BLIS 1.0

This release contains several new features and optimizations related to threaded execution, as well as internal changes that improve maintainability and lay the groundwork for future refactoring. The build system and kernel sets saw lots of new code and tweaks to old code, and of course there were many bugfixes.

Improvements present in 1.0:

Framework:

• Initialize/finalize BLIS via a new bli_pthread_switch_t API. (Field Van Zee, Devin Matthews)
• Revamped bli_init() to use TLS where feasible. (Field Van Zee, Edward Smyth, Minh Quan Ho)
• Implemented support for fat multithreading.
• Implemented tile-level load balancing (tlb), or tile-level partitioning, in jr/ir loops for gemm, gemmt, and trmm macrokernels. (Field Van Zee, Devin Matthews, Leick Robinson, Minh Quan Ho)
• Added padding to thrcomm_t fields to avoid false sharing of cache lines. (Leick Robinson)
• Rewrote/fixed broken tree barrier implementation. (Leick Robinson)
• Refactored some rntm_t management code. (Field Van Zee, Devin Matthews)
• Initialize rntm_t nt/ways fields with 1 (not -1). (Field Van Zee, Jeff Diamond, Leick Robinson, Devin Matthews)
• Defined invscalv, invscalm, invscald operations.
• Added consistent NaN/Inf handling in sumsqv. (Devin Matthews)
• Implemented support for HPX as a threading backend option. (Christopher Taylor, Srinivas Yadav)
• Relocated the pba, sba pool (from the rntm_t), and mem_t (from the cntl_t) to the thrinfo_t object.
• Modified which communicator is associated with a given node of the thrinfo_t tree. (Devin Matthews)
• Refactored level-3 thread decorator into two parts: a thread launcher and a function to pass operands. (Devin Matthews)
• Refactored structure awareness in bli_packm_blk_var1.c. (Devin Matthews)
• Reimplemented bli_l3_determine_kc(). (Devin Matthews)
• Implemented cntx_t pointer caching in gks. (Field Van Zee, Harihara Sudhan S)
• Added const keyword to pointers in kernel APIs. (Field Van Zee, Nisanth M P)
• Migrated all kernel APIs to use void* pointers.
• Defined new global scalar constants: BLIS_ONE_I, BLIS_MINUS_ONE_I, BLIS_NAN. (Devin Matthews)
• Disabled modification of KC in the gemmsup kernels. (Devin Matthews)
• Defined lt, lte, gt, gte operations and other miscellaneous updates.
• Consolidated INSERT_ macro sets via variadic macros. (Devin Matthews)
• De-templatized macrokernels for gemmt, trmm, and trsm to match that of gemm. (Devin Matthews)
• De-templatized bli_l3_sup_var1n2m.c and unified _sup_packm_a/b(). (Devin Matthews)
• Fixed 1m enablement for herk/her2k/syrk/syr2k. (Devin Matthews)
• Fixed trmm[3]/trsm performance bug introduced in cf7d616. (Field Van Zee, Leick Robinson)
• Fixed a 1m optimization bug in right-sided hemm/symm. (Field Van Zee, Nisanth M P)
• Fixed a bug in sup threshold registration. (Devin Matthews, Field Van Zee)
• Fixed brokenness in the small block allocator (sba) when the sba is disabled. (Field Van Zee, John Mather)
• Fixed type bug in bli_cntx_set_ukr_prefs(). (Field Van Zee, Leick Robinson, Devin Matthews, Jeff Diamond)
• Fixed incorrect sizeof(type) in edge case macros. (@moon-chilled)
• Fixed bugs and added sanity check in bli_pool.c. (Devin Matthews)
• Fixed a typo in the macro definition for VEXTRACTF64X2 in bli_x86_asm_macros.h. (Harsh Dave)
• Fixed a typo in bli_type_defs.h where BLIS_BLAS_INT_TYPE_SIZE was misspelled. (Devin Matthews)
• Typecast printf() args in bli_thread_range_tlb.c to avoid compiler warnings. (Lee Killough)
• Minor tweaks to bli_l3_check.c.
• Partial addition of const to all interfaces above the (micro)kernels. (Devin Matthews)
• Fixed a harmless misspelling of xpbys in gemm macrokernel.
• Various internal API renaming/reorganization.
• Various other fixes.

