Data type support

This topic lists the data type support for the hipBLAS library on AMD GPUs for different levels of BLAS operations Level 1, Level 2, and Level 3.

The hipBLAS library functions are also available with ILP64 interfaces. With these interfaces, all int arguments are replaced by the type name int64_t. For more information on these _64 functions, see the ILP64 interfaces section.

The icons representing different levels of support are explained in the following table.

Icon	Definition
NA	Not applicable
❌	Not supported
⚠️	Partial support
✅	Full support

For more information about data type support for the other ROCm libraries, see Data types and precision support page.

Custom types

hipBlas defines the hipblasBfloat16 and uses the HIP types hipComplex and hipDoubleComplex.

The bfloat 16 data type

hipBLAS defines a hipblasBfloat16 data type. This type is exposed as a struct containing 16 bits of data. There is also a C++ hipblasBfloat16 class defined which provides slightly more functionality, including conversion to and from a 32-bit float data type. This class can be used in C++11 or newer by defining HIPBLAS_BFLOAT16_CLASS before including the header file <hipblas.h>.

There is also an option to interpret the API as using the hip_bfloat16 data type. This is provided to avoid casting when using the hip_bfloat16 data type. To expose the API using hip_bfloat16, define HIPBLAS_USE_HIP_BFLOAT16 before including the header file <hipblas.h>.

Note

The hip_bfloat16 data type is only supported on AMD platforms.

Complex data types

hipBLAS uses the HIP types hipComplex and hipDoubleComplex in its API.

Functions data type support

This section collects the data type support tables of BLAS functions and solver functions. The cuBLAS backend does not support all the functions and all the types. For more information, see cuBLAS backend.

Level 1 functions - vector operations

Level-1 functions perform scalar, vector, and vector-vector operations.

Float types

Function	Description	float16	bfloat16	float	double
AMax, AMin, ASum	Finds the first index of the element of minimum or maximum magnitude of a vector x or computes the sum of the magnitudes of elements of a real vector x.	❌	❌	✅	✅
AXPY	Scales a vector and adds it to another: \(y = \alpha x + y\)	✅	❌	✅	✅
Copy	Copies vector x to y: \(y = x\)	❌	❌	✅	✅
Dot	Computes the dot product: \(result = x^T y\)	✅	✅	✅	✅
NRM2	Computes the Euclidean norm of a vector.	❌	❌	✅	✅
Rot, Rotg	Applies and generates a Givens rotation matrix.	❌	❌	✅	✅
Rotm, Rotmg	Applies and generates a modified Givens rotation matrix.	❌	❌	✅	✅
Scal	Scales a vector by a scalar: \(x = \alpha x\)	❌	❌	✅	✅
Swap	Swaps corresponding elements of two vectors x and y: \(x_i \leftrightarrow y_i \quad \text{for} \quad i = 0, 1, 2, \ldots, n - 1\)	❌	❌	✅	✅

Complex types

Function	Description	complex	double complex
AMax, AMin, ASum	Finds the first index of the element of minimum or maximum magnitude of a vector x or computes the sum of the magnitudes of elements of a real vector x.	✅	✅
AXPY	Scales a vector and adds it to another: \(y = \alpha x + y\)	✅	✅
Copy	Copies vector x to y: \(y = x\)	✅	✅
Dot	Computes the dot product: \(result = x^T y\)	✅	✅
NRM2	Computes the Euclidean norm of a vector.	✅	✅
Rot, Rotg	Applies and generates a Givens rotation matrix.	✅	✅
Rotm, Rotmg	Applies and generates a modified Givens rotation matrix.	❌	❌
Scal	Scales a vector by a scalar: \(x = \alpha x\)	✅	✅
Swap	Swaps corresponding elements of two vectors x and y: \(x_i \leftrightarrow y_i \quad \text{for} \quad i = 0, 1, 2, \ldots, n - 1\)	✅	✅

Level 2 functions - matrix-vector operations

Level-2 functions perform matrix-vector operations.

Float types

Function	Description	float16	bfloat16	float	double
GBMV	General band matrix-vector multiplication: \(y = \alpha A x + \beta y\)	❌	❌	✅	✅
GEMV	General matrix-vector multiplication: \(y = \alpha A x + \beta y\)	❌	❌	✅	✅
GER	Generalized rank-1 update: \(A = \alpha x y^T + A\)	❌	❌	✅	✅
GERU and GERC	Generalized rank-1 update for unconjugated or conjugated complex numbers: \(A = \alpha x y^T + A\)	NA	NA	NA	NA
SBMV, SPMV	Symmetric Band matrix-vector multiplication: \(y = \alpha A x + \beta y\)	❌	❌	✅	✅
SPR	Symmetric packed rank-1 update.	❌	❌	✅	✅
SPR2	Symmetric packed rank-2 update.	❌	❌	✅	✅
SYMV	Symmetric matrix-vector multiplication: \(y = \alpha A x + \beta y\)	❌	❌	✅	✅
SYR, SYR2	Symmetric matrix rank-1 or rank-2 update.	❌	❌	✅	✅
TBMV, TBSV	Triangular band matrix-vector multiplication, triangular band solve.	❌	❌	✅	✅
TPMV, TPSV	Triangular packed matrix-vector multiplication, triangular packed solve.	❌	❌	✅	✅
TRMV, TRSV	Triangular matrix-vector multiplication, triangular solve.	❌	❌	✅	✅
HEMV, HBMV, HPMV	Hermitian matrix-vector multiplication.	NA	NA	NA	NA
HER, HER2	Hermitian rank-1 and rank-2 update.	NA	NA	NA	NA
HPR, HPR2	Hermitian packed rank-1 and rank-2 update of packed.	NA	NA	NA	NA

