rocBLAS-2.26.0 for ROCm 3.7.0

New Features

• Improvements to User Guide and Design Document
• L1 dot function optimized to utilize shuffle instructions ( improvements on bf16, f16, f32 data types )
• L1 dot function added x dot x optimized kernel
• Standardization of L1 rocblas-bench to use device pointer mode to focus on GPU memory bandwidth
• Adjustments for hipcc (hip-clang) compiler as standard build compiler and Centos8 support
• Added Fortran interface for all rocBLAS functions
• Improvements to rocblas_Xgemm_batched performance for small m, n, k.
• Improvements to rocblas_Xgemv_batched and rocblas_Xgemv_strided_batched performance for small m (QMCPACK use).
• Improvements to rocblas_Xdot (batched and non-batched) performance when both incx and incy are 1
• Improvements to FP32 ONNX BERT performance for MI50
• Significant improvements to FP32 Resnext, Inception Convolution performance for gfx908
• Slight improvements to FP32 DLRM Terabyte performance for gfx908
• Significant improvements to FP32 BDAS performance for gfx908
• Significant improvements to FP32 BDAS performance for MI50 and MI60
• Added substitution method for small trsm sizes with m <= 64 && n <= 64. Increases performance drastically for small batched trsm.

Known Issues

• None

rocBLAS-2.22.0 for ROCm 3.5.0

Changelist

• add geam complex, geam_batched, and geam_strided_batched
• add dgmm, dgmm_batched, and dgmm_strided_batched

Optimized performance

• ger

  • rocblas_sger, rocblas_dger,
  • rocblas_sger_batched, rocblas_dger_batched
  • rocblas_sger_strided_batched, rocblas_dger_strided_batched

• geru

  • rocblas_cgeru, rocblas_zgeru
  • rocblas_cgeru_batched, rocblas_zgeru_batched
  • rocblas_cgeru_strided_batched, rocblas_zgeru_strided_batched

• gerc

  • rocblas_cgerc, rocblas_zgerc
  • rocblas_cgerc_batched, rocblas_zgerc_batched
  • rocblas_cgerc_strided_batched, rocblas_zgerc_strided_batched

• symv

  • rocblas_ssymv, rocblas_dsymv, rocblas_csymv, rocblas_zsymv,
  • rocblas_ssymv_batched, rocblas_dsymv_batched, rocblas_csymv_batched, rocblas_zsymv_batched,
  • rocblas_ssymv_strided_batched, rocblas_dsymv_strided_batched, rocblas_csymv_strided_batched, rocblas_zsymv_strided_batched,

• sbmv

  • rocblas_ssbmv, rocblas_dsbmv,
  • rocblas_ssbmv_batched, rocblas_dsbmv_batched,
  • rocblas_ssbmv_strided_batched, rocblas_dsbmv_strided_batched,

• spmv

  • rocblas_sspmv, rocblas_dspmv,
  • rocblas_sspmv_batched, rocblas_dspmv_batched,
  • rocblas_sspmv_strided_batched, rocblas_dspmv_strided_batched,

• improved documentation

• Fix argument checking in functions to match legacy BLAS

• Fixed conjugate-transpose version of geam

Known failures

• Compilation for GPU Targets
  • When using the install.sh script for "all" GPU Targets, which is the default, you must first set an environment variable HCC_AMDGPU_TARGET listing the GPU targets, e.g. HCC_AMDGPU_TARGET=gfx803,gfx900,gfx906,gfx908
  • If building for a specific architecture(s) using the -a | --architecture flag, you should also set the environment variable HCC_AMDGPU_TARGET to match.
  • Mismatching the environment variable to the -a flag architectures creates builds that may result in SEGFAULTS when running on GPUs which weren't specified.

rocBLAS-2.24.0 for ROCm 3.6.0

New Features

• Improvements to User Guide and Design Document
• L1 dot function optimized to utilize shuffle instructions ( improvements on bf16, f16, f32 data types )
• L1 dot function added x dot x optimized kernel
• Standardization of L1 rocblas-bench to use device pointer mode to focus on GPU memory bandwidth
• Adjustments for hipcc (hip-clang) compiler as standard build compiler and Centos8 support
• Added Fortran interface for all rocBLAS functions

Known Issues

• None

rocBLAS-2.2.0

Changelist:

• Fix compilation of TRSV, IAMAX, IAMIN
• Add TRSM test sizes
• Fix false negative precision failures for f16_r gemm_ex tests
• Improvements to documentation and addition of sample for i8_r/i32_r gemm_ex
• Tuning for i8_r/i32_r gemm_ex for MIOpen
• Add gtest ConfigurableEventListner to reduce Jenkins log file size
• Initial refactorization of rocblas-bench
• rocblas_dgemm NT tuning

rocBLAS-2.1.0

Changelist:

• Refactor rocBLAS test framework
• Improved performance of i8_r/i32_r rocblas_gemm_ex on gfx906
• Addition of simple trsv implementation using trsm
• Improved performance of trsm
• Tuning improvements for resnet50 problems
• Update tuning to use new Tensile solution selection logic
• rocblas_gemm_ex performance improvement when ldd == lcc and strideD == strideC
• Bug fixes for IAMIN and TRSV
• Add sphinx based readthedoc documentation

rocBLAS-2.0.0 for ROCm 2.0

Changelist:

• improved performance of fp16/fp32 rocblas_gemm_ex on gfx906
• support for i8/i32 rocblas_gemm_ex
• update vega-10 resnet50 tuning
• refactor testing to be data driven
• change gemm-ex API solution index from uint32_t to int32_t
• disable gemm and gemm_ex chunking
• fix gemv argument checking
• add performance script for p1b1 benchmark sizes
• refactor gemm code to reduce use of macros
• trsm performance regression fix

rocBLAS-14.3.0 for ROCm1.9

Changelist:

• add rocblas_gemm_strided_batched_ex for mixed precision support
• tested on ROCm1.9
• fix chunking of A and B matrices
• expand testing of rocblas_gemm
• sgemm and hgemm tuning on gfx906 for Resnet50 from Tensile V4.6.0

Known failures:

• known dgemm failures for m,n < 16

enable gfx906 support

A small incremental release to enable gfx906 support. To get gfx906 support, ROCm 1.9 or later must be used to build rocBLAS.

rocBLAS-14.1.2 for ROCm1.8.2

Changelist:

• Add initial rocblas_gemm_ex for mixed precision support and foundation for future capabilities
• use Tensile 4.5.0 for bug fixes and performance improvements
• separate tests into quick, pre_checkin, and nightly
• add sweep tests for gemm

rocBLAS 14.1.1 for ROCm 1.8.2

Changelist:

• update hgemm asm_full YAML file for performance; re-train hgemm hip_lite YAML file
• new YAML files with PreciseBoundsCheck disabled
• update hgemm asm_full YAML file, source and VW=2 for m,n,k <= 32
• update hgemm asm_full YAML file, source and VW=1 for m,n,k == 1
• add strided_batched tests for hgemm
• correct gemm test matrix initialization
• change cmake and source files to support hip-clang
• change from __fp16 to _Float16