Data type support

This topic lists the supported data types for the hipBLASLt GEMM operation, which is performed by hipblasLtMatmul().

The hipDataType enumeration defines data precision types and is primarily used when the data reference itself does not include type information, such as in void* pointers. This enumeration is mainly utilized in BLAS libraries.

The hipBLASLt input and output types are listed in the following table.

hipDataType	hipBLASLt type	Description
HIP_R_8I	hipblasLtInt8	8-bit real signed integer.
HIP_R_32I	hipblasLtInt32	32-bit real signed integer.
HIP_R_4F_E2M1	N/A	4-bit real float4 precision floating-point
HIP_R_6F_E2M3	N/A	6-bit real float6 precision floating-point
HIP_R_6F_E3M2	N/A	6-bit real bfloat6 precision floating-point
HIP_R_8F_E4M3_FNUZ	hipblaslt_f8_fnuz	8-bit real float8 precision floating-point
HIP_R_8F_E5M2_FNUZ	hipblaslt_bf8_fnuz	8-bit real bfloat8 precision floating-point
HIP_R_8F_E4M3	hipblaslt_f8	8-bit real float8 precision floating-point
HIP_R_8F_E5M2	hipblaslt_bf8	8-bit real bfloat8 precision floating-point
HIP_R_16F	hipblasLtHalf	16-bit real half precision floating-point
HIP_R_16BF	hipblasLtBfloat16	16-bit real bfloat16 precision floating-point
HIP_R_32F	hipblasLtFloat	32-bit real single precision floating-point

Note

The hipblaslt_f8_fnuz and hipblaslt_bf8_fnuz data types are only supported on the gfx942 platform. The hipblaslt_f8 and hipblaslt_bf8 data types are only supported on the gfx950 and gfx12 platforms.

The hipBLASLt compute modes are listed in the following table.

hipDataType	Description
HIPBLAS_COMPUTE_32I	32-bit integer compute mode.
HIPBLAS_COMPUTE_16F	16-bit half precision floating-point compute mode.
HIPBLAS_COMPUTE_32F	32-bit single precision floating-point compute mode.
HIPBLAS_COMPUTE_64F	64-bit double precision floating-point compute mode.
HIPBLAS_COMPUTE_32F_FAST_16F	Enables the library to utilize Tensor Cores with automatic down-conversion and 16-bit half-precision computation for 32-bit float-precision input and output matrices.
HIPBLAS_COMPUTE_32F_FAST_16BF	Enables the library to utilize Tensor Cores with automatic down-conversion and 16-bit bfloat16-precision computation for 32-bit float-precision input and output matrices.
HIPBLAS_COMPUTE_32F_FAST_TF32	Enables the library to utilize Tensor Cores with TF32 computation (on the gfx942 and gfx950 platforms) or emulated TF32 computation (on the gfx950 platform) for matrices with 32-bit input and output.

Note

For information on how to override certain compute types, see the environmental variables documentation.

Data type combinations

hipBLASLt supports various combinations of input (A, B), accumulation (C), output (D), and compute data types for GEMM operations. The library enables mixed-precision operations, allowing you to use lower precision inputs with higher precision compute for optimal performance while maintaining accuracy where needed.

The GEMM operation follows this equation:

\[D = Activation(alpha \cdot op(A) \cdot op(B) + beta \cdot op(C) + bias)\]

Where \(op( )\) refers to in-place operations, such as transpose and non-transpose, and \(alpha\) and \(beta\) are scalars.

For complete details on supported data type combinations, including specific compute types, scale types, and bias configurations, see the hipBLASLt API reference page.

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