2 ====================================================================
8 the newly added BFLOAT16 functions were renamed to use the
9 letter "B" instead of "H" to avoid potential confusion with
10 the IEEE "half precision float" type, i.e. the 0.3.10
11 SHGEMM is now SBGEMM and the corresponding build option
12 was changed from "BUILD_HALF" to "BUILD_BFLOAT16".
13 * Reduced the default BLAS3_MEM_ALLOC_THRESHOLD (used as an upper
14 limit for placing temporary arrays on the stack) to be compatible
15 with a stack size of 1mb (as imposed by the JAVA runtime library)
16 * Added mixed-precision dot function SBDOT and utility functions
17 shstobf16, shdtobf16, sbf16tos and dbf16tod to convert between
18 single or double precision float arrays and bfloat16 arrays
19 * Fixed prototypes of LAPACK_?ggsvp and LAPACK_?ggsvd functions
21 * Fixed underflow and rounding errors in LAPACK SLANV2 and DLANV2
22 (causing miscalculations in e.g. SHSEQR/DHSEQR, LAPACK issue #263)
23 * Fixed workspace calculation in LAPACK ?GELQ (LAPACK issue #415)
24 * Fixed several bugs in the LAPACK testsuite
25 * Improved performance of TRMM and TRSM for certain problem sizes
26 * Fixed infinite recursions and workspace miscalculations in ReLAPACK
27 * CMAKE builds no longer require pkg-config for creating the .pc file
28 * Makefile builds no longer misread NO_CBLAS=0 or NO_LAPACK=0 as
29 enabling these options
30 * Fixed detection of gfortran when invoked through an mpi wrapper
31 * Improve thread reinitialization performance with OpenMP xafter a fork
32 * Added support for building only the subset of the library required
33 for a particular precision by specifying BUILD_SINGLE, BUILD_DOUBLE
34 * Optional function name prefixes and suffixes are now correctly
35 reflected in the generated cblas.h
36 * Added CMAKE build support for the LAPACK and multithreading tests
39 * Added optimized support for POWER10
40 * Added support for compiling for POWER8 in 32bit mode
41 * Added support for compilation with LLVM/clang
42 * Added support for compilation with NVIDIA/PGI compilers
43 * Fixed building on big-endian POWER8
44 * Fixed miscompilation of ZDOTC by gcc10
45 * Fixed alignment errors in the POWER8 SAXPY kernel
46 * Improved CPU detection on AIX
47 * Supported building with older compilers on POWER9
50 * Added support for Intel Cooperlake
51 * Added autodetection of AMD Renoir/Matisse/Zen3 cpus
52 * Added autodetection of Intel Comet Lake cpus
53 * Reimplemented ?sum, ?dot and daxpy using universal intrinsics
54 * Reset the fpu state before using the fpu on Windows as a workaround
55 for a problem introduced in Windows 10 build 19041 (a.k.a. SDK 2004)
56 * Fixed potentially undefined behaviour in the dot and gemv_t kernels
57 * Fixed a potential segmentation fault in DYNAMIC_ARCH builds
58 * Fixed building for ZEN with PGI/NVIDIA and AMD AOCC compilers
61 * Fixed cpu detection on BSD-like systems
64 * Added preliminary support for Apple Vortex cpus
65 * Added support for the Cavium ThunderX3T110 cpu
66 * Fixed cpu detection on BSD-like systems
67 * Fixed compilation in -std=C18 mode
71 * Added support for compiling with the clang compiler
72 * Improved GEMM performance on Z14
74 ====================================================================
79 * Improved thread locking behaviour in blas_server and parallel getrf
80 * Imported bugfix 394 from LAPACK (spurious reference to "XERBL"
81 due to overlong lines)
82 * Imported bugfix 403 from LAPACK (compile option "recursive" required
83 for correctness with Intel and PGI)
84 * Imported bugfix 408 from LAPACK (wrong scaling in ZHEEQUB)
85 * Imported bugfix 411 from LAPACK (infinite loop in LARGV/LARTG/LARTGP)
86 * Fixed mismatches between BUFFERSIZE and GEMM_UNROLL parameters that
87 could lead to crashes at large matrix sizes
88 * Restored internal soname in dynamic libraries on FreeBSD and Dragonfly
89 * Added API (openblas_setaffinity) to set the thread affinity on Linux
90 * Added initial infrastructure for half-precision floating point
91 (bfloat16) support with a generic implementation of SHGEMM
92 * Added CMAKE build system support for building the cblas_Xgemm3m
94 * Fixed CMAKE support for building in a path with embedded spaces
95 * Fixed CMAKE (non)handling of NO_EXPRECISION and MAX_STACK_ALLOC
96 * Fixed GCC version detection in the Makefiles
97 * Allowed overriding the names of AR, AS and LD in Makefile builds
100 * Fixed big-endian POWER8 ELFv2 builds on FreeBSD
101 * Fixed GCC version checks and DYNAMIC_ARCH builds on POWER9
102 * Fixed CMAKE build support for POWER9
103 * fixed a potential race condition in the thread buffer allocation
104 * Worked around LAPACK test failures on PPC G4
107 * Fixed a potential race condition in the thread buffer allocation
108 * Added support for MIPS 24K/24KE family based on P5600 kernels
111 * fixed a potential race condition in the thread buffer allocation
112 * Added TARGET=GENERIC
115 * Fixed a race condition in the thread buffer allocation
118 * Fixed a race condition in the thread buffer allocation
119 * Fixed zero initialisation in the assembly for SGEMM and DGEMM BETA
120 * Improved performance of the ThunderX2 DAXPY kernel
121 * Added an optimized SGEMM kernel for Cortex A53
122 * Fixed Makefile support for INTERFACE64 (8-byte integer)
125 * Fixed a syntax error in the CMAKE setup for SkylakeX
126 * Improved performance of STRSM on Haswell, SkylakeX and Ryzen
127 * Improved SGEMM performance on SGEMM for workloads with ldc a
129 * Improved DGEMM performance on Skylake X
130 * Fixed unwanted AVX512-dependency of SGEMM in DYNAMIC_ARCH
131 builds created on SkylakeX
132 * Removed data alignment requirement in the SSE2 copy kernels
133 that could cause spurious crashes
134 * Added a workaround for an optimizer bug in AppleClang 11.0.3
135 * Fixed LAPACK test failures due to wrong options for Intel Fortran
136 * Fixed compilation and LAPACK test results with recent Flang
138 * Fixed DYNAMIC_ARCH builds with CMAKE on OS X
139 * Fixed missing exports of cblas_i?amin, cblas_i?min, cblas_i?max,
140 cblas_?sum, cblas_?gemm3m in the shared library on OS
141 * Fixed reporting of cpu name in DYNAMIC_ARCH builds (would sometimes
142 show the name of an older generation chip supported by the same kernels)
145 * Improved performance of SGEMM/STRMM and DGEMM/DTRMM on Z14
147 ====================================================================
152 * Fixed a miscompilation of the GETRF functions with CMAKE
153 * Imported bugfix 390 from LAPACK (missing NaN propagation in xCOMBSSQ)
154 * The size of the memory buffer used for splitting GEMM tasks across
155 multiple threads can now be configured in the build system.
