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:
- Renamed various files that were previously named according to a
"with context" or "without context" convention. For example, the
following files in frame/3 were renamed:
frame/3/bli_l3_oapi_woc.c -> frame/3/bli_l3_oapi_ba.c
frame/3/bli_l3_oapi_wc.c -> frame/3/bli_l3_oapi_ex.c
frame/3/bli_l3_tapi_woc.c -> frame/3/bli_l3_tapi_ba.c
frame/3/bli_l3_tapi_wc.c -> frame/3/bli_l3_tapi_ex.c
Here, the "ba" is for "basic" and "ex" is for "expert". This new
naming scheme will make more sense especially if/when additional
expert parameters are added to the expert APIs (typed and object).
Details:
- Split existing typed APIs into two subsets of interfaces: one for use
with expert parameters, such as the cntx_t*, and one without. This
separation was already in place for the object APIs, and after this
commit the typed and object APIs will have similar expert and non-
expert APIs. The expert functions will be suffixed with "_ex" just as
is the case for expert interfaces in the object APIs.
- Updated internal invocations of typed APIs (functions such as
bli_?setm() and bli_?scalv()) throughout BLIS to reflect use of the
new explictly expert APIs.
- Updated example code in examples/tapi to reflect the existence (and
usage) of non-expert APIs.
- Bumped the major soname version number in 'so_version'. While code
compiled against a previous version/commit will likely still work
(since the old typed function symbol names still exist in the new API,
just with one less function argument) the semantics of the function
have changed if the cntx_t* parameter the application passes in is
non-NULL. For example, calling bli_daxpyv() with a non-NULL context
does not behave the same way now as it did before; before, the
context would be used in the computation, and now the context would
be ignored since the interace for that function no longer expects a
context argument.
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:
- Added logic to level-1v, -1d, -1f, -1m, -2, and -3 operations' _check()
functions to ensure that all operands are of the same datatype. There
are some exceptions that were left out, such as the _check() function
for the various norm operations since they have a different idea of
datatype consistency (ie: the norm object must be the real projection
of the primary input vector/matrix object).
- Query pack schemas in level-3 bli_*_front() functions and store those
values in the schema bitfields of the correponding obj_t's when the
cntx's method is not BLIS_NAT. (When method is BLIS_NAT, the default
native schemas are stored to the obj_t's.)
- In bli_l3_cntl_create_if(), query the schemas stored to the obj_t's in
bli_*_front(), clear the schema bitfields, and pass the queried values
into bli_gemm_cntl_create() and bli_trsm_cntl_create().
- Updated APIs for bli_gemm_cntl_create() and bli_trsm_cntl_create() to
take schemas for A and B, and use these values to initialize the
appropriate control tree nodes. (Also cpp-disabled the panel-block cntl
tree creation variant, bli_gemmpb_cntl_create(), as it has not been
employed by BLIS in quite some time.)
- Simplified querying of schema in bli_packm_init() thanks to above
changes.
- Updated openmp and pthreads definitions of bli_l3_thread_decorator()
so that thread-local aliases of matrix operands are guaranteed, even
if aliasing is disabled within the internal back-end functions (e.g.
bli_gemm_int.c). Also added a comment to bli_thrcomm_single.c
explaining why the extra aliasing is not needed there.
- Change bli_gemm() and level-3 friends so that the operation's ind()
function is called only if all matrix operands have the same datatype,
and only if that datatype is complex. The former condition is needed
in preparation for work related to mixed domain operands, while the
latter helps with readability, especially for those who don't want to
venture into frame/ind.
- Reshuffled arguments in bli_cntx_set_thrloop_from_env() to be
consistent with BLIS calling conventions (modified argument(s) are
last), and updated all invocations in the level-3 _front() functions.
- Comment updates to bli_cntx_set_thrloop_from_env().
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:
- Added missing 'restrict' keyword to cntx_t* argument of function
signatures corresponding to level-1v, level-1f, and level-1m kernels.
