I'm generalizing the framework code
in advance.
- Defined two new functions in bli_cntx.c:
bli_cntx_l3_ukr_prefers_rows_dt()
bli_cntx_l3_ukr_prefers_cols_dt()
which return bool_t's based on the current micro-kernel's storage
preferences. For induced methods, the preference of the underlying
real domain microkernel is returned.
- Updated definition of bli_cntx_l3_ukr_dislikes_storage_of(), and
by proxy bli_cntx_l3_ukr_prefers_storage_of(), to be in terms of
the above functions, rather than querying the preferences of the
native microkernel directly (which did the wrong thing for induced
methods).
Details:
- Previously, rs_ct and cs_ct, the strides of the temporary microtile used
primarily in the macrokernels' edge case handling, were unconditionally
set to 1 and MR, respectively. However, Devin Matthews noted that this
ought to be changed so that the strides of ct were in agreement with the
strides of C. (That is, if C was row-stored, then ct should be accessed
as by rows as well.) The implicit assumption is that the strides of C
have already been adjusted, via induced transposition, if the storage
preference of the microkernel is at odds with the storage of C. So, if
the microkernel prefers row storage, the macrokernel's interior cases
would present row-stored (ideal) microkernel subproblems to the
microkernel, but for edge cases, it would still see column-stored
subproblems (not ideal). This commit fixes this issue. Thanks to Devin
for his suggestion.
Details:
- Altered control tree node struct definitions so that all nodes have the
same struct definition, whose primary fields consist of a blocksize id,
a variant function pointer, a pointer to an optional parameter struct,
and a pointer to a (single) sub-node. This unified control tree type is
now named cntl_t.
- Changed the way control tree nodes are connected, and what computation
they represent, such that, for example, packing operations are now
associated with nodes that are "inline" in the tree, rather than off-
shoot braches. The original tree for the classic Goto gemm algorithm was
expressed (roughly) as:
blk_var2 -> blk_var3 -> blk_var1 -> ker_var2
| |
-> packb -> packa
and now, the same tree would look like:
blk_var2 -> blk_var3 -> packb -> blk_var1 -> packa -> ker_var2
Specifically, the packb and packa nodes perform their respective packing
operations and then recurse (without any loop) to a subproblem. This means
there are now two kinds of level-3 control tree nodes: partitioning and
non-partitioning. The blocked variants are members of the former, because
they iteratively partition off submatrices and perform suboperations on
those partitions, while the packing variants belong to the latter group.
(This change has the effect of allowing greatly simplified initialization
of the nodes, which previously involved setting many unused node fields to
NULL.)
- Changed the way thrinfo_t tree nodes are arranged to mirror the new
connective structure of control trees. That is, packm nodes are no longer
off-shoot branches of the main algorithmic nodes, but rather connected
"inline".
- Simplified control tree creation functions. Partitioning nodes are created
concisely with just a few fields needing initialization. By contrast, the
packing nodes require additional parameters, which are stored in a
packm-specific struct that is tracked via the optional parameters pointer
within the control tree struct. (This parameter struct must always begin
with a uint64_t that contains the byte size of the struct. This allows
us to use a generic function to recursively copy control trees.) gemm,
herk, and trmm control tree creation continues to be consolidated into
a single function, with the operation family being used to select
among the parameter-agnostic macro-kernel wrappers. A single routine,
bli_cntl_free(), is provided to free control trees recursively, whereby
the chief thread within a groups release the blocks associated with
mem_t entries back to the memory broker from which they were acquired.
- Updated internal back-ends, e.g. bli_gemm_int(), to query and call the
function pointer stored in the current control tree node (rather than
index into a local function pointer array). Before being invoked, these
function pointers are first cast to a gemm_voft (for gemm, herk, or trmm
families) or trsm_voft (for trsm family) type, which is defined in
frame/3/bli_l3_var_oft.h.
