MIOpen precision support

This topic lists the data type support for the MIOpen library on AMD GPUs.

This page lists the data types supported by the library itself and does not indicate hardware support. A type listed here is only usable if the GPU architecture also supports it; otherwise it is unsupported. For data type support across the other ROCm libraries and by GPU architecture, see the Data types and precision support page.

Supported data types overview

The following table summarizes the input and output data types supported by MIOpen. For the full miopenDataType_t enumeration and per-type support notes, see the section that follows.

Icon	Definition
✅	Fully supported as both an input and output type.
⚠️	Partially supported as an input or output type.

Data types not listed in the table below are not supported.

Data type	Support
int8	⚠️
int32	⚠️
float8 (E4M3)	⚠️
float8 (E5M2)	⚠️
float16	✅
bfloat16	⚠️
float32	✅
float64	⚠️

Datatype enumeration

MIOpen contains several datatypes at different levels of support. The enumerated datatypes are:

typedef enum {
    miopenHalf     = 0,
    miopenFloat    = 1,
    miopenInt32    = 2,
    miopenInt8     = 3,
    /* Value 4 is reserved. */
    miopenBFloat16 = 5,
    miopenDouble   = 6,
    miopenFloat8   = 7,
    miopenBFloat8  = 8
} miopenDataType_t;

Of these types only miopenFloat and miopenHalf are fully supported across all layers in MIOpen. Refer to the individual Modules in the API library for specific datatype support and limitations.

Type descriptions:

• miopenHalf: 16-bit floating point

• miopenFloat: 32-bit floating point

• miopenInt32: 32-bit integer, used primarily for int8 convolution outputs

• miopenInt8: 8-bit integer; supported by int8 convolution forward path, tensor set, tensor copy, tensor cast, tensor transform, tensor transpose, and im2col

• miopenBFloat16: brain float fp-16 (8-bit exponent, 7-bit fraction); supported by convolutions, tensor set, and tensor copy

• miopenDouble: 64-bit floating point; supported by reduction, layerNorm, and batchNorm

• miopenFloat8: 8-bit floating point (layout 1.4.3, exponent bias 7); supported by convolutions

• miopenBFloat8: 8-bit floating point (layout 1.5.2, exponent bias 15); supported by convolutions

Note

Convolution APIs currently only support uniform input/output datatypes and tensor layouts.

In addition to these standard datatypes, pooling also contains its own indexing datatypes:

typedef enum {
    miopenIndexUint8  = 0,
    miopenIndexUint16 = 1,
    miopenIndexUint32 = 2,
    miopenIndexUint64 = 3,
} miopenIndexType_t;