This affected bli_l1v_ker_prot.h, bli_l1f_ker_prot.h, and
bli_l1m_ker_prot.h. (The 'restrict' was already being used to
qualify cntx_t* arguments for kernels defined in bli_l3_ker_prot.h.)
- Added comments to bli_l1v_ker.h, bli_l1f_ker.h, bli_l1m_ker.h, and
bli_l3_ukr.h that help explain how those headers function to produce
kernel prototypes using the prototype macros defined in the files
mentioned above.
Details:
- Merged contributions made by AMD via 'amd' branch (see summary below).
Special thanks to AMD for their contributions to-date, especially with
regard to intrinsic- and assembly-based kernels.
- Added column storage output cases to microkernels in
bli_gemm_zen_asm_d6x8.c and bli_gemmtrsm_l_zen_asm_d6x8.c. Even with
the extra cost of transposing the microtile in registers, this is
much faster than using the general storage case when the underlying
matrix is column-stored.
- Added s and d assembly-based zen gemmtrsm_u microkernel (including
column storage optimization mentioned above).
- Updated zen sub-configuration to reflect presence of new native
kernels.
- Temporarily reverted zen sub-configuration's level-3 cache blocksizes
to smaller haswell values.
- Temporarily disabled small matrix handling for zen configuration
family in config/zen/bli_family_zen.h.
- Updated zen CFLAGS according to changes in 1e4365b.
- Updated haswell microkernels such that:
- only one vzeroupper instruction is called prior to returning
- movapd/movupd are used in leiu of movaps/movups for double-real
microkernels. (Note that single-real microkernels still use
movaps/movups.)
- Added kernel prototypes to kernels/zen/bli_kernels_zen.h, which is
now included via frame/include/bli_arch_config.h.
- Minor updates to bli_amaxv_ref.c (and to inlined "test" implementation
in testsuite/src/test_amaxv.c).
- Added early return for alpha == 0 in bli_dotxv_ref.c.
- Integrated changes from f07b176, including a fix for undefined
behavior when executing the 1m method under certain conditions.
- Updated config_registry; no longer need haswell kernels for zen
sub-configuration.
- Tweaked marginal and pass thresholds for dotxf.
- Reformatted level-1v, -1f, and -3 amd kernels and inserted additional
comments.
- Updated LICENSE file to explicitly mention that parts are copyright
UT-Austin and AMD.
- Added AMD copyright to header templates in build/templates.
Summary of previous changes from 'amd' branch.
- Added s and d assembly-based zen gemm microkernels (d6x8 and d8x6) and
s and d assembly-based zen gemmtrsm_l microkernels (d6x8).
- Added s and d intrinsics-based zen kernels for amaxv, axpyv, dotv, dotxv,
and scalv, with extra-unrolling variants for axpyv and scalv.
- Added a small matrix handler to bli_gemm_front(), with the handler
implemented in kernels/zen/3/bli_gemm_small_matrix.c.
- Added additional logic to sumsqv that first attempts to compute the
sum of the squares via dotv(). If there is a floating-point exception
(FE_OVERFLOW), then the previous (numerically conservative) code is
used; otherwise, the result of dotv() is square-rooted and stored as
the result. This new implementation is only enabled when FE_OVERFLOW
is #defined. If the macro is not #defined, then the previous
implementation is used.
- Added axpyv and dotv standalone test drivers to test directory.
- Added zen support to old cpuid_x86.c driver in build/auto-detect/old.
- Added thread-local and __attribute__-related macros to bli_macro_defs.h.
Details:
- Added a new configure option, --[en|dis]able-packbuf-pools, which will
enable or disable the use of internal memory pools for managing buffers
used for packing. When disabled, the function specified by the cpp
macro BLIS_MALLOC_POOL is called whenever a packing buffer is needed
(and BLIS_FREE_POOL is called when the buffer is ready to be released,
usually at the end of a loop). When enabled, which was the status quo
prior to this commit, a memory pool data structure is created and
managed to provide threads with packing buffers. The memory pool
minimizes calls to bli_malloc_pool() (i.e., the wrapper that calls
BLIS_MALLOC_POOL), but does so through a somewhat more complex
mechanism that may incur additional overhead in some (but not all)
situations. The new option defaults to --enable-packbuf-pools.
