import hipdnn
import torch


def build_instancenorm_graph(
    hipdnn_handle,
    torch_tensor_x,
    torch_tensor_scale,
    torch_tensor_bias,
    torch_tensor_epsilon,
    mode,
    eps,
    hipdnn_data_type,
):
    # Create graph
    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="instancenorm_inference",
    )

    # Create hipdnn tensors
    hipdnn_tensor_x = graph.tensor_like(torch_tensor_x)
    hipdnn_tensor_scale = graph.tensor_like(torch_tensor_scale)
    hipdnn_tensor_bias = graph.tensor_like(torch_tensor_bias)
    hipdnn_tensor_epsilon = graph.tensor_like(torch_tensor_epsilon)
    hipdnn_tensor_epsilon.set_value(eps)

    # Create op
    hipdnn_tensor_y, hipdnn_tensor_mean, hipdnn_tensor_inv_var = graph.instancenorm(
        mode,
        hipdnn_tensor_x,
        hipdnn_tensor_scale,
        hipdnn_tensor_bias,
        hipdnn_tensor_epsilon,
        hipdnn.data_type.FLOAT,
        name="instancenorm",
    )
    hipdnn_tensor_y.set_output(True)
    graph.build(hipdnn_handle)

    return (graph, hipdnn_tensor_x, hipdnn_tensor_scale, hipdnn_tensor_bias, hipdnn_tensor_y)


if __name__ == "__main__":
    # Input dimensions
    n = 16  # Batch size
    c = 32  # Number of input channels
    h = 64  # Height
    w = 64  # Width
    mode = hipdnn.norm_forward_phase.INFERENCE  # Mode
    eps = 1e-5

    hipdnn_data_type = hipdnn.data_type.FLOAT
    torch_data_type = torch.float32

    torch_tensor_x = torch.rand((n, c, h, w), dtype=torch_data_type, device="cuda")
    torch_tensor_scale = torch.rand((1, c, 1, 1), dtype=torch_data_type, device="cuda")
    torch_tensor_bias = torch.rand((1, c, 1, 1), dtype=torch_data_type, device="cuda")
    torch_tensor_epsilon = torch.full(
        (1, 1, 1, 1), eps, dtype=torch.float32, requires_grad=False, device="cpu"
    )

    hipdnn_handle = hipdnn.create_handle()

    graph, hipdnn_tensor_x, hipdnn_tensor_scale, hipdnn_tensor_bias, hipdnn_tensor_y = (
        build_instancenorm_graph(
            hipdnn_handle,
            torch_tensor_x,
            torch_tensor_scale,
            torch_tensor_bias,
            torch_tensor_epsilon,
            mode,
            eps,
            hipdnn_data_type,
        )
    )

    torch_tensor_y = torch.empty(hipdnn_tensor_y.get_dim(), dtype=torch_data_type, device="cuda")
    variant_pack = {
        hipdnn_tensor_x: torch_tensor_x.data_ptr(),
        hipdnn_tensor_scale: torch_tensor_scale.data_ptr(),
        hipdnn_tensor_bias: torch_tensor_bias.data_ptr(),
        hipdnn_tensor_y: torch_tensor_y.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("instancenorm graph execution complete.")