Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Refactor the control tree and other infrastructure (#710)
Details:
1. A "plugin" architecture.
- Users are now able to register new kernels, kernel preferences, and
blocksizes at runtime, directly from user applications.
- Plugins can be created, configured, and built using only an installed
version of BLIS -- no source or source code changes required.
- Plugins support both reference and optimized kernels, as well as
custom configuration-to-kernel-set mappings.
- Building plugins (including reference and relevant optimized kernels)
for enabled architectures or architecture families is automated, as is
linking into the final library.
- The configure script is now installed as 'configure-plugin'. In this
mode, it can be used to initialize a plugin from a template including
optional example code, and prepare a build system for compiling the
plugin into a shared or static library.
- Additional configuration files, templates, and build system components
are also installed to '%prefix%/share/blis'.
- The cntx_t struct now has extensible data structures for holding
kernels, preferences, and blocksizes. These are based on a "stack"
structure which contains a list of fixed-size data blocks. Adding a
new entry (which may require allocating a new block or reallocating
the block pointer array) requires locking, but looking up entries is
lock-free and takes O(1) time.
- Kernels can depend on either 1 or 2 type parameters (e.g.
mixed-precision packing requires 2). The func2_t struct supports
the latter, but can be implicitly cast to func_t if only "diagonal"
entries are needed. The number of type parameters can be inferred from
the kernel ID for type safety.
- Functions have been added to register new kernels, preferences, and
blocksizes with the global kernel structure (gks). This creates
corresponding entries in each allocated context and returns the next
available ID. Plugins use this API to register user kernels, although
the user is responsible for tracking the returned IDs for later
lookup. Setting newly-registered reference kernels, as well as
overriding these with optimized kernels is done in exactly the same
manner as in bli_cntx_init_ref() and bli_cntx_init_<subconfig>().
2. Restructuring of the control and thread control trees.
- The control tree has been substantially restructured to support more
flexibility.
- The "default" control trees for gemm (also used for
hemm/symm/herk/her2k/syrk/syr2k/trmm/trmm3) and trsm are now
represented as a single structure containing all necessary control
tree nodes and parameters.
- An API has been added to modify the default gemm/trsm control trees.
- This same API is used by the framework and packm/gemm/trsm variants
to access specific control tree nodes.
- Users can alternatively create a custom control tree from scratch.
- The blocksizes are now encoded directly in the control tree, rather
than via loop IDs. The logic for adjusting blocksizes for certain
operations has been moved to the control tree initialization.
- Type information is encoded in the control tree to drive proper
selection of packing and computational kernels provided by the user.
- The packing microkernel now receives an opaque "params" struct which
is user-definable and can be used to pass additional information
through the call stack.
- The auxinfo_t struct has been updated with a .params field for
opaque user data as well as the global offsets of the current
microtile.
- The packm and gemm variants can be overridden by the user, and also
receive an opaque params struct via the associated control tree
node.
- The structure-aware packing kernel bli_packm_struc_cxk() is no longer
hard-coded to be called from the default packm variant, but can be
overridden by the user. It also supports mixed-precision/mixed-domain
natively now.
- The thread control tree (thrinfo_t) is now created entirely up-front
by inspecting the control tree. The required number of threads at each
level is encoded in the control tree via loop IDs (actually a bitfield
of loop IDs), although the ordering and number of such IDs is
arbitrary. The logic for adjusting the number of threads at each level
based on operation type (e.g. trmm) is now in the control tree
initialization and expressed by combining loop IDs from multiple
levels into a single level.
- The mem_t object containing the pack buffer pointer has been moved
from the control tree to the thread control tree. NOTE: **The control
tree is now strictly const throughout the operation, and only a
single copy is shared by all threads.**
- The thread control tree node for packing has been changed so that
there is no longer a "fake" node indicating a team of single threads.
Instead, the number of threads and thread IDs in the "normal" thread
control tree node are used. This change has also been made to the
gemmsup thread control tree and packing variants, as well as to the
gemmlike sandbox.