- Removed the reinitialization of the memory pools from the level-3
front-ends and replaced it with automatic reinitialization within the
pool API's implementation. This required an extra argument to
bli_pool_checkout_block() in the form of a requested size, but hides
the complexity entirely from BLIS. And since bli_pool_checkout_block()
is only ever called within a critical section, this change fixes a
potential race condition in which threads using contexts with different
cache blocksizes--most likely a heterogeneous environment--can check
out pool blocks that are too small for the submatrices it wishes to
pack. Thanks to Nisanth Padinharepatt for reporting this potential
issue.
- Removed several functions in light of the relocation of pool reinit,
including bli_membrk_reinit_pools(), bli_memsys_reinit(),
bli_pool_reinit_if(), and bli_check_requested_block_size_for_pool().
- Updated the testsuite to print whether the memory pools are enabled or
disabled.
Details:
- Defined two new functions in bli_init.c: bli_init_once() and
bli_finalize_once(). Each is implemented with pthread_once(), which
guarantees that, among the threads that pass in the same pthread_once_t
data structure, exactly one thread will execute a user-defined function.
(Thus, there is now a runtime dependency against libpthread even when
multithreading is not enabled at configure-time.)
- Added calls to bli_init_once() to top-level user APIs for all
computational operations as well as many other functions in BLIS to
all but guarantee that BLIS will self-initialize through the normal
use of its functions.
- Rewrote and simplified bli_init() and bli_finalize() and related
functions.
- Added -lpthread to LDFLAGS in common.mk.
- Modified the bli_init_auto()/_finalize_auto() functions used by the
BLAS compatibility layer to take and return no arguments. (The
previous API that tracked whether BLIS was initialized, and then
only finalized if it was initialized in the same function, was too
cute by half and borderline useless because by default BLIS stays
initialized when auto-initialized via the compatibility layer.)
- Removed static variables that track initialization of the sub-APIs in
bli_const.c, bli_error.c, bli_init.c, bli_memsys.c, bli_thread, and
bli_ind.c. We don't need to track initialization at the sub-API level,
especially now that BLIS can self-initialize.
- Added a critical section around the changing of the error checking
level in bli_error.c.
- Deprecated bli_ind_oper_has_avail() as well as all functions
bli_<opname>_ind_get_avail(), where <opname> is a level-3 operation
name. These functions had no use cases within BLIS and likely none
outside of BLIS.
- Commented out calls to bli_init() and bli_finalize() in testsuite's
main() function, and likewise for standalone test drivers in 'test'
directory, so that self-initialization is exercised by default.
Details:
- Fixed implicit typecasting of conj_t to trans_t in bli_[un]packm_cxk.c.
- Properly typecast integer arguments to match format specifier in various
calls to printf() in bli_l3_thrinfo.c, bli_cntx.c, bli_pool.c, and
bli_util_oapi.c.
- Fixed "unsigned less-than-comparison with zero" checks in bli_check.c,
bli_cntx.h.
- Fixed mis-typed enums in bli_cntx.c (e.g., l1mkr_t that should have been
l1fkr_t or l1vkr_t).
- Fixed instances of opid_t value BLIS_GEMM that should have been l3ukr_t
value BLIS_GEMM_UKR in bli_cntx_ref.c.
- NOTE: These issues were identified via compiler warnings when building
BLIS with clang on a rather old installation of OS X:
$ clang --version
Apple LLVM version 5.0 (clang-500.2.79) (based on LLVM 3.3svn)
Target: x86_64-apple-darwin15.2.0
Thread model: posix
Details:
- Fixed a bug in bli_l3_packm() that caused cntl_t-cached packed mem_t
entries to be released and then re-acquired unnecessarily. (In essence,
the "<" operands in the conditional that guards the
release-and-reacquire code block simply needed to be swapped.) The bug
should have only affected performance (rather than the computed result).
Thanks to Minh Quan for identifying and reporting the bug.