- Retired herk and trmm internal back-ends, since all execution now flows
through gemm or trsm blocked variants.
- Merged forwards- and backwards-moving variants by querying the direction
from routines as a function of the variant's matrix operands. gemm and
herk always move forward, while trmm and trsm move in a direction that
is dependent on which operand (a or b) is triangular.
- Added functions bli_thread_get_range_mdim(), bli_thread_get_range_ndim(),
each of which takes additional arguments and hides complexity in managing
the difference between the way ranges are computed for the four families
of operations.
- Simplified level-3 blocked variants according to the above changes, so that
the only steps taken are:
1. Query partitioning direction (forwards or backwards).
2. Prune unreferenced regions, if they exist.
3. Determine the thread partitioning sub-ranges.
<begin loop>
4. Determine the partitioning blocksize (passing in the partitioning
direction)
5. Acquire the curren iteration's partitions for the matrices affected
by the current variants's partitioning dimension (m, k, n).
6. Call the subproblem.
<end loop>
- Instantiate control trees once per thread, per operation invocation.
(This is a change from the previous regime in which control trees were
treated as stateless objects, initialized with the library, and shared
as read-only objects between threads.) This once-per-thread allocation
is done primarily to allow threads to use the control tree as as place
to cache certain data for use in subsequent loop iterations. Presently,
the only application of this caching is a mem_t entry for the packing
blocks checked out from the memory broker (allocator). If a non-NULL
control tree is passed in by the (expert) user, then the tree is copied
by each thread. This is done in bli_l3_thread_decorator(), in
bli_thrcomm_*.c.
- Added a new field to the context, and opid_t which tracks the "family"
of the operation being executed. For example, gemm, hemm, and symm are
all part of the gemm family, while herk, syrk, her2k, and syr2k are
all part of the herk family. Knowing the operation's family is necessary
when conditionally executing the internal (beta) scalar reset on on
C in blocked variant 3, which is needed for gemm and herk families,
but must not be performed for the trmm family (because beta has only
been applied to the current row-panel of C after the first rank-kc
iteration).
- Reexpressed 3m3 induced method blocked variant in frame/3/gemm/ind
to comform with the new control tree design, and renamed the macro-
kernel codes corresponding to 3m2 and 4m1b.
- Renamed bli_mem.c (and its APIs) to bli_memsys.c, and renamed/relocated
bli_mem_macro_defs.h from frame/include to frame/base/bli_mem.h.
- Renamed/relocated bli_auxinfo_macro_defs.h from frame/include to
frame/base/bli_auxinfo.h.
- Fixed a minor bug whereby the storage-to-ukr-preference matching
optimization in the various level-3 front-ends was not being applied
properly when the context indicated that execution would be via an
induced method. (Before, we always checked the native micro-kernel
corresponding to the datatype being executed, whereas now we check
the native micro-kernel corresponding to the datatype's real projection,
since that is the micro-kernel that is actually used by induced methods.
- Added an option to the testsuite to skip the testing of native level-3
complex implementations. Previously, it was always tested, provided that
the c/z datatypes were enabled. However, some configurations use
reference micro-kernels for complex datatypes, and testing these
implementations can slow down the testsuite considerably.
Details:
- These changes constitute the first set of changes in preparation to
revamping the structure and use of control trees in BLIS. Modifications
in this commit don't affect the control tree code yet, but rather lay
the groundwork.
- Defined wrappers for the following functions, where the the wrappers
each take a direction parameter of a new enumerated type (BLIS_BWD or
BLIS_FWD), dir_t, and executes the correct underlying function.
- bli_acquire_mpart_*() and _vpart_*()
- bli_*_determine_kc_[fb]()
- bli_thread_get_range_*() and bli_thread_get_range_weighted_*()
- Consolidated all 'f' (forwards-moving) and 'b' (backwards-moving)
blocked variants for trmm and trsm, and renamed gemm and herk variants
accordingly. The direction is now queried via routines such as
bli_trmm_direct(), which deterines the direction from the implied side
and uplo parameters. For gemm and herk, it is uncondtionally BLIS_FWD.
- Defined wrappers to parameter-specific macrokernels for herk, trmm, and
trsm, e.g. bli_trmm_xx_ker_var2(), that execute the correct underlying
macrokernel based on the implied parameters. The same logic used to
choose the dir_t in _direct() functions is used here.
- Simplified the function pointer arrays in _int() functions given the
consolidation and dir_t querying mentioned above.
- Function signature (whitespace) reformatting for various functions.
- Removed old code in various 'old' directories.
Details:
- Reorganized code and renamed files defining APIs related to multithreading.
All code that is not specific to a particular operation is now located in a
new directory: frame/thread. Code is now organized, roughly, by the
namespace to which it belongs (see below).
- Consolidated all operation-specific *_thrinfo_t object types into a single
thrinfo_t object type. Operation-specific level-3 *_thrinfo_t APIs were
also consolidated, leaving bli_l3_thrinfo_*() and bli_packm_thrinfo_*()
functions (aside from a few general purpose bli_thrinfo_*() functions).
- Renamed thread_comm_t object type to thrcomm_t.
- Renamed many of the routines and functions (and macros) for multithreading.
We now have the following API namespaces:
- bli_thrinfo_*(): functions related to thrinfo_t objects
- bli_thrcomm_*(): functions related to thrcomm_t objects.
- bli_thread_*(): general-purpose functions, such as initialization,
finalization, and computing ranges. (For now, some macros, such as
bli_thread_[io]broadcast() and bli_thread_[io]barrier() use the
bli_thread_ namespace prefix, even though bli_thrinfo_ may be more
appropriate.)
- Renamed thread-related macros so that they use a bli_ prefix.
- Renamed control tree-related macros so that they use a bli_ prefix (to be
consistent with the thread-related macros that were also renamed).
- Removed #undef BLIS_SIMD_ALIGN_SIZE from dunnington's bli_kernel.h. This
#undef was a temporary fix to some macro defaults which were being applied
in the wrong order, which was recently fixed.
Details:
- Retrofitted a new data structure, known as a context, into virtually
all internal APIs for computational operations in BLIS. The structure
is now present within the type-aware APIs, as well as many supporting
utility functions that require information stored in the context. User-
level object APIs were unaffected and continue to be "context-free,"
however, these APIs were duplicated/mirrored so that "context-aware"
APIs now also exist, differentiated with an "_ex" suffix (for "expert").
These new context-aware object APIs (along with the lower-level, type-
aware, BLAS-like APIs) contain the the address of a context as a last
parameter, after all other operands. Contexts, or specifically, cntx_t
object pointers, are passed all the way down the function stack into
the kernels and allow the code at any level to query information about
the runtime, such as kernel addresses and blocksizes, in a thread-
friendly manner--that is, one that allows thread-safety, even if the
original source of the information stored in the context changes at
run-time; see next bullet for more on this "original source" of info).
(Special thanks go to Lee Killough for suggesting the use of this kind
of data structure in discussions that transpired during the early
planning stages of BLIS, and also for suggesting such a perfectly
appropriate name.)
- Added a new API, in frame/base/bli_gks.c, to define a "global kernel
structure" (gks). This data structure and API will allow the caller to
initialize a context with the kernel addresses, blocksizes, and other
information associated with the currently active kernel configuration.
The currently active kernel configuration within the gks cannot be
changed (for now), and is initialized with the traditional cpp macros
that define kernel function names, blocksizes, and the like. However,
in the future, the gks API will be expanded to allow runtime management
of kernels and runtime parameters. The most obvious application of this
new infrastructure is the runtime detection of hardware (and the
implied selection of appropriate kernels). With contexts in place,
kernels may even be "hot swapped" at runtime within the gks. Once
execution enters a level-3 _front() function, the memory allocator will
be reinitialized on-the-fly, if necessary, to accommodate the new
kernels' blocksizes. If another application thread is executing with
another (previously loaded) kernel, it will finish in a deterministic
fashion because its kernel information was loaded into its context
before computation began, and also because the blocks it checked out
from the internal memory pools will be unaffected by the newer threads'
reinitialization of the allocator.
- Reorganized and streamlined the 'ind' directory, which contains much of
the code enabling use of induced methods for complex domain matrix
multiplication; deprecated bli_bsv_query.c and bli_ukr_query.c, as
those APIs' functionality is now mostly subsumed within the global
kernel structure.
- Updated bli_pool.c to define a new function, bli_pool_reinit_if(),
that will reinitialize a memory pool if the necessary pool block size
has increased.
- Updated bli_mem.c to use bli_pool_reinit_if() instead of
bli_pool_reinit() in the definition of bli_mem_pool_init(), and placed
usage of contexts where appropriate to communicate cache and register
blocksizes to bli_mem_compute_pool_block_sizes().
- Simplified control trees now that much of the information resides in
the context and/or the global kernel structure:
- Removed blocksize object pointers (blksz_t*) fields from all control
tree node definitions and replaced them with blocksize id (bszid_t)
values instead, which may be passed into a context query routine in
order to extract the corresponding blocksize from the given context.
- Removed micro-kernel function pointers (func_t*) fields from all
control tree node definitions. Now, any code that needs these function
pointers can query them from the local context, as identified by a
level-3 micro-kernel id (l3ukr_t), level-1f kernel id, (l1fkr_t), or
level-1v kernel id (l1vkr_t).
- Removed blksz_t object creation and initialization, as well as kernel
function object creation and initialization, from all operation-
specific control tree initialization files (bli_*_cntl.c), since this
information will now live in the gks and, secondarily, in the context.
- Removed blocksize multiples from blksz_t objects. Now, we track
blocksize multiples for each blocksize id (bszid_t) in the context
object.
- Removed the bool_t's that were required when a func_t was initialized.
These bools are meant to allow one to track the micro-kernel's storage
preferences (by rows or columns). This preference is now tracked
separately within the gks and contexts.
- Merged and reorganized many separate-but-related functions into single
files. This reorganization affects frame/0, 1, 1d, 1m, 1f, 2, 3, and
util directories, but has the most obvious effect of allowing BLIS
to compile noticeably faster.
- Reorganized execution paths for level-1v, -1d, -1m, and -2 operations
in an attempt to reduce overhead for memory-bound operations. This
includes removal of default use of object-based variants for level-2
operations. Now, by default, level-2 operations will directly call a
low-level (non-object based) loop over a level-1v or -1f kernel.
- Converted many common query functions in blk_blksz.c (renamed from
bli_blocksize.c) and bli_func.c into cpp macros, now defined in their
respective header files.
- Defined bli_mbool.c API to create and query "multi-bools", or
heterogeneous bool_t's (one for each floating-point datatype), in the
same spirit as blksz_t and func_t.
- Introduced two key parameters of the hardware: BLIS_SIMD_NUM_REGISTERS
and BLIS_SIMD_SIZE. These values are needed in order to compute a third
new parameter, which may be set indirectly via the aforementioned
macros or directly: BLIS_STACK_BUF_MAX_SIZE. This value is used to
statically allocate memory in macro-kernels and the induced methods'
virtual kernels to be used as temporary space to hold a single
micro-tile. These values are now output by the testsuite. The default
value of BLIS_STACK_BUF_MAX_SIZE is computed as
"2 * BLIS_SIMD_NUM_REGISTERS * BLIS_SIMD_SIZE".
- Cleaned up top-level 'kernels' directory (for example, renaming the
embarrassingly misleading "avx" and "avx2" directories to "sandybridge"
and "haswell," respectively, and gave more consistent and meaningful
names to many kernel files (as well as updating their interfaces to
conform to the new context-aware kernel APIs).
- Updated the testsuite to query blocksizes from a locally-initialized
context for test modules that need those values: axpyf, dotxf,
dotxaxpyf, gemm_ukr, gemmtrsm_ukr, and trsm_ukr.
- Reformatted many function signatures into a standard format that will
more easily facilitate future API-wide changes.
- Updated many "mxn" level-0 macros (ie: those used to inline double loops
for level-1m-like operations on small matrices) in frame/include/level0
to use more obscure local variable names in an effort to avoid variable
shaddowing. (Thanks to Devin Matthews for pointing these gcc warnings,
which are only output using -Wshadow.)
- Added a conj argument to setm, so that its interface now mirrors that
of scalm. The semantic meaning of the conj argument is to optionally
allow implicit conjugation of the scalar prior to being populated into
the object.
- Deprecated all type-aware mixed domain and mixed precision APIs. Note
that this does not preclude supporting mixed types via the object APIs,
where it produces absolutely zero API code bloat.
Details:
- Fixed a family of bugs in the triangular level-3 operations for
certain complex implementations (3m1 and 4m1a) that only manifest if
one of the register blocksizes (PACKMR/PACKNR, actually) is odd:
- Fixed incorrect imaginary stride computation in bli_packm_blk_var2()
for the triangular case.
- Fixed the incorrect computation of imaginary stride, as stored in
the auxinfo_t struct in trmm and trsm macro-kernels.
- Fixed incorrect pointer arithmetic in the trsm macro-kernels in the
cases where the the register blocksize for the triangular matrix is
odd. Introduced a new byte-granular pointer arithmetic macro,
bli_ptr_add(), that computes the correct value.
- Added cpp macro to bli_macro_defs.h for typeof() operator, defined in
terms of __typeof__, which is used by bli_ptr_add() macro.
- Disabled the row- vs. column-storage optimization in bli_trmm_front()
for singleton problems because the inherent ambiguity of whether a
scalar is row-stored or column-stored causes the wrong parameter
combination code to be executed (by dumb luck of our checking for
row storage first).
- Added commented-out debugging lines to 3m1/4m1a and reference
micro-kernels, and trsm_ll macro-kernel.
Details:
- Defined a new "3ms" (separated 3m) pack schema and added appropriate
support in packm_init(), packm_blk_var2().
- Generalized packm_struc_cxk_3mi to take the imaginary stride (is_p)
as an argument instead of computing it locally. Exception: for trmm,
is_p must be computed locally, since it changes for triangular
packed matrices. Also exposed is_p in interface to dt-specific
packm_blk_var2 (and _var1, even though it does not use imaginary
stride).
- Renamed many functions/variables from _3mi to _3mis to indicate that
they work for either interleaved or separated 3m pack schemas.
- Generalized gemm and herk macro-kernels to pass in imaginary stride
rather than compute them locally.
- Added support for 3m2 and 3m3 algorithms to frame/ind, including 3m2-
and 3m3-specific virtual micro-kernels.
- Added special gemm macro-kernels to support 3m2 and 3m3.
- Added support for 3m2 and 3m3 to testsuite.
- Corrected the type of the panel dimension (pd_) in various macro-
kernels from inc_t to dim_t.
- Renamed many functions defined in bli_blocksize.c.
- Moved most induced-related macro defs from frame/include to
frame/ind/include.
- Updated the _ukernel.c files so that the micro-kernel function pointers
are obtained from the func_t objects rather than the cpp macros that
define the function names.
- Updated test/3m4m driver, Makefile, and run script.
Details:
- Renamed all remaining 3m/4m packing files and symbols to 3mi/4mi
('i' for "interleaved"). Similar changes to 3M/4M macros.
- Renamed all 3m/4m files and functions to 3m1/4m1.
- Whitespace changes.
Details:
- Renamed several variables and macros from 3m/4m to 3mi/4mi. This is
because those packing schemas were always implicitly "interleaved".
This new naming scheme will make way for new schemas that separate
instead of interleve the real and imaginary (and summed) parts.
- Expanded the pack format sub-field of the pack schema field of the
info_t to 4 bits (from 3). This will allow for more schema types
going forward.
- Removed old _cntl.c files for herk3m, herk4m, trmm3m, trmm4m.
Details:
- Fixed a bug that only affected trmm3 when performed via 3mh or 4mh,
whereby micro-panels of the triangular matrix were packed with "dead
space" between them due to failing to adjust for the fact that pointer
arithmetic was occurring in units of complex elements while the data
being packed consisted of real elements. It turns out that the macro-
kernel suffered from the same bug, meaning the panels were actually
being packed and read consistently. The only way I was able to
discover the bug in the first place was because the packed block of A
was overflowing into the beginning of the packed row panel of B using
the sandybridge configuration.
Details:
- Modified bli_packm_blk_var1.c and _var2.c to increase the triangular
case's panel increment by 1 if it would otherwise be odd. This is
particularly necessary in _var2.c when handling the interleaved 3m
or ro/io/rpi pack schemas, since division of an odd number by 2 can
happen if both the panel length and the panel packing dimension
(register packing blocksize) are odd, thus making their product odd.
- Modified bli_packm_init.c so that panel strides are increased by 1
if they would otherwise be odd, even for non-3m related packing.
- Modified the trmm and trsm macro-kernels so that triangular packed
micro-panels are traversed with this new "increment by 1 if odd"
policy.
- Added sanity checks in trmm and trsm macro-kernels that would result
in an abort() if the conditions that would lead to a "divide odd
integer by 2" scenario ever manifest.
- Defined bli_is_odd(), _is_even() macros in bli_scalar_macro_defs.h.
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:
- 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:
- Modified macro-kernels to pass the pack_t schema values for matrices
A and B into the datatype-specific functions, where they are now
inserted into a newly-expanded auxinfo_t struct. This gives gives the
micro-kernels access to the pack_t schema values embedded in the
control trees, which determine the precise format into which the
matrix elements are packed.
- Updated a call to bli_packm_init_pack() in src/test_libblis.c to
remove densify argument. Meant to include this in commit c472993b.
Details:
- Instead of inferring the storage format of the micro-panels from within
the packm variants, we now pass in a bool_t value that denotes whether
the packed matrix contains row-stored column panels or column-stored
row panels. This value can then be tested more easily inside the main
packm variant loop.
- Renumbered pack_t schema values in bli_type_defs.h so that there are
now five bits, each with different meaning:
- 4: packed or not packed?
- 3: packed for 3m?
- 2: packed for 4m?
- 1: packed to panels?
- 0: stored by rows or columns?
- Added new macros that test for status of above bits in schema bit
subfield, and renamed some existing macros related to 4m/3m.
Details:
- Updated copyright headers to include "at Austin" in the name of the
University of Texas.
- Updated the copyright years of a few headers to 2014 (from 2011 and
2012).
Should work fine for small number of threads (up to 8 or maybe even 16).
However, performance is yet untested.
This parallelizes the "JR" loop for the left sided cases
and the "IR" loop for the right sided cases.
Future work is to parallelize the outer loops as well.
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:
- 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:
- 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:
- Updated (and fixed some errors in) the "Assumptions/assertions" comment
section of macro-kernels.
- Changed register blocksizes of reference configuration to MR = 8 and
NR = 4. It's always good for MR != NR in the reference configuration
since it may help uncover bugs related to non-square micro-kernels.
Details:
- Modified the interfaces to the datatype-specific macro-kernels so that:
- pd_a and pd_b are passed in (which contain the panel dimensions of
packed panels of a and b).