158 * Fixed several compilation problems related to endianness
159 and ELF version on POWER8 and POWER9
160 * Fixed use of the absolute value IAMIN/IAMAX instead of IMIN/IMAX
161 * Fixed a race condition in the level3 blas code
164 * Fixed use of the absoltute value IAMIN/IAMAX instead of IMIN/IMAX
167 * Fixed a race condition in the level3 blas code
168 * Fixed compilation on Android
170 * Added support for Ampere EMAG8180
171 * Added support for Neoverse N1
172 * Improved performance of the blas_lock function
173 * Fixed a race condition in the level3 blas code
174 * Fixed a performance regression on TSV110-based servers
177 * Fixed a long-standing error with undeclared register overwrites
178 in the DSCAL microkernel for HASWELL,SKYLAKEX and ZEN
179 * Fixed a long-standing bug in the SSE implementation of IAMAX
180 * Fixed a CMAKE build failure with DYNAMIC_ARCH
181 * Fixed cpu autodetection of Goldmont+, Cannon Lake and Ice Lake
182 * Fixed a compilation failure on OSX with compiler name containing dash
183 * Fixed compilation with MinGW on SkylakeX
184 * Improved speed of the AVX512 GEMM3M kernel on SkylakeX
185 * Added an AVX512 STRMM kernel for SkylakeX
186 * Improved GEMM performance on Haswell and Zen
189 * fixed compilation of the DYNAMIC_ARCH code
191 ====================================================================
196 ` * LAPACK has been updated to 3.9.0 (plus patches up to
198 * CMAKE support has been improved in several areas including
200 * a thread race condition in the GEMM3M kernels was resolved
201 * the "generic" (plain C) gemm beta kernel used by many targets
203 * an optimized version of the LAPACK trtrs functions has been added
204 * an incompatibilty between the LAPACK tests and the OpenBLAS
205 implementation of XERBLA was resolved, removing the numerous
206 warnings about wrong error exits in the former
207 * support for NetBSD has been added
208 * support for compilation with g95 and non-GNU versions of ld
210 * support for compilation with (upcoming) gcc 10 has been added
213 * worked around miscompilation of several POWER8 and POWER9
214 kernels by older versions of gcc
215 * added support for big-endian POWER8 and for compilation on AIX
216 * corrected bugs in the big-endian support for PPC440 and PPC970
217 * DYNAMIC_ARCH support is now available in CMAKE builds as well
220 * performance of DGEMM_BETA and SGEMM_NCOPY has been improved
221 * compilation for 32bit works again
222 * performance of the RPCC function has been improved
223 * improved performance on small systems
224 * DYNAMIC_ARCH support is now available in CMAKE builds as well
225 * cross-compilation from OSX to IOS was simplified
228 * a new AVX512 DGEMM kernel was added and the AVX512 SGEMM kernel
229 was significantly improved
230 * optimized AVX512 kernels for CGEMM and ZGEMM have been added
231 * AVX2 kernels for STRMM, SGEMM, and CGEMM have been significantly
232 sped up and optimized CGEMM3M and ZGEMM3M kernels have been added
233 * added support for QEMU virtual cpus
234 * a compilation problem with PGI and SUN compilers was fixed
235 * Intel "Goldmont plus" is now autodetected
236 * a potential crash on program exit on MS Windows has been fixed
239 * an unwanted case sensitivity in the implementation of LSAME
240 on older 32bit AMD cpus was fixed
243 * Z15 is now supported as Z14
244 * DYNAMIC_ARCH is now available on ZARCH as well
246 ====================================================================
251 * having the gmake special variables TARGET_ARCH or TARGET_MACH
252 defined no longer causes build failures in ctest or utest
253 * defining NO_AFFINITY or USE_TLS to 0 in gmake builds no longer
254 has the same effect as setting them to 1
255 * a new test program was added to allow checking the library for
257 * a new option USE_LOCKING was added to ensure thread safety when
258 OpenBLAS itself is built without multithreading but will be
259 called from multiple threads.
260 * a build failure on Linux with glibc versions earlier than 2.5
262 * a runtime error with CPU enumeration (and NO_AFFINITY not set)
263 on glibc 2.6 was fixed
264 * NO_AFFINITY was added to the CMAKE options (and defaults to being
265 active on Linux, as in the gmake builds)
268 * the build-time logic for detection of AVX512 availability in
269 the processor and compiler was fixed
270 * gmake builds on OSX now set the internal name of the library to
271 libopenblas.0.dylib (consistent with CMAKE)
272 * the Haswell DGEMM kernel received a significant speedup through
273 improved prefetch and load instructions
274 * performance of DGEMM, DTRMM, DTRSM and ZDOT on Zen/Zen2 was markedly
275 increased by avoiding vpermpd instructions
276 * the SKYLAKEX (AVX512) DGEMM helper functions have now been disabled
277 to fix remaining errors in DGEMM, DSYMM and DTRMM
280 * added support for building on FreeBSD/powerpc64 and FreeBSD/ppc970
281 * added optimized kernels for POWER9 SGEMM and STRMM
284 * fixed the softfp implementations of xAMAX and IxAMAX
285 * removed the predefined -march= flags on both ARMV5 and ARMV6 as
286 they were appropriate for only a subset of platforms
288 ====================================================================
293 * the build tools now check that a given cpu TARGET is actually valid
294 * the build-time check of system features (c_check) has been made
295 less dependent on particular perl features (this should mainly
296 benefit building on Windows)
297 * several problem with the ReLAPACK integration were fixed,
298 including INTERFACE64 support and building a shared library
299 * building with CMAKE on BSD systems was improved
300 * a non-absolute SUM function was added based on the
301 existing optimized code for ASUM
302 * CBLAS interfaces to the IxMIN and IxMAX functions were added
303 * a name clash between LAPACKE and BOOST headers was resolved
304 * CMAKE builds with OpenMP failed to include the appropriate getrf_parallel
306 * a crash on thread (key) deletion with the USE_TLS=1 memory management
308 * restored several earlier fixes, in particular for OpenMP performance,
309 building on BSD, and calling fork on CYGWIN, which had inadvertently
310 been dropped in the 0.3.3 rewrite of the memory management code.