Details:
- Removed the family field inside the cntx_t struct and re-added it to the
cntl_t struct. Updated all accessor functions/macros accordingly, as well
as all consumers and intermediaries of the family parameter (such as
bli_l3_thread_decorator(), bli_l3_direct(), and bli_l3_prune_*()). This
change was motivated by the desire to keep the context limited, as much
as possible, to information about the computing environment. (The family
field, by contrast, is a descriptor about the operation being executed.)
- Added additional functions to bli_blksz_*() API.
- Added additional functions to bli_cntx_*() API.
- Minor updates to bli_func.c, bli_mbool.c.
- Removed 'obj' from bli_blksz_*() API names.
- Removed 'obj' from bli_cntx_*() API names.
- Removed 'obj' from bli_cntl_*(), bli_*_cntl_*() API names. Renamed routines
that operate only on a single struct to contain the "_node" suffix to
differentiate with those routines that operate on the entire tree.
- Added enums for packm and unpackm kernels to bli_type_defs.h.
- Removed BLIS_1F and BLIS_VF from bszid_t definition in bli_type_defs.h.
They weren't being used and probably never will be.
Details:
- Defined bli_gemmbp_cntl_create(), bli_gemmpb_cntl_create(), with the
body of bli_gemm_cntl_create() replaced with a call to the former.
- Defined bli_cntl_free_w_thrinfo(), bli_cntl_free_wo_thrinfo(). Now,
bli_cntl_free() can check if the thread parameter is NULL, and if so,
call the latter, and otherwise call the former.
- Defined bli_gemm1mbp_cntx_init(), bli_gemm1mpb_cntx_init(), both in
terms of bli_gemm1mxx_cntx_init(), which behaves the same as
bli_gemm1m_cntx_init() did before, except that an extra bool parameter
(is_pb) is used to support both bp and pb algorithms (including to
support the anti-preference field described below).
- Added support for "anti-preference" in context. The anti_pref field,
when true, will toggle the boolean return value of routines such as
bli_cntx_l3_ukr_eff_prefers_storage_of(), which has the net effect of
causing BLIS to transpose the operation to achieve disagreement (rather
than agreement) between the storage of C and the micro-kernel output
preference. This disagreement is needed for panel-block implementations,
since they induce a transposition of the suboperation immediately before
the macro-kernel is called, which changes the apparent storage of C. For
now, anti-preference is used only with the pb algorithm for 1m (and not
with any other non-1m implementation).
- Defined new functions,
bli_cntx_l3_ukr_eff_prefers_storage_of()
bli_cntx_l3_ukr_eff_dislikes_storage_of()
bli_cntx_l3_nat_ukr_eff_prefers_storage_of()
bli_cntx_l3_nat_ukr_eff_dislikes_storage_of()
which are identical to their non-"eff" (effectively) counterparts except
that they take the anti-preference field of the context into account.
- Explicitly initialize the anti-pref field to FALSE in
bli_gks_cntx_set_l3_nat_ukr_prefs().
- Added bli_gemm_ker_var1.c, which implements a panel-block macro-kernel
in terms of the existing block-panel macro-kernel _ker_var2(). This
technique requires inducing transposes on all operands and swapping
the A and B.
- Changed bli_obj_induce_trans() macro so that pack-related fields are
also changed to reflect the induced transposition.
- Added a temporary hack to bli_l3_3m4m1m_oapi.c that allows us to easily
specify the 1m algorithm (block-panel or panel-block).
- Renamed the following cntx_t-related macros:
bli_cntx_get_pack_schema_a() -> bli_cntx_get_pack_schema_a_block()
bli_cntx_get_pack_schema_b() -> bli_cntx_get_pack_schema_b_panel()
bli_cntx_get_pack_schema_c() -> bli_cntx_get_pack_schema_c_panel()
and updated all instantiations. Also updated the field names in the
cntx_t struct.
- Comment updates.
Details:
- Implemented the 1m method for inducing complex domain matrix
multiplication. 1m support has been added to all level-3 operations,
including trsm, and is now the default induced method when native
complex domain gemm microkernels are omitted from the configuration.