Complex types

Function	Description	complex	double complex
GBMV	General band matrix-vector multiplication: \(y = \alpha A x + \beta y\)	✅	✅
GEMV	General matrix-vector multiplication: \(y = \alpha A x + \beta y\)	✅	✅
GER	Generalized rank-1 update: \(A = \alpha x y^T + A\)	NA	NA
GERU and GERC	Generalized rank-1 update for unconjugated or conjugated complex numbers: \(A = \alpha x y^T + A\)	✅	✅
SBMV, SPMV	Symmetric Band matrix-vector multiplication: \(y = \alpha A x + \beta y\)	❌	❌
SPR	Symmetric packed rank-1 update.	✅	✅
SPR2	Symmetric packed rank-2 update.	❌	❌
SYMV	Symmetric matrix-vector multiplication: \(y = \alpha A x + \beta y\)	✅	✅
SYR, SYR2	Symmetric matrix rank-1 or rank-2 update.	✅	✅
TBMV, TBSV	Triangular band matrix-vector multiplication, triangular band solve.	✅	✅
TPMV, TPSV	Triangular packed matrix-vector multiplication, triangular packed solve.	✅	✅
TRMV, TRSV	Triangular matrix-vector multiplication, triangular solve.	✅	✅
HEMV, HBMV, HPMV	Hermitian matrix-vector multiplication.	✅	✅
HER, HER2	Hermitian rank-1 and rank-2 update.	✅	✅
HPR, HPR2	Hermitian packed rank-1 and rank-2 update.	✅	✅

Level 3 functions - matrix-matrix operations

Level-3 functions perform matix-matrix operations. hipBLAS calls the AMD Tensile and hipBLASLt libraries for Level-3 GEMMs (matrix matrix multiplication).

Float types

Function	Description	float16	bfloat16	float	double
GEMM	General matrix-matrix multiplication: \(C = \alpha A B + \beta C\)	✅	❌	✅	✅
SYMM	Symmetric matrix-matrix multiplication: \(C = \alpha A B + \beta C\)	❌	❌	✅	✅
SYRK, SYR2K	Update symmetric matrix with one matrix product or by using two matrices.	❌	❌	✅	✅
SYRKX	SYRKX adds an extra matrix multiplication step before updating the symmetric matrix.	❌	❌	✅	✅
TRMM	Triangular matrix-matrix multiplication.	❌	❌	✅	✅
TRSM	Triangular solve with multiple right-hand sides.	❌	❌	✅	✅
HEMM	Hermitian matrix-matrix multiplication.	NA	NA	NA	NA
HERK, HER2K	Update Hermitian matrix with one matrix product or by using two matrices.	NA	NA	NA	NA
HERKX	HERKX adds an extra matrix multiplication step before updating the Hermitian matrix.	NA	NA	NA	NA
TRTRI	Triangular matrix inversion.	❌	❌	✅	✅
DGMM	Diagonal matrix matrix multiplication.	❌	❌	✅	✅

Complex types

Function	Description	complex	double complex
GEMM	General matrix-matrix multiplication: \(C = \alpha A B + \beta C\)	✅	✅
SYMM	Symmetric matrix-matrix multiplication: \(C = \alpha A B + \beta C\)	✅	✅
SYRK, SYR2K	Update symmetric matrix with one matrix product or by using two matrices.	✅	✅
SYRKX	SYRKX adds an extra matrix multiplication step before updating the symmetric matrix.	✅	✅
TRMM	Triangular matrix-matrix multiplication.	✅	✅
TRSM	Triangular solve with multiple right-hand sides.	✅	✅
HEMM	Hermitian matrix-matrix multiplication.	✅	✅
HERK, HER2K	Update Hermitian matrix with one matrix product or by using two matrices.	✅	✅
HERKX	HERKX adds an extra matrix multiplication step before updating the Hermitian matrix.	✅	✅
TRTRI	Triangular matrix inversion.	✅	✅
DGMM	Diagonal matrix matrix multiplication.	✅	✅

Extensions

The extension functions data type support is listed separately for every function for the different backends in the rocBLAS extensions and cuBLAS extensions documentation.

SOLVER API

Solver API is for solving linear systems, computing matrix inverses, and performing matrix factorizations.

Float types

Function	Description	float16	bfloat16	float	double
GETRF	Compute the LU factorization of a general matrix using partial pivoting with row interchanges.	❌	❌	✅	✅
GETRS	Solve a system of linear equations \(AxX=B\) after performing an LU factorization using GETRF.	❌	❌	✅	✅
GETRI	Compute the inverse of a matrix using its LU factorization.	❌	❌	⚠️ [1]	⚠️ [1]
GEQRF	QR factorization of a general matrix.	❌	❌	✅	✅
GELS	Solve overdetermined or underdetermined linear systems using the QR factorization of a matrix.	❌	❌	✅	✅

Complex types

Function	Description	complex	double complex
GETRF	Compute the LU factorization of a general matrix using partial pivoting with row interchanges.	✅	✅
GETRS	Solve a system of linear equations \(AxX=B\) after performing an LU factorization using GETRF.	✅	✅
GETRI	Compute the inverse of a matrix using its LU factorization.	⚠️ [1]	⚠️ [1]
GEQRF	QR factorization of a general matrix.	✅	✅
GELS	Solve overdetermined or underdetermined linear systems using the QR factorization of a matrix.	✅	✅

Footnotes

[1] (1,2,3,4)

Only the batched GETRI functions are supported.