Compatibility:

• Implemented [cz]symv_(), [cz]syr_(), [cz]rot_(). (Field Van Zee, James Foster)
• Fixed compilation errors when BLIS_DISABLE_BLAS_DEFS is defined. (Field Van Zee, Edward Smyth, Devin Matthews)
• Include bli_config.h before bli_system.h in cblas.h so that BLIS_ENABLE_SYSTEM is defined in time for proper OS detection. (Edward Smyth)

Kernels:

• Updated ARMv8a kernels to fix two prefetching issues and re-enable general stride IO. (Jeff Diamond)
• Restored general storage case to armsve kernels. (RuQing Xu)
• Added arm64 dgemmsup with extended MR and NR. (RuQing Xu)
• Reorganized the way packm kernels are stored within the cntx_t so that BLIS only stores two packm kernels per datatype: one for MRxk upanels and one for kxNR upanels. (Devin Matthews)
• Fixed bugs in scal2v reference kernel when alpha == 1.
• Fixed out-of-bounds read in haswell gemmsup kernels. (Daniël de Kok, Bhaskar Nallani, Madeesh Kannan)
• Fixed k = 0 edge case in power10 microkernels. (Nisanth M P)
• Disabled power10 kernels other than sgemm, dgemm. (Nisanth M P)
• Fixed bli_gemm_small() prototype mismatch. (Jeff Diamond)

Extras:

• Use the conventional level-3 sup thread decorator within the gemmlike sandbox.
• Fixed type-mismatch errors in power10 sandbox. (Nisanth M P)
• Fixed gemmlike sandbox bug that stems from reuse of bli_thrinfo_sup_grow().

Build system:

• Added two arm64 subconfigs: altra and altramax. (Jeff Diamond, Leick Robinson)
• Added support for RISC-V configuration targets. (Angelika Schwarz, Lee Killough)
• Auto-detect the RISC-V ABI of the compiler and use -mabi= during RISC-V builds. (Lee Killough)
• Added sifive_x280 subconfig and kernel set. (Aaron Hutchinson, Lee Killough, Devin Matthews, and Angelika Schwarz)
• Added AddressSanitizer (--enable-asan) option to configure. (Devin Matthews)
• Added option to disable thread-local storage via --disable-tls. (Field Van Zee, Nick Knight)
• Exclude -lrt on Android with Bionic libraries. (Lee Killough)
• Omit -fPIC option when shared library build is disabled. (Field Van Zee, Nick Knight)
• Move -fPIC option insertion to subconfigs' make_defs.mk files. (Field Van Zee, Nick Knight)
• Install one-line helper headers to INCDIR prefix so that user can #include "blis.h" instead of #include <blis/blis.h> and/or "cblas.h" instead of <blis/cblas.h> if CBLAS is enabled). (Field Van Zee, Jed Brown, Devin Matthews, Mo Zhou)
• Enhanced detection of Fortran compiler when checking the version string for the purposes of determining a default return convention for complex domain values. (Bart Oldeman)
• Added detection of the NVIDIA nvhpc compiler (nvc) in configure. (Ajay Panyala)
• Updated zen3 subconfig to support NVHPC compilers. (Abhishek Bagusetty)
• Use kernel CFLAGS for kernels subdirs in addons. (AMD, Mithun Mohan)
• Created power umbrella configuration family (which currently includes power9 and power10 subconfigs). (Nisanth M P)
• Defined BLIS_VERSION_STRING in blis.h instead of via command line argument during compilation. (Field Van Zee, Mohsen Aznaveh, Tim Davis)
• Rewrote regen-symbols.sh as gen-libblis-symbols.sh. (Field Van Zee)
• Support clang targetting MinGW. (Isuru Fernando)
• Added autodetection (via /proc/cpuinfo) for POWER7, POWER9 and POWER10 microarchitectures. (Alexander Grund)
• Added #line directives to flattened blis.h to facilitate easier debugging. (Devin Matthews)
• Added --nosup and --sup shorthand options to configure.
• Use here-document syntax for configure --help output. (Lee Killough)
• Updated configure to pass all shellcheck checks. (Lee Killough)
• Tweaks to .dir-locals.el to enhance emacs formatting of C files. (Lee Killough)
• Removed buggy cruft from power10 subconfig. (Field Van Zee, Nicholai Tukanov)
• Added missing #include <io.h> for Windows. (@h-vetinari)
• Fixed hardware auto-detection for firestorm (Apple M1) subconfig. (Devin Matthews)
• Fixed bug in detection of Fortran compiler vendor. (Devin Matthews)
• Fixed version check for znver3, which needs gcc >= 10.3. (Jed Brown)
• Fixed typo in configure --help text. (Lee Killough)
• Fixed warning about regular expressions with stray backslashes as the result of recent changes to grep.
• Added output.testsuite to .gitignore.
• Minor changes to .gitignore and LICENSE files. (Jeff Diamond)
• Minor decluttering of top-level directory.
• Very minor tweaks to common.mk.

