Mar 6, 2025: version 10.0.1 * (60) bug fix: one of the GrB_mxm kernels (saxpy4) can hit a segfault for user-defined types, in the JIT kernels. Caught by Gabe Gomez. Mar 1, 2025: version 10.0.0 * 32/64 bit matrices and vectors: the GrB_Matrix and GrB_Vector now exploit 32-bit integers when possible. New methods added to pass 32-bit integer arrays to/from GrB_build, extract, assign, subassign, and extractTuples. New object, the GxB_Container added for fast import/export of matrices/vectors with arbitrary integer content. * GrB_Field: this enum is strongly deprecated, and replaced with 'typedef int GrB_Field'. This is an upward-compatible change to the API, and will allow the creation of a future mathematical field object in GraphBLAS. This type should not be used; use an int instead. It will be removed or replaced in a future version of GraphBLAS. * enum parameters: replaced all enum parameters with int, to simplify future updates to enum parameters, including the GrB_Field. * GxB_JIT_ERROR: added in 9.4.x, changed value to avoid conflict with LAGraph error codes. * pack/unpack: these are declared historical; they still work but use 64-bit integers only. Use the new GxB_Container methods instead. * GxB_Matrix_iso and GxB_Vector_iso: declared historical; use GrB_get with the new GxB_ISO enum. * GxB_Matrix_type, GxB_Vector_type, GxB_Scalar_type: no longer historical; added back to the user guide. * Summary: the API is upward-compatible with 9.4.x, but only after the user application is recompiled with GraphBLAS v10.0.0. As a result, the SO version must increase from 9 to 10. Feb 20, 2025: version 9.4.5 * (59) bug fix: the GxB_NO_INT32 and GxB_NOT_INT64 flags in GB_control.h did not completely remove some of the INT32 and INT64 factory kernels. Caught by Erik Welch, NVIDIA. * AppleClang compiler bug: On the Mac, the Source/mask/GB_masker.c file triggers a bug in AppleClang 16.0.0 with -O3 in (MacOS 14.6.1 (23G93), Xcode 16.2, Apple clang version 16.0.0, clang-1600.0.26.6). It also fails in MacOS 15.2 (Target: arm64-apple-darwin23.6.0). The bug is triggered by these tests in LAGraph (v1.2 branch, unreleased, Jan 4, 2025): 39 - LAGraphX_BF (SIGTRAP) 40 - LAGraphX_Coarsen_Matching (Failed) 41 - LAGraphX_FastGraphletTransform (SIGTRAP) 49 - LAGraphX_PageRankGX (SIGTRAP) 54 - LAGraphX_SquareClustering (SIGTRAP) 61 - LAGraphX_msf (Failed) When using clang, optimization is turned off for this file. This has no impact on performance since GB_masker.c is very simple, consisting of a single sequence of calls to other methods. Feb 17, 2025: version 9.4.4 * (58) bug fix: semirings with user-defined monoids and the GrB_ONEB or GxB_PAIR operators caused a JIT compiler error. Caught by Gabe Gomez. * (57) bug fix: GraphBLAS.h header: remove duplicate definitions of GxB_MAX_FIRST_* semirings (incompletely moved to 'historical' section in 9.4.2). Nov 18, 2024: version 9.4.2 * clarified User Guide: regarding when the hyper-hash is built * JIT: reduced JIT kernel encodings * also: includes the updates from 9.4.0.beta and 9.4.1.beta listed below Nov 15, 2024: version 9.4.1 (only released as BETA) * new JIT kernels added: for GrB_assign, GrB_extract, GxB_sort, GrB_kronecker, stand-alone mask phase, and utilities (convert to/from sparse/bitmap, unjumble). All kernels formerly tagged "JIT: needed" are now finished. * removed Factory kernels for: types int8 and uint8, and semirings: max_min, max_plus, max_times, min_max, min_times, plus_min, plus_max, non-Boolean land/lor/lxor/lxnor, and integer times_first/second, to reduce size of compiled library. JIT kernels will be used instead for these types and semirings. * GxB_IndexBinaryOp: finalized and named as GxB_*. Oct 15, 2024: version 9.4.0 (only released as BETA) * new operator and associated methods: added the draft G*B_IndexBinaryOp * JIT error-handling behavior changed: if a compiler error occurs in the JIT, GxB_JIT_ERROR is now returned. Previously, GraphBLAS would fall back to a generic method if such an error occurred. Aug 12, 2024: version 9.3.1 * (56) bug fix: wrong type for fgetc return value in JITpackage; leads to infinite loop on some systems when building GraphBLAS. Aug 2, 2024: version 9.3.0 * code restructuring: Source folder split into many subfolders, and some files and internal functions renamed. No visible external change. * (55) bug fix: GrB_apply with user-defined index-unary op would fail if A was iso and the JIT was disabled (failure in generic case). Caught by Christina Koutsou. * (54) bug fix: reducing a huge iso full matrix to a scalar resulted in integer overflow if nrows*ncols was larger than about 2^60. * reduced size of compiled library: int16 and uint16 types and operators for FactoryKernels are disabled in GB_control.h. The JIT will always be used instead. May 22, 2024: version 9.2.0 * Added graphblas_install.m: for a simpler method of compiling the MATLAB/Octave interface for GraphBLAS. * JIT: sanitizing the JIT cache path, better burble for compiler errors * GrB_get/GrB_set: better handling of concurrent get/set between different user threads Mar 22, 2024: version 9.1.0 * minor updates to build system * C11 complex type detection: this is now detected and configured by cmake, instead of using an #if ... in the GraphBLAS.h header. This change was required to port GraphBLAS to the clang-cl compiler on Windows when it simulates the MSVC compiler. Also added a new feature (thus the minor version update to 9.1.0): GxB_HAVE_COMPLEX* to GraphBLAS.h to indicate which kind of complex data types are available in C11 or