313 * the AVX512 DGEMM kernel has been disabled again due to unsolved problems
314 * building with old versions of MSVC was fixed
315 * it is now possible to build a static library on Windows with CMAKE
316 * accessing environment variables on CYGWIN at run time was fixed
317 * the CMAKE build system now recognizes 32bit userspace on 64bit hardware
318 * Intel "Denverton" atom and Hygon "Dhyana" zen CPUs are now autodetected
319 * building for DYNAMIC_ARCH with a DYNAMIC_LIST of targets is now supported
321 * building for DYNAMIC_ARCH with GENERIC as the default target is now supported
322 * a buffer overflow in the SSE GEMM kernel for Intel Nano targets was fixed
323 * assembly bugs involving undeclared modification of input operands were fixed
324 in the AXPY, DOT, GEMV, GER, SCAL, SYMV and TRSM microkernels for Nehalem,
325 Sandybridge, Haswell, Bulldozer and Piledriver. These would typically cause
326 test failures or segfaults when compiled with recent versions of gcc from 8 onward.
327 * a similar bug was fixed in the blas_quickdivide code used to split workloads
329 * a bug in the IxMIN implementation for the GENERIC target made it return the result of IxMAX
330 * fixed building on SkylakeX systems when either the compiler or the (emulated) operating
331 environment does not support AVX512
332 * improved GEMM performance on ZEN targets
335 * build failures caused by the recently added checks for AVX512 were fixed
336 * an inline assembly bug involving undeclared modification of an input argument was
337 fixed in the blas_quickdivide code used to split workloads in most functions
338 * a bug in the IMIN implementation for the GENERIC target made it return the result of IMAX
341 * a bug in the IMIN implementation made it return the result of IMAX
344 * single precision BLAS1/2 functions have received optimized POWER8 kernels
345 * POWER9 is now a separate target, with an optimized DGEMM/DTRMM kernel
346 * building on PPC970 systems under OSX Leopard or Tiger is now supported
347 * out-of-bounds memory accesses in the gemm_beta microkernels were fixed
348 * building a shared library on AIX is now supported for POWER6
349 * DYNAMIC_ARCH support has been added for POWER6 and newer
352 * corrected xDOT behaviour with zero INC_X or INC_Y
353 * a bug in the IMIN implementation made it return the result of IMAX
356 * added support for HiSilicon TSV110 cpus
357 * the CMAKE build system now recognizes 32bit userspace on 64bit hardware
358 * cross-compilation with CMAKE now works again
359 * a bug in the IMIN implementation made it return the result of IMAX
360 * ARMV8 builds with the BINARY=32 option are now automatically handled as ARMV7
363 * optimized microkernels for single precicion BLAS1/2 functions have been added
366 ====================================================================
371 * loop unrolling in TRMV has been enabled again.
372 * A domain error in the thread workload distribution for SYRK
374 * gmake builds will now automatically add -fPIC to the build
375 options if the platform requires it.
376 * a pthreads key leakage (and associate crash on dlclose) in
377 the USE_TLS codepath was fixed.
378 * building of the utest cases on systems that do not provide
379 an implementation of complex.h was fixed.
382 * the SkylakeX code was changed to compile on OSX.
383 * unwanted application of the -march=skylake-avx512 option
384 to the common code parts of a DYNAMIC_ARCH build was fixed.
385 * improved performance of SGEMM for small workloads on Skylake X.
386 * performance of SGEMM and DGEMM was improved on Haswell.
389 * a configuration error that broke the CNRM2 kernel was corrected.
390 * compilation of the GEMM kernels with CMAKE was fixed.
391 * DYNAMIC_ARCH builds are now available with CMAKE as well.
392 * using CMAKE for cross-compilation to the new cpu TARGETs
393 introduced in 0.3.4 now works.
396 * a problem in cpu autodetection for AIX has been corrected.
398 ====================================================================
403 * the new, experimental thread-local memory allocation had
404 inadvertently been left enabled for gmake builds in 0.3.3
405 despite the announcement. It is now disabled by default, and
406 single-threaded builds will keep using the old allocator even
407 if the USE_TLS option is turned on.
408 * OpenBLAS will now provide enough buffer space for at least 50
410 * The output of openblas_get_config() now contains the version
412 * A serious thread safety bug in GEMV operation with small M and
413 large N size has been fixed.
414 * The code will now automatically call blas_thread_init after a
415 fork if needed before handling a call to openblas_set_num_threads
416 * Accesses to parallelized level3 functions from multiple callers
417 are now serialized to avoid thread races (unless using OpenMP).
418 This should provide better performance than the known-threadsafe
419 (but non-default) USE_SIMPLE_THREADED_LEVEL3 option.
420 * When building LAPACK with gfortran, -frecursive is now (again)
421 enabled by default to ensure correct behaviour.
422 * The OpenBLAS version cblas.h now supports both CBLAS_ORDER and
423 CBLAS_LAYOUT as the name of the matrix row/column order option.
424 * Externally set LDFLAGS are now passed through to the final compile/link
425 steps to facilitate setting platform-specific linker flags.