- Updated _cntx_init() operations to take a datatype parameter. This was
needed for the corresponding function for 1m (because 1m requires us
to choose between column-oriented or row-oriented execution, which
requires us to query the context for the storage preference of the
gemm microkernel, which requires knowing the datatype) but I decided
that it made sense for consistency to add the parameter to all other
cntx initialization functions as well, even though those functions
don't use the parameter.
- Updated bli_cntx_set_blkszs() and bli_gks_cntx_set_blkszs() to take
a second scalar for each blocksize entry. The semantic meaning of the
two scalars now is that the first will scale the default blocksize
while the second will scale the maximum blocksize. This allows scaling
the two independently, and was needed to support 1m, which requires
scaling for a register blocksize but not the register storage
blocksize (ie: "packdim") analogue.
- Deprecated bli_blksz_reduce_dt_to() and defined two new functions,
bli_blksz_reduce_def_to() and bli_blksz_reduce_max_to(), for reducing
default and maximum blocksizes to some desired blocksize multiple.
These functions are needed in the updated definitions of
bli_cntx_set_blkszs() and bli_gks_cntx_set_blkszs().
- Added support for the 1e and 1r packing schemas to packm, including
1e/1r packing kernels.
- Added a minor optimization to bli_gemm_ker_var2() that allows, under
certain circumstances (specifically, real domain beta and row- or
column-stored matrix C), the real domain macrokernel and microkernel
to be called directly, rather than using the virtual microkernel
via the complex domain macrokernel, which carries a slight additional
amount of overhead.
- Added 1m support to the testsuite.
- Added 1m support to Makefile and runme.sh in test/3m4m. Also simplified
some code in test_gemm.c driver.
Details:
- Removed the vast majority of directories named "old", which contained
deprecated code that I wasn't quite ready to jettison from the source
tree.
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:
- Reworked the build system around a configuration registry file, named
config_registry', that identifies valid configuration targets, their
constituent sub-configurations, and the kernel sets that are needed by
those sub-configurations. The build system now facilitates the building
of a single library that can contains kernels and cache/register
blocksizes for multiple configurations (microarchitectures). Reference
kernels are also built on a per-configuration basis.
- Updated the Makefile to use new variables set by configure via the
config.mk.in template, such as CONFIG_LIST, KERNEL_LIST, and KCONFIG_MAP,
in determining which sub-configurations (CONFIG_LIST) and kernel sets
(KERNEL_LIST) are included in the library, and which make_defs.mk files'
CFLAGS (KCONFIG_MAP) are used when compiling kernels.
- Reorganized 'kernels' directory into a "flat" structure. Renamed kernel
functions into a standard format that includes the kernel set name
(e.g. 'haswell'). Created a "bli_kernels_<kernelset>.h" file in each
kernels sub-directory. These files exist to provide prototypes for the
kernels present in those directories.
- Reorganized reference kernels into a top-level 'ref_kernels' directory.
This directory includes a new source file, bli_cntx_ref.c (compiled on
a per-configuration basis), that defines the code needed to initialize
a reference context and a context for induced methods for the
microarchitecture in question.
- Rewrote make_defs.mk files in each configuration so that the compiler
variables (e.g. CFLAGS) are "stored" (renamed) on a per-configuration
basis.
- Modified bli_config.h.in template so that bli_config.h is generated with
#defines for the config (family) name, the sub-configurations that are
associated with the family, and the kernel sets needed by those
sub-configurations.
- Deprecated all kernel-related information in bli_kernel.h and transferred
what remains to new header files named "bli_arch_<configname>.h", which
are conditionally #included from a new header bli_arch.h. These files
are still needed to set library-wide parameters such as custom
malloc()/free() functions or SIMD alignment values.
- Added bli_cntx_init_<configname>.c files to each configuration directory.
The files contain a function, named the same as the file, that initializes
a "native" context for a particular configuration (microarchitecture). The
idea is that optimized kernels, if available, will be initialized into
these contexts. Other fields will retain pointers to reference functions,
which will be compiled on a per-configuration basis. These bli_cntx_init_*()
functions will be called during the initialization of the global kernel
structure. They are thought of as initializing for "native" execution, but
they also form the basis for contexts that use induced methods. These
functions are prototyped, along with their _ref() and _ind() brethren, by
prototype-generating macros in bli_arch.h.