MSVC. Contributed by Markus Mützel. * port to clang-cl: fixing the GxB_get and GxB_set macro * (53) bug fix: eWiseAdd C=A+B when M, A, and B are all hypersparse; access to M was incorrect (also affects C+=T for any operation, if M and T are both hypersparse). Caught by Roi Lipman. Mar 1, 2024: version 9.0.3 * (52) performance bug fix: JIT kernels since v8.3.1 were not compiled with OpenMP. Feb 26, 2024: version 9.0.2 * (51) bug fix: GraphBLAS/Makefile "make static" was incorrect. Jan 20, 2024: version 9.0.1 * minor updates to build system Jan 10, 2024: version 9.0.0 * GrB_get/GrB_set: new functions from the v2.1 C API. * GrB_Type_new, GrB_UnaryOp_new, GrB_IndexUnaryOp_new: no longer macros, since GrB_set can be used to set the names of the operators. These methods no longer extract the name, so the default name is now the empty string. This is because GrB_get/set can only set these names once. This is a non-compatible change of behavior for these 3 methods, so SuiteSparse:GraphBLAS must become v9.0.0. * historical methods: many methods are replaced by GrB_get and GrB_set. They remain in SuiteSparse:GraphBLAS but have been declared historical. Terse prototypes exist in GraphBLAS.h, and any discussion is removed from the User Guide: GxB_get, GxB_set, and the methods they call, and many more. Use GrB_get/set in place those methods, and for: GxB_*type_name, GxB_*type, GxB_Monoid_operator, GxB_Monoid_identity, GxB_Monoid_terminal, GxB_Semiring_add, GxB_Semiring_multiply. Use GrB_STORAGE_ORIENTATION_HINT in place of GxB_FORMAT. * hyper_hash: constructed only if the # of non-empty vectors in a hypersparse matrix is large (> 1024, by default) * minor updates to build system: *.pc files for pkgconfig Dec 30, 2023: version 8.3.1 * major change to build system: by Markus Mützel * GraphBLAS.h: remove "undef I" Oct 7, 2023: version 8.2.1 * (50) bug fix: remove undefined behavior in Source/GB_add.c with Cnvec_nonempty, caught by Yves Vandreissche, Intel * (49) bug fix: GrB_mxm saxpy4 and saxpy5 had incorrectly handling of typecasting in v8.0.0 to v8.2.0 (caught by Erik Welch) * cross-compiler support: replace check_c_source_runs with _compiles, for GraphBLAS and SuiteSparse_config, and remove check for getenv("HOME"). * cmake update: add "None" build type, from Antonio Rojas, for Arch Linux Sept 8, 2023: version 8.2.0 * cmake updates: SuiteSparse:: namespace by Markus Muetzel * GB_VLA_MAXSIZE: increased to 1024; only affects compilers not supporting variable-length arrays (MSVC) Version 8.0.2, June 16, 2023 * added -DJITINIT=option: use -DJITINIT to set the initial state of the GxB_JIT_CONTROL (4:on, 3:load, 2:run, 1:pause, 0:off). The default is 4 (on) if the JIT is enabled, or 2 (run) if -DNJIT=1 is set. * xxHash: upgraded to latest version as of June 16, 2023 Version 8.0.1, May 27, 2023 * (48) bug fix: GrB_*_nvals returned UINT64_MAX ('infinity') for a GrB_Vector of size n-by-2^60; it should return 2^60. Caught by Erik Welch, NVIDIA. * added GxB_Context_error and GxB_Context_wait * C++: changed complex typedefs for C++ that #include GraphBLAS.h; update from Markus Muetzel Version 8.0.0, May 18, 2023 * version 8: This version is a major SO version increase, since it removes a few minor user-visible features from SuiteSparse:GraphBLAS: the GrB_Descriptor no longer supports threading control, and some features of the GxB_SelectOp are removed (see below). Enum values have been changed for compatibility with the upcoming GrB_set/get features in the V2.1 C API. * The JIT: GraphBLAS v8.0.0 now includes a JIT for the CPU kernels, which can compile kernels at run time. Added GxB_set/get options and environment variables to control the JIT. The GxB_*Op_new methods can accept NULL function pointers, if the strings are provided and valid. * GxB_Type_new: the size of the type can be given as zero, in which case the size is determined via a JIT kernel. * GxB_UnaryOp_new, GxB_BinaryOp_new, and GxB_IndexUnaryOp_new: the function pointer can be given as NULL, in which case the function is created by the JIT. * math kernels: revised for CUDA JIT. More accurate complex floating-point for Mac OS on Apple Silicon. * Demo/wildtype_demo: change to double so that CPU and GPU versions compute the same result. * GxB_get: can return malloc/calloc/realloc/free functions * GxB_Context: an object for controlling computational resources: # of OpenMP threads, the chunk factor, and (draft) GPU id. * GrB_Descriptor: removed ability to control # of OpenMP threads from the descriptor (a rarely used feature). Replaced with the GxB_Context object. * GxB_SelectOp: GraphBLAS no longer supports user-defined GxB_SelectOps. Use a GrB_IndexUnaryOp instead. The GxB_SelectOp_new and GxB_SelectOp_free functions are removed entirely. The built-in GxB_SelectOps, GxB_Matrix_select, GxB_Vector_select, and GxB_select still work. However, the GxB_EQ_THUNK, GxB_EQ_ZERO, GxB_NE_THUNK, and GxB_NE_ZERO operators no longer work on user-defined types, as they did in v7.4.4 and earlier. Create a user-defined GrB_IndexUnaryOp to compute these operations instead, and use GrB_select. * alternative/Makefile: removed; incompatible with the JIT * zstd: upgraded to v1.5.5 (Apr 4, 2023) Version 7.4.4, Mar 25, 2023 * (47) bug fix: OpenMP atomics require seq_cst on the ARM. Revised GB_atomics.h accordingly, and added them for all architectures (caught by Gabor Szarnyas). Version 7.4.3, Jan 20, 2023 * debug: turned on in GrB_Matrix_removeElement by mistake. Version 7.4.2, Jan 17, 2023 * minor change to build system: for SuiteSparse v7.0.0 * deprecation notice: in GraphBLAS v8.0.0, the ability to set the # of threads, and chunk size, in the descriptor will be removed. It still appears in v7.x, but will be replaced by a Context object in v8.0.0. Version 7.4.1, Jan 9, 2023 * global free pool: disabled. Benefit for single-thread user applications was modest, and it causes too much contention in a critical section when the user application is multi-threaded. * GrB_mxm: revised task creation heuristics for sparse-times-sparse for better performance. "chunk" and "very_costly" had been tuned in v7.3.x for sparse-times-dense but this slowed down sparse-times-sparse. The new heuristic keeps the chunk & very_costly parameters for sparse-times-dense but restores the prior values for sparse-times-sparse from v7.2.0 (Aug 8, 2022). See Source/GB_AxB_saxpy3_slice_balanced.c. Version 7.4.0, Dec 23, 2022 * added non-va_arg methods: va_arg-based GxB_get/set methods are C11 but cause issues for cffi in Python. As a temporary workaround, new methods have been added that do not use va_arg. The existing GxB_get/set methods are not changed. The new methods are not in the user guide, since all of the GxB_get/set methods will be superceded with GrB_get/set in the v2.1 C API. At that point, all GxB_get/set methods will become historical (kept, not deprecated, but removed from the user guide). Version 7.3.3, Dec 9, 2022 * minor change to build system * stdatomic.h: using #include and atomic_compare_exchange_weak instead of GCC/clang/icx __atomic_* variants. Added -latomic if required. * chunk factor for C=A*B (saxpy3 method): revised for non-builtin-semirings Version 7.3.2, Nov 12, 2022 * cmake_modules: minor revision to build system, to sync with SuiteSparse v6.0.0 * Added option -DNOPENMP=1 to disable OpenMP parallelism Version 7.3.1, Oct 21, 2022 * workaround for Microsoft Visual Studio bug : MSC 19.2x (in vs2019) encounters a compiler bug when compiling the FIRST_FC32 and SECOND_FC32 binary operators. This version adds a flag that disables those methods. The operators still work, they are just handled by the slower generic methods. Thanks to Erik Welch, H. Vetinari, and others in the conda-forge community for tracking this down. Version 7.3.0, Oct 14, 2022 * GrB_Matrix: changes to the internal data structure * minor internal changes: A->nvals for sparse/hypersparse * more significant changes: added A->Y hyper_hash for hypersparse case, speeds up GrB_mxm (C=A*B, C=A*B' (for CSR A'*B for CSC)), C=A+B (eWiseAdd), CA.*B (eWiseMult), some cases for GrB_assign, for hypersparse matrices. * added GxB_unpack_HyperHash and GxB_pack_HyperHash: to pack/unpack the A->Y hyper_hash * @GrB MATLAB/Octave interface: changed license to Apache-2.0 * MATLAB library: renamed to libgraphblas_matlab.so * performance: faster C=A*B when using a single thread and B is a sparse vector with many entries Version 7.2.0, Aug 8, 2022 * added ZSTD as a compression option for serialize/deserialize: Version 1.5.3 by Yann Collet, https://github.com/facebook/zstd.git Copyright (c) 2016-present, Facebook, Inc. All rights reserved. Included in SuiteSparse:GraphBLAS via its BSD-3-clause license. The default method is now ZSTD, level 1. * added GxB_Matrix_reshape and GxB_Matrix_reshapeDup * MATLAB interface: faster C(:)=A, C=A(:), and reshape. Better error messages. Version 7.1.2, July 8, 2022 * MATLAB interface: linear indexing added for C(:)=A, C=A(:), and single-output I=find(C). Faster bandwidth, istriu, istril, isbanded, isdiag. C(I,J)=A can now grow the size of C. Version 7.1.1, June 17, 2022 * minor updates to documentation and error messages * MATLAB interface: minor revision of GrB.deserialize Version 7.1.0, May 20, 2022 * added cube root: GxB_CBRT_FP* unary operators * added GxB_Matrix_isStoredElement and GxB_Vector_isStoredElement Version 7.0.4, Apr 25, 2022 * (46) bug fix: user-defined type size was incorrectly limited to 128 bytes. Caught by Erik Welch. Version 7.0.3, Apr 8, 2022 * faster transpose when using 2 threads Version 7.0.2, Apr 6, 2022 * (45) bug fix: vector iterator was broken for iterating across a vector in bitmap format. Caught by Erik Welch. Version 7.0.1, Apr 3, 2022 * revised ACM TOMS submission: Doc/toms_parallel_grb2.pdf Version 7.0.0, Apr 2, 2022 * (44) spec bug: GrB_Matrix_diag was implemented in v5.2.x and v6.x with the wrong signature. This fix requires the major release to change, from v6.x to v7.x, since the change means that the revised v7 is not backward compatible with v6. * performance for GrB_mxm: auto selection for saxpy method (Hash vs Gustavson) revised. * (43) performance bug fix for GrB_assign: better performance for C(i,j)=scalar and C(i,j)+=scalar when i and j have length 1 (scalar assigment with no scalar expansion). The prior code worked but this is a performance bug introduced in the parallel GraphBLAS. The bug likely doesn't appear in v2.x (the sequential version of SS:GrB). This affects method 1 of Demo/Source/wathen.c. Caught by a reviewer of the ACM TOMS paper on the parallel GraphBLAS. Version 6.2.5, Mar 14, 2022 * For SuiteSparse v5.11.0 Version 6.2.4, Mar 8, 2022 * (42) bug fix: GrB_mxm with 0-by-0 iso full matrices: Caught by Henry Amuasi in the Python grblas interface, then triaged and isolated by Erik Welch. Version 6.2.3, Mar 5, 2022 * minor update to documentation in GrB.build: no change to any code Version 6.2.2 Feb 28, 2022 * revised output of GxB_*_sort: to return newly created matrices C and P as full or bitmap matrices, as appropriate, instead of sparse/hypersparse, following their sparsity control settings Version 6.2.1 Feb 