AOCL-BLAS 5.2.2 Release Notes

Overview

AOCL-BLAS 5.2.2 is an incremental release building on the 5.2 GA release, delivering performance optimizations, bug fixes, improved threading stability, and expanded test coverage.

Performance Optimizations

• Optimized SGEMM rd kernels on Zen3
• Improved SGEMM rd kernel on Zen4/Zen5
• SGEMM tiny path tuning for Zen4 and Zen5
• Added tiny path for SGEMM
• Added fast path for single-threaded AVX512 DGEMV kernel
• Tuned decision logic for DGEMV multithreading for skinny sizes
• Tuned DGEMV no-transpose thresholds
• Re-tuned GEMV thresholds
• Optimal rerouting of GEMV inputs to avoid packing
• Tuned input threshold for tiny DGEMM interface
• Tuned ZGEMM threshold for Zen5
• Changed ZGEMM SUP threshold logic for Zen5 to fix performance regression
• Replaced intrinsics with inline assembly for bli_saxpyv_zen4_int and bli_saxpyf_zen_int_5
• Improved fringe case handling for AXPYV kernel
• Disabled small_gemm for Zen4/Zen5 and added single-thread check for tiny path
• Disabling GEMV(M1) rerouting in BF16 APIs (AVX512)

Bug Fixes

• Fixed memory leak in DGEMV kernel
• Fixed extreme values handling in GEMV
• Fixed integer division in GEMV that was supposed to be a double operation
• Fixed Integer Overflow issue in TPSV
• Fixed out-of-bound access in F32 matrix add/mul ops
• Bugfix: BF16 to F32 conversion in AVX2 F32 codepath
• Bug fix in BF16 AVX2 conversion path
• Fix for F32 to BF16 conversion and AVX512 ISA support checks
• Fixed cblas_ctrmm invalid diag handling
• Coverity issue fix for ZTRSM
• Fixed Coverity static analysis issue in DTRSM
• Fixed high priority Coverity issues in LPGEMM
• Fixed Coverity issues with CID: 23269 and CID: 137049
• Resolved operator precedence warning in Zen5 DCOMPLEX threshold logic
• Modified AXPY kernel to ensure consistency of numerical results
• GCC 15 SUP kernel workaround (2)
• Disabled no post-ops path in LPGEMM F32 kernels for certain GCC versions
• Updated guards in s8s8s32of32_sym_quant framework

Threading & Stability

• Fixed data race in native code-path
• Add OpenMP barrier before releasing threadinfo & global communicator to avoid race
• Replaced OMP barrier with bli_thread_barrier and added similar fixes
• Global communicator is now freed outside the parallel region
• Thread: free global communicator after parallel region completes
• Initialize mem_t structures safely and handle NULL communicator in threading
• Fix DTL dynamic thread logging in BLAS operations
• Added dynamic threads and actual threads in the DTL log of SAXPY
• Enabled disable-sba-pools feature in AOCL-BLAS

Build System & Infrastructure

• Updates to the build systems (CMake and Make) for LPGEMM compilation
• CMake: Adding targets and aliases so that BLIS works with FetchContent
• Set security flags default enable
• DTL Windows getpid support
• Add compiler information to make showconfig and bench_getlibraryInfo
• Make all bench applications consistent
• Standardize Zen kernel names

Test Suite (GTestSuite)

• Added Banded API tests: gbmv, hbmv, sbmv, tbmv, tbsv
• Added Packed API tests: hpmv, spmv, tpmv, tpsv, hpr, hpr2, spr, spr2
• Added conjugate dot and ger IIT_ERS tests
• Added data pool support
• Moved data generator definitions to a cpp file
• Computediff improvements
• Fix in swap
• Break up tests for better organization
• Multiple miscellaneous test fixes

Documentation

• Fixing doc about building bench
• Add external PR integration process and flowchart to CONTRIBUTING.md
• Updated LICENSE and NOTICES files for AOCL-5.2 release

Compiler Warnings

• Multiple rounds of compiler warnings fixes
• Adding bli_print_msg before bli_abort() for bli_thrinfo_sup_create_for_cntl
• DTL log updates
• Code tidying