- Added a new typedef enum in bli_type_defs.h to define an arch_t, which
identifies the various sub-configurations.
- Redesigned the global kernel structure (gks) around a 2D array of cntx_t
structures (pointers to cntx_t, actually). The first dimension is indexed
over arch_t and the inner dimension is the ind_t (induced method) for
each microarchitecture. When a microarchitecture (configuration) is
"registered" at init-time, the inner array for that configuration in the
2D array is initialized (and allocated, if it hasn't been already). The
cntx_t slot for BLIS_NAT is initialized immediately and those for other
induced method types are initialized and cached on-demand, as needed. At
cntx_t registration, we also store function pointers to cntx_init functions
that will initialize (a) "reference" contexts and (b) contexts for use with
induced methods. We don't cache the full contexts for reference contexts
since they are rarely needed. The functions that initialize these two kinds
of contexts are generated automatically for each targeted sub-configuration
from cpp-templatized code at compile-time. Induced method contexts that
need "stage" adjustments can still obtain them via functions in
bli_cntx_ind_stage.c.
- Added new functions and functionality to bli_cntx.c, such as for setting
the level-1f, level-1v, and packm kernels, and for converting a native
context into one for executing an induced method.
- Moved the checking of register/cache blocksize consistency from being cpp
macros in bli_kernel_macro_defs.h to being runtime checks defined in
bli_check.c and called from bli_gks_register_cntx() at the time that the
global kernel structure's internal context is initialized for a given
microarchitecture/configuration.
- Deprecated all of the old per-operation bli_*_cntx.c files and removed
the previous operation-level cntx_t_init()/_finalize() invocations.
Instead, we now query the gks for a suitable context, usually via
bli_gks_query_cntx().
- Deprecated support for the 3m2 and 3m3 induced methods. (They required
hackery that I was no longer willing to support.)
- Consolidated the 1e and 1r packm kernels for any given register blocksize
into a single kernel that will branch on the schema and support packing
to both formats.
- Added the cntx_t* argument to all packm kernel signatures.
- Deprecated the local function pointer array in all bli_packm_cxk*.c files
and instead obtain the packm kernel from the cntx_t.
- Added bli_calloc_intl(), which serves as the calloc-equivalent to to
bli_malloc_intl(). Useful when we wish to allocate and initialize to
zero/NULL.
- Converted existing cpp macro functions defined in bli_blksz.h, bli_func.h,
bli_cntx.h into static functions.
Details:
- Fixed a bug in bli_l3_packm() that caused cntl_t-cached packed mem_t
entries to be released and then re-acquired unnecessarily. (In essence,
the "<" operands in the conditional that guards the
release-and-reacquire code block simply needed to be swapped.) The bug
should have only affected performance (rather than the computed result).
Thanks to Minh Quan for identifying and reporting the bug.
Details:
- Removed the family field inside the cntx_t struct and re-added it to the
cntl_t struct. Updated all accessor functions/macros accordingly, as well
as all consumers and intermediaries of the family parameter (such as
bli_l3_thread_decorator(), bli_l3_direct(), and bli_l3_prune_*()). This
change was motivated by the desire to keep the context limited, as much
as possible, to information about the computing environment. (The family
field, by contrast, is a descriptor about the operation being executed.)
- Added additional functions to bli_blksz_*() API.
- Added additional functions to bli_cntx_*() API.
- Minor updates to bli_func.c, bli_mbool.c.
- Removed 'obj' from bli_blksz_*() API names.
- Removed 'obj' from bli_cntx_*() API names.
- Removed 'obj' from bli_cntl_*(), bli_*_cntl_*() API names. Renamed routines
that operate only on a single struct to contain the "_node" suffix to
differentiate with those routines that operate on the entire tree.
- Added enums for packm and unpackm kernels to bli_type_defs.h.
