OpenBLAS/docs/extensions.md

OpenBLAS for the most part contains implementations of the reference (Netlib) BLAS, CBLAS, LAPACK and LAPACKE interfaces. A few OpenBLAS-specific functions are also provided however, which mostly can be seen as "BLAS extensions". This page documents those non-standard APIs.

BLAS-like extensions

Routine	Data Types	Description
?axpby	s,d,c,z	like axpy with a multiplier for y
?gemm3m	c,z	gemm3m
?imatcopy	s,d,c,z	in-place transposition/copying
?omatcopy	s,d,c,z	out-of-place transposition/copying
?geadd	s,d,c,z	ATLAS-like matrix add B = α*A+β*B
?gemmt	s,d,c,z	gemm but only a triangular part updated
cblas_?gemm_batch	s,d,c,z,b	gemm with several groups of input data
cblas_?gemm_batch_strided	s,d,c,z,b	gemm with groups of data stored at fixed offsets in the input arrays

bfloat16 functionality

BLAS-like and conversion functions for bfloat16 (available when OpenBLAS was compiled with BUILD_BFLOAT16=1):

• void cblas_sbstobf16 converts a float array to an array of bfloat16 values by rounding
• void cblas_sbdtobf16 converts a double array to an array of bfloat16 values by rounding
• void cblas_sbf16tos converts a bfloat16 array to an array of floats
• void cblas_dbf16tod converts a bfloat16 array to an array of doubles
• float cblas_sbdot computes the dot product of two bfloat16 arrays
• void cblas_sbgemv performs the matrix-vector operations of GEMV with the input matrix and X vector as bfloat16
• void cblas_sbgemm performs the matrix-matrix operations of GEMM with both input arrays containing bfloat16
• void cblas_bgemv performs the matrix-vector operations of GEMV with the input matrix, X vector and result as bfloat16
• void cblas_bgemm performs the matrix-matrix operations of GEMM with both input arrays containing bfloat16 and the output being bfloat16 as well

half-precision float or fp16 functionality

BLAS-like and conversion functions for hfloat16 (available when OpenBLAS was compiled with BUILD_HFLOAT16=1):

• void cblas_shgemm performs the matrix-matrix operations of GEMM with both input arrays containing hfloat16

Utility functions

• openblas_get_num_threads
• openblas_set_num_threads
• int openblas_get_num_procs(void) returns the number of processors available on the system (may include "hyperthreading cores")
• int openblas_get_parallel(void) returns 0 for sequential use, 1 for platform-based threading and 2 for OpenMP-based threading
• char * openblas_get_config() returns the options OpenBLAS was built with, something like NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Haswell
• int openblas_set_affinity(int thread_index, size_t cpusetsize, cpu_set_t *cpuset) sets the CPU affinity mask of the given thread to the provided cpuset. Only available on Linux, with semantics identical to pthread_setaffinity_np.
• openblas_set_thread_callback_function overrides the default multithreading backend with the provided argument
• openblas_set_xerbla(openblas_xerbla_handler handler) installs a process-wide XERBLA error handler and returns the previous handler. Passing NULL restores the default handler. Handlers must be thread-safe because they may be called concurrently. Replacing a handler is thread-safe but does not wait for in-progress calls to return, so the previous handler's code must remain loaded until those calls finish. The name and info pointers are valid only during the callback. name points to name_length valid bytes and need not be NUL-terminated; the length stops before the first NUL but otherwise includes trailing spaces. Handler registration works normally on ELF platforms. For backward compatibility, an application-provided strong xerbla symbol still overrides the OpenBLAS dispatcher and bypasses the registered handler.