426 * A potential race condition during the build of LAPACK (that would
427 usually manifest itself as a failure to build TESTING/MATGEN) has been
429 * xHEMV has been changed to stay single-threaded for small input sizes
430 where the overhead of multithreading exceeds any possible gains
431 * CSWAP and ZSWAP have been limited to a single thread except on ARMV8 or
432 ThunderX hardware with sizable input.
433 * Linker flags for the PGI compiler have been updated
434 * Behaviour of AXPY with zero increments is now handled in the C interface,
435 correcting the result on at least Intel Atom.
436 * The result matrix from calling SGELSS with an all-zero input matrix is
437 now zeroed completely.
440 * Autodetection of AMD Ryzen2 has been fixed (again).
441 * CMAKE builds now support labeling of an INTERFACE64=1 build of
442 the library with the _64 suffix.
443 * AVX512 version of DGEMM has been added and the AVX512 SGEMM kernel
444 has been sped up by rewriting with C intrinsics
445 * Fixed compilation on RHEL5/CENTOS5 (issue with typename __WAIT_STATUS)
448 * added support for building on AIX (with gcc and GNU tools from AIX Toolbox).
449 * CPU type detection has been implemented for AIX.
450 * CPU type detection has been fixed for NETBSD.
453 * AXPY on LOONGSON3A has been corrected to pass "zero increment" utest.
454 * DSDOT on LOONGSON3A has been fixed.
455 * the SGEMM microkernel has been hardened against potential data loss.
458 * DYNAMic_ARCH support is now available for 64bit ARM
459 * cross-compiling for ARMV8 under iOS now works.
460 * cpu-specific code has been rearranged to make better use of both
461 hardware commonalities and model-specific compiler optimizations.
462 * XGENE1 has been removed as a TARGET, superseded by the improved generic
466 * Older assembly mnemonics have been converted to UAL form to allow
467 building with clang 7.0
468 * Cross compiling LAPACKE for Android has been fixed again (broken by
469 update to LAPACK 3.7.0 some while ago).
471 ====================================================================
476 * thread memory allocation has been switched back to the method
477 used before version 0.3.1 due to unexpected problems caused by
478 the new code under some circumstances. A new compile-time option
479 USE_TLS has been added to enable the new code, and it is hoped
480 that this can become the default again in the next version.
481 * LAPAck PR272 has been integrated, which fixes spurious errors
482 in DSYEVR and related functions caused by missing conversion
483 from ILAENV to ILAENV_2STAGE in several _2stage routines.
484 * the cmake-generated OpenBLASConfig.cmake now uses correct case
485 for the name of the library
486 * added support for Haiku OS
489 * added AVX512 implementations of SDOT, DDOT, SAXPY, DAXPY,
490 DSCAL, DGEMVN and DSYMVL
491 * added a workaround for a cygwin issue that prevented compilation
495 * added autodetection of Z14
496 * fixed TRMM errors in the generic target
498 ====================================================================
503 * fixes for regressions caused by the rewrite of the thread
504 initialization code in 0.3.1
507 * fixed cpu autodetection for the BSDs
510 * fixed utest errors in AXPY, DSDOT, ROT and SWAP
513 * added autodetection of AMD Ryzen 2
514 * fixed build with older versions of MSVC
516 ====================================================================
521 * rewritten thread initialization code with significantly reduced overhead
522 * added CBLAS interfaces to the IxAMIN BLAS extension functions
523 * fixed the lapack-test target
524 * CMAKE builds now create an OpenBLASConfig.cmake file
525 * ZAXPY now uses a single thread for small input sizes
526 * the LAPACK code was updated from Reference-LAPACK/lapack#253
527 (fixing LAPACKE interfaces to Aasen's functions)
530 * corrected CROT and ZROT behaviour with zero INC_X
533 * corrected xDOT behaviour with zero INC_X or INC_Y
536 * retired some older targets of DYNAMIC_ARCH builds to a new option DYNAMIC_OLDER,
537 this affects PENRYN,DUNNINGTON,OPTERON,OPTERON_SSE3,BOBCAT,ATOM and NANO
538 (which will still be supported via the slower PRESCOTT kernels when this option is not set)
539 * added an option DYNAMIC_LIST that (used in conjunction with DYNAMIC_ARCH) allows to
540 specify the list of x86_64 targets to include. Any target not on the list will be supported
541 by the Sandybridge or Nehalem kernels if available, or by Prescott.
542 * improved SWITCH_RATIO on Haswell for increased GEMM throughput
543 * added initial support for Intel Skylake X, including an AVX512 SGEMM kernel
544 * added autodetection of Intel Cannon Lake series as Skylake X
545 * added a default L2 cache size for hypervisors that return zero here (Chromebook)
546 * fixed a name clash with recent Windows10 headers that broke the build with (at least)
547 recent mingw from MSYS2
548 * fixed a link error in mixed clang/gfortran builds with OpenMP
549 * updated the OSX deployment target to 10.8
550 * switched on parallel make for builds on MS Windows by default
553 * fixed SSWAP and DSWAP behaviour with zero INC_X and INC_Y
555 ====================================================================
560 * fixed some more thread race and locking bugs
561 * added preliminary support for calling an OpenMP build of the library from multiple threads
562 * removed performance impact of thread locks added in 0.2.20 on OpenMP code
563 * general code cleanup
564 * optimized DSDOT implementation
565 * improved thread distribution for GEMM
566 * corrected IMATCOPY/OMATCOPY implementation
567 * fixed out-of-bounds accesses in the multithreaded xBMV/xPMV and SYMV implementations
568 * cmake build improvements
569 * pkgconfig file now contains build options
570 * openblas_get_config() now reports USE_OPENMP and NUM_THREADS settings used for the build
571 * corrections and improvements for systems with more than 64 cpus
572 * LAPACK code updated to 3.8.0 including later fixes