16, 2022 * (41) bug fix: GxB_Iterator_get used (void *) + siz arithmetic Version 6.2.0 Feb 14, 2022 * added row/column/entry/vector iterators * @GrB C=A*B: revised sparse-times-full rule * static_header: no longer used if CUDA enabled * GB_BURBLE: removed. Burble can no longer be disabled at compile-time. Version 6.1.4, Jan 13, 2022 * added section to User Guide: how to get the best performance out of algorithms based on GraphBLAS * cpu_features: no longer built as a separate library, but built directly into libgraphblas.so and libgraphblas.a. Added compile-time flags to optionally disable the use of cpu_features completely. * Octave 7: port to Apple Silicon (thanks to Gabor Szarnyas) * min/max monoids: real case (FP32 and FP64) no longer terminal * @GrB interface: overloaded C=A*B syntax where one matrix is full always results in a full matrix C, which is faster and matches the Octave/ MATLAB policy. Version 6.1.3, Jan 1, 2022 * performance: task creation for GrB_mxm (saxpy method) didn't account for any work for A(:,k)*B(k,j) when nnz(A(:,k))==0, but this takes O(1) work to examine B(k,j). Performance improvement of up to 10x when nnz(A)<=Z not returning C as iso if Z iso and C initialy empty. Caught by Erik Welch, Anaconda. * performance improvements: C=A*B: sparse/hyper times bitmap/full, and visa versa, including C += A*B when C is full Version 5.1.10, Oct 27, 2021 * (35) bug fix: GB_selector; A->plen and C->plen not updated correctly. Caught by Jeffry Lovitz, Redis. Version 5.1.9, Oct 26, 2021 * (34) bug fix: in-place test incorrect for C+=A'*B using dot4 * (32) bug fix: disable free pool if OpenMP not available Version 5.1.8, Oct 5, 2021 * (32) bug fix: C=A*B when A is sparse and B is iso and bitmap. Caught by Mark Blanco, CMU. Version 5.1.7, Aug 23, 2021 * (31) bug fix: GrB_apply, when done in-place and matrix starts non-iso and becomes iso, gave the wrong iso result. Caught by Fabian Murariu. Version 5.1.6, Aug 16, 2021 * one-line change to C=A*B: faster symbolic analysis when a vector C(:,j) is dense (for CSC) or C(i,:) for CSR. * port to MacOSx 11.3.1 (Big Sur): cmake 3.13 or later now required. Version 5.1.5, July 15, 2021 * ACM TOMS submission: rename some functions to simplify discussion in ACM TOMS article Version 5.1.4, July 6, 2021 * faster Octave interface, and a bug fix for 1-based printing * (30) bug fix: 1-based printing not enabled for pending tuples. Caught by Will Kimmerer. Version 5.1.3, July 3, 2021 * added GxB_Matrix_iso and GxB_Vector_iso: to query if a matrix or vector is held as iso-valued. * (29) bug fix: Matrix_pack_*R into a matrix previously held by column, or pack*C into a matrix by row, would flip the dimensions. Caught by Erik Welch, Anaconda. Version 5.1.2, June 30, 2021 * (28) bug fix: kron(A,B) with iso input matrices A and B fixed. Caught by Michel Pelletier, Graphegon. Version 5.1.1, June 29, 2021 * (27) bug fix: v5.1.0 had a wrong version of a file (GB_ops.c, which wouldn't compile) posted by mistake. Caught by Michel Pelletier, Graphegon. Version 5.1.0, June 28, 2021 * iso matrices and vectors added * GxB_*_build_Scalar added: builds an iso matrix or vector * import/export: can now import/export iso matrices and vectors * GxB_*_pack/unpack methods added * GrB.argmin/argmax: added to @GrB interface * GxB_PRINT_1BASED: added to the global settings * GxB_*_memoryUsage added * Source/Generated: split into Source/Generated1 and Source/Generated2 for faster compilation when GBCOMPACT is enabled * port to Octave: gbmake and gbtest work in Octave7 to build and test the @GrB interface to GraphBLAS. Octave 7.0.0 is required. Version 5.0.6, May 24, 2021 * bfs and triangle counting demos: removed. See LAGraph for these algorithms. Version 5.0.5, May 17, 2021 * (26) performance bug fix: reduce-to-vector where A is hypersparse CSR with a transposed descriptor (or CSC with no transpose), and some cases for GrB_mxm/mxv/vxm when computing C=A*B with A hypersparse CSC and B bitmap/full (or A bitmap/full and B hypersparse CSR), the wrong internal method was being selected via the auto-selection strategy, resulting in a significant slowdown in some cases. Version 5.0.4, May 13, 2021 * @GrB interface: changed license from Apache-2.0 to GNU GPLv3 or later. Version 5.0.3, May 12, 2021 * (25) bug fix: disabling ANY_PAIR semirings in Source/GB_control.h would cause a segfault if those disabled semirings were used. * demos: no longer built by default * (24) bug fix: GB_PUBLIC added to the new functions in GraphBLAS.h * GrB_Matrix_reduce_BinaryOp: reinstated from v4.0.3; same limit on built-in ops that correspond to known monoids. * Source/GB_control.h: more semirings disabled by default to reduce compile time and binary library size. The semirings still work, but are not as fast since they now use the generic methods. Version 5.0.2, May 5, 2021 * GrB_Matrix_reduce_BinaryOp: removed. Use a monoid instead, with GrB_reduce or GrB_Matrix_reduce_Monoid. * (23) bug fix: GrB_Matrix_apply_BinaryOp1st and 2nd were using the wrong descriptors for GrB_INP0 and GrB_INP1. Caught by Erik Welch, Anaconda. * memory pool added for faster memory allocation * calloc no longer used: using malloc+memset(0) instead * realloc option: if not available, uses malloc+memcpy+free * @GrB interface ported to R2021a. * GxB_Matrix_diag: construct a diagonal matrix from a vector * GxB_Vector_diag: extract a diagonal from a matrix * concat/split: added methods to concatenate and split matrices * import/export: size of arrays now in bytes, not entries. A new parameter, is_uniform, has been