15.25. __nv_fp8x2_ue5m3

struct __nv_fp8x2_ue5m3

This structure implements the datatype for storage and operations on the vector of two fp8 values of ue5m3 kind each: with no sign, 5 exponent, 1 implicit and 3 explicit mantissa bits. The encoding doesn’t support Infinity. NaNs are limited to 0xFF values.

Public Functions

__nv_fp8x2_ue5m3() = default: Constructor by default.

__host__ __device__ inline explicit __nv_fp8x2_ue5m3(const __half2 f)

Constructor from __half2 data type, relies on __NV_SATFINITE behavior for large input values and cudaRoundNearest for rounding.

See also

__nv_cvt_halfraw_to_ue5m3 for further details

__host__ __device__ inline explicit __nv_fp8x2_ue5m3(const __nv_bfloat162 f)

Constructor from __nv_bfloat162 data type, relies on __NV_SATFINITE behavior for large input values and cudaRoundNearest for rounding.

See also

__nv_cvt_bfloat16raw_to_ue5m3 for further details

__host__ __device__ inline explicit __nv_fp8x2_ue5m3(const double2 f)

Constructor from double2 data type, relies on __NV_SATFINITE behavior for large input values and cudaRoundNearest for rounding.

See also

__nv_cvt_double_to_ue5m3 for further details

__host__ __device__ inline explicit __nv_fp8x2_ue5m3(const float2 f)

Constructor from float2 data type, relies on __NV_SATFINITE behavior for large input values and cudaRoundNearest for rounding.

See also

__nv_cvt_float_to_ue5m3 for further details

__host__ __device__ inline explicit operator __half2() const: Conversion operator to __half2 data type.

__host__ __device__ inline explicit operator __nv_bfloat162() const: Conversion operator to __nv_bfloat162 data type.

__host__ __device__ inline explicit operator float2() const: Conversion operator to float2 data type.

Public Members

__nv_fp8x2_storage_t __x: Storage variable contains the vector of two fp8 floating-point data values.