573 * added ReLAPACK, a recursive implementation of several LAPACK functions
574 * Rewrote ROTMG to handle cases that the netlib code failed to address
575 * Disabled (broken) multithreading code for xTRMV
576 * corrected prototypes of complex CBLAS functions to make our cblas.h match the generally accepted standard
577 * shared memory access failures on startup are now handled more gracefully
578 * restored utests from earlier releases (and made them pass on all affected systems)
581 * several fixes for cpu autodetection
584 * corrected vector register overwriting in several Power8 kernels
585 * optimized additional BLAS functions
588 * added support for CortexA53 and A72
589 * added autodetection for ThunderX2T99
590 * made most optimized kernels the default for generic ARMv8 targets
593 * parallelized DDOT kernel for Haswell
594 * changed alignment directives in assembly kernels to boost performance on OSX
595 * fixed register handling in the GEMV microkernels (bug exposed by gcc7)
596 * added support for building on OpenBSD and Dragonfly
597 * updated compiler options to work with Intel release 2018
598 * support fully optimized build with clang/flang on Microsoft Windows
599 * fixed building on AIX
602 * added optimized BLAS 1/2 functions
605 * fixed cpu autodetection helper code
606 * added mips32 1004K cpu (Mediatek MT7621 and similar SoC)
607 * added mips64 I6500 cpu
609 ====================================================================
614 * Improved CMake support
615 * Fixed several thread race and locking bugs
616 * Fixed default LAPACK optimization level
617 * Updated LAPACK to 3.7.0
618 * Added ReLAPACK (https://github.com/HPAC/ReLAPACK, make BUILD_RELAPACK=1)
621 * Optimizations for Power9
622 * Fixed several Power8 assembly bugs
625 * New optimized Vulcan and ThunderX2T99 targets
626 * Support for ARMV7 SOFT_FP ABI (make ARM_SOFTFP_ABI=1)
627 * Detect all cpu cores including offline ones
628 * Fix compilation with CLANG
629 * Support building a shared library for Android
632 * Fixed several threading issues
633 * Fix compilation with CLANG
636 * Detect Intel Bay Trail and Apollo Lake
637 * Detect Intel Sky Lake and Kaby Lake
638 * Detect Intel Knights Landing
639 * Detect AMD A8, A10, A12 and Ryzen
640 * Support 64bit builds with Visual Studio
641 * Fix building with Intel and PGI compilers
642 * Fix building with MINGW and TDM-GCC
643 * Fix cmake builds for Haswell and related cpus
644 * Fix building for Sandybridge with CLANG 3.9
645 * Add support for the FLANG compiler
648 * New target z13 with BLAS3 optimizations
650 ====================================================================
654 * Improved cross compiling.
655 * Fix the bug on musl libc.
658 * Optimize BLAS on Power8
659 * Fixed Julia+OpenBLAS bugs on Power8
662 * Optimize BLAS on MIPS P5600 and I6400 (Thanks, Shivraj Patil, Kaustubh Raste)
665 * Improved on ARM Cortex-A57. (Thanks, Ashwin Sekhar T K)
668 ====================================================================
672 * If you set MAKE_NB_JOBS flag less or equal than zero,
673 make will be without -j.
676 * Support building Visual Studio static library. (#813, Thanks, theoractice)
677 * Fix bugs to pass buidbot CI tests (http://build.openblas.net)
680 * Provide DGEMM 8x4 kernel for Cortex-A57 (Thanks, Ashwin Sekhar T K)
683 * Optimize S and C BLAS3 on Power8
684 * Optimize BLAS2/1 on Power8
686 ====================================================================
690 * Enable BUILD_LAPACK_DEPRECATED=1 by default.
692 ====================================================================
696 * Avoid potential getenv segfault. (#716)
697 * Import LAPACK svn bugfix #142-#147,#150-#155
700 * Optimize c/zgemv for AMD Bulldozer, Piledriver, Steamroller
701 * Fix bug with scipy linalg test.
704 * Improve DGEMM for ARM Cortex-A57. (Thanks, Ashwin Sekhar T K)
707 * Optimize D and Z BLAS3 functions for Power8.
709 ====================================================================
713 * Upgrade LAPACK to 3.6.0 version.
714 Add BUILD_LAPACK_DEPRECATED option in Makefile.rule to build
715 LAPACK deprecated functions.
716 * Add MAKE_NB_JOBS option in Makefile.
717 Force number of make jobs.This is particularly
718 useful when using distcc. (#735. Thanks, Jerome Robert.)
719 * Redesign unit test. Run unit/regression test at every build (Travis-CI and Appveyor).
720 * Disable multi-threading for small size swap and ger. (#744. Thanks, Jerome Robert)
721 * Improve small zger, zgemv, ztrmv using stack alloction (#727. Thanks, Jerome Robert)
722 * Let openblas_get_num_threads return the number of active threads.
723 (#760. Thanks, Jerome Robert)
724 * Support illumos(OmniOS). (#749. Thanks, Lauri Tirkkonen)
725 * Fix LAPACK Dormbr, Dormlq bug. (#711, #713. Thanks, Brendan Tracey)
726 * Update scipy benchmark script. (#745. Thanks, John Kirkham)
729 * Optimize trsm kernels for AMD Bulldozer, Piledriver, Steamroller.
730 * Detect Intel Avoton.
731 * Detect AMD Trinity, Richland, E2-3200.
732 * Fix gemv performance bug on Mac OSX Intel Haswell.
733 * Fix some bugs with CMake and Visual Studio
736 * Support and optimize Cortex-A57 AArch64.
737 (#686. Thanks, Ashwin Sekhar TK)
738 * Fix Android build on ARMV7 (#778. Thanks, Paul Mustiere)
739 * Update ARMV6 kernels.
742 * Fix detection of POWER architecture
743 (#684. Thanks, Sebastien Villemot)
745 ====================================================================
749 * Support cmake on x86/x86-64. Natively compiling on MS Visual Studio.
750 (experimental. Thank Hank Anderson for the initial cmake porting work.)
752 On Linux and Mac OSX, OpenBLAS cmake supports assembly kernels.
757 On Windows MS Visual Studio, OpenBLAS cmake only support C kernels.
758 (OpenBLAS uses AT&T style assembly, which is not supported by MSVC.)
759 e.g. cmake -G "Visual Studio 12 Win64" .
760 Open OpenBLAS.sln and build.
762 * Enable MAX_STACK_ALLOC flags by default.
763 Improve ger and gemv for small matrices.
764 * Improve gemv parallel with small m and large n case.