- Removed BLIS_1F and BLIS_VF from bszid_t definition in bli_type_defs.h.
They weren't being used and probably never will be.
Details:
- Renamed all level-3 induced method files to use the "_vir.c" suffix
instead of "_ref.c". Also renamed functions within these files
accordingly.
- Renamed cpp macro definitions in frame/ind/include according to the
above changes.
- Removed frame/3/old.
Details:
- Defined bli_gemmbp_cntl_create(), bli_gemmpb_cntl_create(), with the
body of bli_gemm_cntl_create() replaced with a call to the former.
- Defined bli_cntl_free_w_thrinfo(), bli_cntl_free_wo_thrinfo(). Now,
bli_cntl_free() can check if the thread parameter is NULL, and if so,
call the latter, and otherwise call the former.
- Defined bli_gemm1mbp_cntx_init(), bli_gemm1mpb_cntx_init(), both in
terms of bli_gemm1mxx_cntx_init(), which behaves the same as
bli_gemm1m_cntx_init() did before, except that an extra bool parameter
(is_pb) is used to support both bp and pb algorithms (including to
support the anti-preference field described below).
- Added support for "anti-preference" in context. The anti_pref field,
when true, will toggle the boolean return value of routines such as
bli_cntx_l3_ukr_eff_prefers_storage_of(), which has the net effect of
causing BLIS to transpose the operation to achieve disagreement (rather
than agreement) between the storage of C and the micro-kernel output
preference. This disagreement is needed for panel-block implementations,
since they induce a transposition of the suboperation immediately before
the macro-kernel is called, which changes the apparent storage of C. For
now, anti-preference is used only with the pb algorithm for 1m (and not
with any other non-1m implementation).
- Defined new functions,
bli_cntx_l3_ukr_eff_prefers_storage_of()
bli_cntx_l3_ukr_eff_dislikes_storage_of()
bli_cntx_l3_nat_ukr_eff_prefers_storage_of()
bli_cntx_l3_nat_ukr_eff_dislikes_storage_of()
which are identical to their non-"eff" (effectively) counterparts except
that they take the anti-preference field of the context into account.
- Explicitly initialize the anti-pref field to FALSE in
bli_gks_cntx_set_l3_nat_ukr_prefs().
- Added bli_gemm_ker_var1.c, which implements a panel-block macro-kernel
in terms of the existing block-panel macro-kernel _ker_var2(). This
technique requires inducing transposes on all operands and swapping
the A and B.
- Changed bli_obj_induce_trans() macro so that pack-related fields are
also changed to reflect the induced transposition.
- Added a temporary hack to bli_l3_3m4m1m_oapi.c that allows us to easily
specify the 1m algorithm (block-panel or panel-block).
- Renamed the following cntx_t-related macros:
bli_cntx_get_pack_schema_a() -> bli_cntx_get_pack_schema_a_block()
bli_cntx_get_pack_schema_b() -> bli_cntx_get_pack_schema_b_panel()
bli_cntx_get_pack_schema_c() -> bli_cntx_get_pack_schema_c_panel()
and updated all instantiations. Also updated the field names in the
cntx_t struct.
- Comment updates.
Details:
- Implemented the 1m method for inducing complex domain matrix
multiplication. 1m support has been added to all level-3 operations,
including trsm, and is now the default induced method when native
complex domain gemm microkernels are omitted from the configuration.
- Updated _cntx_init() operations to take a datatype parameter. This was
needed for the corresponding function for 1m (because 1m requires us
to choose between column-oriented or row-oriented execution, which
requires us to query the context for the storage preference of the
gemm microkernel, which requires knowing the datatype) but I decided
that it made sense for consistency to add the parameter to all other
cntx initialization functions as well, even though those functions
don't use the parameter.
- Updated bli_cntx_set_blkszs() and bli_gks_cntx_set_blkszs() to take
a second scalar for each blocksize entry. The semantic meaning of the
two scalars now is that the first will scale the default blocksize
while the second will scale the maximum blocksize. This allows scaling
the two independently, and was needed to support 1m, which requires
scaling for a register blocksize but not the register storage
blocksize (ie: "packdim") analogue.
