Full documentation for hipBLASLt is available at rocm.docs.amd.com/projects/hipBLASLt.
- Extension APIs:
hipblasltExtAMaxWithScale
GemmTuning
extension parameter to set wgm by user- Support HIPBLASLT_MATMUL_DESC_AMAX_D_POINTER for FP8/BF8 datatype
- Support for FP8/BF8 input, FP32/FP16/BF16/F8/BF8 output (only for gfx94x platform)
- Support HIPBLASLT_MATMUL_DESC_COMPUTE_INPUT_TYPE_A_EXT and HIPBLASLT_MATMUL_DESC_COMPUTE_INPUT_TYPE_B_EXT for FP16 input datatype to use FP8/BF8 mfma
- Support for gfx110x
- Improve library loading time
- Extension APIs:
hipblasltExtSoftmax
hipblasltExtLayerNorm
hipblasltExtAMax
GemmTuning
extension parameter to set split-k by user- Support for mixed-precision datatype: FP16/FP8 in with FP16 out
- Add CMake support for documentation
- Support for datatype: Int8 in Int8 out
- algoGetHeuristic() ext API for GroupGemm will be deprecated in a future release of hipBLASLt
- New
UserArguments
variable forGroupedGemm
- Support for datatype: FP16 in with FP32 out
- Support for datatype: Int8 in Int32 out
- Support for gfx94x platform
- Support for FP8/BF8 datatype (only for gfx94x platform)
- Support scalar A,B,C,D for FP8/BF8 datatype
- Added samples
- Replaced
hipblasDatatype_t
withhipDataType
- Replaced
hipblasLtComputeType_t
withhipblasComputeType_t
- Deprecated
HIPBLASLT_MATMUL_DESC_D_SCALE_VECTOR_POINTER
- Added
getAllAlgos
extension APIs - TensileLite support for new epilogues: gradient gelu, gradient D, gradient A/B, aux
- Added a sample package that includes three sample apps
- Added a new C++ GEMM class in the hipBLASLt extension
- Refactored GroupGemm APIs as C++ class in the hipBLASLt extension
- Changed the scaleD vector enum to
HIPBLASLT_MATMUL_DESC_D_SCALE_VECTOR_POINTER
- Enabled norm check validation for CI
- GSU kernel: wider memory, PGR N
- Updated logic yaml to improve some FP16 NN sizes
- GroupGemm support for GSU kernel
- Added grouped GEMM tuning for aldebaran
- Added CI tests for TensileLite
- Initialized extension group GEMM APIs (FP16 only)
- Added a group GEMM sample app:
example_hipblaslt_groupedgemm
- Fixed incorrect results for the ScaleD kernel
- Tuned equality sizes for the HHS data type
- Reduced host-side overhead for
hipblasLtMatmul()
- Removed unused kernel arguments
- Schedule values setup before first
s_waitcnt
- Refactored TensileLite host codes
- Optimized build time
- Enabled hipBLASLt APIs
- Support for gfx90a
- Support for problem type: FP32, FP16, BF16
- Support activation: relu, gelu
- Support for bias vectors
- Integrated with TensileLite kernel generator
- Added Gtest:
hipblaslt-test
- Added the full function tool
hipblaslt-bench
- Added the sample app
example_hipblaslt_preference
- gridBase solution search algorithm for untuned size
- Tuned 10k sizes for each problem type