765 * Improve ?imatcopy when lda==ldb (#633. Thanks, Martin Koehler)
766 * Add vecLib benchmarks (#565. Thanks, Andreas Noack.)
767 * Fix LAPACK lantr for row major matrices (#634. Thanks, Dan Kortschak)
768 * Fix LAPACKE lansy (#640. Thanks, Dan Kortschak)
769 * Import bug fixes for LAPACKE s/dormlq, c/zunmlq
770 * Raise the signal when pthread_create fails (#668. Thanks, James K. Lowden)
771 * Remove g77 from compiler list.
772 * Enable AppVeyor Windows CI.
775 * Support pure C generic kernels for x86/x86-64.
776 * Support Intel Boardwell and Skylake by Haswell kernels.
777 * Support AMD Excavator by Steamroller kernels.
778 * Optimize s/d/c/zdot for Intel SandyBridge and Haswell.
779 * Optimize s/d/c/zdot for AMD Piledriver and Steamroller.
780 * Optimize s/d/c/zapxy for Intel SandyBridge and Haswell.
781 * Optimize s/d/c/zapxy for AMD Piledriver and Steamroller.
782 * Optimize d/c/zscal for Intel Haswell, dscal for Intel SandyBridge.
783 * Optimize d/c/zscal for AMD Bulldozer, Piledriver and Steamroller.
784 * Optimize s/dger for Intel SandyBridge.
785 * Optimize s/dsymv for Intel SandyBridge.
786 * Optimize ssymv for Intel Haswell.
787 * Optimize dgemv for Intel Nehalem and Haswell.
788 * Optimize dtrmm for Intel Haswell.
791 * Support Android NDK armeabi-v7a-hard ABI (-mfloat-abi=hard)
792 e.g. make HOSTCC=gcc CC=arm-linux-androideabi-gcc NO_LAPACK=1 TARGET=ARMV7
793 * Fix lock, rpcc bugs (#616, #617. Thanks, Grazvydas Ignotas)
795 * Support ppc64le platform (ELF ABI v2. #612. Thanks, Matthew Brandyberry.)
796 * Support POWER7/8 by POWER6 kernels. (#612. Thanks, Fábio Perez.)
798 ====================================================================
802 * Improve OpenBLASConfig.cmake. (#474, #475. Thanks, xantares.)
803 * Improve ger and gemv for small matrices by stack allocation.
804 e.g. make -DMAX_STACK_ALLOC=2048 (#482. Thanks, Jerome Robert.)
805 * Introduce openblas_get_num_threads and openblas_get_num_procs.
806 (#497. Thanks, Erik Schnetter.)
807 * Add ATLAS-style ?geadd function. (#509. Thanks, Martin Köhler.)
808 * Fix c/zsyr bug with negative incx. (#492.)
809 * Fix race condition during shutdown causing a crash in
810 gotoblas_set_affinity(). (#508. Thanks, Ton van den Heuvel.)
813 * Support AMD Streamroller.
816 * Add Cortex-A9 and Cortex-A15 targets.
818 ====================================================================
822 * Add SYMBOLPREFIX and SYMBOLSUFFIX makefile options
823 for adding a prefix or suffix to all exported symbol names
824 in the shared library.(#459, Thanks Tony Kelman)
825 * Provide OpenBLASConfig.cmake at installation.
826 * Fix Fortran compiler detection on FreeBSD.
827 (#470, Thanks Mike Nolta)
831 * Add generic kernel files for x86-64. make TARGET=GENERIC
832 * Fix a bug of sgemm kernel on Intel Sandy Bridge.
833 * Fix c_check bug on some amd64 systems. (#471, Thanks Mike Nolta)
836 * Support APM's X-Gene 1 AArch64 processors.
837 Optimize trmm and sgemm. (#465, Thanks Dave Nuechterlein)
839 ====================================================================
843 * Added CBLAS interface for ?omatcopy and ?imatcopy.
844 * Enable ?gemm3m functions.
845 * Added benchmark for ?gemm3m.
846 * Optimized multithreading lower limits.
847 * Disabled SYMM3M and HEMM3M functions
848 because of segment violations.
851 * Improved axpy and symv performance on AMD Bulldozer.
852 * Improved gemv performance on modern Intel and AMD CPUs.
854 ====================================================================
858 * Added some benchmark codes.
859 * Fix link error on Linux/musl.(Thanks Isaac Dunham)
862 * Improved s/c/zgemm performance for Intel Haswell.
863 * Improved s/d/c/zgemv performance.
864 * Support the big numa machine.(EXPERIMENT)
867 * Fix detection when cpuinfo uses "Processor". (Thanks Isaiah)
869 ====================================================================
873 * Added BLAS extensions as following.
874 s/d/c/zaxpby, s/d/c/zimatcopy, s/d/c/zomatcopy.
875 * Added OPENBLAS_CORETYPE environment for dynamic_arch. (a86d34)
876 * Added NO_AVX2 flag for old binutils. (#401)
877 * Support outputing the CPU corename on runtime.(#407)
878 * Patched LAPACK to fix bug 114, 117, 118.
879 (http://www.netlib.org/lapack/bug_list.html)
880 * Disabled ?gemm3m for a work-around fix. (#400)
882 * Fixed lots of bugs for optimized kernels on sandybridge,Haswell,
883 bulldozer, and piledriver.
884 https://github.com/xianyi/OpenBLAS/wiki/Fixed-optimized-kernels-To-do-List
887 * Improved LAPACK testing.
889 ====================================================================
893 * Improved the result for LAPACK testing. (#372)
894 * Installed DLL to prefix/bin instead of prefix/lib. (#366)
895 * Build import library on Windows.(#374)
897 * To improve LAPACK testing, we fallback some kernels. (#372)
898 https://github.com/xianyi/OpenBLAS/wiki/Fixed-optimized-kernels-To-do-List
900 ====================================================================
904 * Added OPENBLAS_VERBOSE environment variable.(#338)
905 * Make OpenBLAS thread-pool resilient to fork via pthread_atfork.
906 (#294, Thank Olivier Grisel)
910 * Detect Intel Haswell for new Macbook.
912 ====================================================================
916 * Update LAPACK to 3.5.0 version
917 * Fixed compatiable issues with Clang and Pathscale compilers.
