Details:
- Parallelism within the IC loop (3rd loop around the microkernel) is
now supported within the trsm operation. This is done via a new branch
on each of the control and thread trees, which guide execution of a
new trsm-only subproblem from within bli_trsm_blk_var1(). This trsm
subproblem corresponds to the macrokernel computation on only the
block of A that contains the diagonal (labeled as A11 in algorithms
with FLAME-like partitioning), and the corresponding row panel of C.
During the trsm subproblem, all threads within the JC communicator
participate and parallelize along the JR loop, including any
parallelism that was specified for the IC loop. (IR loop parallelism
is not supported for trsm due to inter-iteration dependencies.) After
this trsm subproblem is complete, a barrier synchronizes all
participating threads and then they proceed to apply the prescribed
BLIS_IC_NT (or equivalent) ways of parallelism (and any BLIS_JR_NT
parallelism specified within) to the remaining gemm subproblem (the
rank-k update that is performed using the newly updated row-panel of
B). Thus, trsm now supports JC, IC, and JR loop parallelism.
- Modified bli_trsm_l_cntl_create() to create the new "prenode" branch
of the trsm_l cntl_t tree. The trsm_r tree was left unchanged, for
now, since it is not currently used. (All trsm problems are cast in
terms of left-side trsm.)
- Updated bli_cntl_free_w_thrinfo() to be able to free the newly shaped
trsm cntl_t trees. Fixed a potentially latent bug whereby a cntl_t
subnode is only recursed upon if there existed a corresponding
thrinfo_t node, which may not always exist (for problems too small
to employ full parallelization due to the minimum granularity imposed
by micropanels).
- Updated other functions in frame/base/bli_cntl.c, such as
bli_cntl_copy() and bli_cntl_mark_family(), to recurse on sub-prenodes
if they exist.
- Updated bli_thrinfo_free() to recurse into sub-nodes and prenodes
when they exist, and added support for growing a prenode branch to
bli_thrinfo_grow() via a corresponding set of help functions named
with the _prenode() suffix.
- Added a bszid_t field thrinfo_t nodes. This field comes in handy when
debugging the allocation/release of thrinfo_t nodes, as it helps trace
the "identity" of each nodes as it is created/destroyed.
- Renamed
bli_l3_thrinfo_print_paths() -> bli_l3_thrinfo_print_gemm_paths()
and created a separate bli_l3_thrinfo_print_trsm_paths() function to
print out the newly reconfigured thrinfo_t trees for the trsm
operation.
- Trival changes to bli_gemm_blk_var?.c and bli_trsm_blk_var?.c
regarding variable declarations.
- Removed subpart_t enum values BLIS_SUBPART1T, BLIS_SUBPART1B,
BLIS_SUBPART1L, BLIS_SUBPART1R. Then added support for two new labels
(semantically speaking): BLIS_SUBPART1A and BLIS_SUBPART1B, which
represent the subpartition ahead of and behind, respectively,
BLIS_SUBPART1. Updated check functions in bli_check.c accordingly.
- Shuffled layering/APIs for bli_acquire_mpart_[mn]dim() and
bli_acquire_mpart_t2b/b2t(), _l2r/r2l().
- Deprecated old functions in frame/3/bli_l3_thrinfo.c.
Details:
- Implemented a sophisticated data structure and set of APIs that track
the small blocks of memory (around 80-100 bytes each) used when
creating nodes for control and thread trees (cntl_t and thrinfo_t) as
well as thread communicators (thrcomm_t). The purpose of the small
block allocator, or sba, is to allow the library to transition into a
runtime state in which it does not perform any calls to malloc() or
free() during normal execution of level-3 operations, regardless of
the threading environment (potentially multiple application threads
as well as multiple BLIS threads). The functionality relies on a new
data structure, apool_t, which is (roughly speaking) a pool of
arrays, where each array element is a pool of small blocks. The outer
pool, which is protected by a mutex, provides separate arrays for each
application thread while the arrays each handle multiple BLIS threads
for any given application thread. The design minimizes the potential
for lock contention, as only concurrent application threads would
need to fight for the apool_t lock, and only if they happen to begin
their level-3 operations at precisely the same time. Thanks to Kiran
Varaganti and AMD for requesting this feature.