Testing:

• Rewrote test/3 drivers to take parameters via command line arguments. (Field Van Zee, Jeff Diamond, Leick Robinson)
• Added arm64 entry to .travis.yml so that Travis CI will compile/test ARM builds. (Field Van Zee, RuQing Xu)
• Test the gemmlike sandbox via AppVeyor. (Jeff Diamond)
• Added -q quiet mode option to testsuite.
• Fixed non-deterministic segfault in standalone test/3 drivers. (Field Van Zee, Leick Robinson)
• Fixed a crash that occurs when either cblat1 or zblat1 are linked with a build of BLIS that was compiled with --complex-return=intel. (Bart Oldeman)
• Other minor fixes/tweaks.

Documentation:

• Added Discord documentation (docs/Discord.md) and logo to README.md.
• Added the mm_algorithm files (for bp and pb) to docs/diagrams.
• Added mention of Wilkinson Prize to README.md.
• Minor fixes and improvements to docs/Multithreading.md.
• Fix typos in docs + example code comments. (Igor Zhuravlov)

BLIS 0.9.0

This release contains a slew of improvements, new kernels and APIs, bugfixes, and more (including lots of code reduction). It also contains foundational support for an exciting new class of expert functionality: creating new operations without the need to duplicate the middleware that sits between the API and kernels.

Improvements present in 0.9.0:

Framework:

• Added various fields to obj_t that relate to storing function pointers to custom packm kernels, microkernels, etc as well as accessor functions to set and query those fields. (Devin Matthews)
• Enabled user-customized packm microkernels and variants via the aforementioned new obj_t fields. (Devin Matthews)
• Moved edge-case handling out of the macrokernel and into the gemm and gemmtrsm microkernels. This also required updating of APIs and definitions of all existing microkernels in kernels directory. Edge-case handling functionality is now facilitated via new preprocessor macros found in bli_edge_case_macro_defs.h. (Devin Matthews)
• Avoid gemmsup thread barriers when not packing A or B. This boosts performance for many small multithreaded problems. (Field Van Zee, AMD)
• Allow the 1m method to operate normally when single and double real-domain microkernels mix row and column I/O preference. (Field Van Zee, Devin Matthews, RuQing Xu)
• Removed support for execution of complex-domain level-3 operations via the 3m and 4m methods.
• Refactored herk, her2k, syrk, syr2k in terms of gemmt. (Devin Matthews)
• Defined setijv and getijv to set/get vector elements.
• Defined eqsc, eqv, and eqm operations to test equality between two scalars, vectors, or matrices.
• Added new bounds checking to setijm and getijm to prevent use of negative indices.
• Renamed membrk files/variables/functions to pba.
• Store error-checking level as a thread-local variable. (Devin Matthews)
• Add err_t* "return" parameter to bli_malloc_*() and friends.
• Switched internal mutexes of the sba and pba to static initialization.
• Changed return value method of bli_pack_get_pack_a(), bli_pack_get_pack_b().
• Fixed a bug that allows bli_init() to be called more than once (without segfaulting). (@lschork2, Minh Quan Ho, Devin Matthews)
• Removed a sanity check in bli_pool_finalize() that prevented BLIS from being re-initialized. (AMD)
• Fixed insufficient pool_t-growing logic in bli_pool.c, and always allocate at least one element in .block_ptrs array. (Minh Quan Ho)
• Cleanups related to the error message array in bli_error.c. (Minh Quan Ho)
• Moved language-related definitions from bli_macro_defs.h to a new header, bli_lang_defs.h.
• Renamed BLIS_SIMD_NUM_REGISTERS to BLIS_SIMD_MAX_NUM_REGISTERS and BLIS_SIMD_SIZE to BLIS_SIMD_MAX_SIZE for improved clarity. (Devin Matthews)
• Many minor bugfixes.
• Many cleanups, including removal of old and commented-out code.