920 * Optimization on Intel Haswell.
921 * Enable optimization kernels on AMD Bulldozer and Piledriver.
924 * Support ARMv6 and ARMv7 ISA.
925 * Optimization on ARM Cortex-A9.
927 ====================================================================
931 * Support Open64 5.0. (#266)
932 * Add executable stack markings. (#262, Thank Sébastien Fabbro)
933 * Respect user's LDFLAGS (Thank Sébastien Fabbro)
936 * Rollback bulldozer and piledriver kernels to barcelona kernels (#263)
937 We will fix the compuational error bug in bulldozer and piledriver kernels.
939 ====================================================================
943 * Support LSB (Linux Standard Base) 4.1.
945 * Include LAPACK 3.4.2 source codes to the repo.
946 Avoid downloading at compile time.
947 * Add NO_PARALLEL_MAKE flag to disable parallel make.
948 * Create openblas_get_parallel to retrieve information which
949 parallelization model is used by OpenBLAS. (Thank grisuthedragon)
950 * Detect LLVM/Clang compiler. The default compiler is Clang on Mac OS X.
951 * Change LIBSUFFIX from .lib to .a on windows.
952 * A work-around for dtrti_U single thread bug. Replace it with LAPACK codes. (#191)
955 * Optimize c/zgemm, trsm, dgemv_n, ddot, daxpy, dcopy on
956 AMD Bulldozer. (Thank Werner Saar)
957 * Add Intel Haswell support (using Sandybridge optimizations).
959 * Add AMD Piledriver support (using Bulldozer optimizations).
960 * Fix the computational error in zgemm avx kernel on
962 * Fix the overflow bug in gemv.
963 * Fix the overflow bug in multi-threaded BLAS3, getrf when NUM_THREADS
964 is very large.(#214, #221, #246).
966 * Support loongcc (Open64 based) compiler for ICT Loongson 3A/B.
969 * Support Power7 by old Power6 kernels. (#220)
971 ====================================================================
975 * Improved OpenMP performance slightly. (d744c9)
976 * Improved cblas.h compatibility with Intel MKL.(#185)
977 * Fixed the overflowing bug in single thread cholesky factorization.
978 * Fixed the overflowing buffer bug of multithreading hbmv and sbmv.(#174)
981 * Added AMD Bulldozer x86-64 S/DGEMM AVX kernels. (Thank Werner Saar)
982 We will tune the performance in future.
983 * Auto-detect Intel Xeon E7540.
984 * Fixed the overflowing buffer bug of gemv. (#173)
985 * Fixed the bug of s/cdot about invalid reading NAN on x86_64. (#189)
989 ====================================================================
993 * Added NO_SHARED flag to disable generating the shared library.
994 * Compile LAPACKE with ILP64 modle when INTERFACE64=1 (#158)
995 * Export LAPACK 3.4.2 symbols in shared library. (#147)
996 * Only detect the number of physical CPU cores on Mac OSX. (#157)
997 * Fixed NetBSD build. (#155)
998 * Fixed compilation with TARGET=GENERIC. (#160)
1000 * Restore the original CPU affinity when calling
1001 openblas_set_num_threads(1) (#153)
1002 * Fixed a SEGFAULT bug in dgemv_t when m is very large.(#154)
1005 ====================================================================
1009 * Upgraded LAPACK to 3.4.2 version. (#145)
1010 * Provided support for passing CFLAGS, FFLAGS, PFLAGS,
1011 FPFLAGS to make. (#137)
1012 * f77blas.h:compatibility for compilers without C99 complex
1013 number support. (#141)
1015 * Added NO_AVX flag. Check OS supporting AVX on runtime. (#139)
1016 * Fixed zdot incompatibility ABI issue with GCC 4.7 on
1017 Windows 32-bit. (#140)
1019 * Fixed the generation of shared library bug.
1020 * Fixed the detection bug on the Loongson 3A server.
1021 ====================================================================
1025 * Fixed LAPACK unstable bug about ?laswp. (#130)
1026 * Fixed the shared library bug about unloading the library on
1028 * Fixed the compilation failure on BlueGene/P (TARGET=PPC440FP2)
1029 Please use gcc and IBM xlf. (#134)
1031 * Supported goto_set_num_threads and openblas_set_num_threads
1032 APIs in Windows. They can set the number of threads on runtime.
1034 ====================================================================
1038 * Fixed exporting DLL functions bug on Windows/MingW
1039 * Support GNU Hurd (Thank Sylvestre Ledru)
1040 * Support kfreebsd kernel (Thank Sylvestre Ledru)
1042 * Support Intel Sandy Bridge 22nm desktop/mobile CPU
1044 * Improve the detection of SPARC (Thank Sylvestre Ledru)
1046 ====================================================================
1051 * Fixed the SEGFAULT bug about hyper-theading
1052 * Support AMD Bulldozer by using GotoBLAS2 AMD Barcelona codes
1054 ====================================================================
1058 * Removed the limitation (64) of numbers of CPU cores.
1059 Now, it supports 256 cores at max.
1060 * Supported clang compiler.
1061 * Fixed some build bugs on FreeBSD
1063 * Optimized Level-3 BLAS on Intel Sandy Bridge x86-64 by AVX instructions.
1064 Please use gcc >= 4.6 or clang >=3.1.
1065 * Support AMD Bobcat by using GotoBLAS2 AMD Barcelona codes.
1067 ====================================================================
1071 * Upgraded LAPACK to 3.4.1 version. (Thank Zaheer Chothia)
1072 * Supported LAPACKE, a C interface to LAPACKE. (Thank Zaheer Chothia)
1073 * Fixed the build bug (MD5 and download) on Mac OSX.
1074 * Auto download CUnit 2.1.2-2 from SF.net with UTEST_CHECK=1.
1075 * Fxied the compatibility issue for compilers without C99 complex number
1076 (e.g. Visual Studio)
1078 * Auto-detect Intel Sandy Bridge Core i7-3xxx & Xeon E7 Westmere-EX.
1079 * Test alpha=Nan in dscale.
1080 * Fixed a SEGFAULT bug in samax on x86 windows.
1082 ====================================================================
1086 * Set soname of shared library on Linux.
