Releases: ROCm/Tensile
Releases · ROCm/Tensile
Tensile 4.24.0 for rocm 3.10.0
Known Issues
- None
Tensile 4.23.0 for rocm 3.9.0
Merge pull request #1160 from zaliu/master ROCm 3.9 merge develop into master
Tensile-4.22.0 for ROCm 3.8.0
New Features
Known Issues
- None
V4.9.0 Performance improvements
Features
- Improve persistent kernel implementation
- Add sequential indexing to tensile merge script
V4.8.0 Performance improvements, bug fixes, add assembly hpa_igemm
NOTE: ABI/API breaking changes introduced in this release, related to the addition of a separate pointer to the output Matrix D.
Features
- new solution selection logic
- add persistant kernel
- enable StaggerU
- Add 6x6 and 6x4 micro-tiles
- Add FUSS kernels
- gfx906 DGEMM NN/NT tuning
- add dot4 instructions for i8_r/i32_r gemm_ex on gfx906
- Matrix D Support.
- GEMM Calls now take in a separate pointer to the output matrix D, replacing matrix C for output (D = aAB + bC).
- Matrix C will now only be used for the input to GEMM calls.
- For this release, Matrix D uses the same stride as Matrix C.
- Previous functionality can be obtained by passing in the same pointer to both Matrix C and D.
- fixes for merge script
- add scripts for tuning automation
- add replacement kernel logic
- Improved Tensile run times for large numbers of solutions
-
V4.7.0 Performance improvements, bug fixes, add assembly hpa_hgemm, initial source hpa_igemm
Features
- add dot2 instructions for fp16/fp32 hpa_hgemm on gfx906
- initial i8/i32 hpa_igemm
- enable fractional loads
- enable precise bounds check
V4.6.0 Performance improvements, Bug fixes, add source hpa_hgemm
Features
- Merge gfx906 code into gfx900/gfx803 code
- Tune hgemm and sgemm for Resnet50 on gfx906
- Add source hpa_hgemm
- Use precise bounds check when possible
- Tested on ROCm 1.9
Make rocBLAS build with pre-ROCm 1.9 compilers work
A small incremental release to make rocBLAS v14.2.5 buildable with pre-ROCm 1.9 compilers.
V4.5.0 Performance improvements, Bug fixes, add hpa_hgemm
Features
- add support for vega20
- add hpa_hgemm assembly and source
- tuning for sgemm and hgemm
- bug fixes for sgemm and hgemm small sizes
- use SGPR for alpha and beta