import hipdnn
import torch


def build_groupnorm_fwd_graph(
    hipdnn_handle,
    torch_tensor_x,
    torch_tensor_scale,
    torch_tensor_bias,
    torch_tensor_epsilon,
    groups,
    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="groupnorm_fwd",
    )

    # 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)

    # Create groupnorm op
    hipdnn_tensor_y, hipdnn_tensor_mean, hipdnn_tensor_inv_var = graph.groupnorm(
        norm_forward_phase=hipdnn.norm_forward_phase.TRAINING,
        input=hipdnn_tensor_x,
        scale=hipdnn_tensor_scale,
        bias=hipdnn_tensor_bias,
        epsilon=hipdnn_tensor_epsilon,
        groups=groups,
        name="groupnorm",
    )
    hipdnn_tensor_y.set_output(True)
    hipdnn_tensor_mean.set_output(True)
    hipdnn_tensor_inv_var.set_output(True)
    graph.build(hipdnn_handle)

    return (
        graph,
        hipdnn_tensor_x,
        hipdnn_tensor_scale,
        hipdnn_tensor_bias,
        hipdnn_tensor_epsilon,
        hipdnn_tensor_y,
        hipdnn_tensor_mean,
        hipdnn_tensor_inv_var,
    )


if __name__ == "__main__":
    # Input dimensions
    batch, channels, height, width = 2, 16, 512, 512

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

    torch_tensor_x = torch.rand(
        batch, channels, height, width, dtype=torch_data_type, device="cuda"
    )
    torch_tensor_scale = torch.rand(channels, dtype=torch_data_type, device="cuda")
    torch_tensor_bias = torch.rand(channels, dtype=torch_data_type, device="cuda")
    torch_tensor_epsilon = torch.full(
        (1, 1, 1, 1), 1e-5, dtype=torch.float32, requires_grad=False, device="cpu"
    )
    groups = 2

    hipdnn_handle = hipdnn.create_handle()

    (
        graph,
        hipdnn_tensor_x,
        hipdnn_tensor_scale,
        hipdnn_tensor_bias,
        hipdnn_tensor_epsilon,
        hipdnn_tensor_y,
        hipdnn_tensor_mean,
        hipdnn_tensor_inv_var,
    ) = build_groupnorm_fwd_graph(
        hipdnn_handle,
        torch_tensor_x,
        torch_tensor_scale,
        torch_tensor_bias,
        torch_tensor_epsilon,
        groups,
        hipdnn_data_type,
    )

    torch_tensor_y = torch.empty(hipdnn_tensor_y.get_dim(), dtype=torch_data_type, device="cuda")
    torch_tensor_mean = torch.empty(
        hipdnn_tensor_mean.get_dim(), dtype=torch.float32, device="cuda"
    )
    torch_tensor_inv_var = torch.empty(
        hipdnn_tensor_inv_var.get_dim(), dtype=torch.float32, 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(),
        hipdnn_tensor_mean: torch_tensor_mean.data_ptr(),
        hipdnn_tensor_inv_var: torch_tensor_inv_var.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("groupnorm fwd graph execution complete.")