- Added a configure option to disable the sba pools, which are enabled
by default; renamed the --[dis|en]able-packbuf-pools option to
--[dis|en]able-pba-pools; and rewrote the --help text associated with
this new option and consolidated it with the --help text for the
option associated with the sba (--[dis|en]able-sba-pools).
- Moved the membrk field from the cntx_t to the rntm_t. We now pass in
a rntm_t* to the bli_membrk_acquire() and _release() APIs, just as we
do for bli_sba_acquire() and _release().
- Replaced all calls to bli_malloc_intl() and bli_free_intl() that are
used for small blocks with calls to bli_sba_acquire(), which takes a
rntm (in addition to the bytes requested), and bli_sba_release().
These latter two functions reduce to the former two when the sba pools
are disabled at configure-time.
- Added rntm_t* arguments to various cntl_t and thrinfo_t functions, as
required by the new usage of bli_sba_acquire() and _release().
- Moved the freeing of "old" blocks (those allocated prior to a change
in the block_size) from bli_membrk_acquire_m() to the implementation
of the pool_t checkout function.
- Miscellaneous improvements to the pool_t API.
- Added a block_size field to the pblk_t.
- Harmonized the way that the trsm_ukr testsuite module performs packing
relative to that of gemmtrsm_ukr, in part to avoid the need to create
a packm control tree node, which now requires a rntm_t that has been
initialized with an sba and membrk.
- Re-enable explicit call bli_finalize() in testsuite so that users who
run the testsuite with memory tracing enabled can check for memory
leaks.
- Manually imported the compact/minor changes from 61441b24 that cause
the rntm to be copied locally when it is passed in via one of the
expert APIs.
- Reordered parameters to various bli_thrcomm_*() functions so that the
thrcomm_t* to the comm being modified is last, not first.
- Added more descriptive tracing for allocating/freeing small blocks and
formalized via a new configure option: --[dis|en]able-mem-tracing.
- Moved some unused scalm code and headers into frame/1m/other.
- Whitespace changes to bli_pthread.c.
- Regenerated build/libblis-symbols.def.
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:
- 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:
- 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:
- Implemented Ricardo Magana's distributed thread info/communicator
management. Rather that fully construct the thrinfo_t structures, from
root to leaf, prior to spawning threads, the threads individually
construct their thrinfo_t trees (or, chains), and do so incrementally,
as needed, reusing the same structure nodes during subsequent blocked
variant iterations. This required moving the initial creation of the
thrinfo_t structure (now, the root nodes) from the _front() functions
to the bli_l3_thread_decorator(). The incremental "growing" of the tree
is performed in the internal back-end (ie: _int()) function, and so
mostly invisible. Also, the incremental growth of the thrinfo_t tree is
done as a function of the current and parent control tree nodes (as well
as the parent thrinfo_t node), further reinforcing the parallel
relationship between the two data structures.
- Removed the "inner" communicator from thrinfo_t structure definition,
as well as its id. Changed all APIs accordingly. Renamed
bli_thrinfo_needs_free_comms() to bli_thrinfo_needs_free_comm().
- Defined bli_l3_thrinfo_print_paths(), which prints the information
in an array of thrinfo_t* structure pointers. (Used only as a
debugging/verification tool.)
- Deprecated the following thrinfo_t creation functions:
bli_packm_thrinfo_create()
bli_l3_thrinfo_create()
because they are no longer used. bli_thrinfo_create() is now called
directly when creating thrinfo_t nodes.
Details:
- Added optional printf() statements to print out thread communicator
info as the thrinfo_t structure is built in bli_l3_thrinfo.c.
- Minor changes to frame/thread/bli_thrinfo.h.
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:
- 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.