- rs_a and cs_b are no longer passed in (they were guaranteed to be 1).
- Modified implementations of datatype-specific macro-kernels so pd_a,
pd_b, cs_a, and rs_b are used instead of cpp macros for MR, NR, PACKMR,
and PACKNR, respectively.
- Declare temporary c matrices (ct) as being maxmr-by-maxnr, which for now
is equivalent to being mr-by-nr. maxmr and maxnr are declared in a new
header file bli_kernel_post_macro_defs.h.
Details:
- Modified all control tree node definitions to include a new field of
type func_t*, which is similar to a blksz_t except that it contains
one function pointer (each typed simply as void*) for each datatype.
We use the func_t* to embed pointers to the micro-kernels to use for
the leaf-level nodes of each control tree. This change is a natural
extension of control trees and will allow more flexibility in the
future.
- Modified all macro-kernel wrappers to obtain the micro-kernel pointers
from the incomming (previously ignored) control tree node and then pass
the queried pointer into the datatype-specific macro-kernel code, which
then casts the pointer to the appropriate type (new typedefs residing
in bli_kernel_type_defs.h) and then uses the pointer to call the micro-
kernel. Thus, the micro-kernel function is no longer "hard-coded" (that
is, determined when the datatype-specific macro-kernel functions are
instantiated by the C preprocessor).
- Added macros to bli_kernel_macro_defs.h that build datatype-specific
base names if they do not exist already, and then uses those to build
datatype-specific micro-kernel function names. This will allow
developers extra flexibility if they wanted to, for example, name each
of their datatype-specific micro-kernels differently (e.g. double
real might be named bli_dgemm_opt_4x4() while double complex might be
named bli_zgemm_opt_2x2()).
- Inserted appropriate code into _cntl_init() functions that allocates
and initializes a func_t object for the corresponding micro-kernels.
The gemm ukernel func_t object is created once, in bli_gemm_cntl_init(),
and then reused via extern wherever possible.
Details:
- Changed the value being stored into the auxinfo_t structure in trmm
and trsm macro-kernels. Whereas before we stored whatever value was
provided to the macro-kernel implementation via ps_a/ps_b, now we
store the stride that will advance to the next variable-length
micro-panel of the triangular matrix A (left) or B (right).
- Whitespace changes to the files affected above.
Details:
- Replaced conditional expressions in macro-kernels related to computing
the addresses a2 and b2 (a_next and b_next) with a preprocessor macro
invocation, bli_is_last_iter(), that tests the same condition.
- Updated gemm_ukr module to use auxinfo_t argument.
- Whitespace changes in test suite ukr modules.
Details:
- Removed a_next and b_next arguments to micro-kernels and replaced them
with a pointer to a new datatype, auxinfo_t, which is simply a struct
that holds a_next and b_next. The struct may hold other auxiliary
information that may be useful to a micro-kernel, such as micro-panel
stride. Micro-kernels may access struct fields via accessor macros
defined in bli_auxinfo_macro_defs.h.
- Updated all instances of micro-kernel definitions, micro-kernel calls,
as well as macro-kernels (for declaring and initializing the structs)
according to above change.
Details:
- Added infrastructure to support a new scalar representation, whereby
every object contains an internal scalar that defaults to 1.0. This
facilitates passing scalars around without having to house them in
separate objects. These "attached" scalars are stored in the internal
atom_t field of the obj_t struct, and are always stored to be the same
datatype as the object to which they are attached. Level-3 variants no
longer take scalar arguments, however, level-3 internal back-ends stll
do; this is so that the calling function can perform subproblems such
as C := C - alpha * A * B on-the-fly without needing to change either
of the scalars attached to A or B.
- Removed scalar argument from packm_int().
- Observe and apply attached scalars in scalm_int(), and removed scalar
from interface of scalm_unb_var1().