Compatibility:

• Expanded BLAS layer to include support for ?axpby_() and ?gemm_batch_(). (Meghana Vankadari, AMD)
• Added gemm3m APIs to BLAS and CBLAS layers. (Bhaskar Nallani, AMD)
• Handle ?gemm_() invocations where m or n is unit by calling ?gemv_(). (Dipal M Zambare, AMD)
• Removed option to finalize BLIS after every BLAS call.
• Updated default definitions of bli_slamch() and bli_dlamch() to use constants from standard C library rather than values computed at runtime. (Devin Matthews)

Kernels:

• Added 512-bit SVE-based a64fx subconfiguration that uses empirically-tuned blocksizes (Stepan Nassyr, RuQing Xu)
• Added a vector-length agnostic armsve subconfig that computes blocksizes via an analytical model. (Stepan Nassyr)
• Added vector-length agnostic d/s/sh gemm kernels for Arm SVE. (Stepan Nassyr)
• Added gemmsup kernels to the armv8a kernel set for use in new Apple Firestorm subconfiguration. (RuQing Xu)
• Added 512-bit SVE dpackm kernels (16xk and 10xk) with in-register transpose. (RuQing Xu)
• Extended 256-bit SVE dpackm kernels by Linaro Ltd. to 512-bit for size 12xk. (RuQing Xu)
• Reorganized register usage in bli_gemm_armv8a_asm_d6x8.c to accommodate clang. (RuQing Xu)
• Added saxpyf/daxpyf/caxpyf kernels to zen kernel set. (Dipal M Zambare, AMD)
• Added vzeroupper instruction to haswell microkernels. (Devin Matthews)
• Added explicit beta == 0 handling in s/d armsve and armv7a gemm microkernels. (Devin Matthews)
• Added a unique tag to branch labels to accommodate clang. (Devin Matthews, Jeff Hammond)
• Fixed a copy-paste bug in the loading of kappa_i in the two assembly cpackm kernels in haswell kernel set. (Devin Matthews)
• Fixed a bug in Mx1 gemmsup haswell kernels whereby the vhaddpd instruction is used with uninitialized registers. (Devin Matthews)
• Fixed a bug in the power10 microkernel I/O. (Nicholai Tukanov)
• Many other Arm kernel updates and fixes. (RuQing Xu)

Extras:

• Added support for addons, which are similar to sandboxes but do not require the user to implement any particular operation.
• Added a new gemmlike sandbox to allow rapid prototyping of gemm-like operations.
• Various updates and improvements to the power10 sandbox, including a new testsuite. (Nicholai Tukanov)

Build system:

• Added explicit support for AMD's Zen3 microarchitecture. (Dipal M Zambare, AMD, Field Van Zee)
• Added runtime microarchitecture detection for Arm. (Dave Love, RuQing Xu, Devin Matthews)
• Added a new configure option --[en|dis]able-amd-frame-tweaks that allows BLIS to compile certain framework files (each with the _amd suffix) that have been customized by AMD for improved performance (provided that the targeted configuration is eligible). By default, the more portable counterparts to these files are compiled. (Field Van Zee, AMD)
• Added an explicit compiler predicate (is_win) for Windows in configure. (Devin Matthews)
• Use -march=haswell instead of -march=skylake-avx512 on Windows. (Devin Matthews, @h-vetinari)
• Fixed configure breakage on MacOSX by accepting either clang or LLVM in vendor string. (Devin Matthews)
• Blacklist clang10/gcc9 and older for armsve subconfig.
• Added a configure option to control whether or not to use @rpath. (Devin Matthews)
• Added armclang detection to configure. (Devin Matthews)
• Use @path-based install name on MacOSX and use relocatable RPATH entries for testsuite binaries. (Devin Matthews)
• For environment variables CC, CXX, FC, PYTHON, AR, and RANLIB, configure will now print an error message and abort if a user specifies a specific tool and that tool is not found. (Field Van Zee, Devin Matthews)
• Added symlink to blis.pc.in for out-of-tree builds. (Andrew Wildman)
• Register optimized real-domain copyv, setv, and swapv kernels in zen subconfig. (Dipal M Zambare, AMD)
• Added Apple Firestorm (A14/M1) subconfiguration, firestorm. (RuQing Xu)
• Added armsve subconfig to arm64 configuration family. (RuQing Xu)
• Allow using clang with the thunderx2 subconfiguration. (Devin Matthews)
• Fixed a subtle substitution bug in configure. (Chengguo Sun)
• Updated top-level Makefile to reflect a dependency on the "flat" blis.h file for the BLIS and BLAS testsuite objects. (Devin Matthews)
• Mark xerbla_() as a "weak" symbol on MacOSX. (Devin Matthews)
• Fixed a long-standing bug in common.mk whereby the header path to cblas.h was omitted from the compiler flags when compiling CBLAS files within BLIS.
• Added a custom-made recursive sed script to build directory.
• Minor cleanups and fixes to configure, common.mk, and others.

Testing:

• Fixed a race condition in the testsuite when the SALT option (simulate application-level threading) is enabled. (Devin Matthews)
• Test 1m method execution during make check. (Devin Matthews)
• Test make install in Travis CI. (Devin Matthews)
• Test C++ in Travis CI to make sure blis.h is C++-compatible. (Devin Matthews)
• Disabled SDE testing of pre-Zen microarchitectures via Travis CI.
• Added Travis CI support for testing Arm SVE. (RuQing Xu)
• Updated SDE usage so that it is downloaded from a separate repository (ci-utils) in our GitHub organization. (Field Van Zee, Devin Matthews)
• Updated octave scripts in test/3 to be robust against missing datasets as well as to fixed a few minor issues.
• Added test_axpbyv.c and test_gemm_batch.c test driver files to test directory. (Meghana Vankadari, AMD)
• Support all four datatypes in her, her2, herk, and her2k drivers in test directory. (Madan mohan Manokar, AMD)

Documentation:

• Added documentation for: setijv, getijv, eqsc, eqv, eqm.
• Added docs/Addons.md.
• Added dedicated "Performance" and "Example Code" sections to README.md.
• Updated README.md.
• Updated docs/Sandboxes.md.
• Updated docs/Multithreading.md. (Devin Matthews)
• Updated docs/KernelHowTo.md.
• Updated docs/Performance.md to report Fujitsu A64fx (512-bit SVE) results. (RuQing Xu)
• Updated docs/Performance.md to report Graviton2 Neoverse N1 results. (Nicholai Tukanov)
• Updated docs/FAQ.md with new questions.
• Fixed typos in docs/FAQ.md. (Gaëtan Cassiers)
• Various other minor fixes.