added to all import/export methods, which indicates that the matrix values are all the same. * (22) bug fix: SIMD vectorization was missing reduction(+,task_cnvals) in GB_dense_subassign_06d_template.c. Caught by Jeff Huang, Texas A&M, with his software package for race-condition detection. Version 4.0.3, Jan 19, 2021 * faster min/max monoids * @GrB: G=GrB(G) converts G from v3 object to v4 Version 4.0.2, Jan 11, 2021 * added ability to load *.mat files saved with GraphBLAS v3. * GrB_wait: added #pragma omp flush after any GrB_wait Version 4.0.1, Jan 4, 2021 * substantial performance gains: compared with v3.3.3: up to 5x faster in BFS (with LAGraph_bfs_parent2), 2x faster in Betweeness-Centrality (in LAgraph_bc_bactch5); BC now faster than GAP for kron, urand, and twitter (the 3 largest matrices in the GAP benchmark). * GrB_wait(), with no inputs: removed * GrB_wait(&object): polymorphic function added * GrB_*_nvals: no longer guarantees completion; use GrB_wait for that * GrB_error: now has two parameters: a string (char **) and an object The no-input GrB_error(void) method has been deprecated and removed. * GrB_Matrix_reduce_BinaryOp: limited to certain built-in ops. Support for user-defined operators and binary ops that do not correspond to built-in monoids has been deprecated and removed. * GrB_*_extractTuples: may return indices out of order * removed internal features: GBI iterator, slice and hyperslice matrices * bitmap/full matrices and vectors added: faster and lower memory usage when many entries present in a matrix or vector * GxB_SPARSITY_CONTROL, GxB_SPARSITY_STATUS added: for matrix get/set * positional operators and semirings: GxB_FIRSTI_INT32 and related ops * jumbled matrices: sort left pending, like zombies and pending tuples * GxB_get/set: added GxB_SPARSITY_* (hyper, sparse, bitmap, or full) * GxB_HYPER: enum renamed to GxB_HYPER_SWITCH * GxB_BITMAP_SWITCH added: for matrix/vector get/set and global get/set * GxB*import/export: API modified * GxB_SelectOp: nrows and ncols removed from function signature * OpenMP tasking removed from mergesort: just as fast; now ports to Windows * GxB_BURBLE added: for diagnostic output * (21) bug fix: A({lo,hi})=scalar in @GrB; A(lo:hi)=scalar was OK Version 3.3.3, July 14, 2020 * (20) bug fix: w=A*u with mask non-empty and u empty (issue #13). affects mxm (A*B when B is a vector), vxm, and mxv. Version 3.3.2, July 3, 2020 * minor changes to build system * GraphBLAS v3.3.2 released as part of SuiteSparse v5.8.0 Version 3.3.1, June 30, 2020 * (19) bug fix: incorrect typecasting when GrB_assign or GxB_subassign is simple: C=A where A is sparse or dense, in GB_dense_subassign_24. Bug introduced in v3.2.0. * revised gbmake: to allow GraphBLAS to be compiled in Octave Version 3.3.0, June 26, 2020 * GrB_wait( ): with no input arguments, has been deprecated. It will be removed in V4.0. * added complex types: GxB_FC32 and GxB_FC64, many unary operators, binary operators, monoids, and semirings * added GrB_*_apply_BinaryOp1st and 2nd: also GxB version with GxB_Scalar * added bitwise operators: and their monoids and semirings * added predefined GrB* monoids and semirings: from the v1.3 spec. * @GrB interface: added complex matrices and operators, bitwise operators; improved performance * changed typecasting rules: for casting floating point types to integers * added GrB_*_wait: wait for specific object to complete * added GrB_*_resize: same as prior GxB_*_resize functions * added GrB_kronecker: same as prior GxB_kron * added version methods: GrB_getVersion, GRB_VERSION, GRB_SUBVERSION * added GrB_*_removeElement * (18) bug fix: fixed typecasting in GB_dense_subassign_23, generic case * (17) bug fix: non-polymorphic GrB_eWiseAdd and eWiseMult functions were misnamed. Version 3.2.2, Apr 2, 2020 * (16) bug fix to @GrB interface: for matrices with dimension > 2^53. * port to Microsoft Visual Studio: including GraphBLAS/Demo and GraphBLAS/Test, which had not been ported in v3.1.2. * remove dependency on m4: end-users no longer need m4; still needed by the Source/codegen*.m scripts but those are for development only. Version 3.2.1, Mar 10, 2020 * GAP benchmarks in *.m in GraphBLAS/GraphBLAS/GAP (all but CC). * atomic read/write: faster atomics for icc (see Source/GB_atomics.h) * install name for Mac: in alternative/Makefile and Tcov/Makefile * enable -DGBCOMPACT from CMake: for faster Travis builds * @GrB interface: better error message in GrB.reduce * @GrB triangle count: added sorting heuristic in GrB.tricount. * bug fix for GraphBLAS/Demo/simple_timer: better selection of which timer to use; simple_timer.[ch] made consistent. * no bug fixes needed in primary GraphBLAS library: just in Demo, test, and alternative/Makefile. Version 3.2.0, Feb 20, 2020 * GxB_*_define for user-defined compile-time objects: removed. Not compatible with the faster kernels for mxm and dense matrices/vectors. Use the GrB_*_new functions in the GraphBLAS C API Specification instead. * faster saxpy-based matrix multiply (about 5x to 10x for mxv and vxm): removed Sauna workspace. Heap method removed. Hash method added. * better performance for dense matrix and vectors: NOTE: see bug (19) above * faster typecast of the mask matrix M: GB_mcast replaces cast_M * added GB_BURBLE: for development diagnostics * changed default chunk size: from 4K to 64K * added the PAIR binary operator: f(x,y)=1 * added the ANY binary operator: f(x,y)=x, or f(x,y)=y ; arbitrary choice * added structural mask: from v1.3 C API Specification * added GrB_DESC_*: predefined descriptor, from