- Renamed the following functions (and corresponding invocations):
bli_obj_init_scalar_copy_of()
-> bli_obj_scalar_init_detached_copy_of()
bli_obj_init_scalar() -> bli_obj_scalar_init_detached()
bli_obj_create_scalar_with_attached_buffer()
-> bli_obj_create_1x1_with_attached_buffer()
bli_obj_scalar_equals() -> bli_obj_equals()
- Defined new functions:
bli_obj_scalar_detach()
bli_obj_scalar_attach()
bli_obj_scalar_apply_scalar()
bli_obj_scalar_reset()
bli_obj_scalar_has_nonzero_imag()
bli_obj_scalar_equals()
- Placed all bli_obj_scalar_* functions in a new file, bli_obj_scalar.c.
- Renamed the following macros:
bli_obj_scalar_buffer() -> bli_obj_buffer_for_1x1()
bli_obj_is_scalar() -> bli_obj_is_1x1()
- Defined new macros to set and copy internal scalars between objects:
bli_obj_set_internal_scalar()
bli_obj_copy_internal_scalar()
- In level-3 internal back-ends, added conditional blocks where alpha and
beta are checked for non-unit-ness. Those values for alpha and beta are
applied to the scalars attached to aliases of A/B/C, as appropriate,
before being passed into the variant specified by the control tree.
- In level-3 blocked variants, pass BLIS_ONE into subproblems instead of
alpha and/or beta.
- In level-3 macro-kernels, changed how scalars are obtained. Now, scalars
attached to A and B are multiplied together to obtain alpha, while beta
is obtained directly from C.
- In level-3 front-ends, removed old function calls meant to provide
future support for mixed domain/precision. These can be added back later
once that functionality is given proper treatment. Also, removed the
creating of copy-casts of alpha and beta since typecasting of scalars
is now implicitly handled in the internal back-ends when alpha and
beta are applied to the attached scalars.
Details:
- Removed restrict declaration from b_cast and c_cast from
bli_trsm_lu_ker_var2.c and bli_trsm_rl_ker_var2.c. Curiously, they
are causing problems for xlc only in those two files and no other
macro-kernels.
- Fixed (hopefully) kernel function parameter type declarations in
kernels/bgq/1f/bli_axpyf_opt_var1.c and kernels/bgq/3/bli_gemm_8x8.c.
Details:
- Fixed sloppy placement of 'restrict' pointer declarations in level-3
macro-kernels. Previously, all restricted pointers were being declared
at the outer-most function scope level. While this violates the C99
standard, very few of the compilers used with BLIS so far have seemed
to care. The lone exception has been IBM's xlc. Thanks to Tyler Smith
for identifying this bug (and suggesting the fix).
Details:
- Removed support for duplication from the gemmtrsm/trsm micro-kernels
and all framework code.
- Updated test suite modules according to above changes.
Details:
- Fixed a bug in trsm_rl_ker_var2() caused by incorrect edge case handling.
- Fixed a bug in trsm_rl_ker_var2() and trsm_ru_ker_var2() whereby k was
incorrectly being adjusted upward by MR, instead of NR. The rl and ru
trmm macro-kernels were updated in a similar fashion.
- Fixed a bug in trsm_ru_ker_var2() that was due to a missing negation on
diagoffb when recomputing k to skip a zero region below where the
diagonal intersects the right side of the block. The corresponding
trmm macro-kernel was also updated.
- Fixed a bug in trsm_ru_ker_var2() where the the adjustment of k (by NR)
needed to be placed AFTER the block that recomputes k to skip the zero
region (if present). The other three trsm macro-kernels, as well as the
trmm macro-kernels, were updated in the same manner, for consistency.
- Fixed a bug in trmm_lu_ker_var2() in which the wrong dimension (n) was
being updated to skip a zero region to the left of where the diagonal
of A intersects the top edge of the block.
- Comment updates to all trsm and trmm macro-kernels.
- Comment updates to bli_packm_init.c.