- Parameters controlling packing (e.g. inversion of the diagonal,
direction, schema) are not stored directly in the control tree but in
the opaque params struct. The packing control tree node and its
default params struct are stored together in the "combined"
gemm/trsm control tree structure and initialized as a unit. Users can
update these parameters individually or substitute a custom packm
variant and params struct.
- The "target" and "execution" datatypes has been removed from the obj_t
struct and replaced by type information in the control tree.
- The "sub-node" and "sub-prenode" of a control tree node have been
replaced by an arbitrary number of sub-nodes accessed by index. There
is a hard cap on the number of sub-nodes (currently 2). Sub-nodes are
added during control tree initialization, *after*
creation/initialization of the parent node through an updated API.
- The level-3 thread decorator has been significantly simplified and
directly calls bli_l3_int(). The control tree is created externally,
and it is no longer necessary to alias matrices or set object pack
schemas. Also, the rntm_t passed in may be NULL. Finally, family
and scalar information is no longer needed here.
- bli_l3_int() is now a simple inline function which extracts the next
control tree node and variant and calls it.
- bli_*_front() have been removed and inlined into the expert object
API with significant simplification.
- 1m (or other induced method) no longer uses an alternative cntx_t.
- The .pack_fn/.ker_fn pointers and associated params fields on the
obj_t were removed in favor of the present solution.
3. Overhaul of variable substitution in configure script.
- The configure script has been somewhat re-written to use a
centralized mechanism for substituting variables into build system and
other configuration files.
- All substitution variables go through the same pathway now, which
necessitated some variable naming changes for variables which were
named the same in e.g. Makefile and bli_config.h but with
different definitions.
- CC and CXX variables can now contain spaces, e.g. 'g++ -std=c++17'.
This provides better support for integration with build tooling such
as autotools.
4. Overhaul of packing kernels.
- Previously there were two packing kernels referenced in the cntx_t
structure for MRxk and NRxk shaped micropanels, respectively. These
have now been merged into one kernel which is responsible for packing
any dense rectangular portion of either A or B.
- The packing kernel now receives information about the register
blocksize (cdim_max) and duplication factor (the "broadcast-B"
format, although this can also apply to the A matrix).
- The structure-aware packing kernel (bli_packm_struc_cxk(), which is
now user-overridable) also receives global offsets of the current
micropanel within A or B.
- Explicit kernels for packing the diagonal blocks of
triangular/symmetric/Hermitian matrices have been added to the
cntx_t. This means that the bli_packm_struc_ckx() "kernel" no longer
needs to directly touch data (except to zero out some regions).
- bli_packm_struc_cxk() has also been updated to work only in terms of
fundamental elements (i.e., real datatypes) when computing offsets and
when zeroing data, which greatly simplifies mixed-domain/1m packing.
- bli_packm_scalar() has been updated to better support complex scalars
in mixed-domain operations.
- Pack schemas for PACKED_ROW_PANELS* and PACKED_COL_PANELS* have
been merged into simply PACKED_PANELS*. This reflects the merging of
the packing kernels into a single generic kernel. There were only a
very few places which needed the row/column information and this is
now supplied by alternative means.
- Packing variants always behave "as if" the A matrix were being packed
(i.e. the code assumes packing column-stored row panels). Packing of B
is handled by applying an implicit or explicit transpose before
packing. This change also applies to gemmsup.
5. Improved MD/MP support.
- All level-3 operations (except trsm) now support full
mixed-domain/mixed-precision operation.
- Explicit 1m packing kernels have been added in the cntx_t.
- An explicit 1m microkernel wrapper has been added to the cntx_t.
- An extra packing kernel for the "ro" format has been added, along with
the pack_t enumeration value. This supports the packing for
real*complex -> real, including potential scaling by a complex alpha,
support for structured matrices, etc.
- Extra microkernel wrappers for mixed-domain operations have been added
to support the 'ccr' (and by extension, 'crc'), 'rcc', and 'crr'
cases. Notably this includes full support for general stride storage
and complex alpha/beta.
- Packing kernels and gemm microkernels are now "templated" based on two
type parameters rather than one. For packing this allows direct
optimization of mixed-precision kernels, and for gemm microkernels
this allows direct optimization of mixed-precision without writing to
a temporary buffer. Reference packing kernels are directly
instantiated for all mixes of precisions, while by default
mixed-precision gemm microkernels are supported via a microkernel
wrapper. The "old" way of specifying optimized kernels using a single
type parameter works unchanged.
- alpha and beta are typecast appropriately to the computational or
output datatype, respectively, and **always** to the complex domain.
Scalar typecasting has also been added to gemmsup for safety.
- The gemm macrokernel doesn't have to do any typecasting anymore, as a
microkernel wrapper or optimized mixed-precision/mixed-domain kernel
now handles this.
- 1m and mixed-domain operations now always use a microkernel wrapper,
rather than adjusting parameters in the gemm macrokernel.
- The gemmt macrokernel **does** still have to handle explicit
write-back of microtiles which intersect the diagonal, although
typecasting has already been performed.
- The gemmt_x_ker_var2(), trmm_xx_ker_var2(), and trsm_xx_ker_var2()
functions have been removed. The appropriate macrokernel pointer is
selected during control tree initialization.
- Real domain MR/NR are checked for even-ness based on the gemm
microkernel's row preference in order to guarantee proper 1m and
mixed-domain operation.
- Full range of mixed-domain/mixed-precision functionality tested in the
testsuite ('input.*.mixed').
6. Other changes:
- The build system has been updated to support C++ source files
throughout the framework. While the intent is not to add such files to
BLIS itself, this supports plugins written in C++.
- Many instances of configuration-specific code have been simplified by
introducing an INSERT_GENTCONF macro which instantiates a block of
code for each enabled sub-configuration. The ConfigurationHowTo.md
document has been updated accordingly.
- PASTEMAC?/PASTECH?/PASTEF77? have been removed in favor of
variadic macros which accept any number of arguments (up to a
reasonable limit).
- The INSERT_GENTFUNC* macros have been updated to clean up
mixed-precision and mixed-domain instantiations.
- bli_align_dim_to_mult() has been updated to support rounding either up
or down based on a flag.
- Checking for empty matrices and other early exits (level-3 only) has
been consolidated into a single utility function.
- The auxinfo_t struct is always passed as const.
- The new function bli_obj_alias_submatrix() aliases a matrix while also
resetting the root to NULL, offsets to zero (while adjusting the
buffer), and applying any implicit transpose.
- Level-3 pruning functions now only check matrix structure to see what
to do, not the operation family.
- gemmsup packing has been updated to use the "normal" pack buffer
allocation routines.
- Remove duplicate checks for early return from gemmsup handler.
- bli_determine_blocksize() has been significantly simplified.
- Partitioning packed panels is no longer allowed.
- Added bli_xxsame macros.
- Automated the calculation of info bit shifts and masks based on
predefined bit sizes for various flags. This greatly simplifies
reordering, adding, or removing flags from the info/info2 bitfields.
- Moved more BLIS_NUM_* macros into the corresponding enums as the
last entry so that the value is automatically computed.
- Better const-correctness in some level0 scalar macros.
- Better mixed-precision support in some level0 scalar macros.
- Added a bli_axpbys_mxn() macro.
- bli_thread_range_sub() takes explicit thread ID and number of threads
rather than a thrinfo_t node.
- "De-templated" BLIS gemmlike sandbox (specifically, bls_gemm_bp_var1()
and bls_packm_var1()).
- Combined bls_l3_packm_[ab]() into one function with thin wrappers.
- Deleted bls_packm_var[23]().
- Add a "termination tag" to the testsuite output so that
'make check-blis' can accurately check for successful completion.
- Add a new function to centrally compute FLOPs for level-3 operations
in the testsuite.- Loading branch information
1 parent
a316d2c
commit a49238e