The MIGraphX gpu-driver is used to test the performance of GPU kernels produced in MIGraphX. Example usage:

gpu-driver kernel_test.json

The input json file describes the gpu kernel and the input data settings. Random data is passed into the gpu kernel and the kernel is run a set number of iterations (1000 in the example below) and timed for performance.

• settings:

  • iterations: the number of iterations to run the kernel
  • lens: the dimensions for the input data shape
  • strides: strides for the input dimension, optional for standard shapes
  • type: data type

• compile_op:

  • name: name of the gpu operation to compile
  • lambda: lambda function
  • inputs: input shapes into the kernel, need 1 more than lambda function input for output buffer

TODO: many other possible settings

Example hjson file that tests a pointwise GELU approximation (note this is hjson and needs to be converted to json before usage):

# sigmoid GELU approximation
{
    settings: {
        iterations: 1000
        lens: [10, 384, 3072]
        type: "float"
    },
    compile_op: {
        name: "pointwise"
        lambda:
        '''
        [](auto x)
        {
            using x_type = decltype(x);
            x_type one = 1.;
            x_type fit_const = 2.;
            return x / (one + exp(-fit_const * x));
        }
        '''
        inputs: [{}, {}]
    }
}

To convert the hjson file to a json file you can use hjson -j. To install hjson: pip install hjson