import hipdnn
import torch


def build_soft_margin_loss_graph(
    hipdnn_handle, torch_tensor_input, torch_tensor_target, hipdnn_data_type
):
    graph = hipdnn.pygraph(
        handle=hipdnn_handle,
        io_data_type=hipdnn_data_type,
        intermediate_data_type=hipdnn.data_type.FLOAT,
        compute_data_type=hipdnn.data_type.FLOAT,
        name="soft_margin_loss_forward",
    )
    hipdnn_tensor_input = graph.tensor_like(torch_tensor_input)
    hipdnn_tensor_target = graph.tensor_like(torch_tensor_target)
    hipdnn_tensor_output = graph.soft_margin_loss(
        input=hipdnn_tensor_input,
        target=hipdnn_tensor_target,
        reduction=hipdnn.reduction_mode.AVG,
        name="soft_margin_loss",
    )
    hipdnn_tensor_output.set_output(True)
    graph.build(hipdnn_handle)
    return (graph, hipdnn_tensor_input, hipdnn_tensor_target, hipdnn_tensor_output)


if __name__ == "__main__":
    batch, num_classes = 16, 10
    hipdnn_data_type = hipdnn.data_type.HALF
    torch_data_type = torch.float16
    torch_tensor_input = torch.rand(batch, num_classes, dtype=torch_data_type, device="cuda")
    torch_tensor_target = torch.rand(batch, num_classes, dtype=torch_data_type, device="cuda")
    hipdnn_handle = hipdnn.create_handle()
    graph, hipdnn_tensor_input, hipdnn_tensor_target, hipdnn_tensor_output = (
        build_soft_margin_loss_graph(
            hipdnn_handle,
            torch_tensor_input,
            torch_tensor_target,
            hipdnn_data_type,
        )
    )
    torch_tensor_output = torch.empty(1, dtype=torch_data_type, device="cuda")
    variant_pack = {
        hipdnn_tensor_input: torch_tensor_input.data_ptr(),
        hipdnn_tensor_target: torch_tensor_target.data_ptr(),
        hipdnn_tensor_output: torch_tensor_output.data_ptr(),
    }
    workspace = torch.empty(graph.get_workspace_size(), dtype=torch.uint8, device="cuda")
    graph.exec(variant_pack=variant_pack, workspace=workspace.data_ptr())
    print("soft_margin_loss graph execution complete.")