1087 * Added LIBNAMESUFFIX flag in Makefile.rule. The user can use
1088 this flag to control the library name, e.g. libopenblas.a,
1089 libopenblas_ifort.a or libopenblas_omp.a.
1090 * Added GEMM_MULTITHREAD_THRESHOLD flag in Makefile.rule.
1091 The lib use single thread in GEMM function with small matrices.
1093 * Used GEMV SSE/SSE2 kernels on x86 32-bit.
1094 * Exported CBLAS functions in Windows DLL.
1096 * Completed Level-3 BLAS optimization on Loongson 3A CPU.
1097 * Improved GEMV performance on Loongson 3A CPU.
1098 * Improved Level-3 BLAS performance on Loongson 3B CPU. (EXPERIMENT)
1100 ====================================================================
1101 Version 0.1 alpha2.5
1104 * Fixed missing "#include <sched.h>" bug on Mac OS X.
1105 Thank Mike Nolta for the patch.
1106 * Upgraded LAPACK to 3.4.0 version
1107 * Fixed a bug on Mac OS X. Don't require SystemStubs on OS X.
1108 SystemStubs does not exist on Lion. Thank Stefan Karpinski.
1109 * Improved README with using OpenMP. Check the internal threads
1110 count less than or equal to omp_get_max_threads()
1112 * Auto-detect Intel Core i6/i7 (Sandy Bridge) CPU with Nehalem assembly kernels
1113 * Fixed some bugs on MingW 64-bit including zgemv, cdot, zdot.
1115 ====================================================================
1116 Version 0.1 alpha2.4
1119 * Fixed a bug about installation. The header file "fblas77.h"
1121 * Fixed #61 a building bug about setting TARGET and DYNAMIC_ARCH.
1122 * Try to handle absolute path of shared library in OSX. (#57)
1123 Thank Dr Kane O'Donnell.
1124 * Changed the installation folder layout to $(PREFIX)/include and
1128 * Fixed #58 zdot/xdot SEGFAULT bug with GCC-4.6 on x86. According
1129 to i386 calling convention, The callee should remove the first
1130 hidden parameter.Thank Mr. John for this patch.
1132 ====================================================================
1133 Version 0.1 alpha2.3
1137 * Added DTB_ENTRIES into dynamic arch setting parameters. Now,
1138 it can read DTB_ENTRIES on runtime. (Refs issue #55 on github)
1140 ====================================================================
1141 Version 0.1 alpha2.2
1145 * Fixed a building bug when DYNAMIC_ARCH=1 & INTERFACE64=1.
1146 (Refs issue #44 on github)
1148 ====================================================================
1149 Version 0.1 alpha2.1
1153 * Stop the build and output the error message when detecting
1154 fortran compiler failed. (Refs issue #42 on github)
1156 ====================================================================
1161 * Fixed blasint undefined bug in <cblas.h> file. Other software
1162 could include this header successfully(Refs issue #13 on github)
1163 * Fixed the SEGFAULT bug on 64 cores. On SMP server, the number
1164 of CPUs or cores should be less than or equal to 64.(Refs issue #14
1166 * Support "void goto_set_num_threads(int num_threads)" and "void
1167 openblas_set_num_threads(int num_threads)" when USE_OPENMP=1
1168 * Added extern "C" to support C++. Thank Tasio for the patch(Refs
1169 issue #21 on github)
1170 * Provided an error message when the arch is not supported.(Refs
1171 issue #19 on github)
1172 * Fixed issue #23. Fixed a bug of f_check script about generating link flags.
1173 * Added openblas_set_num_threads for Fortran.
1174 * Fixed #25 a wrong result of rotmg.
1175 * Fixed a bug about detecting underscore prefix in c_check.
1176 * Print the wall time (cycles) with enabling FUNCTION_PROFILE
1177 * Fixed #35 a build bug with NO_LAPACK=1 & DYNAMIC_ARCH=1
1178 * Added install target. You can use "make install". (Refs #20)
1182 * Fixed #28 a wrong result of dsdot on x86_64.
1183 * Fixed #32 a SEGFAULT bug of zdotc with gcc-4.6.
1184 * Fixed #33 ztrmm bug on Nehalem.
1185 * Work-around #27 the low performance axpy issue with small imput size & multithreads.
1188 * Fixed #28 a wrong result of dsdot on Loongson3A/MIPS64.
1189 * Optimized single/double precision BLAS Level3 on Loongson3A/MIPS64. (Refs #2)
1190 * Optimized single/double precision axpy function on Loongson3A/MIPS64. (Refs #3)
1192 ====================================================================
1197 * Support "make NO_LAPACK=1" to build the library without
1199 * Fixed randomly SEGFAULT when nodemask==NULL with above Linux 2.6.34.
1200 Thank Mr.Ei-ji Nakama providing this patch. (Refs issue #12 on github)
1201 * Added DEBUG=1 rule in Makefile.rule to build debug version.
1202 * Disable compiling quad precision in reference BLAS library(netlib BLAS).
1203 * Added unit testcases in utest/ subdir. Used CUnit framework.
1204 * Supported OPENBLAS_* & GOTO_* environment variables (Pleas see README)
1205 * Imported GotoBLAS2 1.13 BSD version
1208 * On x86 32bits, fixed a bug in zdot_sse2.S line 191. This would casue
1209 zdotu & zdotc failures. Instead, work-around it. (Refs issue #8 #9 on github)
1210 * Modified ?axpy functions to return same netlib BLAS results
1211 when incx==0 or incy==0 (Refs issue #7 on github)
1212 * Modified ?swap functions to return same netlib BLAS results
1213 when incx==0 or incy==0 (Refs issue #6 on github)
1214 * Modified ?rot functions to return same netlib BLAS results
1215 when incx==0 or incy==0 (Refs issue #4 on github)
1216 * Detect Intel Westmere,Intel Clarkdale and Intel Arrandale
1217 to use Nehalem codes.
1218 * Fixed a typo bug about compiling dynamic ARCH library.
1220 * Improve daxpy performance on ICT Loongson 3A.
1221 * Supported ICT Loongson 3A CPU (Refs issue #1 on github)
1222 ====================================================================