#3.0.1 #3.1.1 #3.2.0 3.2.4 #5745:37f59e65eb6c 5891:d8652709345d # introduce AVX #5893:24b4dc92c6d3 # merge 5895:997c2ef9fc8b # introduce FMA #5904:e1eafd14eaa1 # complex and AVX 5908:f8ee3c721251 # improve packing with ptranspose #5921:ca808bb456b0 # merge #5927:8b1001f9e3ac 5937:5a4ca1ad8c53 # New gebp kernel handling up to 3 packets x 4 register-level blocks #5949:f3488f4e45b2 # merge #5969:e09031dccfd9 # Disable 3pX4 kernel on Altivec #5992:4a429f5e0483 # merge before-evaluators #6334:f6a45e5b8b7c # Implement evaluator for sparse outer products #6639:c9121c60b5c7 #6655:06f163b5221f # Properly detect FMA support on ARM #6677:700e023044e7 # FMA has been wrongly disabled #6681:11d31dafb0e3 #6699:5e6e8e10aad1 # merge default to tensors #6726:ff2d2388e7b9 # merge default to tensors #6742:0cbd6195e829 # merge default to tensors #6747:853d2bafeb8f # Generalized the gebp apis 6765:71584fd55762 # Made the blocking computation aware of the l3 cache; Also optimized the blocking parameters to take into account the number of threads used for a computation 6781:9cc5a931b2c6 # generalized gemv 6792:f6e1daab600a # ensured that contractions that can be reduced to a matrix vector product #6844:039efd86b75c # merge tensor 6845:7333ed40c6ef # change prefetching in gebp #6856:b5be5e10eb7f # merge index conversion 6893:c3a64aba7c70 # clean blocking size computation 6899:877facace746 # rotating kernel for ARM only #6904:c250623ae9fa # result_of 6921:915f1b1fc158 # fix prefetching change for ARM 6923:9ff25f6dacc6 # prefetching 6933:52572e60b5d3 # blocking size strategy 6937:c8c042f286b2 # avoid redundant pack_rhs 6981:7e5d6f78da59 # dynamic loop swapping 6984:45f26866c091 # rm dynamic loop swapping, adjust lhs's micro panel height to fully exploit L1 cache 6986:a675d05b6f8f # blocking heuristic: block on the rhs in L1 if the lhs fit in L1. 7013:f875e75f07e5 # organize a little our default cache sizes, and use a saner default L1 outside of x86 (10% faster on Nexus 5) 7015:8aad8f35c955 # Refactor computeProductBlockingSizes to make room for the possibility of using lookup tables 7016:a58d253e8c91 # Polish lookup tables generation 7018:9b27294a8186 # actual_panel_rows computation should always be resilient to parameters not consistent with the known L1 cache size, see comment 7019:c758b1e2c073 # Provide a empirical lookup table for blocking sizes measured on a Nexus 5. Only for float, only for Android on ARM 32bit for now. 7085:627e039fba68 # Bug 986: add support for coefficient-based product with 0 depth. 7098:b6f1db9cf9ec # Bug 992: don't select a 3p GEMM path with non-vectorizable scalar types, this hits unsupported paths in symm/triangular products code 7591:09a8e2186610 # 3.3-alpha1 7650:b0f3c8f43025 # help clang inlining 7708:dfc6ab9d9458 # Improve numerical accuracy in LLT and triangular solve by using true scalar divisions (instead of x * (1/y)) #8744:74b789ada92a # Improved the matrix multiplication blocking in the case where mr is not a power of 2 (e.g on Haswell CPUs) 8789:efcb912e4356 # Made the index type a template parameter to evaluateProductBlockingSizes. Use numext::mini and numext::maxi instead of std::min/std::max to compute blocking sizes 8972:81d53c711775 # Don't optimize the processing of the last rows of a matrix matrix product in cases that violate the assumptions made by the optimized code path 8985:d935df21a082 # Remove the rotating kernel. 8988:6c2dc56e73b3 # Bug 256: enable vectorization with unaligned loads/stores. 9148:b8b8c421e36c # Relax mixing-type constraints for binary coefficient-wise operators 9174:d228bc282ac9 # merge 9175:abc7a3600098 # Include the cost of stores in unrolling 9212:c90098affa7b # Fix performance regression introduced in changeset 8aad8f35c955 9213:9f1c14e4694b # Fix performance regression in dgemm introduced by changeset 81d53c711775 9361:69d418c06999 # 3.3-beta2 9445:f27ff0ad77a3 # Optimize expression matching "d?=a-b*c" as "d?=a; d?=b*c;" 9583:bef509908b9d # 3.3-rc1 9593:2f24280cf59a # Bug 1311: fix alignment logic in some cases of (scalar*small).lazyProduct(small) 9722:040d861b88b5 # Disabled part of the matrix matrix peeling code that's incompatible with 512 bit registers 9792:26667be4f70b # 3.3.0 9891:41260bdfc23b # Fix a performance regression in (mat*mat)*vec for which mat*mat was evaluated multiple times. 9942:b1d3eba60130 # Operators += and -= do not resize! 9943:79bb9887afd4 # Ease compiler job to generate clean and efficient code in mat*vec 9946:2213991340ea # Complete rewrite of column-major-matrix * vector product to deliver higher performance of modern CPU. 9955:630471c3298c # Improve performance of row-major-dense-matrix * vector products for recent CPUs. (this is the next changeset fixing a typo) 9975:2eeed9de710c # Revert vec/y to vec*(1/y) in row-major TRSM 10442:e3f17da72a40 # Bug 1435: fix aliasing issue in exressions like: A = C - B*A; 10735:6913f0cf7d06 # Adds missing EIGEN_STRONG_INLINE to support MSVC properly inlining small vector calculations 10943:4db388d946bd # Bug 1562: optimize evaluation of small products of the form s*A*B by rewriting them as: s*(A.lazyProduct(B)) to save a costly temporary. Measured speedup from 2x to 5x. 10961:5007ff66c9f6 # Introduce the macro ei_declare_local_nested_eval to help allocating on the stack local temporaries via alloca, and let outer-products makes a good use of it. 11083:30a528a984bb # Bug 1578: Improve prefetching in matrix multiplication on MIPS. 11533:71609c41e9f8 # PR 526: Speed up multiplication of small, dynamically sized matrices 11535:6d348dc9b092 # Vectorize row-by-row gebp loop iterations on 16 packets as well 11568:efda481cbd7a # Bug 1624: improve matrix-matrix product on ARM 64, 20% speedup 11596:b8d3f548a9d9 # do not read buffers out of bounds 11628:22f9cc0079bd # Implement AVX512 vectorization of std::complex 11638:81172653b67b # Bug 1515: disable gebp's 3pX4 micro kernel for MSVC<=19.14 because of register spilling. 11659:b500fef42ced # Artificially increase l1-blocking size for AVX512. +10% speedup with current kernels.