Skip to content

[QST] INT8 GEMM with FP16 output #767

@gsujankumar

Description

@gsujankumar

Hey I have noticed INT8 gemm with the following settings where we have float bias and float output works okay:

using ElementAccumulator = int32_t;                   // <- data type of accumulator
using ElementComputeEpilogue = float;  // <- data type of epilogue operations
using ElementInputA = int8_t;              // <- data type of elements in input matrix A
using ElementInputB = int8_t;              // <- data type of elements in input matrix B
using ElementOutput = float;                        // <- data type of elements in output matrix D

But setting the outputs in cutlass::half does not?

using ElementAccumulator = int32_t;                   // <- data type of accumulator
using ElementComputeEpilogue = float;  // <- data type of epilogue operations
using ElementInputA = int8_t;              // <- data type of elements in input matrix A
using ElementInputB = int8_t;              // <- data type of elements in input matrix B
using ElementOutput = cutlass::half;                        // <- data type of elements in output matrix D

Is there any specific restriction preventing it? How can this be enabled?

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions