StructuredBuffer<float4x4> _PerInstanceData;

void GetInstanceDataMatrix_float(float instanceID, out float4x4 mat)
{
    uint id = (uint) instanceID;
    mat = _PerInstanceData[id];
}

void ApplyInstanceMatrix_float(float3 position, float4x4 mat, out float3 worldPosition)
{
    float4 pos = float4(position, 1);
    worldPosition = mul(mat, pos).xyz;
}