- Deprecated bli_blksz_reduce_dt_to() and defined two new functions,
bli_blksz_reduce_def_to() and bli_blksz_reduce_max_to(), for reducing
default and maximum blocksizes to some desired blocksize multiple.
These functions are needed in the updated definitions of
bli_cntx_set_blkszs() and bli_gks_cntx_set_blkszs().
- Added support for the 1e and 1r packing schemas to packm, including
1e/1r packing kernels.
- Added a minor optimization to bli_gemm_ker_var2() that allows, under
certain circumstances (specifically, real domain beta and row- or
column-stored matrix C), the real domain macrokernel and microkernel
to be called directly, rather than using the virtual microkernel
via the complex domain macrokernel, which carries a slight additional
amount of overhead.
- Added 1m support to the testsuite.
- Added 1m support to Makefile and runme.sh in test/3m4m. Also simplified
some code in test_gemm.c driver.
Details:
- Disabled the implementation of trsm_r that allows the right-hand matrix
B to be trianglar, and switched to the implementation that simply
transposes the operation (and thus the storage of C) in order to recast
the operation as trsm_l. This avoids the need to use trsm_rl and trsm_ru
macrokernels, which require an awkward swapping of MR and NR. For now,
the support for trsm_r macrokernels, via separate control trees, remains.
- Modified bli_config_macro_defs.h so that BLIS_RELAX_MCNR_NCMR_CONSTRAINTS
is defined by default. This is mostly a safety precaution in case someone
tries to switch back to the previous trsm_r implementation, but also
serves as a convenience on some systems where one does not naturally
choose blocksizes in a way that satisfies MC % NR = 0 and NC % MR = 0.
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:
- Changed a cpp macro that was meant to prevent using certain trsm_r code
if BLIS_RELAX_MCNR_NCMR_CONSTRAINTS was defined. It was actually coded
incorrectly at first. I've now fixed its location and changed its
consequence to a compile-time #error message.
- Number of threads is determined by BLIS_NUM_THREADS or OMP_NUM_THREADS, but can be overridden by BLIS_XX_NT as before.
- Threads are assigned to loops (ic, jc, ir, and jc) automatically by weighted partitioning and heuristics, both of which are tunable via bli_kernel.h.
- All level-3 BLAS covered.
Details:
- Added cpp guards around the constraints in bli_kernel_macro_defs.h
that enforce MC % NR = 0 and NC % MR = 0. These constraints are ONLY
needed when handling right-side trsm by allowing the matrix on the
right (matrix B) to be triangular, because it involves swapping
register, but not cache, blocksizes (packing A by NR and B by MR)
and then swapping the operands to gemmtrsm just before that kernel
is called. It may be useful to disable these constraints if, for
example, the developer wishes to test the configuration with
a different set of cache blocksizes where only MC % MR = 0 and
NC % NR = 0 are enforced.
- In summary, #defining BLIS_RELAX_MCNR_NCMR_CONSTRAINTS will bypass
the enforcement of MC % NR = 0 and NC % MR = 0.
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:
- Added cpp-guarded code to bli_thrcomm_openmp.c that allows a curious
developer to print the contents of the thrinfo_t structures of each
thread, for verification purposes or just to study the way thread
information and communicators are used in BLIS.
- Enabled some previously-disabled code in bli_l3_thrinfo.c for freeing
an array of thrinfo_t* values that is used in the new, cpp-guarde code
mentioned above.
- Removed some old commented lines from bli_gemm_front.c.
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:
- Removed thread barriers from the end of the loop bodies of
bli_gemm_blk_var1(), bli_gemm_blk_var2(), bli_trsm_blk_var1(),
and bli_trsm_blk_var2().
- Moved the thread barrier at the end of bli_packm_int() to the
end of bli_l3_packm(), and added missing barriers to that function.
- Removed the no longer necessary (and now incorrect) ochief guard
in bli_gemm3m3_packa() on the bli_obj_scalar_reset() on C.
- Thanks to Tyler Smith for help with these changes.
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.