v1.3 C API Specification * many atomics added: for the faster matrix multiply. These changes have not been ported to Microsoft Visual Studio, which only supports a subset of "#pragma omp atomic" statements. This will be resolved in a future version; in the meantime, use v3.1.2 with MS Visual Studio instead of v3.2.0. * no bug fixes Version 3.1.2, Dec 16, 2019 * (15) bug fix in parallel matrix-matrix multiply: could occur when # of threads is larger than the # of columns of B, for C=A*B, and when B is a standard matrix with one or more empty vectors. * (14) bug fix for @GrB interface: A(58) and A(:) were conflated. * compiler port: changes to allow GraphBLAS to be compiled with older compilers (pre C11), and Microsoft Visual Studio. Version 3.1.1, Oct 21, 2019 * minor edits: user guide and comments in code Version 3.1.0, Oct 2, 2019 * added @GrB interface: GraphBLAS/GraphBLAS is new. In Source/, added global pointer to printf for mexPrintf, pointer to mexMakeMemoryPersistent for Sauna workspace. Changed how GraphBLAS objects are printed with GxB_print. Changed how duplicate indices are handled in assign and extract, to match the built-in stadard. Added helper functions. Code size: @GrB is 9.7KLOC, test/ is 4.5KLOC). * bug counter added to this ChangeLog: to count # of bugs that appeared in formal releases that affect production code. The count excludes bug fixes for test code, bugs in the demo codes, bugs introduced in beta versions that were fixed before any formal release, and bugs prior to version 1.0. Code size of Source/, Include/ excluding Source/Generated, and Config/*.m4, is 42,659 lines (not including the new @GrB interface). 13 bugs / 42K lines is a bug rate of 0.3 bugs per KLOC, much lower than most commercial software, but higher than UMFPACK, CHOLMOD, etc (with about 0.1 bug per KLOC). GraphBLAS is a much more complex library, from the external view, than solving Ax=b. If UMFPACK has a bug, then Ax-b is typically large; there is no 'residual' to check for GraphBLAS. Also, in GraphBLAS, the test suite has about the same size as the main library (32K lines in Test/ and Tcov/). For UMFPACK, etc, the test suite is always about 1/3 the size of the library itself. * 'make dox': for doxygen removed (not really that useful) * (13) bug fix to GB_reduce_to_vector: to avoid integer divide-by-zero for a matrix with n=0 columns. * (12) bug fix to GB_accum_mask: when C+=T if C has no entries except pending tuples * (11) bug fix to GB_resize: when pending tuples exist and vdim is growing from vdim <= 1 to vdim > 1, matrix must be finished first. * (10) bug fix to GB_subref_phase1: "int nI" parameter should be int64_t. Version 3.0.1, July 26, 2019 * version number: Three changes to the user-visible API are not backward-compatible with V2.x: the added parameters to GxB_init and GxB_SelectOp_new, and the change in the type of the Thunk argument for GxB_select. Thus, the SO version of SuiteSparse:GraphBLAS is now 3, no longer 2. This change only affects SuiteSparse:GraphBLAS GxB_* extenstions, not any GrB_* functions or definitions. * added GxB_Scalar: acts like a GrB_Vector of length 1. * OpenMP parallelism: added nthreads and chunk parameters to GxB_set/get. * added parameter to GxB_init: bool user_malloc_is_thread_safe, for the @GrB mexFunction interface, or any other malloc library that might not be thread-safe. mxMalloc is not thread-safe. This change is not backward compatible with Version 2.x. * changed thunk parameter of GxB_select: was (void *), now GxB_Scalar. This change is not backward compatible with Version 2.x. * added parameter to GxB_SelectOp_new: to specify the type of the Thunk. This change is not backward compatible with Version 2.x. * added options to GxB_get: determine if a matrix is hypersparse or not, global library, API information, nthreads, and chunk. * added options to GxB_set: nthreads, and chunk. * new operators and semirings: RDIV (f(x,y)=y/x) and RMINUS (y-x) binary operators. Version 2.3.5, July 20, 2019 * minor update: fix for tests in R2019b Version 2.3.4, July 4, 2019 * Collected Algorithm of the ACM * (9) minor bug fix to GrB_*_extractTuples: nvals was left unchanged if the matrix or vector had no entries (now returned as zero). Version 2.3.3, May 27, 2019 * minor changes to the documentation: no change to the code Version 2.3.2, Mar 18, 2019 * GxB_Matrix_Option_set: removed spurious check of input matrix Version 2.3.1, Mar 9, 2019 * _Generic: further modifications for 'int' vs 'const int', needed for icc 19.0.0. * (8) bug fix in user_def1.m4: compile-time GxB_Semring_define with terminal monoids Version 2.3.0, Feb 25, 2019 * early exit: for integer TIMES monoids (terminal value is 0); cannot use zero for floating-point, to preserve NaN behavior * GrB_assign / GxB_subassign: further exploit of aliased inputs * Test/: @GrB mexFunctions now link against the dynamic -lgraphblas * bug fix for Demo/mis*: random number generator could create duplicate scores, which leads to a stall * terminal values for user monoids: added GxB_Monoid_terminal_new, GxB_Monoid_terminal_define, and GxB_Monoid_terminal * added GxB_init: GrB_init with malloc/calloc/realloc/free redefinition * _Generic: compilers differ on _Generic(x,int:,const int:). gcc, clang, icc 18.0.0 treat int and const int the same; icc 18.0.3 treats them as different. This is a result of an ambiguity in the C11 specification. Minor update so that the use of _Generic(...) is compatible with all compilers. * early exit: for GrB_reduce Version 2.3.0 (BETA1), Feb 16, 2019 * simplified GBI_*_iterators * towards OpenMP: PARALLEL comments added to most files (all upper case) where parallelism will be added in the future. * early exit: for monoids: MIN, MAX, AND, OR, based on terminal values * Demo/bfs*: much faster (prior versions had performance bug) * GrB_assign for accum/mask step: faster, zombies and pending tuples exploited by C=accum(C,T) for any GraphBLAS method, when either the mask M or accum are present. * moved persistent thread-workspace from the matrix to a global array * added isequal function to Demos/ * added import/export method * added nthreads: to descriptor, the matrix, and a global nthreads setting, and added Context as parameter in many internal functions, to pass in Context->nthreads for future parallelism * defined all enum constants explicitly: no change to their values, this just ensures they remain fixed in future versions Version 2.2.3, Dec 30, 2018 * added doxygen documentation: minor changes to the code comments to accomodate doxygen. Added Doxyfile settings file. Version 2.2.2, Dec 18, 2018 * GxB_print: better output for compile-time user objects * (7) bug fix: for GxB_Type_define, for GrB_*_build. * (6) bug fix: for GxB_Semiring_define, which was unable to access the predefined GxB_*_BOOL_MONOIDs. Version 2.2.0, Nov 29, 2018 * identical to 2.2.0 (BETA4): date changed, and "BETA4" removed. Demo/choldemo removed (testing artifact, not needed). Version 2.2.0 (BETA4), Nov 28, 2018 * minor change to GxB_get, GxB_set (all functions using va_arg): const removed from parameter definitions to silence spurious warnings from clang version 10.0.0, Mac OSX 10.14. * added Doc/CONTRIBUTOR-LICENSE.txt * added Doc/README_Doc.txt: includes instructions for OpenMP on the Mac * minor changes to comments of GB_subref_template: to match the ACM TOMS paper discussion of GrB_extract * included the GraphBLAS Extras: triangle counting, k-truss, and Kronecker matrix generator * minor changes for gcc -Wall -Wpedantic -Wextra, and icc -w3: to fix or silence warnings. In a few cases, prior versions were relying on extensions to the C11 standard in gcc, clang, xlc. The extensions are widely supported but best avoided anway. Version 2.2.0 (BETA3), Nov 23, 2018 * removed persistent thread-local O(n) workspace: the Mark, Flag, and Work arrays in GB_thread_local in prior versions are removed. Replaced with temporary workspace, except for C=A*B using Gustavson's method, which now uses the C->Sauna workspace, held in the output matrix. * better support for user-application threads: GrB_finalize is now fully compliant with the spec. User applications can use POSIX or OpenMP threads, or no threading, and use GraphBLAS in a thread-safe manner. Added OpenMP and POSIX pthread demos. * bug fix: freeing an object (monoid, matrix, or vector) that is dependent on another user-defined objects (an operator or type) that has already been freed led to the reading of an invalid value: the size of the object. This only affects the memory usage statistics for testing only, not production use. * beta bug fix: memory leak in pagerank_demo program (bug first appears in 2.2.0BETA2, not in any non-BETA final release Version 2.2.0 (BETA2), Nov 15, 2018 * added additional pagerank example in Demo/ * beta bug fix: GxB_Monoid_define: bug fix for C struct types. bug first appears in Version 2.2.0 (BETA1), not in any non-BETA final release. Version 2.2.0 (BETA1), Nov 2, 2018 * added GxB_*_define: to create user-defined objects at compile-time. All internal macros renamed to add the GB_ prefix, to help avoid name conflict with possible user-defined objects and macros. * changed default format to GxB_BY_ROW * performance improvements to GrB*assign* and GrB*extract * added PageRank demos * added GxB_*print: to display all GraphBLAS objects Version 2.1.0, Oct 14, 2018 * identical to V2.1.0 BETA2: except for a minor update to the user guide, and date of release Version 2.1.0 (BETA2), Sept 17, 2018 * user function z=f(x,y): passed to GrB_BinaryOp_new, may now be called as z=f(z,y), with x and z aliased. This is permitted by the API specification, and greatly improves performance of user-defined semirings. Version 2.1.0 (BETA1), Sept 11, 2018 * major update: added hypersparse and CSR formats, colon notation, and user selection of the method used by C=A*B. All changes to the API are backward compatible with Version 2.0, so the SO version does not increase (it is still 2.x). * added GxB_AxB_METHOD: for user selection of C=A*B method * added GxB_RANGE, GxB_STRIDE and GxB_BACKWARDS: for colon notation: begin:end and begin:inc:end. * added GxB_*Option* functions: to control hypersparsity and CSR/CSC format * added GxB_set and GxB_set: to set/get the descriptor and other options * (5) bug fix: if the accum operator z=f(x,y) was user-defined with different built-in types for x, y, z, the typecasting was wrong * GXB_* macros renamed GxB_*: old ones kept for backward compatibility Version 2.0.3: July 4, 2018 * (4) bug fix to GxB_resize * improvements to CMake script: version now defined only in CMakeLists.txt, and then cmake configures Include/GraphBLAS.h, README, and User Guide * bfs5 demo: asymptotically faster using sparse vector v but somewhat slower if entire graph is traversed Version 2.0.2: June 25, 2018 * added -lm to library dependency * minor changes to demos and user guide * removed unused code from Generated/* semiring functions Version 2.0.1: Mar 15, 2018 * (3) bug fix in GxB_kron: computed C incorrectly when A and B had pending updates on input Version 2.0.0: Mar 12, 2018 * moved opaque object definitions