mixed precision casues failure to benchmark

[NevesLucas]

setting up the problem type as:

    - # ProblemType
      OperationType: ConvolutionForward
      HighPrecisionAccumulate: False
      DataType: h
      ComputeDataType: h
      DestDataType: s
      TransposeA: False
      TransposeB: False
      UseBeta: False
      Batched: True
      ConvolutionConfig:
        - TensorAFormat: NCHW
        - TensorBFormat: KCYX
        - TensorDFormat: NCHW
        - Filter: 1x1
        - Stride: 1x1
        - Dilation: 1x1
        - GroupCount: 1
        - PackedSpatialDims: 1
        - PackedFilterDims: 1
        - UnrollOnChannel: 1

Causes valid kernels to be generated, but throws
Invalid combination of data types: a: Half, b: Half, c: Float, d: Float, alpha: Half, beta: Half
when benchmarking begins.

Is it even possible to assign different datatypes as above? if not a more graceful exit message would be helpful, waiting until all kernels are generated before erroring out can take a long time.

[babakpst]

Hi @NevesLucas , It is possible to have mixed datatypes. Not all combinations are admissible, though. To enable mixed-precision types HighPrecisionAccumulate must be True (yours is False). Please let me know if that solves the error message.

[babakpst]

The current data types you have is not working, you can modify it as follows:
DataType: h
ComputeDataType: s
DestDataType: s