from Include/GraphBLAS.h to Source/GB.h * modifications to meet changes in the Nov 14, 2017, API specification: GrB_Monoid_TYPE_new changed to GrB_Monoid_new_TYPE. GrB_Type_new modified: 2nd argument no longer a type, but sizeof(type). This change is not backward compatible with Version 1.x, so this required an increase in the SO version, so the version number increased from 1.x to 2.x. * to allow non-C applications to access all GraphBLAS functions: These were just macros, now defined as both macros and functions: GrB_Type_new, GrB_UnaryOp_new, GrB_BinaryOp_new, and GxB_SelectOp_new. * added GxB_kron, GxB_resize (GxB_Matrix_resize and GxB_Vector_resize) * (2) bug fix to mask: GrB_anymethod (C, C, ... , r_desc) where r_desc.outp is 'replace', and C is aliased to the Mask, was incorrect. Version 1.1.2: Dec 28, 2017 * build issues with cmake * performance improvement of dot product method, C=A'*B and v=u'*A, particularly when A, B and u have dense columns, or nearly dense columns. Version 1.1.1: Dec 17, 2017 * (1) bug fix to GrB_assign and GxB_subassign: result was incorrect when output C was the same as the input A or Mask. * performance improvement in C=A*B and GrB_reduce to scalar * split AxB built-in semirings into Source/Generated/* for faster compilation * added memory usage statistics Version 1.1.0: Dec 1, 2017 * add new feature: GxB_SelectOp operator and GxB_select operation * to clarify which features of SuiteSparse:GraphBLAS are extensions to the spec, the names of functions and objects not in the spec have changed. They are now GxB_* (GXB_ for macros). * modified how the global matrix queue is accessed: for thread safety in case the user application is multi-threaded Version 1.0.0: Nov 25, 2017 * transition to CMake * tested on many compilers: gcc 4.9 to 7.2 on Linux, clang 8 and gcc 6.2.0 on the Mac, xlc 13.01 on IBM Power) * changed printf formats in Demo/ to silence warnings from older compilers * complex.h in gcc-6.2.0 on the Mac doesn't #define CMPLX; added to usercomplex.h * mergesort removed, quicksort performance improved * performance improvement for masked matrix-matrix multiply * bug fix for masked matrix-matrix multiply. In 0.5.0 only the Mask structure was used, not the values. The spec requires the values of the Mask. This is a bug but into only affects masked matrix multiply when the Mask has explicit zero entries. Version 0.5.0: Nov 13, 2017 * resolved icc -O3 segfault: Intel compiler bug fixed icc version 18.0 * added triangle-counting demo: in C (Demo/tricount.c) and Test/tricount.m. See also Test/test69, Test/test70. and Test/test71. * performance enhancements to GrB_eWiseMult, GrB_assign, GrB_subassign, GrB_extract, and GrB_mxm, reducing time and memory. mxm now exploits the mask during computation, to save memory and often saving time. This is essential for good performance for triangle counting. * reduced Test/testall.m and Tcov/testcov.m testing time from 2 hours to 15 minutes, maintained 100% test coverage Version 0.4.0: Oct 24, 2017 * SPEC: SuiteSparse:GraphBLAS 0.4.0 fully conforms to 1.1.0 of the spec * GrB_assign functions added; all now conform to the 1.1.0 spec * renamed all prior GrB_*_assign functions to GrB_*_subassign Version 0.3.0: Oct 14, 2017 * changes for spec 1.1.0 * performance improvement of GrB_mxm, GrB_mxv, transpose, GrB_reduce; added workspace kept internally, reused for subsequent calls to GraphBLAS * performance improvement of bfs in Demo (v held as dense vector) * performance improvement of GrB_assign when A is a scalar * performance improvement of C= ... mask operation * GrB_Column_extract and GrB_Column_assign renamed GrB_Col_* to match changes in the spec * changed GrB_*_setElement and GrB_*_extractTuples to match the spec * added GrB_INSUFFICIENT_SPACE error code for GrB_*_extractTuples * added GrB_INVALID_HANDLE * GrB_assign still does not match the spec (in progress). It differs on how it treats entries outside C(I,J), and it differs in the size and interpretation of the Mask. Version 0.2.0: Sept 5, 2017 * No significant change to the code; but preparing for wider release * content of GrB_stddef.h moved into Include/GraphBLAS.h and file removed * User Guide updated (SPEC discussion) * GraphBLAS.h modified (SPEC discussions) * Licensed under Apache 2.0; versions 0.1.x were limited releases * removed wathen_mod.m Version 0.1.2: Aug 30, 2017 * added (double) to fpclassify in CAST macro, to silence spurious errors in gcc * minor typos * icc -O3 causes a segfault when compiling GB_AxB_builtin.c, so Lib/GraphBLAS_config.mk now uses -O2 with icc. * some versions of gcc have trouble finding libgraphblas.so.0 when compiling the Demo (if GraphBLAS is not installed with 'make install'). Instead of requiring a 'make install' static linking is now used for the Demo programs. This has no effect on the GraphBLAS library itself. * changes for GraphBLAS spec 1.0.2 (Aug 10, 2017 version): * GrB_Vector_reduce_Monoid renamed to GrB_Matrix_reduce_Monoid * GrB_Vector_reduce_BinaryOp renamed to GrB_Matrix_reduce_BinaryOp * GrB_Row_extract was renamed in the spec to GrB_Col_extract, but this version keeps the name GrB_Column_extract since it parallels the name GrB_Column_assign. Version 0.1.1: Aug 28, 2017 * minor changes to Makefiles and GraphBLAS_config.mk * port to Linux with gcc 4.9.3, 5.40, and 6.3.0 * minor code changes to silence spurious gcc warnings; (&Ax[i]) changed to (Ax +(i)). * bug fix to GB_assign1.c (apply_pending_updates variable) Version 0.1.0: July 31, 2017 * first version released; only to GraphBLAS API Committee